Advances in Swarm Intelligence 2014

Ying Tan Yuhui Shi
Carlos A. Coello Coello (Eds.)

LNCS 8795
Advances
in Swarm Intelligence
5th International Conference, ICSI 2014
Hefei, China, October 17–20, 2014
Proceedings, Part II
123
Lecture Notes in Computer Science 8795
Commenced Publication in 1973
Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen
Editorial Board
David Hutchison
Lancaster University, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Alfred Kobsa
University of California, Irvine, CA, USA
Friedemann Mattern
ETH Zurich, Switzerland
John C. Mitchell
Stanford University, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
Oscar Nierstrasz
University of Bern, Switzerland
C. Pandu Rangan
Indian Institute of Technology, Madras, India
Bernhard Steffen
TU Dortmund University, Germany
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max Planck Institute for Informatics, Saarbruecken, Germany
Ying Tan Yuhui Shi
Carlos A. Coello Coello (Eds.)
Advances
in Swarm Intelligence
5th International Conference
ICSI 2014, Hefei, China, October 17-20, 2014
Proceedings, Part II
13
Volume Editors
Ying Tan
Peking University
Key Laboratory of Machine Perception (MOE)
School of Electronics Engineering and Computer Science
Department of Machine Intelligence
Beijing 100871, China
E-mail: ytan@pku.edu.cn
Yuhui Shi
Xi’an Jiaotong-Liverpool University
Department of Electrical and Electronic Engineering
Suzhou 215123, China
E-mail: yuhui.shi@xjtlu.edu.cn
Carlos A. Coello Coello
CINVESTAV-IPN
Investigador Cinvestav 3F, Depto. de Computación
México, D.F. 07300, Mexico
E-mail: ccoello@cs.cinvestav.mx
ISSN 0302-9743 e-ISSN 1611-3349

ISBN 978-3-319-11896-3 e-ISBN 978-3-319-11897-0
DOI 10.1007/978-3-319-11897-0
Springer Cham Heidelberg New York Dordrecht London
Library of Congress Control Number: 2014949260
LNCS Sublibrary: SL 1 – Theoretical Computer Science and General Issues

© Springer International Publishing Switzerland 2014
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered and
executed on a computer system, for exclusive use by the purchaser of the work. Duplication of this publication
or parts thereof is permitted only under the provisions of the Copyright Law of the Publisher’s location,
in ist current version, and permission for use must always be obtained from Springer. Permissions for use
may be obtained through RightsLink at the Copyright Clearance Center. Violations are liable to prosecution
under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of publication,
neither the authors nor the editors nor the publisher can accept any legal responsibility for any errors or
omissions that may be made. The publisher makes no warranty, express or implied, with respect to the
material contained herein.
Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)
Preface
This book and its companion volume, LNCS vols. 8794 and 8795, constitute the
proceedings of the fifth International Conference on Swarm Intelligence (ICSI
2014) held during October 17–20, 2014, in Hefei, China. ICSI 2014 was the
fifth international gathering in the world for researchers working on all aspects
of swarm intelligence, following the successful and fruitful Harbin event (ICSI
2013), Shenzhen event (ICSI 2012), Chongqing event (ICSI 2011) and Beijing
event (ICSI 2010), which provided a high-level academic forum for the partici-
pants to disseminate their new research findings and discuss emerging areas of
research. It also created a stimulating environment for the participants to inter-
act and exchange information on future challenges and opportunities in the field
of swarm intelligence research.
ICSI 2014 received 198 submissions from about 475 authors in 32 countries
and regions (Algeria, Australia, Belgium, Brazil, Chile, China, Czech Repub-
lic, Finland, Germany, Hong Kong, India, Iran, Ireland, Italy, Japan, Macao,
Malaysia, Mexico, New Zealand, Pakistan, Romania, Russia, Singapore, South
Africa, Spain, Sweden, Taiwan, Thailand, Tunisia, Turkey, United Kingdom,
United States of America) across six continents (Asia, Europe, North Amer-
ica, South America, Africa, and Oceania). Each submission was reviewed by at
least two reviewers, and on average 2.7 reviewers. Based on rigorous reviews
by the Program Committee members and reviewers, 105 high-quality papers
were selected for publication in this proceedings volume with an acceptance rate
of 53.03%. The papers are organized in 18 cohesive sections, 3 special sessions
and one competitive session, which cover all major topics of swarm intelligence
research and development.
As organizers of ICSI 2014, we would like to express sincere thanks to Univer-
sity of Science and Technology of China, Peking University, and Xi’an Jiaotong-
Liverpool University for their sponsorship, as well as to the IEEE Computational
Intelligence Society, World Federation on Soft Computing, and
International Neural Network Society for their technical co-sponsorship. We ap-
preciate the Natural Science Foundation of China for its financial and logistic
support. We would also like to thank the members of the Advisory Committee
for their guidance, the members of the International Program Committee and
additional reviewers for reviewing the papers, and the members of the Publi-
cations Committee for checking the accepted papers in a short period of time.
Particularly, we are grateful to Springer for publishing the proceedings in the
prestigious series of Lecture Notes in Computer Science. Moreover, we wish to
express our heartfelt appreciation to the plenary speakers, session chairs, and
student helpers. In addition, there are still many more colleagues, associates,
VI Preface
friends, and supporters who helped us in immeasurable ways; we express our

sincere gratitude to them all. Last but not the least, we would like to thank all
the speakers, authors, and participants for their great contributions that made
ICSI 2014 successful and all the hard work worthwhile.
July 2014 Ying Tan

Yuhui Shi
Carlos A. Coello Coello
Organization
General Chairs
Russell C. Eberhart Indiana University-Purdue University, USA
Ying Tan Peking University, China
Programme Committee Chairs

Yuhui Shi Xi’an Jiaotong-Liverpool University, China
Carlos A. Coello Coello CINVESTAV-IPN, Mexico
Advisory Committee Chairs

Gary G. Yen Oklahoma State University, USA
Hussein Abbass University of New South Wales, ADFA,
Australia
Xingui He Peking University, China
Technical Committee Chairs

Xiaodong Li RMIT University, Australia
Andries Engelbrecht University of Pretoria, South Africa
Ram Akella University of California, USA
M. Middendorf University of Leipzig, Germany
Kalyanmoy Deb Indian Institute of Technology Kanpur, India
Ke Tang University of Science and Technology of China,
China
Special Sessions Chairs

Shi Cheng The University of Nottingham, Ningbo, China
Meng-Hiot Lim Nanyang Technological University, Singapore
Benlian Xu Changshu Institute of Technology, China
Competition Session Chairs

Jane J. Liang Zhengzhou University, China
Junzhi Li Peking University, China
VIII Organization
Publications Chairs
Radu-Emil Precup Politehnica University of Timisoara, Romania
Haibin Duan Beihang University, China
Publicity Chairs
Yew-Soon Ong Nanyang Technological University, Singapore
Juan Luis Fernandez Martinez University of Oviedo, Spain
Hideyuki Takagi Kyushu University, Japan
Qingfu Zhang University of Essex, UK
Suicheng Gu University of Pittsburgh, USA
Fernando Buarque University of Pernambuco, Brazil
Ju Liu Shandong University, China
Finance and Registration Chairs

Chao Deng Peking University, China
Andreas Janecek University of Vienna, Austria
Local Arrangement Chairs

Wenjian Luo University of Science and Technology of China,
China
Bin Li University of Science and Technology of China,
China
Program Committee
Kouzou Abdellah University of Djelfa, Algeria
Ramakrishna Akella University of California at Santa Cruz, USA
Rafael Alcala University of Granada, Spain
Peter Andras Newcastle University, UK
Esther Andrés INTA, USA
Sabri Arik Istanbul University, Turkey
Helio Barbosa Laboratório Nacional de Computação
Cientı́fica, Brazil
Carmelo J.A. Bastos Filho University of Pernambuco, Brazil
Christian Blum Technical University of Catalonia, Spain
Salim Bouzerdoum University of Wollongong, Australia
Xinye Cai Nanhang University, China
David Camacho Universidad Autonoma de Madrid, Spain
Bin Cao Tsinghua University, China
Kit Yan Chan DEBII, Australia
Organization IX
Mu-Song Chen Da-Yeh University, Taiwan

Walter Chen National Taipei University of Technology,
China
Shi Cheng The University of Nottingham Ningbo, China
Leandro Coelho Pontifı́cia Universidade Católica do Parana,
Brazil
Chenggang Cui Shanghai Advanced Research Institute, Chinese
Academy of Sciences, China
Chaohua Dai Southwest Jiaotong University, China
Arindam K. Das University of Washington, USA
Prithviraj Dasgupta University of Nebraska at Omaha, USA
Kusum Deep Indian Institute of Technology Roorkee, India
Mingcong Deng Tokyo University of Agriculture and
Technology, Japan
Yongsheng Ding Donghua University, China
Madalina-M. Drugan Vrije University, The Netherlands
Mark Embrechts RPI, USA
Andries Engelbrecht University of Pretoria, South Africa
Fuhua Fan Electronic Engineering Institute, China
Zhun Fan Technical University of Denmark, Denmark
Komla Folly University of Cape Town, South Africa
Shangce Gao University of Toyama, Japan
Ying Gao Guangzhou University, China
Shenshen Gu Shanghai University, China
Suicheng Gu University of Pittsburgh, USA
Ping Guo Beijing Normal University, China
Haibo He University of Rhode Island, USA
Ran He National Laboratory of Pattern Recognition,
China
Marde Helbig CSIR: Meraka Institute, South Africa
Mo Hongwei Harbin Engineering University, China
Jun Hu Chinese Academy of Sciences, China
Xiaohui Hu Indiana University Purdue University
Indianapolis, USA
Guangbin Huang Nanyang Technological University, Singapore
Amir Hussain University of Stirling, UK
Hisao Ishibuchi Osaka Prefecture University, Japan
Andreas Janecek University of Vienna, Austra
Changan Jiang RIKEN-TRI Collaboration Center for
Human-Interactive Robot Research, Japan
Mingyan Jiang Shandong University, China
Liu Jianhua Fujian University of Technology, China
Colin Johnson University of Kent, USA
Farrukh Khan FAST-NUCES Islamabad, Pakistan
Arun Khosla National Institute of Technology, Jalandhar,
India
X Organization
Franziska Klügl Örebro University, Sweden

Thanatchai
Kulworawanichpong Suranaree University of Technology, Thailand
Germano Lambert-Torres Itajuba Federal University, Brazil
Xiujuan Lei Shaanxi Normal University, China
Bin Li University of Science and Technology of China,
China
Xiaodong Li RMIT University, Australia
Xuelong Li Chinese Academy of Sciences, China
Yangmin Li University of Macau, China
Jane-J. Liang Zhengzhou University, China
Andrei Lihu Politehnica University of Timisoara, Romania
Fernando B. De Lima Neto University of Pernambuco, Brazil
Ju Liu Shandong University, China
Wenlian Lu Fudan University, China
Wenjian Luo University of Science and Technology of China,
China
Jinwen Ma Peking University, China
Chengying Mao Jiangxi University of Finance and Economics,
China
Michalis Mavrovouniotis De Montfort University, UK
Bernd Meyer Monash University, Australia
Martin Middendorf University of Leipzig, Germany
Sanaz Mostaghim Institute IWS, Germany
Jonathan Mwaura University of Pretoria, South Africa
Pietro S. Oliveto University of Sheffield, UK
Feng Pan Beijing Institute of Technology, China
Bijaya Ketan Panigrahi IIT Delhi, India
Sergey Polyakovskiy Ufa State Aviation Technical University, USA
Thomas Potok ORNL, USA
Radu-Emil Precup Politehnica University of Timisoara, Romania
Kai Qin RMIT University, Australia
Quande Qin Shenzhen University, China
Boyang Qu Zhongyuan University of Technology, China
Robert Reynolds Wayne State University, USA
Guangchen Ruan Indiana University Bloomington, USA
Eugene Santos Dartmouth College, USA
Gerald Schaefer Loughborough University, USA
Kevin Seppi Brigham Young University, USA
Zhongzhi Shi Institute of Computing Technology, CAS,
China
Pramod Kumar Singh ABV-IIITM Gwalior, India
Ponnuthurai Suganthan Nanyang Technological University, Singapore
Mohammad Taherdangkoo Shiraz University, Iran
Hideyuki Takagi Kyushu University, Japan
Ying Tan Peking University, China
Organization XI
Ke Tang University of Science and Technology of China,

China
Peter Tino University of Birmingham, UK
Mario Ventresca Purdue University, USA
Cong Wang Northeasten University, China
Guoyin Wang Chongqing University of Posts and
Telecommunications, China
Jiahai Wang Sun Yat-sen University, China
Jun Wang Peking University, China
Lei Wang Tongji University, China
Ling Wang Tsinghua University, China
Lipo Wang Nanyang Technological University, Singapore
Qi Wang Xi’an Institute of Optics and Precision
Mechanics of CAS, China
Zhenzhen Wang Jinling Institute of Technology, China
Man Leung Wong Lingnan University, China
Shunren Xia Zhejiang University, China
Bo Xing University of Johannesburg, South Africa
Ning Xiong Mälardalen University, Sweden
Benlian Xu Changsu Institute of Technology, China
Bing Xue Victoria University of Wellington, New Zealand
Pei Yan University of Aziz, Japan
Yingjie Yang De Montfort University, UK
Wei-Chang Yeh National Tsing Hua University, China
Gary Yen Oklahoma State University, USA
Peng-Yeng Yin National Chi Nan University, Taiwan
Ivan Zelinka FEI VSB-Technical University of Ostrava,
Czech Republic
Zhi-Hui Zhan Sun Yat-sen University, China
Jie Zhang Newcastle University, UK
Jun Zhang Waseda University, Japan
Junqi Zhang Tongji University, China
Lifeng Zhang Renmin University of China, China
Mengjie Zhang Victoria University of Wellington, New Zealand
Qieshi Zhang Waseda University, Japan
Qingfu Zhang University of Essex, UK
Zhenya Zhang Anhui University of Architecture, China
Qiangfu Zhao The University of Aizu, Japan
Wenming Zheng Southeast University, China
Yujun Zheng Zhejiang University of Technology, China
Cui Zhihua Complex System and Computational
Intelligence Laboratory, China
Aimin Zhou East China Normal University, China
Zexuan Zhu Shenzhen University, China
Xingquan Zuo Beijing University of Posts and
Telecommunications, China
XII Organization
Additional Reviewers
Alves, Felipe Rakitianskaia, Anna

Bi, Shuhui Singh, Garima
Chalegre, Marlon Singh, Pramod
Cheng, Shi Wang, Aihui
Dong, Xianguang Wen, Shengjun
Gonzalez-Pardo, Antonio Wenbo, Wan
Haifeng, Sima Wu, Peng
Hu, Weiwei Wu, Zhigang
Jun, Bo Xiao, Xiao
Lacerda, Marcelo Xin, Cheng
Lee, Jie Yang, Wankou
Li, Yexing Yang, Zhixiang
Ling, Haifeng Yu, Czyujian
Menéndez, Héctor Zheng, Zhongyang
Pei, Yan
Table of Contents – Part II
Classification Methods
Semi-supervised Ant Evolutionary Classification . . . . . . . . . . . . . . . . . . . . . 1
Ping He, Xiaohua Xu, Lin Lu, Heng Qian, Wei Zhang, and
Kanwen Li
Evolutionary Ensemble Model for Breast Cancer Classification . . . . . . . . . 8
R.R. Janghel, Anupam Shukla, Sanjeev Sharma, and
A.V. Gnaneswar
Empirical Analysis of Assessments Metrics for Multi-class Imbalance
Learning on the Back-Propagation Context . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Juan Pablo Sánchez-Crisostomo, Roberto Alejo,
Erika López-González, Rosa Marı́a Valdovinos, and
J. Horacio Pacheco-Sánchez
A Novel Rough Set Reduct Algorithm to Feature Selection Based on
Artificial Fish Swarm Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
Fei Wang, Jiao Xu, and Lian Li
Hand Gesture Shape Descriptor Based on Energy-Ratio and Normalized
Fourier Transform Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Wenjun Tan, Zijiang Bian, Jinzhu Yang, Huang Geng,
Zhaoxuan Gong, and Dazhe Zhao
A New Evolutionary Support Vector Machine with Application to
Parkinson’s Disease Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
Yao-Wei Fu, Hui-Ling Chen, Su-Jie Chen, LiMing Shen, and
QiuQuan Li
GPU-Based Methods
Parallel Bees Swarm Optimization for Association Rules Mining Using
GPU Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Youcef Djenouri and Habiba Drias
A Method for Ripple Simulation Based on GPU . . . . . . . . . . . . . . . . . . . . . 58
Xianjun Chen, Yanmei Wang, and Yongsong Zhan
cuROB: A GPU-Based Test Suit for Real-Parameter Optimization . . . . . 66
Ke Ding and Ying Tan
XIV Table of Contents – Part II
Scheduling and Path Planning

A Particle Swarm Optimization Based Pareto Optimal Task Scheduling
in Cloud Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
A.S. Ajeena Beegom and M.S. Rajasree
Development on Harmony Search Hyper-heuristic Framework for

Examination Timetabling Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Khairul Anwar, Ahamad Tajudin Khader,
Mohammed Azmi Al-Betar, and Mohammed A. Awadallah
Predator-Prey Pigeon-Inspired Optimization for UAV

Three-Dimensional Path Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
Bo Zhang and Haibin Duan
Research on Route Obstacle Avoidance Task Planning Based on

Differential Evolution Algorithm for AUV . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Jian-Jun Li, Ru-Bo Zhang, and Yu Yang
Wireless Sensor Network

An Improved Particle Swarm Optimization-Based Coverage Control
Method for Wireless Sensor Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Huimin Du, Qingjian Ni, Qianqian Pan, Yiyun Yao, and Qing Lv
An Improved Energy-Aware Cluster Heads Selection Method for

Wireless Sensor Networks Based on K-means and Binary Particle
Swarm Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Qianqian Pan, Qingjian Ni, Huimin Du, Yiyun Yao, and Qing Lv
Power System Optimization

Comparison of Multi-population PBIL and Adaptive Learning Rate
PBIL in Designing Power System Controller . . . . . . . . . . . . . . . . . . . . . . . . . 135
Komla A. Folly
Vibration Adaptive Anomaly Detection of Hydropower Unit in Variable

Condition Based on Moving Least Square Response Surface . . . . . . . . . . . 146
Xueli An and Luoping Pan
Capacity and Power Optimization for Collaborative Beamforming with

Two Relay Clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Bingbing Lu, Ju Liu, Chao Wang, Hongji Xu, and Qing Wang
Solving Power Economic Dispatch Problem Subject to DG Uncertainty

via Bare-Bones PSO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
Yue Jiang, Qi Kang, Lei Wang, and Qidi Wu
Table of Contents – Part II XV
Other Applications
Extracting Mathematical Components Directly from PDF Documents
for Mathematical Expression Recognition and Retrieval . . . . . . . . . . . . . . . 170
Botao Yu, Xuedong Tian, and Wenjie Luo
An Efficient OLAP Query Algorithm Based on Dimension Hierarchical

Encoding Storage and Shark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Shengqiang Yao and Jieyue He
The Enhancement and Application of Collaborative Filtering in

e-Learning System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
Bo Song and Jie Gao
A Method to Construct a Chinese-Swedish Dictionary via English

Based on Comparable Corpora . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
Fang Li, Guangda Shi, and Yawei Lv
The Design and Implementation of the Random HTML Tags and

Attributes-Based XSS Defence System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
Heng Lin, Yiwen Yan, Hongfei Cai, and Wei Zhang
Special Session on Swarm Intelligence in Image and

Video Processing
DWT and GA-PSO Based Novel Watermarking for Videos Using Audio
Watermark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
Puja Agrawal and Aleefia Khurshid
Application and Comparison of Three Intelligent Algorithms in 2D

Otsu Segmentation Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Lianlian Cao, Sheng Ding, Xiaowei Fu, and Li Chen
A Shape Target Detection and Tracking Algorithm Based on the Target

Measurement Intensity Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
Weifeng Liu, Chenglin Wen, and Shuyu Ding
Multi-cell Contour Estimate Based on Ant Pheromone Intensity

Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
Qinglan Chen, Benlian Xu, Yayun Ren, Mingli Lu, and Peiyi Zhu
A Novel Ant System with Multiple Tasks for Spatially Adjacent Cell
State Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
Mingli Lu, Benlian Xu, Peiyi Zhu, and Jian Shi
A Cluster Based Method for Cell Segmentation . . . . . . . . . . . . . . . . . . . . . . 253

Fei Wang, Benlian Xu, and Mingli Lu
XVI Table of Contents – Part II
Research on Lane Departure Decision Warning Methods Based on

Machine Vision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
Chuncheng Ma, Puheng Xue, and Wanping Wang
Searching Images in a Textile Image Database . . . . . . . . . . . . . . . . . . . . . . . 267

Yin-Fu Huang and Sheng-Min Lin
IIS: Implicit Image Steganography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275

K. Jithesh and P. Babu Anto
Humanized Game Design Based on Augmented Reality . . . . . . . . . . . . . . . 284

Yanhui Su, Shuai Li, and Yongsong Zhan
Special Session on Applications of Swarm Intelligence

to Management Problems
A Hybrid PSO-DE Algorithm for Smart Home Energy Management . . . . 292
Yantai Huang, Lei Wang, and Qidi Wu
A Multiobjective Large Neighborhood Search for a Vehicle Routing

Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
Liangjun Ke and Laipeng Zhai
A Self-adaptive Interior Penalty Based Differential Evolution Algorithm

for Constrained Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
Cui Chenggang, Yang Xiaofei, and Gao Tingyu
A Novel Hybrid Algorithm for Mean-CVaR Portfolio Selection with

Real-World Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
Quande Qin, Li Li, and Shi Cheng
A Modified Multi-Objective Optimization Based on Brain Storm

Optimization Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
Lixia Xie and Yali Wu
Modified Brain Storm Optimization Algorithm for Multimodal

Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
Xiaoping Guo, Yali Wu, and Lixia Xie
Special Session on Swarm Intelligence for Real-World

Application
Classification of Electroencephalogram Signals Using Wavelet
Transform and Particle Swarm Optimization . . . . . . . . . . . . . . . . . . . . . . . . 352
Nasser Omer Ba-Karait, Siti Mariyam Shamsuddin, and
Rubita Sudirman
Table of Contents – Part II XVII
FOREX Rate Prediction Using Chaos, Neural Network and Particle

Swarm Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
Dadabada Pradeepkumar and Vadlamani Ravi
Path Planning Using Neighborhood Based Crowding Differential

Evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
Boyang Qu, Yanping Xu, Dongyun Wang, Hui Song, and
Zhigang Shang
Neural Network Based on Dynamic Multi-swarm Particle Swarm

Optimizer for Ultra-Short-Term Load Forecasting . . . . . . . . . . . . . . . . . . . . 384
Jane Jing Liang, Hui Song, Boyang Qu, Wei Liu, and Alex Kai Qin
Dynamic Differential Evolution for Emergency Evacuation

Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
Shuzhen Wan
Centralized Charging Strategies of Plug-in Electric Vehicles on Spot

Pricing Based on a Hybrid PSO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
Jiabao Wang, Qi Kang, Hongjun Tian, Lei Wang, and Qidi Wu
A New Multi-region Modified Wind Driven Optimization Algorithm

with Collision Avoidance for Dynamic Environments . . . . . . . . . . . . . . . . . 412
Abdennour Boulesnane and Souham Meshoul
Special Session on ICSI 2014 Competition on Single

Objective Optimization
Evaluating a Hybrid DE and BBO with Self Adaptation on ICSI 2014
Benchmark Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422
Yu-Jun Zheng and Xiao-Bei Wu
The Multiple Population Co-evolution PSO Algorithm . . . . . . . . . . . . . . . . 434

Xuan Xiao and Qianqian Zhang
Fireworks Algorithm and Its Variants for Solving ICSI2014 Competition

Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
Shaoqiu Zheng, Lang Liu, Chao Yu, Junzhi Li, and Ying Tan
Performance of Migrating Birds Optimization Algorithm on Continuous

Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
Ali Fuat Alkaya, Ramazan Algin, Yusuf Sahin,
Mustafa Agaoglu, and Vural Aksakalli
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461

Table of Contents – Part I
Novel Swarm-Based Search Methods

Comparison of Different Cue-Based Swarm Aggregation Strategies . . . . . 1
Farshad Arvin, Ali Emre Turgut, Nicola Bellotto, and Shigang Yue
PHuNAC Model: Emergence of Crowd’s Swarm Behavior . . . . . . . . . . . . . 9
Olfa Beltaief, Sameh El Hadouaj, and Khaled Ghedira
A Unique Search Model for Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
A.S. Xie
Improve the 3-flip Neighborhood Local Search by Random Flat Move
for the Set Covering Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Chao Gao, Thomas Weise, and Jinlong Li
The Threat-Evading Actions of Animal Swarms without Active Defense
Abilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
Qiang Sun, XiaoLong Liang, ZhongHai Yin, and YaLi Wang
Approximate Muscle Guided Beam Search for Three-Index Assignment
Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
He Jiang, Shuwei Zhang, Zhilei Ren, Xiaochen Lai, and Yong Piao
Novel Optimization Algorithm

Improving Enhanced Fireworks Algorithm with New Gaussian
Explosion and Population Selection Strategies . . . . . . . . . . . . . . . . . . . . . . . 53
Bei Zhang, Minxia Zhang, and Yu-Jun Zheng
A Unified Matrix-Based Stochastic Optimization Algorithm . . . . . . . . . . . 64
Xinchao Zhao and Junling Hao
Chaotic Fruit Fly Optimization Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Xiujuan Lei, Mingyu Du, Jin Xu, and Ying Tan
A New Bio-inspired Algorithm: Chicken Swarm Optimization . . . . . . . . . . 86
Xianbing Meng, Yu Liu, Xiaozhi Gao, and Hengzhen Zhang
A Population-Based Extremal Optimization Algorithm with
Knowledge-Based Mutation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Junfeng Chen, Yingjuan Xie, and Hua Chen
A New Magnetotactic Bacteria Optimization Algorithm Based on
Moment Migration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Hongwei Mo, Lili Liu, and Mengjiao Geng
XX Table of Contents – Part I
A Magnetotactic Bacteria Algorithm Based on Power Spectrum for

Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Hongwei Mo, Lili Liu, and Mengjiao Geng
Particle Swarm Optimization

A Proposal of PSO Particles’ Initialization for Costly Unconstrained
Optimization Problems: ORTHOinit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
Matteo Diez, Andrea Serani, Cecilia Leotardi, Emilio F. Campana,
Daniele Peri, Umberto Iemma, Giovanni Fasano, and Silvio Giove
An Adaptive Particle Swarm Optimization within the Conceptual

Framework of Computational Thinking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
Bin Li, Xiao-lei Liang, and Lin Yang
Topology Optimization of Particle Swarm Optimization . . . . . . . . . . . . . . . 142

Fenglin Li and Jian Guo
Fully Learned Multi-swarm Particle Swarm Optimization . . . . . . . . . . . . . 150

Ben Niu, Huali Huang, Bin Ye, Lijing Tan, and Jane Jing Liang
Using Swarm Intelligence to Search for Circulant Partial Hadamard

Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Frederick Kin Hing Phoa, Yuan-Lung Lin, and Tai-Chi Wang
Ant Colony Optimization for Travelling Salesman

Problem
High Performance Ant Colony Optimizer (HPACO) for Travelling
Salesman Problem (TSP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Sudip Kumar Sahana and Aruna Jain
A Novel Physarum-Based Ant Colony System for Solving the

Real-World Traveling Salesman Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
Yuxiao Lu, Yuxin Liu, Chao Gao, Li Tao, and Zili Zhang
Three New Heuristic Strategies for Solving Travelling Salesman

Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Yong Xia, Changhe Li, and Sanyou Zeng
Artificial Bee Colony Algorithms

A 2-level Approach for the Set Covering Problem: Parameter Tuning of
Artificial Bee Colony Algorithm by Using Genetic Algorithm . . . . . . . . . . 189
Broderick Crawford, Ricardo Soto, Wenceslao Palma,
Franklin Johnson, Fernando Paredes, and Eduardo Olguı́n
Table of Contents – Part I XXI
Hybrid Guided Artificial Bee Colony Algorithm for Numerical Function

Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
Habib Shah, Tutut Herawan, Rashid Naseem, and Rozaida Ghazali
Classification of DNA Microarrays Using Artificial Bee Colony (ABC)

Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Beatriz Aurora Garro, Roberto Antonio Vazquez, and
Katya Rodrı́guez
Crowding-Distance-Based Multiobjective Artificial Bee Colony

Algorithm for PID Parameter Optimization . . . . . . . . . . . . . . . . . . . . . . . . . 215
Xia Zhou, Jiong Shen, and Yiguo Li
Artificial Immune System

An Adaptive Concentration Selection Model for Spam Detection . . . . . . . 223
Yang Gao, Guyue Mi, and Ying Tan
Control of Permanent Magnet Synchronous Motor Based on Immune

Network Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
Hongwei Mo and Lifang Xu
Adaptive Immune-Genetic Algorithm for Fuzzy Job Shop Scheduling

Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
Beibei Chen, Shangce Gao, Shuaiqun Wang, and Aorigele Bao
Evolutionary Algorithms
A Very Fast Convergent Evolutionary Algorithm for Satisfactory
Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
Xinchao Zhao and Xingquan Zuo
A Novel Quantum Evolutionary Algorithm Based on Dynamic

Neighborhood Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
Feng Qi, Qianqian Feng, Xiyu Liu, and Yinghong Ma
Co-evolutionary Gene Expression Programming and Its Application in

Wheat Aphid Population Forecast Modelling . . . . . . . . . . . . . . . . . . . . . . . . 275
Chaoxue Wang, Chunsen Ma, Xing Zhang, Kai Zhang, and
Wumei Zhu
Neural Networks and Fuzzy Methods

Neural Network Intelligent Learning Algorithm for Inter-related Energy
Products Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
Haruna Chiroma, Sameem Abdul-Kareem, Sanah Abdullahi Muaz,
Abdullah Khan, Eka Novita Sari, and Tutut Herawan
XXII Table of Contents – Part I
Data-Based State Forecast via Multivariate Grey RBF Neural Network

Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
Yejun Guo, Qi Kang, Lei Wang, and Qidi Wu
Evolving Flexible Neural Tree Model for Portland Cement Hydration

Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
Zhi-feng Liang, Bo Yang, Lin Wang, Xiaoqian Zhang,
Lei Zhang, and Nana He
Hybrid Self-configuring Evolutionary Algorithm for Automated Design

of Fuzzy Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
Maria Semenkina and Eugene Semenkin
The Autonomous Suspending Control Method for Underwater

Unmanned Vehicle Based on Amendment of Fuzzy Control Rules . . . . . . 318
Pengfei Peng, Zhigang Chen, and Xiongwei Ren
How an Adaptive Learning Rate Benefits Neuro-Fuzzy Reinforcement

Learning Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
Takashi Kuremoto, Masanao Obayashi, Kunikazu Kobayashi, and
Shingo Mabu
Hybrid Methods
Comparison of Applying Centroidal Voronoi Tessellations and
Levenberg-Marquardt on Hybrid SP-QPSO Algorithm for High
Dimensional Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
Ghazaleh Taherzadeh and Chu Kiong Loo
A Hybrid Extreme Learning Machine Approach for Early Diagnosis of

Parkinson’s Disease . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
Yao-Wei Fu, Hui-Ling Chen, Su-Jie Chen, Li-Juan Li,
Shan-Shan Huang, and Zhen-Nao Cai
A Hybrid Approach for Cancer Classification Based on Particle Swarm

Optimization and Prior Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
Fei Han, Ya-Qi Wu, and Yu Cui
Multi-objective Optimization
Grover Algorithm for Multi-objective Searching with Iteration
Auto-controlling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
Wanning Zhu, Hanwu Chen, Zhihao Liu, and Xilin Xue
Pareto Partial Dominance on Two Selected Objectives MOEA on

Many-Objective 0/1 Knapsack Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Jinlong Li and Mingying Yan
Table of Contents – Part I XXIII
Analysis on a Multi-objective Binary Disperse Bacterial Colony

Chemotaxis Algorithm and Its Convergence . . . . . . . . . . . . . . . . . . . . . . . . . 374
Tao Feng, Zhaozheng Liu, and Zhigang Lu
Multi-objective PSO Algorithm for Feature Selection Problems with
Unreliable Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
Yong Zhang, Changhong Xia, Dunwei Gong, and Xiaoyan Sun
Convergence Enhanced Multi-objective Particle Swarm Optimization
with Introduction of Quorum-Sensing Inspired Turbulence . . . . . . . . . . . . 394
Shan Cheng, Min-You Chen, and Gang Hu
Multiobjective Genetic Method for Community Discovery in Complex
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404
Bingyu Liu, Cuirong Wang, and Cong Wang
A Multi-objective Jumping Particle Swarm Optimization Algorithm for
the Multicast Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
Ying Xu and Huanlai Xing
Multi-agent Systems
A Physarum-Inspired Multi-Agent System to Solve Maze . . . . . . . . . . . . . 424
Yuxin Liu, Chao Gao, Yuheng Wu, Li Tao, Yuxiao Lu, and
Zili Zhang
Consensus of Single-Integrator Multi-Agent Systems at a Preset
Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
Cong Liu, Qiang Zhou, and Yabin Liu
Representation of the Environment and Dynamic Perception in
Agent-Based Software Evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
Qingshan Li, Hua Chu, Lihang Zhang, and Liang Diao
Evolutionary Clustering Algorithms

Cooperative Parallel Multi Swarm Model for Clustering in Gene
Expression Profiling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
Zakaria Benmounah, Souham Meshoul, and Mohamed Batouche
Self-aggregation and Eccentricity Analysis: New Tools to Enhance
Clustering Performance via Swarm Intelligence . . . . . . . . . . . . . . . . . . . . . . 460
Jiangshao Gu and Kunmei Wen
DNA Computation Based Clustering Algorithm . . . . . . . . . . . . . . . . . . . . . . 470
Zhenhua Kang, Xiyu Liu, and Jie Xue
Clustering Using Improved Cuckoo Search Algorithm . . . . . . . . . . . . . . . . . 479
Jie Zhao, Xiujuan Lei, Zhenqiang Wu, and Ying Tan
XXIV Table of Contents – Part I
Sample Index Based Encoding for Clustering Using Evolutionary

Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489
Xiang Yang and Ying Tan
Data Mining Tools Design with Co-operation of Biology Related

Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
Shakhnaz Akhmedova and Eugene Semenkin
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507

Semi-supervised Ant Evolutionary Classification
Ping He , Xiaohua Xu , Lin Lu, Heng Qian, Wei Zhang, and Kanwen Li
Department of Computer Science, Yangzhou University, Yangzhou 225009, China

{angeletx,arterx,linlu60}@gmail.com
Abstract. In this paper, we propose an ant evolutionary classification model,

which treats different classes as ant colonies to classify the unlabeled instances. In
our model, each ant colony sends its members to propagate its unique pheromone
on the unlabeled instances. The unlabeled instances are treated as unlabeled ants.
They are assigned to different ant colonies according to the pheromone that dif-
ferent colonies leave on it. Next, the natural selection is carried out to maintain
the history colony information as well as the scale of swarms. Theoretical anal-
ysis and experimental results show the effectiveness of our proposed model for
evolutionary data classification.
Keywords: Evolutionary classification, Ant colony, K-nearest neighbor.
1 Introduction
Evolutionary data comes from many application fields, such as topics in weblogs and
locations in GPS sensors. Evolutionary data mining can be classified into the two
categories, evolutionary clustering and evolutionary classification. Among them, evo-
lutionary classification refers to the situation where some instances in the data flow
are attached with known labels, and the target is to classify the unlabeled data in the
real-time.
Various evolutionary classification methods [3]–[13] have been proposed from dif-
ferent aspects, including concept drifts, class distribution and temporal smoothness.
However, the assumption of entire labeled data availability is often violated in the real-
world problems, because labels may be scare or not readily available. As a result, semi-
supervised evolutionary learning methods have been recently put forward. Yangging
Jia et al. [14] proposed a semi-supervised classification algorithm for dynamic mail
post categorization. They carried out temporal smoothness assumption using temporal
regularizers defined in the Hilbert space, and then derived the online algorithm that
efficiently finds the closed-form solution to the target function. Later, H. Borchani et
al. [15] proposed a new semi-supervised learning approach for concept-drifting data
streams. They aim to take advantage of unlabeled data to detect possible concept drifts
and, if necessary, update the classifier over time even if only a few labeled data are
available. However, both the previous works assumes that at any time stamp, at least
one labeled instance for each class should be provided, which can be easily violated in
the real-world applications.

Corresponding author.
Y. Tan et al. (Eds.): ICSI 2014, Part II, LNCS 8795, pp. 1–7, 2014.
c Springer International Publishing Switzerland 2014
2 P. He et al.
In this paper, we propose a semi-supervised ant evolutionary classification model,

which only require users to specify the number of labels and provide at least one labeled
sample in the beginning. In our work, we treat each data instance as an ant and each
class of labeled instances as an ant colony. The whole swarm, i.e., the whole dataset,
is composed of all the different colonies and the unlabeled ants. They evolve with time
based on the simulation of natural selection. Therefore, our proposed algorithm is ’self-
training’ in nature. Compared to the previous research, our method can be applied to a
more generalized scenario, where the class distribution is arbitrary and the number of
labeled instances is unfixed (even down to 0) at each time step.
The rest of this paper is organized as follows: Section 2 describes our ant evolution-
ary classification model in detail. Section 3 presents some simulation results to demon-
strate its classification performance. Finally, Section 4 concludes the paper.
2 Ant Evolutionary Classification Model

In semi-supervised evolutionary classification, each data is associated with not only
a label y but also a time stamp t ∈ {1, . . . , T }. Given a set of data subsetsX =
{X 1 , X 2 , . . . , X T }, where X t represents the data at time step t, X t = Xm t
Xut ,
t
Xm = (xti )i=1...|Xm t | is labeled, the corresponding label subset is Y
t
m = (y t
) t |,
i i=1...|Xm
and Xut = (xti )|Xm t
t |+1...|X t | is unlabeled, the goal is to predict the label of X , i.e.,
u
Yut = (yit )|Xm t |+1...|X t | , in the real time.
To solve this problem, we propose an Ant Evolutionary Classification (AEC) model.

It treats each class as an ant colony, respectively denoted as Al=1...c , where c is the
number of classes or labels. Particularly, we let the unlabeled dataset Xu form a special
colony with unknown class, represented by A0 = Xu . The lth colony at time step t
is denoted as Atl . The ith member of Atl is denoted by atli , atli is labeled if l > 0 and
unlabeled otherwise. Therefore, the swarm at time t is composed by c + 1 ant colonies,
i.e., At = {At0 , At1 , . . . , Atc }.
We assume that each ant colony possesses a unique pheromone. In order to expand
the territory, each colony has to recruit new ants by spreading its pheromone onto the
unlabeled ones. The new members joining the ant colony Atl>0 is composed of two
groups: 1) the labeled data provided at time step t, 2) the unlabeled data assigned to
Atl>0 at the time step t.
Instead of recording the pheromone left by each ant individual, we record the phero-
mone left each ant colony. We define the pheromone matrix at time step t as a |X t | × c
matrix τ t . Each column of τ t records the pheromone left by one colony on all the ants.
t
The element τ(j+|X indicates the pheromone left by ants from colony Atl on the ant
m |)l
t
a0j at time step t. The matrix τ t is divided into two blocks. The first block with size
|Xm t
| × c records the pheromone left on the labeled ants, the second block with size
|Xut | × c records the pheromone left on the unlabeled ants. In this paper, we fix the
first block of τ t unchanged and only update its second block, whose element is τij . The
initial value of τ at time step 0 is set τij0 = 1 if and only if yi belongs to the j th class,
otherwise τij0 = 0.
Semi-supervised Ant Evolutionary Classification 3
Since labeled and unlabeled data is provided at each time, the pheromone matrix
needs to be updated accordingly. We define τ ts as the pheromone matrix in the sth
iteration of the pheromone update at time step t. Without loss of generality, given an
ant at0i and a nest Atl (l > 0), the updated pheromone intensity on at0i is
|Xu
t
| |Xm
t
|
ts+1

τ(i+|X ← t
ηi(j+|X τ ts t
m |) (j+|Xm |)l
+ t t0
ηik τkl (1)
m |)l
t t
j=1 k=1
η t = (ηij
t
)|Xut |×|X t | is the heuristic value matrix (or similarty matrix),
= e−dij
t
t
ηij (2)
dtij th th
is the distance between the i and j ants at time step t,
⎧ t t −1 t
⎨(ali − a0j ) (Σl ) (ali − a0j ) if l > 0
⎪ t T t
||at0i −at0j ||2

dtij = if at0i , at0j ∈ At0 (3)
⎪
⎩ 2σ2
∞ otherwise
Σlt is the covariance matrix of the lth colony at time step t, and σ is a spread parameter.
Note that we define the distance between labeled and unlabeled ants as Mahalanobis
distance so as to utilize the prior class distribution of labeled data, and define the dis-
tance among unlabeled ants as Euclidean distance due to the lack of class information.
Similar to the partition of τ t , we also divide η t into two parts. The first block with size
|Xut | × |Xm t
| records the similarity between unlabeled ants and labeled ants, and the
second block with size |Xut | × |Xut | records the similarity among unlabeled ants.
To interprete eq. (1), we view as two parts, corresponding to the two blocks of η t
and τ t . In the first term, ηi(j+|X
t
represents the similarity between the ith and j th
m |)
t
ts
unlabeled ant, τ(j+|X is the pheromone on the j th unlabeled ant left by the lth
m |)l
t
colony in the sth iteration at time step t. Therefore, the first term computes the sum of
pheromone indirectly propagated from the labeled ants via the unlabeled ants. In the
t
second term, ηik is the similarity between the ith ant and the k th labeled ant, τkl 0
is
the initial pheromone on the k th labeled ant. Hence the second term computes the sum
t0
of the pheromone directly propagated from the labeled ants. The reason for using τkl
ts
instead of τkl is because we keep the pheromone on the labeled ants unchanged to avoid
concept drifting.
After the convergence of the pheromone matrix τ t = τ t∞ , we predict the label of
each unlabeled ant at0i according to the amount of pheromone that different ant colonies
leave on it.
yit = arg max τilt (4)
l
To determine whether an unlabeled data should be included in its predicted ant
colony, we need to further evaluate its fitness to the colony. Given a colony Atl at time
step t, we define the fitness of atli ∈ Atl as
1
(Σlt )−1 (atli −atlj )
e−(ali −alj )
t t T
f itness(atli ) =
|Atl | (5)
atlj ∈Atl ,i=j
4 P. He et al.
where |Atl | is the size of Atl , Σlt is the covariance matrix of Atl . Based on the fitness
evaluation, the evolution of ant colonies are composed of two steps. 1) Member Addi-
tion. For each unlabeled ant, if its fitness to its predicted class is higher than a thresh-
old β ∈ (0, 1], then it will be included in the target colony, used as the training set
for the label prediction at next time step. 2) Member Deletion. To avoid class im-
balance and allow member change, we set a maximum for the size of an ant colony
(M axColonySize). Once this maximum is reached, the members with the lowest fit-
ness in that colony will be removed.
3 Experiments
We test our algorithm on three datasets, whose details are summarized in Table 1.
Twomoons is a synthetic dataset including two classes of intertwining moons. Mush-
room and Hyperplane datasets from the UCI repository are used to simulate the concept
drift problem.
Table 1. Summarization of the test datasets
Dataset #Size #Attributes #Classes

Twomoons 2000 2 2
Mushroom 8124 22 2
Hyperplane 10000 10 2
3.1 Synthetic Dataset

We first test our algorithm on the synthetic dataset Twomoons dataset for illustration. At
first, the dataset is divided into T = 100 time intervals. Fig. 1 shows the the evolutionary
classification process in five ascending time steps. The red crosses and blue circles
respectively denote the two different classes of data, and the black dots represent the
unlabeled instance. The subfigures at the left side depict the input data including the
previous classification result, while the subfigures at the right side depict the predicted
labels of those black dots in the left subfigures. As we can see, when t = 1, only two
labeled instances are provided and we cannot recognize the intrinsic structure of the
input data. Later, after more instances are provided, our algorithm gradually assigns
labels to the unlabeled data points and then discovers the manifold structure of two
moons.
3.2 Real-World Dataset

We use the 1% labeled ratio to generate the training data and randomly distribute them
into T = 100 time blocks. Therefore, the provided labeled data may vary with different
times blocks, and even maybe absent. For the evaluation of evolutionary classification
1.5 1.5
0 1
1 2
2
1 1
0.5 0.5
0 0
−0.5 −0.5
−1 −1
−0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
1
a. Input dataset X (left) and the classification result (right) at t = 1
1.5 1.5
0 1
1 2
2
1 1
0.5 0.5
0 0
−0.5 −0.5
−1 −1
−1.5 −1 −0.5 0 0.5 1 1.5 2 −1.5 −1 −0.5 0 0.5 1 1.5 2
5
b. Input dataset X (left) and the classification result (right) at t = 5
1.5 1.5
0 1
1 2
2
1 1
0.5 0.5
0 0
−0.5 −0.5
−1 −1
−1.5 −1 −0.5 0 0.5 1 1.5 2 2.5 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
30
c. Input dataset X (left) and the classification result (right) at t = 30
1.5 1.5
0 1
1 2
2
1 1
0.5 0.5
0 0
−0.5 −0.5
−1 −1
−1.5 −1 −0.5 0 0.5 1 1.5 2 2.5 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
50
d. Input dataset X (left) and the classification result (right) at t = 50
1.5 1.5
0 1
1 2
2
1 1
0.5 0.5
0 0
−0.5 −0.5
−1 −1
−1.5 −1 −0.5 0 0.5 1 1.5 2 2.5 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
100
e. Input dataset X (left) and the classification result (right) at t = 100
Fig. 1. Illustration of AEC Model on Twomoons dataset

6 P. He et al.

(a) Mushroom (b) Hyperplane
Fig. 2. Average block accuracy with different M axColonySize on two real-world datasets
Table 2. Average overall accuracy with different M axColonySize
50 100 150 200 500

Mushroom 92.76 95.01 96.26 96.7 96.6
Hyperplane 74.41 75.50 75.65 74.86 75.19
performance, we adopt both overall classification accuracy and local classification ac-
curacy, which refers to the classification accuracy within each time block. To give a
reliable result, 50 runs of random simulation are carried out to produce an average
overall classification accuracy.
In our algorithm, we set a ceiling for the size of ant colonies, i.e., M axColonySize.
We adopt five values (50, 100, 150, 200, 500) as the max colony size. Fig. 2 illustrates
the relationship between the block accuracy and parameter M axColonySize on the
two real-world datasets. We can see that M axColonySize exerts obvious influence on
Mushroom dataset.
In Table 2, each row shows the average overall accuracy on one dataset with five dif-
ferent M axColonySize values. We note that Twomoons and Hyperplane datasets per-
form best at size 150. It indicates that 150 might be a good choice for M axColonySize.
In addition, the setting of this parameter should also take into account of the physical
memory and the runtime cost.
4 Conclusion
In this paper, we present an ant classification model for dynamic semi-supervised clas-
sification. It simulates a swarm containing varied ant colonies that will evolve with time
under the rule of natural selection. Meanwhile, each generation of unlabeled instances
are classified into these colonies using our proposed swarm classification method. Ex-
perimental results on a synthetic dataset demonstrate the effectiveness of our method. In
the future work, we will investigate AEC and compare it with other classifiers on real-
world datasets. Another interesting future line of research is to consider the scenario
where labeled and unlabeled data possibly come from different distributions.
Acknowledgment. This work was supported in part by the Chinese National Natural
Science Foundation under Grant nos. 61402395, 61003180, 61379066 and 61103018,
Natural Science Foundation of Education Department of Jiangsu Province under con-
tracts 13KJB520026 and 09KJB20013, Natural Science Foundation of Jiangsu Province
under contracts BK2010318 and BK20140492, and the New Century Talent Project of
Yangzhou University.
References
1. Chakrabarti, D., Kumar, R., Tomkins, A.: Evolutionary clustering. In: The ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, pp. 554–560 (2006)
2. Chi, Y., Song, X., Zhou, D., Hino, K., et al.: Evolutionary spectral clustering by incorporat-
ing temporal smoothness. In: The ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, pp. 53–62 (2007)
3. Wang, P., Wang, H.X., Wu, X.C., et al.: A low-granularity classifier for data streams with
concept drifts and biased class distribution. IEEE Transactions on Knowledge and Data En-
gineering 19(9), 1202–1213 (2007)
4. Gao, J., Ding, B.L., Han, J.W., et al.: Classifying Data Streams with Skewed Class Distribu-
tions and Concept Drifts. IEEE Internet Computing 12(6), 37–49 (2008)
5. Anagnostopoulos, C., Tasoulis, D.K., Adams, N.M., et al.: Temporally adaptive estimation
of logistic classifiers on data streams. Advances in Data Analysis and Classification 3(3),
243–261 (2009)
6. Kuncheva, L.I., Zliobaite, I.: On the window size for classification in changing environments.
Intelligent Data Analysis 13(6), 861–872 (2009)
7. Peng, Z., Xingquan, Z., Jianlong, T., et al.: Classifier and Cluster Ensembles for Mining
Concept Drifting Data Streams. In: IEEE 10th International Conference on Data Mining
(ICDM), pp. 1175–1180 (2010)
8. Zhang, P., Li, J., Wang, P., et al.: Enabling fast prediction for ensemble models on data
streams. In: The 17th ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, San Diego, CA, USA, pp. 177–185 (2011)
9. Peng, Z., Gao, B.J., Xingquan, Z., et al.: Enabling Fast Lazy Learning for Data Streams.
In: IEEE 11th International Conference on Data Mining (ICDM), Vancouver, Canada, pp.
932–941 (2011)
10. Zhang, P., Gao, B.J., Liu, P., et al.: A framework for application-driven classification of data
streams. Neurocomput. 92, 170–182 (2012)
11. Lines, J., Davis, L.M., Hills, J., et al.: A shapelet transform for time series classification.
In: The 18th ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, Beijing, China, pp. 289–297 (2012)
12. Masud, M.M., Woolam, C., Gao, J., et al.: Facing the reality of data stream classification:
coping with scarcity of labeled data. Knowledge and Information Systems 33(1), 213–244
(2012)
13. Li, L.J., Zou, B., Hu, Q.H., et al.: Dynamic classifier ensemble using classification confi-
dence. Neurocomput. 99(1), 581–591 (2013)
14. Jia, Y., Yan, S., Zhang, C.: Semi-Supervised Classification on Evolutionary Data. In: The
21st International Joint Conference on Artifical Intelligence, pp. 1083–1088 (2009)
15. Borchani, H., Larranaga, P., Bielza, C.: Classifying evolving data streams with partially la-
beled data. Intelligent Data Analysis 15(5), 655–670 (2011)
Evolutionary Ensemble Model for Breast
Cancer Classification
R.R. Janghel1,*, Anupam Shukla2, Sanjeev Sharma2, and A.V. Gnaneswar2

1
Sagar Institute of Research Technology and Science, Bhopal, India
rrj.iiitm@gmail.com
2
ABV- Indian Institute of Information Technology and Management Gwalior, India
{dranupamshukla,sanjeev.sharma1868,gnani0826}@gmail.com
Abstract. A major problem in medical science is attaining the correct diagnosis

of disease in precedence of its treatment. For the ultimate diagnosis, many tests
are generally involved. Too many tests could complicate the main diagnosis
process so that even the medical experts might have difficulty in obtaining the
end results from those tests. A well-designed computerized diagnosis system
could be used to directly attain the ultimate diagnosis with the aid of artificial
intelligent algorithms and hybrid system which perform roles as classifiers. In
this paper, we describe a Ensemble model which uses MLP, RBF, LVQ models
that could be efficiently solve the above stated problem. The use of the
approach has fast learning time, smaller requirement for storage space during
classification and faster classification with added possibility of incremental
learning. The system was comparatively evaluated using different ensemble
integration methods for breast cancer diagnosis namely weighted averaging,
product, minimum and maximum integration techniques which integrate the
results obtained by modules of ensemble, in this case MLP, RBF and LVQ.
These models run in parallel and results obtained will be integrated to give final
output. The best accuracy, sensitivity and specificity measures are achieved
while using minimum integration technique.
Keywords: Breast Cancer, Medical Diagnostics, Pattern Recognition,

Ensemble Approach, Neural Networks, MLP, Multilayer perceptron, RBF,
Radial Basis Function Network, LVQ, Learning Vector Quantization.
1 Introduction
Many real life applications are so complex that they cannot be solved by the
application of a single algorithm. This necessitated the need for development of
algorithms by mixing two or more of the studied algorithms. The choice of algorithms
depends upon the needs and characteristics of the problem. This further helps in
solving the problem to a reasonably good extent and achieving higher performances.
*
Evolutionary Ensemble Model for Breast Cancer Classification 9
In this paper, we have concentrated our efforts towards solving the problem of breast
cancer diagnosis. Every year in many countries, number of woman died from breast
cancer is increasing. Breast cancer is the most common cancer in women in many
countries in the world. One out of eight women wills diagnosis and prognosis of breast
cancer in this country. Early detection is one of the best defenses against cancer [1].
The database used in analysis of the system has been taken from Wisconsin
Diagnostic Breast Cancer (WDBC) from UCI Machine Learning Repository, which
comprises of data vectors from 569 patients. Then, this data is divided into training
and testing data by taking 398 vectors as training data set (about 70% of the total data
set) and rest as testing data set (about 30% of the total data set).
This paper is organized as follows. Section 2 reviews related work done in the
concerned field. Section 3 gives the methodology used in tackling the problem.
Experimental results are presented in Section 4. Conclusion and future work are given
in the last section.
2 Related Work
A classification system is one that actually maps input vectors to a specific class.
Hence, classification is basically the job of learning the procedure that maps the input
data [2]. This has, in turn, has enthused researchers to replicate this success in the
field of medical diagnostics. Their efforts have bore significant gains through the
application of several standards and techniques of pattern recognition to the said
problem [3]. Also it is the most widespread form of cancer among women in the
world. Early detection is one of the best defenses against cancer. According to the
American Cancer Society (ACS), after every thirteen minutes, four American women
develop breast cancer, and one woman dies from breast cancer [1, 4-6].
Yao and Liu et.al described neural network based approaches to breast cancer
diagnosis, which had displayed good generalization. The approach was based on
artificial neural networks. In this approach, a feed forward neural network was
evolved using BP algorithm [12]. Fogel et al. were first to derive technique to model
neural networks for solving breast cancer classification [13].
Rahul et al. used multilayer perceptron neural networks (MLPNNs), radial basis
function network (RBFN), competitive learning network (CL), learning vector
quantization network(LVQ), combined neural networks (CNNs), probabilistic neural
networks(PNNs), and recurrent neural networks (RNNs) for breast cancer diagnosis [14].
The artificial immune system with the GA in one hybrid algorithm which is the clonal
selection algorithm was inspired from the clonal selection principle and affinity
maturation of the human immune responses by hybridizing it with the crossover operator,
which is imported from GAs to increase the exploration of the search space. [13].
Contrary to neural networks, clustering, rule induction and many other machine
learning approaches, Genetic Algorithms (GAs) provide a means to encode and
evolve rule antecedent aggregation operators, different rule semantics, rule base
aggregation operators and defuzzification methods. Therefore, GAs remain today as
one of the few and, in some sense, optimize fuzzy systems with respect to the design
10 R.R. Janghel et al.
decisions, allowing decision makers to decide what components are fixed and which
ones evolve according to the performance measures [14]. Carlos Andres Pena-Reyes
et.al proposed a fuzzy-genetic approach produces systems exhibiting two prime
characteristics: first they attain high classification performance and second the
resulting systems involve a few simple rules and gave 97.50 % classification accuracy
[15]. The goal of Fuzzy CoCo model was to evolve a fuzzy model that describes the
diagnostic decision and the classification performance was 98.98%. [16].
F A good collection of methods and applications can be found in the books by
Mellin and Castillo [10], and Bunke and Kendel [11, 12].
3 Methodology
Pattern Recognition and Machine Learning field have established research work on
the combination of multiple classifiers (also known as ensemble of classifiers,
Mixture of experts). Overall predictive accuracy can be increased by the use of
multiple classifiers instead of a single classifier. The ensemble procedure constitute
two steps mainly module formation and then integration of results of modules. Firstly
we need to formulate the number of modules to be used, that constitute the entire
ensemble architecture. Decision towards the model and architectural parameters of
each of the module is made. All the networks may be initialized in this mechanism.
Next the entire ensemble needs to be trained, which means the training up of the
individual models making up the ensemble.
Each of the modules is trained independently and in-parallel by all the training data
present in the system.
The ANNs with BPA still have some shortcomings. It is quite likely that BPA
results in some local minima in place of global minima. Also we need to specify the
initial parameters before the learning starts. These pose restrictions on the use of
ANNs. The GA on the other hand is known for its ability of optimization. In this
section we will fuse this capability of the GA along with the ANNs to train the ANN.
This solution overcomes much of the problems with the ANN training.
The block diagram of the proposed system for breast cancer diagnosis is shown in
Figure 1.
Evolutionary
Testing Integration
MLP
Data Techniques
1. Polling Output
Patient Data Evolutionary 2. Maximum
Collection MLP
3. Minimum
Training 4. Product
Data
Evolutionary
MLP
Fig. 1. Block Diagram of the Proposed System for Breast Cancer Diagnosis
In this paper, we have used the ensemble approach for classifying the inputs as
malignant or benign which has 3 modules. Here, each module has evolutionary ANN
and the difference between them is change in hidden neurons.
Here the GA is supposed to fix the values and the various weights as well as biases
that exist in the neural networks. The GA in other words optimizes the network
parameters for better performance. An ANN is a collection of various neurons. These
neurons are arranged in a layered manner. Any ANN model being used in real life
application normally uses a single hidden layer. The hidden layer has a specified
number of neurons in it. In a fully connectionist approach, every neuron of a forward
layer is connected to every neuron of the forward layer i+1 by some weight. The
hidden and output neuron further have weight is adjusted during training. Besides
every neuron has some bias associated with it. Now we would study the application of
GA in this problem for training. Some biases that need to be optimally set.
The first task is problem encoding. The problem encoding consists of these
parameters in a linear array. This is the phenotype problem representation. The
population may be represented using any of double vector or a bit string
representation. The Genetic Operators include Selection, Crossover, Eliticism,
Mutation, etc. The Genetic Operators ensured creation of good individuals from one
population to the other. Let us assume that there was a single hidden layer consisting
of H neurons. The input and output layers have I an O neurons respectively. In this
system, it may easily be seen that there are I x H weights between the input layer and
the hidden layer and H x O weights between the hidden layer and the output layer,
this makes the total number of weights as W=I x H + H x O. Further the number of
biases is equal to the number of neurons. The total number of biases is H + O. This
means that for a single layer ANN there would be I x H + H x O + H + O parameters
to be optimized.
The fitness of any individual in the population is measured with the help of fitness
function. The fitness function consists of the ANN along with its training data set. In
the fitness function we initialize the ANN by the various parameters that are
generated by GA. These parameters were extracted from the individual and used to
set the weights and biases of the ANN. Then the training data set is passed through
the ANN. The performance of the ANN against this data set is measured. This
performance is the net fitness value of the GA that needs to be maximized (or the
negative performance need to be minimized).Hence every time that the GA demands
the measurement of fitness value of some individual, the ANN is created and the
value is measured by the performance. This interfaces the GA and the ANN while
training.
The neural training by GA possesses a very complex fitness landscape. Hence it is
wise to use a local search strategy that places any ANN or genetic individual at the
closest minima, before its fitness value is reported. This local search strategy assists
the GA in the search or optimization process. In this algorithm we use Back
Propagation Algorithm as the local search method. The epochs, momentum, and
learning rate are kept low as per the requirements of local search.
Once the GA reaches its optimal state and terminates as per the stopping criterion,
we get the final values of the weights and the biases. Then we create the ANN with
these weights and bias values and this is regarded as the most optimal ANN as a result
of the ANN training. We can then use this for the testing purposes. It may be seen
here that validation data is not necessarily required in this type of training.
The net fitness may hence be given by equation
Fit (N) = P (N) – α C (N)
Here N is the genetic individual or ANN, α is the penalty constant, Fit() is the
fitness function, P() is the performance function, C() is the number of connections.
Methods for Response Integration

Here, we use different integration schemes including polling, maximum, minimum,
weighed average. These schemes are used to integrate the outputs from each of the
four networks separately and the resulting detection accuracies measured. In polling
scheme, each network returns the class that it considers the one to which the input
belongs. After taking these classes, voting takes place between the network modules.
The class with the highest votes is taken as the winner. In weighted average scheme,
mean of matching scores of all the networks is taken.
The integrator receives all the probability vectors and does the task of deciding the
final output of the system. For this if probability is greater than 0.5 it is marked as
malignant, otherwise as benign.
4 Simulation Results
Wisconsin Diagnostic Breast Cancer (WDBC) database of UCI Machine Learning

Repository is used for the experimentation of our model. The goal is to classify a
tumor as either benign or malignant based on cell descriptions gathered by FNA
image test. The Breast Cancer data set has vectors with a total of 30 input attributes.
This database contains information about 569 patients with 212 out of 569 having
malignant tumors. Attributes used here are radius mean of distances from center to
points on the perimeter, texture means standard deviation of gray-scale values,
smoothness means local variation in radius lengths, perimeter, area, smoothness (local
variation in radius lengths), compactness (perimeter2 / area - 1.0), concavity (severity
of concave portions of the contour), concave points (number of concave portions of
the contour), symmetry and fractal dimension (coastline approximation - 1).They are
measured for a total of 3 cells. The various integration methods are compared with
respect to their ability to train, learn and generalize the data. One with the best
generalizing capacity will give the best detection efficiency.
First we divide the data set into training and testing sets at 70% and 30% by taking
398 vectors as training data set and rest as testing data set. Then data set is used to
train and the test the ensemble model.
The results are measured against the TP (true positive), TN(true negative), FP(false
positive) and FN (false negative). The various performance measures are summarized
in the table 1.
Table 1. Diagnostic performance measures Breast cancer
Cancer Test Present Absent Total

Positive True Positive [ TP] False Positive [FP] [TP +FP ]
Negative False Negative [FN] True negative [TN] [FN+TN ]
( TP + FN+ TN
Total ( TP + FN) (TN + FP)
+ FP )
Sensitivity TP / (TP + FN)

Specificity TN / (TN + FP)
Accuracy (TP + TN) / (TP + TN + FP + FN)
We run the Evolutionary ANN modules to obtain optimum weights for ANN. We
applied GA for the parameter optimization. The weight matrix consisted of 30*x
weights between input and hidden layer, x*1 between the hidden and the output layer
and a total of 18, 20, 25 hidden layer biases and 1 output layer bias. This made the
total number of variables for the GA as 30*x + x*1 + x + 1. We use 18, 20, 25 hidden
neurons for each module respectively.
In GA, the double vector method of population representation was used. The total
number of individuals in the population was 50. A uniform creation function, rank
based scaling function and stochastic uniform selection methods were used. The elite
count was 2. Single point crossover was used. The program was executed till 100
generations. The crossover rate was 0.7, Best fitness is 98.78 and mean fitness is
96.61.The best performance in terms of sensitivity, specificity, accuracy, false
negative and false positive are 98.70% 97.42%, 98.24%, 4.6% and 0.65% for testing
respectively.
Here the results of GA are exported which are optimized weights and bias of ANN
and ANN is run for 10,000 epochs. The result of this is passed through various
integrators.
We then experiment ensemble model with various integration methods to find an
optimized parameter which gives best performance. After getting the optimized
parameter, the detection procedure is run 20 times for the same configuration. After
this, the mean and standard deviation are computed. The mean is taken as the
performance accuracy of the system for training and testing dataset.
The results show that the maximum accuracy was achieved when using maximum
integrator with Accuracy of 99.07% along with sensitivity, specificity, FPR and FNR
values as 98.79, 99.01%, 1.23%, 0.65% respectively. Figure 2 shows the spread of
values of Evolutionary ANN module.
Fig. 2. Performance of Evolutionary ANN module
Table 2. Diagnostic performance of various integration techniques
Integration S (%) Sp (%) A (%) FPR (%) FNR (%)

Techniques
Maximum 98.79 99.01 99.07 1.23 0.65
Minimum 98.46 99.01 98.79 1.53 0.98
Polling 99.00 96.47 97.36 4.9 1.56
Sum 98.70 97.42 98.24 5.0 0.65
Now we compare our model performance with Ensemble architecture with same
ANN model in all the modules. The only difference in the modules is number of
hidden layers. Table 3 is given with the best results of ensemble with same modules
on various integration techniques.
Here we used hidden layer of 18, 20, 25 for the 3 modules of ensemble while
keeping other parameters same as that in our proposed model. Figure 3 shows the
comparative analysis.
Table 3. Diagnostic performance of various integration techniques
ANN Models Integration Testing performances (%)

Techniques S (%) Sp (%) A (%)
MLP Minimum 99.00 96.47 97.36
RBF Maximum 97.01 98.05 97.17
LVQ Polling 93.22 96.04 95.00
Fig. 3. Comparison of Multi Model ANN with Same ANN ensemble with different integrators
From the experimental results we can show that our proposed method performs
better than that of ensemble of ANN with same ANN modules.
5 Conclusion and Future Work
In this paper we saw the working of different ensemble integration methods with three
modules of Evolutionary ANN model using MLP model for detection of Breast
Cancer. In all the cases we were able to solve the problem with fine accuracies using
the different ensemble integration methods. The results show that, Maximum
integration gives the most favorable performance when compared with other
integration methods.
References
1. American Cancer Society, Cancer Facts and Figures (2011-2012)
2. Weiss, S.I., Kulikowski, C.: Computer Systems That Learn: Classification and Prediction
Methods from Statistics, Neural Networks, Machine Learning and Expert Systems.
Morgan Kaufmann Publishers (1991)
3. Coleman, T.F., Li, Y.: Large Scale Numerical optimization. In: Proceedings of Workshop
on Large Scale Numerical Optimization, Cornell University, New York (1989)
4. Andolina, V.F., Lille, S.L., Willison, K.M.: Mammographic Imaging: A Practical Guide,
New York (1992)
5. Antani, S., Lee, D.J., Long, L.R., Thoma, G.R.: Evaluation of shape similarity
measurement methods for spine X-ray images. Journal of Visual Communication & Image
Representation 15, 285–302 (2004)
6. Clemen, R.: Combining forecasts: A review and annotated bibliography. International
Journal of Forecasting, 559–583 (1989)
7. Yao, X., Liu, Y.: Neural networks for breast cancer diagnosis. In: Proceedings of the
Congress on Evolutionary Computation, vol. 3, pp. 1767–1773 (August 2002)
8. Fogel, D.B., Wasson, E.C., Boughton, E.M., Porto, V.W., Angeline, P.J.: Linear and
neural models for classifying breast masses. IEEE Transactions on Medical Imaging 17(3),
485–488 (1998)
9. Rahul, K., Anupam, S., Ritu, T., Janghel, R.R.: Breast cancer diagnostic system using
artificial neural networks model. In: International Conference on Information Sciences and
Interaction Sciences (ICIS), pp. 89–94 (2010)
10. Melin, P., Castillo, O.: Hybrid Intelligent Systems for Pattern Recognition Using Soft
Computing. STUDFUZZ, vol. 172. Springer, Heidelberg (2005)
11. Bunke, H., Kandel, A. (eds.): Hybrid Methods in Pattern Recognition. World Scientific
(2002)
12. Sivanandan, S.N., Deepa, S.N.: Principle of soft computing. Wiley India Private Limited
(2007)
13. Nabil, E., Badr, A., Farag, I.: An Immuno-Genetic Hybrid Algorithm. Int. Journal of
Computers, Communications & Control IV(4), 374–385 (2009)
14. Alcala, R., Nojima, Y.: Special issue on genetic fuzzy systems: new advances.
Evolutionary Intelligence 2, 1–3 (2009)
15. Pena-Reyes, C.A., Sipper, M.: A fuzzy-genetic approach to breast cancer diagnosis.
Artificial Intelligence in Medicine, 131–155 (1999)
16. Peña Reyes, C.A.: Breast Cancer Diagnosis by Fuzzy CoCo. In: Peña Reyes, C.A. (ed.)
Coevolutionary Fuzzy Modeling. LNCS, vol. 3204, pp. 71–87. Springer, Heidelberg
(2004)
Empirical Analysis of Assessments Metrics
for Multi-class Imbalance Learning
on the Back-Propagation Context
Juan Pablo Sánchez-Crisostomo1, Roberto Alejo1 , Erika López-González1,

Rosa Marı́a Valdovinos2, and J. Horacio Pacheco-Sánchez3
1
Tecnológico de Estudios Superiores de Jocotitlán, Carretera Toluca-Atlacomulco KM. 44.8,
Col. Ejido de San Juan y San Agustı́n, 50700 Jocotitlán, México
2
Faculty of Engineering, Universidad Autónoma del Estado de México Cerro de Coatepec s/n,
Ciudad Universitaria C.P. 50100, Toluca, Estado de México
3
Instituto Tecnológico de Toluca
Av. Tecnológico s/n Ex-Rancho La Virgen, 52140, Metepec, México
Abstract. In this paper we study some of the most common assessment metrics
employed to measure the classifier performance on the multi-class imbalanced
problems. The goal of this paper is empirically analyzing the behavior of these
metrics on scenarios where the dataset contains multiple minority and multiple
majority classes. The experimental results presented in this paper indicate that the
studied metrics might be not appropriate in situations where multiple minority
and multiple majority classes exist.
Keywords: Metrics, Multi-class Imbalance, Multiple Minority and Majority

Classes.
1 Introduction
Class imbalance problems have drawn growing interest recently because of their classi-
fication difficulty caused by the imbalanced class distributions [8]. So, it has been into
the 10 challenging problems identified in data mining research [9]. The class imbalance
problem appears when in a training data set the number of instances in at least one class
is much less than the samples in another class or classes [6]. Much work has been done
in addressing the class imbalance problem [4] but, while two-class imbalance prob-
lem has been widely studied the multi-class imbalance problem has been relatively less
investigated [8].
The multi-class imbalance problems pose new challenges, for instance, in task of
assessments classifier performance it is necessary to apply different metrics than those
used in traditional two-class classification problems [7]. On two-class imbalance prob-
lems, sometimes it is possible to provide more appropriate assessments metrics to
measure the classifier performance, but in the multi-class imbalance problems, it is
extremely difficult to provide realistic assessments of the relative severity of the classi-
fication performance [3].

This work has been partially supported under grants of: PROMEP/103.5/12/4783 from the
Mexican SEP and SDMAIA-010 of the TESJO.
18 J.P. Sánchez-Crisostomo et al.
Often in research where the multi-class imbalance is the focus of study, the authors
use metrics that have been extended from two-class imbalance scenarios to compare
the classifier performance, for example the geometric mean, F-measure or measures of
the area under curve family [2]. However, this situation presents an interesting research
question: The proposed metrics to assessment the classifier performance are appropriate
on scenarios where exist multiple minority and multiple majority classes?.
We are interested in this question, so, in this paper we study the behavior of seven
the most common multi-class assessment metrics on real multi-class databases with
minority and majority multiple classes.
2 Assessments Metrics for Multi-class Imbalance Learning

The most studied metrics for assessment the classifier performing in class imbalance
domains have been focused a two class imbalance problems and some of them have
been modified to accommodate them at the multi-class imbalanced learning problems
[4]. In this section we present some of the most common two-class imbalance metrics
adapted at multi-class imbalance scenarios.
Macro Average Geometric (MAvG): This is defined as the geometric average of the
partial accuracy of each class.

J
1
MAvG = ( ACCi ) J , (1)
i=1
where ACCj = ( correctly classif ied of class j)/( total of samples of classj),
i.e., the accuracy on the class j. J is the number of classes.
Mean F-measure (MFM): This measure has been widely employed in information
retrieval
2 · recall(j) · precision(j)
F − measure(j) = , (2)
recall(j) + precision(j)
where recall(j) = (correctly classif ied positives)/(total positives) and precision
(j) = (correctly classif ied positives)/(total predicted as positives); j is the in-
dex of the class considered as positive. Finally, mean F -measure is defined for multi-
class in Reference [2] as follow:

J
F M easure(j)
MFM = . (3)
j=1
J
Macro Average Arithmetic (MAvA): This is defined as the arithmetic average of the
partial accuracies of each class.
J
ACCi
i=1
MAvA = . (4)
J
One the most widely used techniques for the evaluation of binary classifiers in imbal-
anced domains is the Receiver Operating Characteristic curve (ROC), which is a tool
Empirical Analysis of Assessments Metrics 19
for visualizing, organizing and selecting classifiers based on their trade-offs between
true positive rates and false positive rates. Furthermore, a quantitative representation of
a ROC curve is the area under it, which is known as AUC [1]. The AUC measure has
been adapted at multi-class problems [2] and can be defined as follow.
AUC of each class against each other, using the uniform class distribution (AU1U):
2
AU 1U = AU CR (ji , jk ) , (5)
J(J − 1)
ji ,jk J
where AUCR (ji , jk ) is the AUC for each pair of classes ji and jk .
AUC of each class against each other, using the a priori class distribution (AU1P):
2
AU 1P = p(j)AU CR (ji , jk ) , (6)
J(J − 1)
ji ,jk J
where p(j) is a priori class distribution.

AUC of each class against the rest, using the uniform class distribution (AUNU):
1
AU N U = AU CR (j, restj ) , (7)
J
jJ
where restj gathers together all classes different from class j, i.e., the area under the
ROC curve is computed in the approach one against all.
AUC of each class against the rest, using the a priori class distribution (AUNP):
1
AU N P = p(j)AU CR (j, restj ) , (8)
J
jJ
this measure takes into account the prior probability of each class (p(j)).
3 Experimental Protocols
3.1 Database Description
In this section we describe briefly the two databases (92AV3C and ALL-DATA) used in
our experimentation. 92AV3C corresponds to a hyperspectral image (145 x 145 pixels)
taken over Northwestern Indianas Indian Pines by the AVIRIS sensor1 . For simplicity,
in this paper we use only 38 attributes from the 220 attributes of the original dataset.
The attributes were selected using a common features selection algorithm (Best-First
Search [5] implemented in WEKA.2
ALL-DATA consists of the reflectance values of image pixels that were taken by
the Compact Airborne Spectrographic Imager (CASI) and the Airborne Hyper-spectral
Scanner (AHS) sensors. Corresponding chlorophyll measurements for these pixels were
1
engineering.purdue.edu/biehl/MultiSpec/hyperspectral.html
2
www.cs.waikato.ac.nz/ml/weka/
also performed. CASI set consists of the reflectance values of image pixels that were
taken by the CASI sensor. Corresponding thermal measurements for these pixels were
also made. The CASI sensor reflectance curves are formed by 144 bands between 370
and 1049 nm. AHS images consist of 63 bands between 455 and 2492 nm. Therefore,
the input dimensionality of this dataset is 207 (the sum of the bands corresponding to
the CASI and AHS sensors).
Table 1. The class distribution of the 92AV3C and ALL-DATA datasets is presented in this table.
Size represents the number of samples on each class.
Database Class Size Database Class Size

0 95 1 8258
1 489 2 4588
2 834 3 11346
3 968 4 4751
4 54 5 1123
5 614 6 5762
6 497 ALL-DATA 7 3020
7 1294 8 7013
92AV3C 8 380 9 15
9 26 10 79
10 234 12 82
11 20 13 222
12 1434 14 1733
13 2468 15 3628
14 747 – –
15 212 – –
16 10659 – –
In order to study the multi-class assessment metrics behavior in domains of multiple

minority classes and multiple majority classes we split the original datasets (92AV3C
and ALL-DATA) in subsets (Si, Ai and Bi).
For 92AV3C dataset the Si subsets were integrated in the following way: the first
subset (S1) contains the four more minority classes (3, 7, 12 and 13) and the most
majority class (16) of the original dataset. The next subset (S2) was integrated with the
union of S1 with the next minority class (not used before, class 2). This process finishes
when all minority have been integrated at one subset.
The ALL-DATA was split in subsets as follow: the first subset A1 was integrated
by the more minority classes (9, 10, 12 and 13) of ALL-DATA dataset, the next subset
(A2) corresponds to join of A1 with the class 1, i.e., A1∪1. The rest of subsets were
made of the same way.
The Bi subsets were integrated as follow: B1 contains eight majority classes (1–8)
of ALL-DATA dataset, B2 = B1 ∪ with the next minority class (class 9), and the rest
of Bi subsets were integrated using this process. The Table 2 shows a brief summary of
the integration of Si, Ai and Bi subsets.
In 92AV3C subsets (Si) the 10–fold cross–validation was applied. The datasets were
divided into ten equal parts, using nine folds as training set and the remaining block test
set. ALL-DATA subsets (Ai and Bi) were split in disjoints subsets: training (50% of the
samples) and test (50% of the samples).
Table 2. Classes that ingrate the Si, Ai and Bi subsets
Subset S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 S13

Classes 3,7,12,13,16 S1∪ 2 S2∪14 S3∪5 S4∪6 S5∪1 S6∪8 S7∪10 S8∪15 S9∪0 S10∪4 S11∪9 S12∪11
Subset A1 A2 A3 A4 A5 A6 A7 A8 A9 – – – –
Classes 9,10,12,13 A1∪ 1 A2∪2 A3∪3 A4∪4 A5∪5 A6∪6 A7∪7 A8∪8 – – – –
Subset B1 B2 B3 B4 B5 – – – – – – – –
Classes 1,2,3,4,5,6,7,8 B1∪9 B2∪10 B3∪12 B4∪13 – – – – – – – –
3.2 Neural Network Configuration
In the experimental phase we use the MLP trained with the standard back-propagation
in sequential mode. For each training data set, MLP was initialized ten times with dif-
ferent weights, i.e., the MLP was run ten times with the same training dataset. The
results here included correspond to the average of those accomplished ten different
initialization and of ten partitions for 92AV3C, and only the average of ten different
initializations for ALL-DATA. The learning rate (η) was set at 0.1 and only one hid-
den layer was used. The stop criterion was established at 5000 epoch or an MSE below
to 0.001. The number of neurons (n) for the hidden layer was fixed as n = number of
classes +1, because our goal it is not to find the optimal MLP configuration but to study
the assessment metrics behavior.
4 Experimental Results
In order to assessment the multi-class imbalance metrics: M AvG, AU N P , AU 1P ,

M F M , AU N U , M AvA and AU 1U (see section 2), we have carried out an exper-
imental comparison over twenty seven datasets with multiple minority and multiple
majority classes (see Table 2).
The tables 3 and 4 present the experimental results. The first column represents the
dataset used (Si, Ai or Bi), the next columns exhibit the metrics used and at least one
the number of minority classes do not classified, i.e., ignored by the classifier. The rows
show the values obtained from different metrics on each dataset.
Tables 3 and 4 show an interesting behavior. Observe that in some datasets all metrics
(except M AvG) present very similar results but, in these datasets the classifier does not
classify or ignored different minority classes in each dataset. For example, in Table 3 the
values the AU N P for the S10 and S11 datasets are 0.717557 and 0.718413 respectively,
i.e., they are very similar values. However, in S10 the classifier does not classify two
minority classes and S11 does not classify one class. In other words, the values the
AU N P for S10 and S11 are very similar but in S10 the classifier ignored more classes
Table 3. Classification performance of the subsets (Si) obtained from 92AV3C dataset (see Table
2) measured by the metrics: M AvG, M AvA, AU 1U , AU 1P , AU N U , M F M and AU N P
No. of classes
Metric MAvG AUN P AU1P MF M AUN U MAvA AU1U ignored by the
classifier
S1 0.000000 0.680849 0.680849 0.296549 0.606591 0.349023 0.349023 3
S2 0.000000 0.864575 0.864575 0.298384 0.637522 0.373081 0.373082 2
S3 0.000000 0.843382 0.843382 0.326805 0.700919 0.524812 0.524812 1
S4 0.000000 0.775310 0.775310 0.272290 0.627219 0.442667 0.442667 1
S5 0.000000 0.763563 0.763564 0.260260 0.624413 0.448988 0.448988 1
S6 0.000000 0.755563 0.755563 0.245079 0.645932 0.500241 0.500241 2
S7 0.000000 0.745715 0.745714 0.227627 0.628405 0.478391 0.478391 2
S8 0.000000 0.738600 0.738599 0.212890 0.626571 0.486770 0.486770 3
S9 0.000000 0.731669 0.731668 0.197799 0.628791 0.499175 0.499175 1
S10 0.000000 0.737835 0.737835 0.193418 0.655136 0.549927 0.549927 2
S11 0.000000 0.717557 0.717558 0.170630 0.615831 0.487472 0.487472 2
S12 0.000000 0.718413 0.718413 0.162070 0.626332 0.509208 0.509208 1
Table 4. Classification performance of ALL-DATA dataset measured by the metrics: M AvG,

M AvA, AU 1U , AU 1P , AU N U and AU N P
No. of classes
Metric MAvG AUN P AU1P MF M AUN U MAvA AU1U ignored by the
classifier
A1 0.995583 0.995543 0.995543 0.984598 0.995823 0.995625 0.995625 0
A2 0.601254 0.915319 0.915319 0.300290 0.844969 0.796271 0.796271 0
A3 0.408697 0.775386 0.775386 0.271673 0.752187 0.760755 0.760755 0
A4 0.000000 0.747106 0.747106 0.289039 0.716155 0.682684 0.682684 1
A5 0.000000 0.803038 0.803038 0.336305 0.751786 0.699337 0.699337 1
A6 0.000000 0.833518 0.833518 0.350768 0.741196 0.649088 0.649088 1
A7 0.000000 0.734872 0.734872 0.263121 0.659345 0.580251 0.580251 2
A8 0.000000 0.708736 0.708736 0.234921 0.629486 0.546548 0.546549 2
A9 0.000000 0.706589 0.706589 0.230379 0.596762 0.483704 0.483704 3
B1 0.648743 0.750903 0.750903 0.391792 0.748725 0.744061 0.744061 0

B2 0.000000 0.730440 0.730440 0.320462 0.676222 0.618496 0.618496 1
B3 0.000000 0.735042 0.735042 0.301764 0.667699 0.598593 0.598593 2
B4 0.000000 0.708683 0.708683 0.246963 0.609454 0.506425 0.506425 2
B5 0.000000 0.706589 0.706589 0.230379 0.596762 0.483704 0.483704 3
than S11. Similar situations were observed in AU 1P with S8 and S10, M F M with S1
and S2, AU N U with S4 and S8, M AvA with S8 and S11, and AU 1U with S8 and S11
(see Table 3). On Table 4 this behavior was observed in AU N U and AU 1P with B4
and B5. AU N U with A3 and A5.
A dramatic situation was noticed in Table 4, we observe that in some datasets the
classifier presents better results when does not classify one or more classes that when it
classify all classes. For example, the values for A3 and A6 with AU N P are 0.775386
and 0.833518, respectively, i.e., the result the A6 is better than the A3 result, but in A6
one class is ignored for the classifier meanwhile that in A3 all classes are identified for
it. This behavior was adviced too in the metrics AU 1P and M F M for these datasets
(A3 and A6).
On the other hand, the M AvG could be more appropriate in classification problems
with multiple majority classes and multiple minority classes, because it notice when the
classifier ignores any class (see Table 4).
5 Conclusions
In this paper we study some of the most common metrics employed to measure the
classifier performance on the multi-class imbalanced problems. We focused in problems
with multiple minority classes and multiple majority classes. So, some experiments
have been carried out over twenty seven real data sets using a multilayer perceptron
trained with the back-propagation algorithm.
From the analysis of the experimental results in this work, we might suggest that
the main problem of the assessment metrics studied in this paper (except M AvG), is
that they were designed to provide an average performance of the pairs of classes, so
this metrics, in some cases, do not provide information when one or more classes are
ignored for the classifier.
We think, therefore, that they might not be appropriate when the dataset contains
multiple minority classes and multiple majority classes, in other words these metrics
might not be appropriate in muti-class imbalance context as the accuracy was in two-
class imbalance problems. However, the M AvG could be more appropriate in this sce-
nario because it notice when the classifier ignores any class.
The assessment metrics were developed with different proposes and goals, never-
theless, in the literature the researchers use they to compare the classifier performance,
for this reason we consider is necessary a deeper study about of this problem than the
previous one.
References
1. Fawcett, T.: An introduction to roc analysis. Pattern Recogn. Lett. 27, 861–874 (2006)
2. Ferri, C., Hernández-Orallo, J., Modroiu, R.: An experimental comparison of performance
measures for classification. Pattern Recognition Letter 30(1), 27–38 (2009)
3. Hand, D.J., Till, R.J.: A simple generalisation of the area under the roc curve for multiple class
classification problems. Machine Learning 45, 171–186 (2001)
4. He, H., Garcia, E.: Learning from imbalanced data. IEEE Transactions on Knowledge and
Data Engineering 21(9), 1263–1284 (2009)
5. Kohavi, R., John, G.H.: Wrappers for feature subset selection. Artif. Intell. 97(1-2), 273–324
(1997)
6. Ou, G., Murphey, Y.L.: Multi-class pattern classification using neural networks. Pattern
Recognition 40(1), 4–18 (2007)
7. Tsoumakas, G., Katakis, I.: Multi-label classification: An overview. Int. J. Data Warehousing
and Mining, 1–13 (2007)
8. Wang, S., Yao, X.: Multi-class imbalance problems: Analysis and potential solutions. IEEE
Transactions on Systems, Man and Cybernetics, Part B: Cybernetics 42(4), 1–12 (2012)
9. Yang, Q., Wu, X.: 10 challenging problems in data mining research. International Journal of
Information Technology and Decision Making 5(4), 597–604 (2006)
A Novel Rough Set Reduct Algorithm to Feature
Selection Based on Artificial Fish Swarm Algorithm
Fei Wang, Jiao Xu, and Lian Li
School of Information Science & Engineering, Lanzhou University,

730000 Lanzhou, China
{wangf12,xujiao12}@lzu.edu.cn
Abstract. With the purpose of finding the minimal reduct, this paper proposes a
novel feature selection algorithm based on artificial fish swarm algorithm
(AFSA) hybrid with rough set (AFSARS). The proposed algorithm searches the
minimal reduct in an efficient way to observe the change of the significance of
feature subsets and the number of selected features, which is experimentally
compared with the quick reduct and other hybrid rough set methods such as
genetic algorithm (GA), ant colony optimization (ACO), particle swarm
optimization (PSO) and chaotic binary particle swarm optimization (CBPSO).
Experiments demonstrate that the proposed algorithm could achieve the
minimal reduct more efficiently than the other methods.
Keywords: feature selection, rough set, fish swarm algorithm, ant colony
optimization, chaotic binary particle swarm optimization.
1 Introduction
Feature selection is the process of choosing a good subset of relevant features and
eliminating redundant ones from an original feature set, which can be perceived as a
principal pre-processing tool for solving the classification problem [1]. The main
objective of feature selection is to find a minimal feature subset from a set of features
with high performance in representing the original features [2]. In classification
problems, feature selection is a necessary step due to lots of irrelevant or redundancy
features. By eliminating these features, the dimensionality of feature can be reduced
and the predictive performance can be improved for classification. Feature selection
methods are dimensionality reduction methods often associated to data mining tasks
of classification [3], which provide a reduced subset of the original features while
preserving the representative power of the original features.
Rough set (RS) was proposed by Pawlak, which provides a valid tool that can be
applied for both feature selection and knowledge discovery. It has been proved to be
an effective feature selection approach, which can select a subset of features while
preserving the meaning of the features, therefore it can predict the classification
accuracy as well as the original feature set. The essence of rough set to feature
selection are to find a minimal subset of the original features with the most
A Novel Rough Set Reduct Algorithm to Feature Selection Based on AFSA 25
informative features and remove all other attributes from the feature set with minimal
information loss [4]. Rough set is a powerful mathematical tool to reduce the number
of features based on the degree of dependency between condition attributes and
decision attributes, which has been widely applied in many fields such as machine
learning and data mining. Though rough set has been used as a feature selection
method with much success, it is inadequate at finding optimal reduct because of no
perfect search techniques.
In order to find the optimal reduct and improve the performance, a variety of
search techniques hybrid with rough set are introduced to address feature selection
problems such as genetic algorithm (GA), ant colony optimization (ACO), particle
swarm optimization (PSO). These swarm intelligence based algorithms such as
particle swarm and ant colony optimization have been proved to be competitive in
rough set attribute reduction fields. However, these algorithms have some
disadvantages such as premature convergence in PSO and the performance of the
reduct depending on initial parameters in ACO. In this paper, we propose a novel
feature selection algorithm based on artificial fish swarm algorithm hybrid with rough
set, which is not sensitive to initial parameters, has a strong robustness and has the
faster convergence speed to find the minimal reduct subset.
2 Rough Set Theory
Rough set theory is an extension of traditional set theory that provides approximations
in decision making, in which attribute reduction provides a valid method to extract
knowledge from feature set in a concise way. In this paper, we adopt some relevant
concepts of rough set theory related to our attribute reduction approach in [5] such as
equivalence relation, lower approximation, positive region, and degree of dependency.
Definition of Core. The elements of feature core are those features that cannot be
eliminated. In this paper, the algorithm for finding feature core is as follows:
initialize Core = ∅ ; for every attribute a ∈ C , if μC −{a} (D) < μC ( D) , then attribute a is
one element of feature core, namely Core = Core ∪ {a} . Where μC ( D) represents the
degree of dependency between condition attributes C and decision attribute D.
The quick reduct (QR) algorithm proposed in [6], attempts to obtain a reduct
without exhaustively generating all possible subsets. It starts from an empty set, adds
one attribute at a time until it generates its maximum value for the dataset.
3 Swarm Intelligence Based Rough Set Reduct Algorithm
3.1 Ant Colony Optimization Based Reduct Algorithm (ACORS)

ACO is considered a new meta-heuristic algorithm that is used successfully to solve
many NP-hard combinatorial optimization problems [7]. In ACO, a swarm of artificial
ants cooperate for finding good solutions to optimization problems. Every ant
26 F. Wang, J. Xu, and L. Li
searches for optimal solutions in the problem space, which has a start state and one or
more end conditions. The next move is determined by a probabilistic transition rule
that is a function of locally available pheromone trails. Once ant has constructed a
solution, then it updates the pheromone trial values which depend on the quality of
solutions constructed by the ants. Finally, the ant constructs the optimal solution with
the higher amount of pheromone trails. The algorithm stops iterating when an end
condition is satisfied. The search for the optimal feature subset is a traversal through
the graph where a minimal number of nodes are visited and the end conditions are
satisfied [8]. This algorithm performs as follow: All ants start from feature core, each
ant builds a solution and then the pheromone trials for every ant are updated.
Pheromone Trials and Heuristic Information. In the step, each edge is assigned a
pheromone trail and heuristic information. Firstly, the initial pheromone trial on each
edge is initialized to equal amount of pheromone. Secondly, each ant constructs a
solution; after that, the pheromone of each edge in this solution is updated. In
ACORS, the heuristic information is on the basis of the degree of dependency
between the two attributes and decision attribute. The value of heuristic information η
is limited in this paper, If η(a, b) < ε , then η(a, b) = ε , where ε is set to 0.001.
Formally, for any two attributes a , b ∈ C , the heuristic information is defined as
POS{a,b} ( D)
η ( a, b) = (1)
U
Where U is the cardinality of set U and the POS{a,b} (D) , called positive region,
is defined in [5].
Construction of Feasible Solution. When constructing a solution, each ant should
start from the feature core. Firstly, the ant selects randomly a feature, after that, it
probabilistically selects the second attribute from those unselected attributes. That
probability is calculated by
α β
 τij (t )  ⋅  ηij 
P (t ) = 
k
(2)
  τij (t ) ⋅ ηij 
ij α β
l∈J
Where t and k represent the number of iterations and ants, respectively, J represents
the set of unvisited features of ant k, ηij is heuristic information of choosing feature j
when at feature i, τij (t ) is the amount of pheromone between feature i and feature j at
iteration t. In addition, α and β are two parameters corresponding to the importance
of the pheromone trail and heuristic information. When μR ( D) = μC ( D ) , the
construction process stops, where R is the current solution constructed by an ant.
Pheromone Update. After each ant has constructed its own solution, the pheromone
of only edges along the path visited by the ant is updated as
τij (t +1) = ρτij (t ) + q / Lmin (3)
While for other edges, the pheromone trails are updated according to the following
equation.
τij (t +1) = ρτij (t ) (4)
Where ρ is a decay constant used to simulate the evaporation of pheromone, q is
a given constant and Lmin is the minimal feature reduct at iteration t. In ACORS, if the
maximum iteration is reached, then the algorithm terminates and outputs the minimal
reduct encountered. If not, then the pheromone is updated, a new colony of ants are
created and the process iterates once more.
3.2 Particle Swarm Optimization Based Reduct Algorithm (PSORS)
PSO is an efficient evolutionary computation technique based on swarm intelligence

and originates from the simulation of social behaviors such as birds in a flock or
fishes in a school. In PSO, a particle represents a candidate solution to the problem,
which has its own velocity and position in a given search space. PSO starts with the
stochastic initialization of a population of particles which move in the search space to
find the optimal solution by updating the position of each particle by using its own
experience and its companion’s experience [9]. Assume a swarm includes N particles
which move around in a D-dimensional search space. The velocity of the ith particle
in different space can be represented by vi = (vi1 , vi 2 , , viD ) , and the position for the
ith particle in different space can be noted as xi = ( xi1 , xi 2 , , xiD ) . The positions and
velocities of the particles are restricted to a predefined range, respectively. The
personal best position recording the previous best position of the particle is called
pbest and the best position achieved by all individual is denoted the global best
position and called gbest . Based on pbest and gbest , the velocity and position of
each particle are updated to search for the optimal solutions. When BPSO is applied
to solve the feature selection problem, a binary digit is employed to stand for a
feature, where the bit values 1 and 0 stand for selected and non-selected features,
respectively. The velocity of each particle is updated using (5), while the position of
each particle is updated using (6). The position and velocity of each particle are
updated according to the following equations:
vidt +1 = w × vidt + c1 × r1× ( pbest − xidt ) + c2 × r 2 × ( gbest − xidt ) (5)
1, if r < S (vid ) 1
xid =  S (vid ) = (6)
0, otherwise 1 + e− vid
Where t represents the iteration counter, r1 and r2 are random numbers between 0
and 1, c1 and c2 are learning factors that control how far a particle moves in a single
generation, w is called the inertia weight, and the function S (vid ) is a sigmoid
limiting transformation which is introduced to transform vid to the range of (0, 1), r
is random number selected from a uniform distribution between 0 and 1.
In BPSO, the inertia weight w is the modulus that controls the influence of
previous velocity on the present one, thus balancing the global exploration and local
search ability. It means the appropriate control of inertia weight value is imperative to
search for the optimum solution efficiently and precisely. In this paper, chaos theory
and BPSO are combined into a method called CBPSO to avoid this early
convergence, then CBPSO based RS reduct algorithm (CBPSORS) could be achieved
to find superior reduct. Since logistic maps are the most frequently used chaotic
behavior maps and chaotic sequences have been proven easy and fast to generate and
store, as there is no need for storage of long sequences [10], so logistic map is used to
determine the inertia weight value. The inertia weight value is substituted by
sequences generated by the logistic map according to the following (7).
w(t + 1) = μ × w(t ) × (1 − w(t )) w(t ) ∈ (0,1) (7)
Where μ is a control parameter, which cannot be bigger than 4. When the inertia
weight value is close to 0, CBPSO promotes the local search ability. For inertia
weight values near 1, CBPSO strengthens the global search ability.
During the search process, each individual is evaluated using the fitness.
According to the definition of RS reduct, the reduction solution must ensure that the
decision ability is the same as the original decision table and the number of features in
the feasible solution is kept as less as possible. Therefore, classification quality and
the number of selected features are the two pivotal factors used to design a fitness
function which is used to evaluate each individual. The fitness function is defined as
C −R
Fitness = λ ∗ μR ( D) + ξ ∗ (8)
C
Where μR ( D) represents the classification quality of selected condition attributes R

relative to decision D; R denotes the number of selected feature subset; C denotes
the number of the original feature set; λ and ξ are two parameters which determine
the relative importance of classification quality and the number of features, λ ∈ [0,1]
and λ + ξ = 1 .
4 Artificial Fish Swarm Based Reduct Algorithm (AFSARS)
AFSA is a swarm-intelligence based optimization algorithm that simulates the fish

swarm behaviors such as praying, swarming and following with local search of fish
individual for obtaining the global optimum, which was successfully applied to solve
several combinatorial problems. It is a stochastic and parallel search algorithm.
What’s more, it does not need to know the concrete information of problems; instead
it only needs to compare disadvantages and advantages of the solutions of the
problems [11], then the final global optimum will be displayed in the population
through artificial fish individual behaviors of local optimization, which has strong
robustness, fast speed of convergence and being non-sensitive to initial parameters.
The AFSA has been proved to be an effective global optimization algorithm using the
swarm intelligence in the solution of the combinatorial problem [12].
Due to these characteristics of the AFSA, it is introduced to solve feature selection
problems. Assume a fish swarm includes n particles which move around in a D-
dimensional search space. The artificial fish swarm is represented
as F = { f1 , f 2 , , f n } , where fi is an artificial fish (AF). An AF can represent a subset
of features, and a subset of features can be a binary vector: X = {x1 , x2 , , xD } ,
xi ∈ {0,1}, i = 1, 2, , D , where X is the current state of AF, D is the number of
features and the bit values 1 and 0 stand for selected and non-selected features
respectively. Let Y stand for the food concentration, namely the objective function
value; the visual scope of AF is represented as visual distance. The Hamming distance
is used to calculate the visual distance in AFSARS. The Hamming distance of two
points of equal bits length is the number of positions at which the corresponding bits
are different. Sm is the moving step length, trynumber is the try number and δ is the
crowd factor. The representative behavior is described as follows:
Following Behavior. In the following behavior, when the AF current state is X i , it
will judge the food concentration of all its neighborhood partners. Then it will find the
state X j in the current neighborhood, which has the greatest food concentration Yj . Let
n f represent the number of its neighbors in the current neighborhood and n represent
nf
the total number of AF. If Yi < Yj and < δ , it denotes the state X j has more food
n
and is not crowded, it will moves a step toward the state X j . Otherwise, it performs
the swarming behavior.
Swarming Behavior. In the swarming behavior, when the AF current state is X i , it
will assemble in groups naturally in the moving process. Let X c represent the center
n
position in its visual scope. If Yi < Yc and f < δ , it denotes the center position has
n
higher food concentration and is not crowded. It moves a step toward the center
position. Otherwise, it performs the preying behavior. The center position X c of m
fishes is defined as
 m
m
1,

X
k =1
k (i ) ≥
2
X c (i ) =  m
i = 1, 2,3, , D (9)
0, m
 
k =1
X k (i ) ≤
2
Preying Behavior. In the preying behavior, when the AF current state is X i , it needs
to select a state Yj randomly in its visual scope. If Yi < Yj , it moves forward a step in
this direction. Otherwise, it selects randomly a state X j again in its visual distance,
and it judges whether the forward condition is satisfied. If it can satisfy before
trynumber times, it moves a step toward the state X j , otherwise, it moves a step
randomly. When the AF selects to go forward a step in this direction, the mutation
operation of genetic algorithm is adopted in the proposed AFSARS. One position
mutation is used to create a trial point. If the AF will go forward a step from the
state X i to the state X j , then the number of different bits nb is calculated. Here, if
nb > Sm , then Sm = 3 , otherwise, Sm = nb . Randomly generate a digit nr which
represents the number of mutations, where nr is between 1 and Sm . Here, some
indexes of the positions of mutation are selected and then the bits of selected positions
are changed from 0 to 1 or vice versa.
Fig. 1. Artificial fish swarm algorithm based rough set reduct algorithm
Random Behavior. If the other fish behaviors are not executed, the AF performs the
random behavior. This behavior is related with a random movement for a better
position. The behavior is similar to preying behavior, but the different point is the
position of mutation which can be any position of the state X i . The pseudo-code of
our proposed method is illustrated in Fig.1.
5 Experiments and Results
5.1 Datasets and Parameters Setting

To evaluate the usefulness of the proposed algorithms, we carry out experiments on
six datasets of the UCI machine learning repository. In Dermatology (Der) dataset,
some samples are missed in age feature, so it is removed. In the experiments, the five
algorithms require additional parameter settings for their operations. The parameters
of GA are set as follows: population size P = 20, maximum iteration T = 500, the
default parameters of crossover and mutation are adopted in matlab 7.0. The
parameters of ACORS are set as follows: α = 1 , β = 0.01 , ρ = 0.9 , q = 0.1 and the
initial pheromone is set to 0.5, the number of ants is half the number of features and
the maximum iteration equals 50. The parameters of PSORS are set as follows: the
inertia weight decreases along with the iterations, varying from 1.4 to 0.4 according to
the reference [9], acceleration constants c1 = c2 = 2.0 , population size P = 20,
maximum iteration T = 500, velocity Vmax = 4, Vmin = −4 . The parameters of
CBPSORS are set as follows: the inertia weight w(0) = 0.48, μ = 4 , acceleration
constants c1 = c2 = 2.0 , population size P = 20, maximum iteration T = 500,
velocity Vmax = 4, Vmin = −4 . These parameters are chosen based on the literature [10].
The parameters of AFSARS are set as follows: population size P = 50, maximum
iteration T = 50, trynumber=20, maximum step Sm =3, the visual distance of fish is
half the number of features, crowd factor δ =0.618. The parameter λ of the fitness is
set to 0.9 and ξ = 0.1 according to the reference [5]. The fitness function of GARS,
PSORS, CBPSORS and AFSARS are defined as the equation (9). In CBPSORS, the
core of feature set needs to compute, after that the population is initialized, and the
operation is the same as AFSARS. The results achieved from 3 independent runs are
employed in terms of the number of the evolved feature subsets in this paper.
5.2 Results and Analysis
Table 1 shows the reduct results of the various methods on the 6 UCI datasets.
According to the experimental results, we find that AFSARS, CBPSORS and PSORS
have similar efficiency and they are more effective than ACORS and GARS when
dealing with datasets having less than 30 features. However, when dealing with
datasets with over 30 features, PSORS is easy to fall into premature convergence,
which means PSORS is not suitable to find the optimal reduct in most cases.
Comparing with PSORS, AFSARS and ACORS become much effective and find
successfully the global optimum in limited number of iterations on datasets with over
30 features. For those datasets having many features such as Dermatology and Lung,
AFSARS and ACORS are more effective than PSORS and CBPSORS. Furthermore,
PSORS hardly finds the optimal reduction until the maximum iteration is reached
when it deals with datasets with many features. The performances of PSORS and
CBPSORS are not improved after we change their generations to 1000. Apart from
these, we find that the performance of CBPSORS is similar to the performance of
PSORS when the maximum generation is 500, but when we run the two methods
many times, we find that the result of CBPSO is more stable and better. On the whole,
it seems to be the case that AFSARS outperforms the other methods in terms of the
number of the minimal reducts. But compared to the other methods, AFSARS spends
more time to find the optimum reducts.
Table 1. The experimental results of the different algorithms
Dataset #Features QR GARS ACORS PSORS CBPSORS AFSARS

Momk1 6 4 3 3 3 3 3
Tic-tac 9 8 8 8 7 7 7
Zoo 16 5 6-7 5 5 5 5
Vote 16 10 8-10 9 8 8 8
Der 33 10 11-12 9-10 10 9-10 9
Lung 56 5 12-13 5 9-10 9 5
6 Conclusion
This paper starts with the concepts of rough set theory and the QR algorithm, but this
technique often fails to find optimal reducts because of no perfect search strategy.
Therefore, the swarm intelligence methods have been introduced to guide RS method
to find the minimal reducts. Here, we have discussed four different computational
intelligence based reducts: GARS, ACORS, PSORS and CBPSORS. These methods
perform well on some datasets, but sometimes they cannot find the optimal solution in
the limited number of iteration. In this paper, we propose a novel feature selection
algorithm based on artificial fish swarm algorithm hybrid with rough set (AFSARS),
which is non-sensitive to initial values, has a strong robustness and has the faster
convergence speed to find the minimal reducts. Experimental results on real datasets
have demonstrate our proposed method can provide competitive solutions in
generating short reducts more efficiently than the other methods.
Acknowledgments. The authors would like to thank to the Natural Science

Foundation of the People Republic of China (61073193, 61300230), the Key
science and technology Foundation of Gansu Province (1102FKDA010), Natural
Science Foundation of Gansu Province (1107RJZA188), science and technology
support program of Gansu Province (1104GKCA037) for supporting this research.
References
1. Wang, X., Yang, J., Teng, X., Xia, W., Jensen, R.: Feature Selection Based on Rough Sets
and Particle Swarm Optimization. Pattern Recogn. Lett. 28, 459–471 (2007)
2. Suguna, N., Thanushkodi, D.K.: A Novel Rough Set Reduct Algorithm for Medical
Domain Based on Bee Colony Optimization. J. Comput. 6, 49–54 (2010)
3. Inbarani, H.H., Azar, A.T., Jothi, G.: Supervised Hybrid Feature Selection Based on PSO
and Rough Sets for Medical Diagnosis. Comput. Meth. Prog. Bio. 113, 175–185 (2014)
4. Arafat, H., Elawady, R.M., Barakat, S., Elrashidy, N.M.: Using Rough Set and Ant Colony
Optimization in Feature Selection. Int. J. Emerg. Trends Technol. Comput. Sci. 2, 148–155
(2013)
5. Bae, C., Yeh, W., Chung, Y.Y., Liu, S.: Feature Selection with Intelligent Dynamic Swarm
and Rough Set. Expert Syst. Appl. 37, 7026–7032 (2010)
6. Velayutham, C., Thangavel, K.: Unsupervised Quick Reduct Algorithm using Rough Set
Theory. J. Electron. Sci. Technol. 9, 193–201 (2011)
7. Qablan, T., Al-Radaideh, Q.A., Shuqeir, S.A.: A Reduct Computation Approach Based on
Ant Colony Optimization. Basic Sci. Eng. 21, 29–40 (2012)
8. Chen, Y., Miao, D., Wang, R.: A Rough Set Approach to Feature Selection Based on Ant
Colony Optimization. Pattern Recogn. Lett. 31, 226–233 (2010)
9. Xue, B., Zhang, M., Browne, W.N.: Particle swarm optimisation for feature selection in
classification: Novel initialisation and updating mechanisms. Appl. Soft Comput. 18, 261–
276 (2014)
10. Chuang, L., Yang, C., Li, J.: Chaotic Maps Based on Binary Particle Swarm Optimization
for Feature Selection. Appl. Soft Comput. 11, 239–248 (2011)
11. Farzi, S.: Efficient Job Scheduling in Grid Computing with Modified Artificial Fish
Swarm Algorithm. Int. J. Comput. Theor. Eng. 1, 13–18 (2009)
12. Liu, T., Qi, A., Hou, Y., Chang, X.: Feature Optimization Based on Artificial Fish-swarm
Algorithm in Intrusion Detections. In: International Conference on Networks Security,
Wireless Communications and Trusted Computing, vol. 1, pp. 542–545. IEEE Press,
Wuhan (2009)
Hand Gesture Shape Descriptor Based on Energy-Ratio
and Normalized Fourier Transform Coefficients
Wenjun Tan1,2,*, Zijiang Bian1, Jinzhu Yang1,2,

Huang Geng1, Zhaoxuan Gong1, and Dazhe Zhao1,2
1
Medical Image Computing Laboratory of Ministry of Education,
Northeastern University, 110819, Shenyang, China
2
College of Information Science and Engineering,
Northeastern University, 110819, Shenyang, China
{tanwenjun,bianzijian,yangjinzhu,genghuan,
gongzhaoxuan,zhaodzh}@mail.neu.edu.cn
Abstract. The hand gesture shape is the most remarkable feature for gesture
recognition system. Since hand gesture is diversity, polysemy, complex
deformation and spatio-temporal difference, the hand gesture shape descriptor
is a challenging problem for gesture recognition. This paper presents a hand
gesture shape describing method based on energy-ratio and normalized Fourier
descriptors. Firstly, the hand gesture contour of the input image is extracted by
YCb'Cr' ellipse skin color model. Secondly, the Fourier coefficients of the
contour are calculated to transform the point sequence of the contour to
frequency domain. Then the Fourier coefficients are normalized to meet the
rotation, translation, scaling and curve origin point invariance. Finally, the
items of normalized Fourier coefficients are selected by calculating energy-ratio
information as the hand shape descriptors. For validating the shape descriptors
performance, the hand gestures 1-10 are recognized with the template matching
method and the shape descriptor method, respectively. The experiment results
show that the method can well describe the hand shape information and
are higher recognition rate.
Keywords: Hand gesture, Shape descriptor, Fourier descriptors, Skin color.
1 Introduction
Since hand gesture has the characteristics of diversity, polysemy, complex

deformation and spatio-temporal difference, hand gesture recognition is one of the
current topics of new generation human-computer interaction techniques. Hand
gesture recognition system based machine vision recognizes gestures by segmentation
and feature extraction from 2D image sequences from camera. This hand gesture
recognition system has many good performances such as simple inputs, low device
requirement, freedom from interference and so on[1]. The goal of hand gesture
*
Energy-Ratio and Normalized Fourier Transform Coefficients 35
segmentation is to seperate hand region from complex background and to retain

gestures in foregrounds. Because the independent from data gloves or color
landmarks, naturally interact and fast detect, the skin color-based model has become
the most mature method at present[2-3]. Gesture feature extraction is the key
recognition procedure, in which shape information is the most remarkable features
and proper information of gesture recognition system[4-5].
Shape is defined as a function describing position, direction, and surrounded region
of a closed curve in 2D image space. Regular descriptors originate from point
coordinates on closed curve of corresponding target outline. According to the
expression methods, shape descriptors are commonly divided into two classes:
region-expression and outline-expression. Region descriptor focuses on global
geometry features, including area, perimeter, axis directions, compactness, solid
degrees and so on[6]. Usually several global geometry features are adopted in shape
matching and recognition, which is slow in descripting calculating complex shapes
and loses information seriously in feature extraction. These defects lead to low
resolution to express shape inaccurately. Besides, the region descriptors use inter-
target texture distribution statistics, such as the 7 invariant moment[7]. Contour
descriptor includes skeleton-based morphology[8], neural network[9], fractal
method[10], which usually requires complex geometry relevant function to meet the
invariance limits in translation, rotation and zoom, leading to complex calculation.
Fourier descriptor is proposed by Zhan C T and Roskies R Z[11], and improved by
Persoon[12]. The method regards contour as a closed curve formed by a sequence of
end-to-end discrete points, and transforms the point sequence to frequency domain.
The Fourier coefficients are defined as Fourier descriptor. The high-frequency
components of Fourier transformation corresponds to detail information and the low-
frequency components corresponds to overall shape, thus the low-frequency
components could be selected to descript object's shape information. Meanwhile,
Fourier inverse transform is able to restore the shape information expressed by the
descriptor, which is unable by other descriptors. However, the Fourier transformation
coefficients are concerned with target's scale, direction curve origin. So the classical
Fourier transformation method is difficult to express object's shape invariance
accurately. Besides, Fourier coefficients corresponds with contour sequence point
number, which is usually numerous. Thus, if the whole coefficients are calculated, the
heavy burden for gesture classifiers and low recognition rate would be lead to.
On the basis of the gesture region and contour extraction method in the our
previous work, this paper presents an accurate and effective gesture shape expression
method based on energy-ratio and normalized Fourier descriptors, focusing on shape
invariance and effective coefficients selection of the descriptors.
2 Contour of Hand Gesture Extraction
Hand gesture segmentation is the first step of the hand contour extraction. This means
segmenting the region of gestures from a complex background, leaving them alone in
the foreground. Skin color is so good clustering property to be able to separate
‘complexion’ and ‘non-complexion’ region.
36 W. Tan et al.
Hsu R L proposed a way to use ellipse model to describe the skin color distribution
nonlinearly transformed to YCb'Cr' in the Cb'Cr' region, and apply it to face detection,
obtaining better result[14]. Hsu R L put forward that color values always have
nonlinear dependence relation to the luminance value Y in YCbCr color space.
Just through the calculation that whether the pixel is in the ellipse can we detect
whether it belongs to the skin color. So it has a fast computing speed and high
detection accuracy. But it may not describe accurately for the particular imaging
equipment. This skin model easily leads to some non-skin color point being included
to cause skin point over detection. Meanwhile, the model may not include all color
regions, causing the skin color detection incomplete.
Aim at this problem, we presents a YCb'Cr' space ellipse fitting under the skin
modeling method based on the specific statistical properties of the color distribution
to segment hand gesture[15]. So the hand regions are segmented by the method in this
work. Then the contours of hand gestures are easily extracted by 8-neighbourhood
tracing algorithm.
3 Shape Descriptor of Hand Gesture
3.1 Normalized Fourier Transform Coefficients

Let gesture contour be a closed curve represented by coordinate sequence
s(k ) = [ x(k ), y (k )] , in which [ x(k ), y (k )] donates the coordinate pairs starting
from ( x0 , y0 ) and going contour clockwise along the curve. The complex number is
p (l ) = x(l ) + jy (l ) ， (l = 0,1,, n −1), j = −1 , whose discrete Fourier coefficient
is expressed as:
2 π lk
1 n −1 −j
z (k ) = 
n l =0
p ( l ) e n
(1)
where (k = 0,1 , n − 1) , z (k ) is the Fourier transformation of p(l ) and the

expression of point sequence in frequency domain. z ( k ) correlates with object’s
shape. After the transformation in Eq(1), the object’s contour is simplified from 2D
to 1D space. The high-frequency Fourier coefficients are able to describe contour
details and low-frequency ones identify overall shape information. As a result, the
low-frequency Fourier coefficients could be selected to express object’s shape
information. The inverse Fourier transformation z ( k ) is defined as:
n −1 2 π lk
p (l ) =  z ( k ) e
j
n
(2)
k =0
Object’s shape information could be restored with the inverse transform from Eq
(2). Table.1 is the properties of Fourier descriptor of contour sequence p(l ) for
rotation, translation, zooming and origin moving procedures[16], where Δ xy is
defined to be Δxy = (Δx + jΔy) = ( x0 + jy0 ) .
The Fourier descriptors got by Eq(1) transformed were concerned with shape’s
scale, direction and curve origin point, thus the descriptor should be processed to meet
the requirement of shape characteristic invariance. For the transformation property of
translation, only the k=0 becomes impulse function and other properties are no
changes. So the z (0) only change the centroid position of the object and don’t
change the object’s shape. The z (0) can be set as 0 for the shape descriptors. The
Eq(9) could be derived from Eq(1) expressed as[17]:
2 π kl0
j
α e jθ e n
z (k )
z ′( k ) z(k )
d (k ) = = = (3)
z ′(1) jθ
j
2π l0
z (1)
α e e n
z (1)
The Eq(3) shows the change of Fourier coefficients of module and phase of
object’s rotation, scaling and origin position. The d (k ) in Eq(3) is called
normalized Fourier descriptor, k = (2, , n − 1) , which conforms the rotation,
scaling, translation and origin position invariance.
3.2 Fourier Coefficients Selection Based on Energy-Ratio

Usually the high-frequency Fourier coefficients could explain shape details well, but
the low-frequency ones decide object’s overall shape. Thus partial low-frequency
Fourier coefficients could be selected to express gesture shape information. However,
the coefficients are few selected, the corresponding details will be lost so that the
shape information will be difficult to express accurately. So it is important to select
proper descriptor numbers for the gesture shape description.
Assume the first p Fourier coefficients express gesture shape information but not
all the ones, so k > p − 1 , z (k ) = 0 in Eq(1), and the other ones remain unchanged,
the curve p (l ) will be defined as:
p −1 2π lk
pˆ (l ) =  z (k ) e
j
n
(4)
k =0
where, l = (0,1 , n − 1) . Though p Fourier coefficients could obtain each

components p ˆ (l ) , the range of l is still from 0 to n-1. That is to say, similar
boundaries have the same numbers of points, but these needn’t to be so much Fourier
coefficients.
Thus the Fourier descriptors are defined as each item coefficient, and the descriptor
numbers selection corresponds with Fourier coefficient items. Fourier transformation
could translate gesture shape from spatial to frequency domain, and the frequency
corresponds to energy information with amplitude values. The energy information
shows corresponding coefficient ratio in shape expression. Thus the Fourier
38 W. Tan et al.
coefficients are selected by calculating Fourier coefficients energy-ratio information.

The Fourier coefficient energy E (l ) of curve p (l ) could be expressed as follows:
l
E (l ) =  z (k ) (5)
k =0
The energy ratio of the first p Fourier coefficient is:
p
E ( p)  z (k )
e( p ) = = k =0
l (6)

E (l )
z (k )
k =0
From Eq(1), the Fourier transformation in translation is converted to impulse
function when k=0, that is, the value of z (0) varies a lot, which make significant
influence of the calculation of e( p) . Thus, the coefficient energy of z (0) is
ignored in this paper and only the energy ration information is calculated of
l = (1, n − 1) . e( p) is iteratively calculated by increasing the value of p, when
e( p) is greater than a threshold or the difference between e( p) and e( p + 1) is
relatively small, the p is regarded as the final Fourier coefficient number. The
normalized Fourier descriptor d (k − 1) , k = (2, p) , is the descriptor of
gesture shape.
4 Experiments and Discussion

To verify the shape descriptor and the items selecting method of the Fourier coefficient,
the four gesture images are adopted in this work(Fig.1). The contour of gestures is
extracted by the method in section 2 and is expressed with red line in Fig.1. Fig.2(a-d)
shows the amplitude scattergram of the Fourier transformation coefficient
corresponding of these gesture contours. To analyze accurately, Fig.2(a-d) retained the
coefficient of z (0) . It is shown that the amplitude of z (0) had a greater difference
with other coefficients, which is as same as the analysis in section 3.2. When the
coefficient number was bigger than 15, the amplitude varied little. The energy ratio
information of different Fourier items were calculated as Eq(6) and shown in Fig.2(e).
To verify the Fourier coefficients selection method in this paper, the results of the
coefficient numbers from 9 to 15 are compared. Fig.3 is the experiment results of
hand gesture 1-8, in which the green lines express the contour curves through Fourier
inverse transform. The Fourier descriptor details lose seriously of gesture 3,4,5,6,8
from the figure and could not express the whole shape information with 9 coefficient
numbers. And the contour boundary information could be basically expressed by 13
items; and the contour boundary information of 15 descriptors is as same as 13
descriptors. According to the energy-ratio of the Fourier descriptors and analyzing
dimensions and detail lose situations, the 13 item Fourier coefficients were adopted to
express shape characters in this paper. The shape descriptors d (k -1) is calculated as
Eq(3), where k = 12 . Because the z(0)=0, so there are 11 items of Fourier coefficient
as hand gesture shape descriptors in this work.
Fig. 1. The test hand gesture: (a) hand gesture 1; (b) hand gesture 2; (c) hand gesture 4; (d)
hand gesture 5
Fig. 2. Amplitude of Fourier coefficient of hand gesture contour: (a) hand gesture 1; (b) hand
gesture 2; (c) hand gesture 4; (d) hand gesture 5; (e) energy ratio map of Fourier coefficient
Fig.3. Comparison of the reconstruction outlines with Fourier descriptors: (a1-h1) items 9; (a2-
h2) items 11; (a3-h3) items 13; (a4-h4) items 15.
The hand gestures 1-10 are recognized with the shape descriptor method in this
paper to verify the descriptor validity. The template matching method is a common
target recognition method, which is widely applied in pattern recognition system. For
40 W. Tan et al.
identification accuracy-ratio and time performance, the template matching method

and this work are compared in this paper. In order to compare the consistency of the
results of two methods, the Euclidean distance method is adopted to recognize the
gesture categories. Table 1 is the recognition results of the two methods. The results
show that the average recognition rate of the template matching method and our
method is 79% and 90%, respectively.
Table 1. Recognition results of hand gesture with template matching method and this method
Template matching method This method

Hand gesture Correct Recognition Correct Recognition
number rate number rate
Gesture 1 35 87.50% 37 92.50%
Gesture 2 36 90.00% 34 85.00%
Gesture 3 32 80.00% 35 87.50%
Gesture 4 30 75.00% 36 90.00%
Gesture 5 29 72.50% 38 95.00%
Gesture 6 35 87.50% 39 97.50%
Gesture 7 27 67.50% 34 85.00%
Gesture 8 34 85.00% 38 95.00%
Gesture 9 28 70.00% 33 82.50%
Gesture 10 30 75.00% 36 90.00%
5 Conclusions
The hand gesture shape expression methods based on energy-ratio and normalized
Fourier descriptor is presented in this work. The hand gesture contour is the input data
for the shape descriptor, which is extracted by our previous work. Then the Fourier
coefficients of the hand contour are transformed to express the point sequence of hand
gesture contour in frequency domain. For meeting the rotation, translation, zooming
and curve origin point invariance, the Fourier coefficients are normalized. The items
of Fourier coefficients are selected by calculating energy-ratio information. Finally, to
verify the shape descriptors and selection items method of the Fourier coefficients, the
hand gestures 1-10 are recognized with the shape descriptor method in this paper. The
experiment results show that the method can well describe the hand shape information
and access higher recognition rate.
Acknowledgment. This research was partly supported by National Natural Science

Foundation of China (NSFC) under Grant No. 61302012 and No. 61172002, the
Fundamental Research Funds for the Central Universities under Grant N130418002
and N120518001, and Liaoning Natural Science Foundation under Grant No.
2013020021.
References
1. Elli, A.: Understanding The Color of Human Skin. In: Proceeding of the SPIE Conference
on Human Vision and Electronic Imaging VI (SPIE), pp. 243–251. SPIE Press (2001)
2. Yang, J., Lu, W., Waibel, A.: Skin-Color Modeling and Adaptation. In: Chin, R., Pong, T.-
C. (eds.) ACCV 1998. LNCS, vol. 1352, pp. 687–694. Springer, Heidelberg (1997)
3. Bradski, G., Yeo, B., Minerva, M.Y.: Gesture for Video Current Navigation. In: SPIE
IS&T/SPIE, vol. 36, pp. 230–242 (1999)
4. Kawulok, M.: Fast Propagation-based Skin Regions Segmentation in Color Images. In:
Proc. IEEE FG, pp. 1–7. IEEE Press (2013)
5. Nalepa, J., Grzejszczak, T., Kawulok, M.: Wrist Localization in Color Images for Hand
Gesture Recognition. In: Gruca, A., Czachórski, T., Kozielski, S. (eds.) Man-Machine
Interactions 3. AISC, vol. 242, pp. 81–90. Springer, Heidelberg (2014)
6. Ang, Y.H., Li, Z., Ong, S.H.: Image Retrieval Based on Multidimensional Feature
Properties. In: SPIE, vol. 24, pp. 47–57 (1995)
7. Hu, M.K.: Visual Pattern Recognition by Moment Invariants. IEEE Trans. Informaiton
Theory 8, 179–187 (1962)
8. Xu, J.: Morphological Decomposition of 2-D Binary Shapes Into Modestly Overlapped
Octagonal and Disk Components. IEEE Transactions on Image Processing 16, 337–348
(2007)
9. Gupta, L., Sayeh, M.R., Tammana, R.: Neural Network Approach to Robust Shape
Classification. Pattern Recognition 23, 563–568 (1990)
10. Taylor, R., Lewis, P.H.: 2D Shape Signature Based on Fractal Measurements. IEEE
Proceedings -Vision, Image and Signal Processing 4, 422–430 (1994)
11. Zhan, C.T., Roskies, R.Z.: Fourier Descriptors for Plane Closed Curves. IEEE Trans.
Computer 21, 269–281 (1972)
12. Persoon, E., Fu, K.S.: Shape Discrimination Using Fourier Descripters. IEEE Trans.
System, Man, Cybernetics 7, 170–179 (1997)
13. Drolon, H., Druaux, F., Faure, A.: Particles Shape Analysis and Classification Using The
Wavelet Transform. Pattern Recognition Letters 21, 473–482 (2000)
14. Hsu, R.L., Jain, A.K.: Face Detection in Color Images. IEEE Transactions Pattern Analysis
and Machine Intelligence 24, 696–706 (2002)
15. Tan, W.J., Dai, G.Y., Su, H., Feng, Z.Y.: Gesture Segmentation Based on YCb’Cr’ Color
Space Ellipse Fitting Skin Color Modeling. In: The 24th Chinese Control and Decision
Conference (CCDC), pp. 1905–1908. IEEE Press, Taiyuan (2012)
16. Gonzalez, R.C., Woods, R.E., Eddins, S.L.: Digital image processing using MATLAB.
Publishing House of Electronics Industry, Beijing (2005)
17. Wang, T., Liu, W.Y., Sun, J.G., Zhang, H.J.: Using Fourier Descriptors to Recognize
Object’s Shape. Journal of Computer Research and Development 39, 1714–1719 (2002)
A New Evolutionary Support Vector Machine
with Application to Parkinson’s Disease Diagnosis
Yao-Wei Fu, Hui-Ling Chen*, Su-Jie Chen, LiMing Shen, and QiuQuan Li
College of Physics and Electronic Information,

Wenzhou University, China
chenhuiling.jlu@gmail.com
Abstract. In this paper, we present a bacterial foraging optimization (BFO)

based support vector machine (SVM) classifier, termed as BFO_SVM, and it is
applied successfully to Parkinson’s disease (PD) diagnosis. In the proposed
BFO-SVM, the issue of parameter optimization in SVM is tackled using the
BFO technique. The effectiveness of BFO-SVM has been rigorously evaluated
against the PD Dataset. The experimental results demonstrate that the proposed
approach outperforms the other two counterparts via 10-fold cross validation
analysis. In addition, compared to the existing methods in previous studies, the
proposed system can also be regarded as a promising success with the excellent
classification accuracy of 96.89%.
Keywords: Support vector machines, Parameter optimization, Bacterial

foraging optimization, Parkinson’s disease diagnosis, Medical diagnosis.
1 Introduction
As a primary machine learning technique, support vector machines (SVM) [1] is rooted
in the Vapnik-Chervonenkis theory and structural risk minimization principle. Thanks to
its good properties, SVM has found it’s applications in a wide range of classification
tasks. In particular, SVM has demonstrated excellent performance on many medical
diagnosis tasks. However, there is still much room for improvement of the SVM
classifier. Because it has been proved that proper model parameters setting can improve
the SVM classification accuracy substantially [2]. Values of parameters such as penalty
parameter and the kernel parameter of the kernel function should be carefully chosen in
advance when SVM is applied to the practical problems. Traditionally, these parameters
were handled by the grid-search method and the gradient descent method. However, one
common drawback of theses methods is that they are vulnerable to local optimum.
Recently, biologically inspired metaheuristics such as genetic algorithm and particle
swarm optimization (PSO) have been considered to have a better chance of finding the
global optimum solution than the traditional aforementioned methods. As a relatively
new member of the swarm-intelligence algorithms, BFO has been found to be a
promising technique for real-world optimization problems such as optimal controller
design [3], learning of artificial neural networks [4] and active power filter design [5].
*
A New Evolutionary Support Vector Machine 43
This study attempts to employ BFO to handle the parameter optimization of SVM and
applied the resultant effective model BFO-SVM for effective detection of Parkinson's
disease (PD). The main objective of this study is to explore the maximum generalization
capability of SVM and apply it to PD diagnosis to distinguish patients with PD from the
healthy ones.
The remainder of this paper is organized as follows. The related works on detection
of PD is presented in Section 2. In section 3 the detailed implementation of the BFO-
SVM diagnostic system is presented. Section 4 describes the experimental design.
The experimental results and discussion of the proposed approach are presented in
Section 5. Finally, Conclusions and recommendations for future work are summarized
in Section 6.
2 Related Works on Detection of PD
PD is one kind of degenerative diseases of the nervous system, which is characterized

by a group of conditions called motor system disorders because of the loss of
dopamine-producing brain cells. Till now, the cause of PD is still unknown, however,
it is possible to alleviate symptoms significantly at the onset of the illness in the early
stage [6]. It is claimed that approximately 90% of the patients with PD show vocal
impairment [7], the patients with PD typically exhibit a group of vocal impairment
symptoms, which is known as dysphonia. The dysphonic indicators of PD make
speech measurements an important part of diagnosis. Recently, dysphonic measures
have been proposed as a reliable tool to detect and monitor PD [8].
Various researchers have studied the PD diagnosis problem. Little et al. [8]
conducted a remarkable study about PD identification, they employed an SVM
classifier with Gaussian radial basis kernel functions to predict PD, and also
performed feature selection to select the optimal subset of features from the whole
feature space, and the best accuracy rate of 91.4% was obtained by the best model.
Das [9] presented a comparative study of using Neural Networks (ANN), DMneural,
Regression and Decision Tree for effective diagnosis of PD, the experimental results
have shown that the ANN classifier yielded the best results, the overall classification
score of 92.9% was achieved. Recently, Ozcift et al. [10] combined the correlation
based feature selection (CFS) algorithm with the rotation forest (RF) ensemble
classifiers of 30 machine learning algorithms to identify PD, and the best
classification accuracy of 87.13% was achieved by the proposed CFS-RF model.
Chen et al. [11] employed the fuzzy k-nearest neighbor (FKNN) approach in
combination with the principle component analysis (PCA-FKNN) to diagnose PD,
and the best classification accuracy of 96.07% was obtained by the proposed
diagnosis system.
3 The Proposed BFO-SVM Model
This study proposes a novel BFO-SVM model for parameter optimization problem of
SVM. In the proposed model, the parameter optimization for SVM are dynamically
44 Y.-W. Fu et al.
conducted by implementing BFO algorithm, then the obtained optimal parameters are
taken by the SVM model to perform the classification task. The proposed model is
comprised of two main evaluation procedures:
1) Inner_Parameter_Optimization procedure: Evaluate the performance of each
candidate parameters;
2) Outer_Performance_Estimation procedure: Evaluate the overall performance
of the SVM classifier with the optimal parameter values obtained;
In the Inner_Parameter_Optimization procedure, the parameters C and γ are
dynamically optimized by implementing BFO algorithm. The classification accuracy
is taken into account in designing the fitness:
f = avgACC = (ΣiK=1testACCi ) k . (1)
where variable avgACC in the function f represents the average test accuracy achieved
by the SVM classifier via k-fold CV, where k = 5. Note that here the 5-fold CV is
employed to do the model selection that is different from the outer loop of 10-fold
CV, which is used to do the performance estimation. The pseudo-code of this
procedure is given bellow:
__________________________________________________________________
Pseudo-code for the Inner_Parameter_Optimization procedure
step 1. Initialize parameters p, S, Nc, Ns, Nre, Ned, Ped, θi
where
p: number of dimension of the search space,
S: swarm size of the population,
Nc: number of chemotactic steps,
Ns: swimming length,
Nre: the number of reproduction steps,
Ned: the number of elimination-dispersal events,
Ped: elimination-dispersal probability, and
C(i): the size of step taken in the random direction specified by the tumble.
step 2. Elimination-dispersal loop: l=l+1.
step 3. Reproduction loop: k=k+1.
step 4. Chemotaxis loop: j=j+1.
(a) For i=1,2,…,S, take a chemotactic step for bacterium i as follows.
(b) Train SVM and compute the fitness J(i, j, k, l)
Let, J(i, j, k, l)=J(i, j, k, l)+Jar(θ) where Jar is defined in Eq. (8).
(c) Let Jlast=J(i, j, k, l) to save this value since we may find a better cost
via a run.
∈
(d) Tumble: generate a random vector Δ(i) Rp with each element
Δm(i),m=1,2,…,p, a uniformly distributed random number on [-1, 1].
(f) Move: let
Δ (i )
θ i ( j + 1, k , l , di ) = θ i ( j , k , l , di ) + C ( i )
Δ (i ) Δ (i )
T
(g) Train SVM and compute the fitness J(i,j+1,k,l), and let
J(i,j+1,k,l)=J(i, j, k, l)+Jar(θ).
(h) Swim.
i) Let n=0;
ii) While n<Ns
iii) Let n=n+1;
iv) If J(i,j+1,k,l)<Jlast, let Jlast=J(i,j+1,k,l) and let
Δ (i )
θ i ( j + 1, k , l , di ) = θ i ( j , k , l , di ) + C ( i )
Δ (i ) Δ (i )
T
and use this θi(j+1,k,l) to train SVM, and then compute the
new fitness
J(i, j+1,k, l) as did in (g);
v) Else, let n =Ns.
(i) Go to next bacterium (i+1) if i S. ≠
step 5. If j<Nc, go to step 4.
step 6. Reproduction:
Rank all of the individuals according to the sum of the evaluation results
in this period, and then removes out the last half individuals and duplicates
one copy for each of the rest half.
step 7. If k<Nre, go to step 3.
step 8. Elimination-dispersal:
For i=1,2,…,S with probability Ped, eliminate and disperse each
bacterium.
If l<Ned, then go to step 2; otherwise end.
__________________________________________________________________
In the Outer_Performance_Estimation procedure, SVM model performs the
classification tasks using the obtained optimal parameters via 10-fold CV analysis.
The pseudo-code of this procedure is given bellow:
__________________________________________________________________
Pseudo-code for the Outer_Performance_Estimation procedure
/*performance estimation by using k-fold CV where k = 10*/
Begin
For j = 1:k
Training set = k-1 subsets;
Testing set = remaining subset;
Train the SVM classifier on the training set using the parameters and feature
subsets obtained from Inner_Parameter_Optimization procedure;
Test it on the testing set;
End For;
Return the average classification accuracy rates of SVM over j testing set;
End.
__________________________________________________________________
46 Y.-W. Fu et al.
4 Experimental Setup
In this study, we have performed our conduction on the Parkinson’s data set taken
from UCI machine learning repository.
The BFO-SVM, PSO-SVM and Grid-SVM classification models were
implemented using MATLAB platform. For SVM, LIBSVM implementation was
utilized, which was originally developed by Chang and Lin [12]. We implemented the
BFO, PSO and grid search algorithm from scratch. The computational analysis was
conducted on Windows 7 operating system with AMD Athlon 64 X2 Dual Core
Processor 5000+ (2.6 GHz) and 4GB of RAM. Normalization is firstly employed
before classification, in order to avoid feature values in greater numerical ranges
dominating those in smaller numerical ranges, as well as to avoid the numerical
difficulties during the calculation. In order to guarantee the valid results, the k-fold
CV was employed to evaluate the classification accuracy.
In this study, we designed our experiment using a two-loop scheme, which also
was used in [13]. The detailed parameter setting for BFO-SVM is shown in Table 1.
For PSO-SVM, the number of the iterations and particles are set to 250 and 8,
respectively. vmax is set about 60% of the dynamic range of the variable on each
dimension for the continuous type of dimensions, c1 = 2, c2 = 2, wmax and wmin are set
to 0.9 and 0.4, respectively. The searching ranges of C ∈ [2 ^ (−5), 2 ^ (15)] and
γ ∈ [2 ^ (−5), 2 ^ (15)] for BFO-SVM, PSO-SVM and Grid-SVM were set as the same.
Table 1. Common parameter setup for BFO
S Nc Ns Ned Nre ped datt watt wrepe hrepe C(i)

8 100 20 1 4 0.25 0.1 0.2 10 0.1 0.1
Classification accuracy, sensitivity and specificity were used to test the

performance of the proposed BFO-SVM model.
5 Experimental Results and Discussions
In BFO, the parameter chemotaxis step size C(i) plays an important role in controlling
the search ability of BFO. Thus, we firstly present results from our investigations on
the impacts of C(i) and assign initial values for it. C(i) can be initialized with
biologically motivated values, but a biologically motivated value may not be the best
for specific application [3]. In Table 2, we illustrate the relationship between the
different values of C(i) and the performance of BFO-SVM. The average results are
presented with the standard deviation described in the parenthesis. From the table we
can see that BFO-SVM reaches the best performance at C(i) = 0.1 in terms of
accuracy, sensitivity and specificity. Therefore, we select 0.1 as the parameter value
of C(i) for the proposed BFO-SVM to implement the coming tasks.
Table 2. The detailed results of BFO-SVM with different C(i) on the PD data set
Chemotactic BFO-SVM
step size
Accuracy (%) Sensitivity (%) Specificity (%)
parameter C(i)
0.05 94.84(6.43) 97.41(6.11) 88.00(17.79)
0.1 96.90(4.34) 98.75(3.95) 90.83(16.87)
0.15 95.39(4.53) 97.44(4.33) 90.42(10.77)
0.2 94.42(4.97) 97.57(4.19) 85.50(13.17)
0.25 93.40(6.28) 96.08(5.48) 87.71(17.36)
0.3 94.47(4.97) 96.02(4.82) 90.64(12.32)
To evaluate the effectiveness of the proposed BFO-SVM system for PD, we

conducted experiments on the PD database. Table 3 shows the classification accuracy,
sensitivity, specificity, and optimal pairs of (C, γ) for each fold obtained by BFO-
SVM. The comparison among PSO-SVM, Grid-SVM and BFO-SVM is shown in
Figure 1.
Table 3. The detailed results of BFO-SVM on the PD data set
Fold BFO-SVM
No. Accuracy Sensitivity Specificity C (×104) γ
1 0.8947 0.8750 1.0000 1.4250 3.7726
2 1.0000 1.0000 0.8333 2.2658 3.3197
3 0.9474 1.0000 1.0000 2.2930 3.1821
4 1.0000 1.0000 0.7500 1.2372 4.6147
5 1.0000 1.0000 1.0000 2.8519 3.5251
6 0.9000 1.0000 1.0000 1.4252 3.8452
7 1.0000 1.0000 0.5000 2.0478 3.2553
8 1.0000 1.0000 1.0000 0.1331 3.7350
9 1.0000 1.0000 1.0000 2.2390 3.7168
10 0.9689 1.0000 1.0000 0.8046 4.9414
Avg. 0.0434 0.9875 0.9083 1.6723 3.7908
48 Y.-W. Fu et al.
Fig. 1. The cross-validation accuracy obtained for each fold by PSO-SVM, Grid-SVM and
BFO-SVM
As shown in Figure 1, we can see that BFO-SVM has dominated PSO-SVM and
Grid-SVM in most folds in the process of the 10-fold CV, namely, BFO-SVM has
achieved the high classification accuracy equal to or better than that of the other two
models obtained for 7 folds in the whole 10 folds. The average classification accuracy
of BFO-SVM is 96.89%, while the average classification accuracy of PSO-SVM and
Grid-SVM are 94.89% and 93.87%, respectively.
6 Conclusions and Future Work
This work has explored a new diagnostic system, BFO-SVM, for detection of PD.
The main novelty of this paper lies in the proposed BFO-based approach, which aims
at maximizing the generalization capability of the SVM classifier by exploring the
new swarm intelligence technique for optimal parameter tuning for PD diagnosis. The
empirical experiments on the PD database have demonstrated the superiority of the
proposed BFO-SVM over PSO-SVM and Grid-SVM in terms of classification
accuracy. It indicates that the proposed BFO-SVM system can be used as a viable
alternative solution to PD diagnosis.
Acknowledgments. This work is supported by the open project program of Wenzhou

University Laboratory under Grant No. 13SK29A.
References
1. Vapnik, V.N.: The nature of statistical learning theory. Springer, New York (1995)
2. Keerthi, S., Lin, C.: Asymptotic behaviors of support vector machines with Gaussian
kernel. Neural Computation 15(7), 1667–1689 (2003)
3. Passino, K.M.: Biomimicry of bacterial foraging for distributed optimization and control.
IEEE Control Systems Magazine 22(3), 52–67 (2002)
4. Ulagammai, M., et al.: Application of bacterial foraging technique trained artificial and
wavelet neural networks in load forecasting. Neurocomputing 70(16-18), 2659–2667
(2007)
5. Mishra, S., Bhende, C.N.: Bacterial foraging technique-based optimized active power filter
for load compensation. IEEE Transactions on Power Delivery 22(1), 457–465 (2007)
6. Singh, N., Pillay, V., Choonara, Y.E.: Advances in the treatment of Parkinson’s disease.
Progress in Neurobiology 81(1), 29–44 (2007)
7. Ho, A.K., Iansek, R., Marigliani, C., Bradshaw, J.L., Gates, S.: Speech impairment in a
large sample of patients with Parkinson’s disease. Behavioural Neurology 11(3), 131–137
(1999)
8. Little, M.A., McSharry, P.E., Hunter, E.J., Spielman, J., Ramig, L.O.: Suitability of
dysphonia measurements for telemonitoring of Parkinson’s disease. IEEE Transactions on
Biomedical Engineering 56(4), 1015–1022 (2009)
9. Das, R.: A comparison of multiple classification methods for diagnosis of Parkinson
disease. Expert Systems with Applications 37(2), 1568–1572 (2010)
10. Ozcift, A., Gulten, A.: Classifier ensemble construction with rotation forest to improve
medical diagnosis performance of machine learning algorithms. Computer Methods and
Programs in Biomedicine 104(3), 443–451 (2011)
11. Chen, H.L., Huang, C.C., Yu, X.G., Xu, X., Sun, X., Wang, G., Wang, S.J.: An efficient
diagnosis system for detection of Parkinson’s disease using fuzzy< i> k</i>-nearest
neighbor approach. Expert Systems with Applications 40(1), 263–271 (2013)
12. Chang, C.C., Lin, C.J.: LIBSVM: a library for support vector machines. (2001),
http://www.csie.ntu.edu.tw/cjlin/libsvm
13. Chen, H.L., Yang, B., Wang, G., Liu, J., Xu, X., Wang, S.J., Liu, D.Y.: A novel
bankruptcy prediction model based on an adaptive fuzzy< i> k</i>-nearest neighbor
method. Knowledge-Based Systems 24(8), 1348–1359 (2011)
Parallel Bees Swarm Optimization
for Association Rules Mining
Using GPU Architecture
Youcef Djenouri and Habiba Drias
USTHB LRIA, Algiers, Algeria

y.djenouri@gmail.com, hdrias@hotmail.fr
Abstract. This paper addresses the problem of association rules min-

ing with large scale data sets using bees behaviors. The bees swarm op-
timization method have been successfully running on small and medium
data size. Nevertheless, when dealing with large benchmark, it is bluntly
blocked. Additionally, graphic processor units are massively threaded
providing highly intensive computing and very usable by the optimiza-
tion research community. The parallelization of such method on GPU
architecture can be deal large data sets as the case of WebDocs in real
time. In this paper, the evaluation process of the solutions is parallelized.
Experimental results reveal that the suggested method outperforms the
sequential version at the order of ×70 in most data sets, furthermore,
the WebDocs benchmark is handled with less than forty hours.
Keywords: bees swarm optimization, association rule mining, parallel

algorithms, GPU architecture.
1 Introduction
Association Rules Mining (ARM) is one of the most important and well studied
techniques of data mining tasks [1]. It aims at extracting frequent patterns,
associations or causal structures among sets of items from a given transactional
database. Formally, the association rule problem is as follows: let T be a set of
transactions {t1 , t2 , . . . , tm } representing a transactional database, and I be a
set of m different items or attributes {i1 , i2 , . . . , im }, an association rule is an
implication of the form X → Y where X ⊂ I, Y ⊂ I, and X ∩ Y = ∅. The
itemset X is called antecedent while the itemset Y is called consequent and the
rule means X implies Y .
Many exacts algorithms have been developed for solving ARM problem. Apri-
ori [2] and FPgrowth [3] are the most used algorithms, Nonetheless, the exacts
algorithms are high time consuming for large data sets.
Swarm intelligence algorithms have been successfully applied for association
rules mining problem like: PSOARM [4], ACOR [7], HBSO-TS [6], and BSO-
ARM [5]. The experiments reveal that the bees swarm optimization outperforms

Parallel Bees Swarm Optimization for Association Rules Mining 51
ACO and PSO in terms of rules quality. However, when dealing with large in-
stances like WebDocs (the huge benchmark among the web), the complexity
time still expensive. In fact, the main challenge in sequential ARM algorithms
is to handle massive data-sets. ARM problem has been parallelized in different
ways. On each approach, the authors took advantage of the used parallel hard-
ware like: CD, DD algorithms for distributed memory system [8], CCPD and
PCCD for shared memory system [9]. However, All these algorithms have been
implemented for traditional parallel architectures (supercomputers, clusters ..)
which are still expensive and not always accessible for every one.
Motivating by the forcefully of graphic processors units, in this paper, we
propose a parallel GPU-based approach for association rules mining problem. It
is an extended version of the bees swarm optimization algorithm. The genera-
tion of solutions and the search process are done on CPU. However, to benefit
from the massive GPU threaded, the evaluation process of the solutions have
been performed concurrently on GPU. The results show the effectiveness of the
parallel approach compared to the sequential one.
The rest of the paper is organized as follows: Section 2 relates the state of the
art of ARM algorithms followed by a brief explanation of BSO-ARM algorithm.
In section 4, we present the proposed algorithm PMES. Section 5 shows the
results of our algorithm using several data sets. In section 6, we conclude by
some remarks and futures perspective of the present work.
2 Related Works
ARM community is many investigating on GPU architecture, for this many
algorithms based on GPU have been developed. The parallel ARM on GPU was
first introduced by Wenbin et al. in [10]. They proposed two parallel versions
of Apriori called PBI (Pure Bitmap Implementation) and TBI (Trie Bitmap
Implementation) respectively. In PBI, the transactions datasets and the itemsets
are represented by a bitmap data structure. Indeed, the itemset structure is a
bitmap (n ∗ m) where n is the number of k itemsets and m is the number of its
items. In this representation, bit (i, j) = 1 if itemsets i contains item j otherwise
0. Similarly, a transaction structure is a bitmap (n ∗ m) where n is the number of
itemsets and m is the number of transactions. Here bit (i, j) = 1 if transaction
j contains itemsets i, 0 otherwise.
In [11], a new algorithm called GPU-FPM is developed. It is based on Apriori-
Like algorithm and uses a vertical representation of the data set because of the
limitation of the memory size in the GPU. The speed up reached during the
reported experimentations varies from ×10 to ×15.
Syed et al. in [12] proposed a new Apriori algorithm for GPU architecture which
performs on two steps. In the first step, the generation of itemsets is done on GPU
host where each thread block computes the support of a set of itemsets. In the sec-
ond step, the generated itemsets are sent back to the CPU in order to generate the
rules corresponding to each itemset and to determine the confidence of each rule.
The main drawback of this algorithm is the cost of the CPU/GPU communications.
The speed up of this algorithm reaches ×20 with a large data set.
52 Y. Djenouri and H. Drias
Another Cuda-Apriori algorithm is proposed in [13]. First, the transactional

dataset is divided among different threads. Then, k-candidate itemsets are gen-
erated and allocated on global memory. Each thread handles one candidate using
only the portion of the dataset assigned to its block. In each iteration, a syn-
chronization between blocks is done in order to compute the global support of
each candidate. In [14], a GPApriori algorithm is proposed using two data struc-
tures in order to accelerate the itemsets counting. For small size datasets this
algorithm reaches a speed up of (×100). Nevertheless, the speed up is reduced
when the number of transactions increases du to thread divergence.
In [15] another work for maximal frequent itemsets mining on GPU is re-
ported. This work is based on the main proprieties of GPU which are data
parallelism and data independence. A tree structure is used to store the fre-
quent itemsets where each node contains an itemsets, its support and a bitmap
of each itemset. A bitmap is a boolean structure that represents all transactions
containing a given itemset. This algorithm works well on large datasets. Nev-
ertheless, it requires a high space memory and many pointer links which are
difficult to manage on GPU architecture.
3 BSO-ARM Algorithm
In [5], we proposed BSO-ARM algorithm. The aim of this algorithm is to find one
part of association rules respecting minimum support and minimum confidence
constraints with reasonable time. Each rule is considered as one solution in the
search space, each of which is represented by a vector S of N bits and their
positions are defined as follows:
1. S[i] = 0 if the item i is not in the solution S.

2. S[i] = 1 if the item i belongs to the antecedent part of the solution S.
3. S[i] = 2 if the item i belongs to the consequent part of the solution S.
The algorithm can be decomposed into four steps (Neighborhood Search Com-
putation, Search Area Determination, Fitness computing and Dancing Step).
Neighborhood Search Computation. The neighborhood search is obtained

by changing from a given solution S one bit in random way. Based on this
simple operation, N neighborhoods are created. Notice that this operation does
not generate non admissible solutions.
Example 1
Consider the given solution: S = {1, 0, 0, 1, 2}
1. change the first bit in S : S1 = {0, 0, 0, 1, 2}

2. change the second bit in S: S2 = {1, 2, 0, 1, 2}
3. change the third bit in S:S3 = {1, 0, 1, 1, 2}
4. change the fourth bit S: S4 = {1, 0, 0, 0, 2}
5. change the fifth bit S: S5 = {1, 0, 0, 1, 0}
Search Area Determination. It determines K search spaces, each one is as-

sociated to a bee. The j th bee builds its own search area by changing successively
in the solution Sref the bits j + i × F lip where i ∈ [0..N − 1], j ∈ [1..K] and
Flip is a given parameter. This strategy can be used if and only if K ≤ FNlip . If
the distance between solutions is the number of different bits, then the distance
between the bees and the solution reference is equal to FNlip .
Fitness Computing. In this step for each generated solution (rule), the entire
transactional database should be scanned. The solution fitness is based on the
support and the confidence measures as:
F itness(s) = α × conf (s) + β × sup(s) (1)

This function should be maximized. For each invalid solution s where Sup(s) <
M insup or Conf (s) < M inConf , the Fitness(s) is set to −1 and the solution
is rejected.
Dancing Step. Each bee puts in the dance table the best rule found among its
search. The communication between bees is done in order to find the best dance
(the best rule) which becomes the reference solution for the next pass.
The general functioning of the algorithm is as follows: First, the solution
reference (Sref ) is initialized randomly so that each element of Sref belongs to
{0,1,2}.
After that, except the Fitness Computing which is applied for each generated
solution, the other steps are repeated in the order until IMAX is reached.
4 Parallel Single Evaluation of Solution Algorithm
The proposed algorithm parallel single evaluation of solution (PSES for short)
is based on the master/slave paradigm. The master is executed on CPU and the
slave is offloaded to the GPU. First, The master initializes randomly the solution
reference. After that, it determines regions of the whole bees by generating the
neighbors of each bee. Unlike BSO-ARM, single solution is evaluated on GPU
in parallel. After, the master receives back the fitness of all rules, each bee
calculates sequentially the best rule and puts it in the table dance. The best
rule of the dance table become the solution reference for the next iteration.
This combined CPU/GPU process is repeated until the maximum number of
iterations is reached. We opted for a mapping in which all threads are mapped
to one rule. Threads of the same block are launched to calculate collaboratively
the fitness of a single rule with one packet of transactions. Therefore, we have
as many threads as transactions. The transactions are subdivided into subsets
and each subset seti is assigned to one bloc so that each thread calculates only
one transaction. After that a sum reduction is applied to aggregate the fitness
value.
Such a strategy allows us to benefit from the massively parallel power of

GPU by launching a large number of threads per rule. Indeed, in PSES, only
the threads must synchronize after each iteration to process sum reduction tech-
nique, thus N × K × IM AX synchronization during all the lifetime of the algo-
rithm. The general algorithm of the GPU kernel is given in Algorithm 1.
Algorithm 1. The GPU kernel

Input: Sol: Single solution
Freq[] Array of integer
recuperate Sol from CPU
initialize Freq to zero
idt = blockIdx.x × blockDim.x + threadIdx.x.
for i=0 to nb transactions do
if Sol∈ tidt then
freq[idt][i]=1.
else
freq[idt][i]=0.
end if
end for
Evaluation(Sol)=Sum Reduction(freq).
cudaMemcpy(Evaluation(Sol), cudaMemcpyDeviceToHost).
First, GPU recuperates the single sol containing the set of solutions generated
on CPU. It initializes freq by zero. Then, each thread evaluates one solution with
one transaction.
5 Performance Evaluation
To evaluate the performance of the proposed approach PSES, several data sets
of different size are considered. The data sets are the well-known scientific
databases that are frequently used in data mining community (Frequent and
Mining Dataset Repository [17] and Bilkent University Function Approximation
Repository[16].
From the smallest benchmark (Bolts data set) to the largest one (WebDocs
data set), the used data sets are divided according to the number of transactions
into three categories (small, average, large). The description of the different used
data sets are presented in Table 1. Notice that the data sets differ according to
the average size of items. On one hand, there are big data sets with a few number
of items per transaction. On the other hand, there are other small data sets
with a significant number of items per transaction. For instance, the number
of transactions on the Connect data set is 100000 and the average items per
transaction is only 10. Whereas, in the IBM data set the number of transactions
is only 1000 and the number of items per transaction is 20.
Table 1. Data sets description
Data Set Data set Name Transactions Item Average

Type Size Size Size
Bolts 40 8 8
Sleep 56 8 8
Basket ball 96 5 5
Small IBM Quest Std. 1000 40 20
Quake 2178 4 4
Chess 3196 75 37
Mushroom 8124 119 23
Pumbs star 40385 7116 50
BMS-WebView-1 59602 497 2.5
Average BMS-WebView-2 77512 3340 5
Korasak 80769 7116 50
retail 88162 16469 10
Connect 100000 999 10
Large BMP POS 515597 1657 2.5
WebDocs 1692082 526765 -
Table 2. Runtime of the proposed approach with different data sets (in Sec)
Data Set Data set BSO-ARM PSES

Type Name
Bolts 9 2.5
Sleep 15 5
Pollution 22 6
Basket ball 35 12
Small IBM Q.S. 618 50
Quake 80 20
Chess 149956 125
Mushroom 28815 350
Pumbs star 144120 1550
BMS-W.V.-1 180524 6200
Average BMS-W.V.-2 249985 1850
Korasak 258451 1948
Retail 299658 2185
Connect 300985 4485
Large BMP POS Stopped After
15 Days 22565
WebDocs Stopped After
15 Days 45965
The suggested approach has been implemented using C-CUDA 4.0 and the
experiments have been carried out using a CPU host coupled with a GPU device.
The CPU host is a 64-bit quad-core Intel Xeon E5520 having a clock speed of
2.27GHz. The GPU device is an Nvidia Tesla C2075 with 448 CUDA cores (14
multiprocessors with 32 cores each), a clock speed of 1.15GHz, a 2.8GB global
memory, a 49.15KB shared memory, and a warp size of 32. Both the CPU and
GPU are used in single precision.
Table 2 presents the execution time of the sequential and parallel version of
BSO-ARM. In order to well exploring the search space, the number of bees K,
respectively the number of iterations IMAX are set to 20, respectively 100.
The parallel version outperforms the sequential one in all cases. Furthermore,
the GPU-based parallelization allowed us to solve two challenging large data sets
(BMP POS and Web Docs) containing more than 1.5 millions of transactions
and more than 0.5 million of items. To the best of our knowledge, these these
two data sets have newer been solved before in the literature. Indeed, BSO-ARM
blocked after 12 days whereas it takes only few hours using PSES.
6 Conclusion
In this paper, we proposed a new algorithm for association rules mining on GPU
architecture. It is based on the bees behaviors, we first generate the solutions on
CPU, then, the evaluation of each solution is performed in parallel using GPU
threaded. The intensive multi-threaded provided on GPU conduct us to perform
the single evaluation of solution at the same time. In fact, each thread is mapped
with one transaction, this permits to accelerate the process of the evaluation.
The experiments show that the parallel approach outperforms the sequential one
in terms of the execution time. The results also reveal that using the massive
threaded in GPU and the intelligent bees, the largest transactions base on the
web is mined in real time.
References
1. Han, J., Kamber, M., Pei, J.: Data mining: concepts and techniques. Morgan Kauf-
mann (2006)
2. Agrawal, R., Imielinski, T., Swami, A.: Mining association rules between sets of
items in large databases. ACM SIGMOD Record 22(2) (1993)
3. Han, J., Pei, J., Yin, Y.: Mining frequent patterns without candidate generation.
ACM SIGMOD Record 29(2) (2000)
4. Kuo, R.J., Chao, C.M., Chiu, Y.T.: Application of particle swarm optimization to
association rule mining. Applied Soft Computing 11(1), 326–336 (2011)
5. Djenouri, Y., Drias, H., Habbas, Z., Mosteghanemi, H.: Bees Swarm Optimization
for Web Association Rule Mining. In: 2012 IEEE/WIC/ACM International Con-
ferences on Web Intelligence and Intelligent Agent Technology (WI-IAT), vol. 3,
pp. 142–146. IEEE (2012)
6. Djenouri, Y., Drias, H., Chemchem, A.: A hybrid Bees Swarm Optimization and
Tabu Search algorithm for Association rule mining. In: 2013 World Congress on
Nature and Biologically Inspired Computing (NaBIC). IEEE (2013)
7. Moslehi, P., et al.: Multi-objective Numeric Association Rules Mining via Ant
Colony Optimization for Continuous Domains without Specifying Minimum Sup-
port and Minimum Confidence. International Journal of Computer Science (2008)
8. Agrawal, R., Shafer, J.C.: Parallel mining of association rules. IEEE Transactions
on Knowledge and Data Engineering 8(6) (1996)
9. Parthasarathy, S., et al.: Parallel data mining for association rules on shared-
memory systems. Knowledge and Information Systems 3(1) (2001)
10. Fang, W., et al.: Frequent itemset mining on graphics processors. In: Proceedings of
the Fifth International Workshop on Data Management on New Hardware. ACM
(2009)
11. Zhou, J., Yu, K.-M., Wu, B.-C.: Parallel frequent patterns mining algorithm on
GPU. In: IEEE International Conference on Systems Man and Cybernetics (SMC).
IEEE (2010)
12. Adil, S.H., Qamar, S.: Implementation of association rule mining using CUDA. In:
International Conference on Emerging Technologies, ICET 2009. IEEE (2009)
13. Cui, Q., Guo, X.: Research on Parallel Association Rules Mining on GPU. In:
Yang, Y., Ma, M. (eds.) Proceedings of the 2nd International Conference on Green
Communications and Networks. LNEE, vol. 224, pp. 215–222. Springer, Heidelberg
(2012)
14. Zhang, F., Zhang, Y., Bakos, J.: Gpapriori: Gpu-accelerated frequent itemset min-
ing. In: IEEE International Conference on Cluster Computing (CLUSTER). IEEE
(2011)
15. Li, H., Zhang, N.: Mining maximal frequent itemsets on graphics processors. In:
Seventh International Conference on Fuzzy Systems and Knowledge Discovery
(FSKD), vol. 3. IEEE (2010)
16. Guvenir, H.A., Uysal, I.: Bilkent university function approximation repository,
2012-03-12 (2000), http://funapp.cs.bilkent.edu.tr/DataSets
17. Goethals, B., Zaki, M.J.: Frequent itemset mining implementations repository. This
site contains a wide-variety of algorithms for mining frequent, closed, and maximal
itemsets (2003), http://fimi.cs.helsinki.fi
A Method for Ripple Simulation Based on GPU
Xianjun Chen1,2, Yanmei Wang3,*, and Yongsong Zhan1

1
Guangxi Key Laboratory of Trusted Software, Guilin University of Electronic Technology,
Guilin, Guangxi, 541004, PR China
2
Information Engineering School, Haikou College of Economics,
Haikou, Hainan, 571127, PR China
3
QiongTai Teachers College, Haikou, Hainan, 571127, PR China
{Yanmei Wang,hingini}@126.com
Abstract. To improve the simulation of ripple on a personal workstation, a

novel vector algebra model based on Graphic Process Unit (GPU) is proposed.
First, the data structures and rules for data operation are established to meet the
needs of vector algebra model. Second, the physical equation governing ripple
motion is transformed discretely for vector multiplication, which will be solved
by the Conjugate Gradient Method. Finally, the simulation of ripple is achieved
from the height map providing normal information used by the calculation of
light reflection and refraction in real time. Experiment results show that the
method is robust and efficient to achieve real-time ripple simulation by making
full use of the excellent computation power of programmable GPU.
Keywords: Ripple simulation, GPU, Vector algebra operation, Conjugate

Gradient Method.
1 Introduction
As one of the most intriguing problems in computer graphics, the simulation of ripple
has drawn the attention of a great sum of researchers. Ripples are everywhere in the
nature, ranging from the streams to the rivers, from the pools to the oceans. In the
applications of computer games and virtual reality, it is necessary to immerse players
into plausible virtual worlds, which shall be constructed by the photorealistic
simulation of natural scenes, such as ripple, smoke, and so on. Also, animators can
also benefit from ripple simulation to achieve realistic effects in real time and
improve the product efficiency of cartoon. With developing computer graphics, there
exist many models that attempt to fake fluid-like effects. However, it is not easy to
mimic the complexities and subtleties of ripple motion in a convincing manner on a
graphic workstation in real time.
In this paper, a novel GPU based vector algebra operation model is proposed to
improve the simulation of ripple, which is physically described by the fluid equation.
First of all, the data structures and rules for data operation are established to meet the
*
A Method for Ripple Simulation Based on GPU 59
needs of vector algebra operation model. Then, the fluid dynamics equation governing
ripple is transformed discretely for vector multiplication, which is to be solved by
Conjugate Gradient Method. Finally, ripple rendering is achieved from the height map
providing normal information used by the calculation of light reflection and
refraction.
The rest of the paper is organized as follows. Related work is discussed in section
2. Section 3 introduces the GPU based vector algebra operation, including the data
structures and rules for data operation. Section 4 gives a description on the Conjugate
Gradient Method and the dynamics equation depicting water wave. Discretization of
the fluid equation for vector multiplication is also included in this section. Section 5
presents the experimental result of a running system instance. Conclusions are drawn
in section 6.
2 Related Work
In the graphics community, the early work for water phenomenon modeling placed
emphasis on representations of the water surface as a parametric function, which
could be animated over time to simulate wave transport[1-3]. But they are unable to
easily deal with complex three-dimensional behaviors such as flow around objects
and dynamically changing boundaries. To obtain water models which could be used
in a dynamic animation environment, researchers turned to use two-dimensional
approximations to the full 3D fluid equations[4]. Kass[5] approximated the 2D
shallow water equations to get a dynamic height field surface that interacted with a
static ground object. A pressure defined height field formulation was used by Chen[6]
in fluid simulations with moving obstacles. O’Brien[7] simulated splashing liquids by
combining a particle system and height field, while Miller[8] used viscous springs
between particles to achieve dynamic flow in 3D. Terzopoulos[9] simulated melting
deformable solids using a molecular dynamics approach to simulate the particles in
the liquid phase. The simulation of complex water effects using the full 3D Navier-
Stokes equations has been based upon the large amount of research done by the
computational fluid dynamics community over the past 60 years. Foster[10]
developed a 3D Navier-Stokes methodology for the realistic animation of liquids.
Stam[11] replaced their finite difference scheme with a semi-Lagrangian method to
achieve significant performance improvements at the cost of increased rotational
damping. Foster[12] made significant contributions to the simulation and control of
three dimensional fluid simulations through the introduction of a hybrid liquid volume
model combining implicit surfaces and massless marker particles, the formulation of
plausible boundary conditions for moving objects in a liquid, the use of an efficient
iterative method to solve for the pressure, and a time step subcycling scheme for the
particle and implicit surface evolution equations.
On the other hand, graphics hardware has undergone a true revolution in the past
ten years. It went from being a simple memory device to a configurable unit and a
fully programmable parallel processor. Although designed for fast polygon rendering,
graphics hardware has been extended to various applications of general-purpose
60 X. Chen, Y. Wang, and Y. Zhan
computations[13-14]. Harris[15] have implemented on the GPU the coupled map

lattice, and have simulated cloud dynamics using partial differential equations.
Goodnight[16] have implemented the multi-grid method on the GPU, and have
applied it to heat transfer and modeling of fluid mechanics. Boltz[17] have also
developed a GPU-based multi-grid solver. In addition, a conjugate-gradient solver is
proposed on the GPU with a sparse matrix representation, which has been applied to
the Navier-Stokes equations. At the same time, Kruger[18] has presented the GPU
implementation of several linear algebra operators, which have been used to solve the
Navier-Stokes equations as well. In all the above works, the computation of fluid
dynamics is translated from the CPU to the GPU.
3 Vector Algebra Operation
As a main problem in the field of applied mathematics, the numerical solution for
differential equations has been of prime importance in many applications of physical
simulation and image procession. Transformed discretely to be linear, the differential
equations are now widely used by 3D graphics applications for natural phenomena
simulation. To solve the linear equations on GPU, the model of vector algebra
operation is proposed to be composed of data structures and rules for data operation,
both of which can be implemented by object-oriented program and extended freely.
3.1 Data Structures

The proposed model consists of three data types: float, vector and matrix. With the
improvement of architecture in these years, the current GPU have been good at
texture access, making it proper to define a float as a texture with the size of 1 × 1 . A
vector can be considered as either a one dimension (1D) texture or a two dimension
(2D) texture. The latter is preferable in our model, for it has a better GPU support
than the former. In practice, most of the current GPU set a limit to the total of 1D
texture, namely 4096, which is not applicable for the complex numerical solution of
differential equation. Furthermore, during the course of linear equation solving, the
median is usually defined as vector, which is to be rendered as 2D texture to achieve
high performance. In most computer graphics applications, matrix is usually
considered as 2D texture, resulting in a high texture usage. As many matrices may be
composed of a great majority of zero values and few non-zero values, it is necessary
to define the matrix according to its attribute. The dense matrix can be divided into a
set of vectors, which shall be considered as 2D textures. In the band matrix, the non-
zero values have a distribution of band, which is considered to be a vector denoted by
2D texture as well. As to the sparse matrix, the non-zero values are distributed
randomly, making it improper to be described by texture. To make full use of the
parallelization operation of GPU, the sparse matrix is defined by a vertex buffer as
figure 1.
Following their priority order in the sparse matrix, all the elements are inputted to
the vertex buffer in turn. Each vertex will be provided with 4 non-zero values. The
rows of sparse matrix are indexed by the vertex coordination, which can be adjusted
by program parameter input. The texture coordination of each matrix element shall be
the same as the final vertex coordination acquired by the procession of model
transformation, view transformation and projection. Obviously, it can be seen from
the above structure that each vertex is equipped with 6 texture coordination values,
where (tu_0, tv_0), (tu_1, tv_1), (tu_2, tv_2), (tu_3, tv_4) are used as indexes for the
4 elements, (val0, val1, val2, val3) for their values, and (posX, posY) for index of the
output. This definition is helpful for the multiplication between matrix and vector.
struct SPARSEMATRIXVERTEX
{
FLOAT x,y,z;
FLOAT tu_0,tv_0;
FLOAT tu_1,tv_1;
FLOAT tu_2,tv_2;
FLOAT tu_3,tv_3;
FLOAT val0,val1,val2,val3;
FLOAT posX,posY;
static const DWORD FVF;
};
Fig. 1. Data structures
3.2 Data Operation

The common data operation in the proposed model includes vector operation, vector
reduction and multiplication of matrix and vector. To achieve vector addition, the
following sequence of steps is performed: (1) a view frustum is set to cover a number
of pixels, which is equal to the number of vector elements, and the target vector is set
as Render Target; (2) a quad is rendered to cover the entire view port; (3) rasterization
is implemented by the vertex program, and a mapping relationship is established
between the vector elements and pixels, which can be accessed by the fragment
program via texture coordination; (4) parallel processing of all the pixels by GPU is
accomplished by fragment program, whose output is the target texture. The other
vector operations, such as vector subtraction, multiplication between vector and scalar,
and etc. can be implemented in the same way.
The operation of vector reduction is to count all the element values. Given a vector
defined as a texture of n × n , reduction can be implemented in log(n) rendering
cycles. In the fragment program, the mean value of 4 neighboring texels is calculated
and written into the Render Target, which is considered as input for the next rendering
cycle. The above operation will repeat till the completion of reduction, as shown in
figure 2.
Fig. 2. Data operation
4 Ripple Rendering
The motion of ripple is controlled by 2D wave equation, which is of prime importance

for the vivid simulation of wave animation. The numerical solution for this physical
equation is as follows: (1) the equation is transformed to be linear by discretization;
(2) the linear equation is solved by the Conjugate Gradient Method, which is
completely based on the proposed vector algebra operation model accelerated by
GPU.
4.1 Conjugate Gradient Method

As an applicable iteration process for solving linear equations, the Conjugate Gradient
Method mainly deals with the cases with symmetric positive definite coefficient
n×n
matrix. Given the linear equations as Ax = b , where A ∈ R , the vector
sequences of iteration and remainder are constructed, and the search directions are
updated. These operations shall be iterated to achieve the ideal computation precision.
The procedure for Conjugate Gradient Method includes a pre-computation and a
main cycle. In the course of pre-computation, the following steps shall be performed:
(1) given ∀x0 ∈ R n ; (2) r0 = b − Ax0 ; (3) p0 = r0 ; (4) a threshold constant ε
is set, and k = 0 .
4.2 Wave Equation

The inherent relationship between wave height and velocity, time and space can be
described by the physical dynamics equation, which is defined as following.
∂2 y ∂2 y ∂2 y
c2 ( + ) =
∂x 2 ∂z 2 ∂t 2
where x-z is the water surface, y is height, t is time and c is wave velocity. To achieve
the numerical solution for wave equation, it shall be transformed to algebraic equation
by Taylor series and Centre Differentia Method, which can be defined as follows.
f ( x) = f ( x + Δh) + f ′( x + Δh) × Δh + Ω(Δh)
 yi +1, j − yi , j
 + Ο(Δh)
 y Δ h
∂y  i , j − yi −1, j
 ( )i, j =  + Ο(Δh)
∂h  Δh
 yi +1, j − yi −1, j + Ο(Δh) 2
 2Δh
∂ y
2
yi +1, j − 2 yi , j + yi −1, j
 ( 2 )i, j = + Ο(Δh) 2
∂x ( Δh) 2
Finally, the 2D discrete wave equation can be denoted as follows, where

c 2 ( Δt ) 2
β=
(Δh) 2 .
yit,+j1 − 2 yit, j + yit,−j1 2 yi +1, j − 2 yi , j + yi −1, j
t t t
yit, j +1 − 2 yit, j + yit, j −1
= c ( + ) x = z =h
( Δt ) 2 ( Δx ) 2 ( Δz ) 2
yit,+j1 − 2 yit, j + yit,−j1 yit+1, j − 2 yit, j + yit−1, j yit, j +1 − 2 yit, j + yit, j −1

⇔ = c2 ( + )
( Δt ) 2 ( Δh ) 2 ( Δh ) 2
⇔ yit,+j1 = β ( yit+1, j + yit−1, j + yit, j +1 + yit, j −1 ) + (2 − 4 β ) yit, j − yit,−j1
5 Experiment
Based on the proposed method, an applicable implemented with C++ computer

program, DirectX 9.0 and the hardware rendering program of HLSL is shown as
figure 3. Furthermore, figure 4 gives a convincing representation of ripple on the
workstation with NVIDIA Quadro Fx540. In this case, the 2D wave surface is
presented discretely as meshes of 512X512, and simulated in real-time with the frame
rate of 16fps.
In order to demonstrate the performance of the proposed model, two common-used
GPU have been employed in the test, including NVIDIA Quadro FX 540 (with the
memory of 256M) and NVIDIA FX 8800 (with the memory of 640M). The resultant
data of table 1-2 show the efficiency and robustness of our method.
Fig. 3. Instance of system implementation Fig. 4. The real-time ripple simulation
Table 1. Test result of NVIDIA Quadro FX 540.
Image Resolution
512 x 512 512 x 256 256 x 256
Vector Reduction(ms) 1.00 0.71 0.62
Vector Addition(ms) 1.44 0.61 0.12
Frame Rate(fps) 17 32 64
Table 2. Test result of NVIDIA FX 8800.
Image Resolution
512 x 512 512 x 256 256 x 256
Vector Reduction(ms) 0.10295 0.10076 0.09052
Vector Addition(ms) 0.01404 0.0139 0.0126
Frame Rate(fps) 295 423 440
6 Conclusion
In this paper, a novel GPU based vector algebra operation model is proposed to
improve the simulation of water surface. The data structures and rules for data
operation are established to meet the needs of vector algebra operation model, and the
fluid dynamics equation governing ripple is transformed discretely for vector
multiplication. Experimental results show the robustness and efficiency of the
proposed method for the real-time simulation of water surface on GPU.
Acknowledgments. This research work is supported by the grant of Guangxi science

and technology development project (No: 1355011-5), the grant of Guangxi Key
Laboratory of Trusted Software of Guilin University of Electronic Technology (No:
kx201309), the grant of Guangxi Education Department (No: SK13LX139) the ，

grant of Guangxi Undergraduate Training Programs for Innovation and
Entrepreneurship (No:20121059519), the grant of Hainan Natural Science Foundation
(No: 613169), and the Universities and colleges Science Research Foundation of
Hainan (No: Hjkj2013-48).
References
1. Fournier, A., Reeves, W.T.: A simple model of ocean waves. In: Proc. SIGGRAPH ACM,
pp. 75–84 (1986)
2. Peachy, D.: Modeling waves and surf. In: Proc. ACM SIGGRAPH, pp. 65–74 (1986)
3. Schachter, B.: Long crested wave models. Computer Graphics and Image Processing 12,
187–201 (1980)
4. Chorin, A.J., Marsden, J.E.: A mathematical introduction to fluid mechanics. Texts in
Applied Mathematics, vol. 4 (1990)
5. Kass, M., Miller, G.: Rapid, stable fluid dynamics for computer graphics. In: Proc. ACM
SIGGRAPH, pp. 49–57 (1990)
6. Chen, J., Lobo, N.: Toward interactive-rate simulation of fluids with moving obstacles
using the Navier-Stokes Equations. Graphical Models and Image Processing 57, 107–116
(1994)
7. O’Brien, J., Hodgins, J.: Dynamic simulation of splashing fluids. In: Proc. Computer
Animation, pp. 198–205 (1995)
8. Miller, G., Pearce, A.: Globular dynamics: a connected particle system for animating
viscous fluids. Computers and Graphics 13, 305–309 (1989)
9. Terzopoulos, D., Platt, J., Fleischer, K.: Heating and melting deformable models (from
goop to glop). In: Proc. Graphics Interface, pp. 219–226 (1989)
10. Foster, N., Metaxas, D.: Realistic animation of liquids. Graphical Models and Image
Processing 58, 471–483 (1996)
11. Stam, J.: Stable fluids. In: Proc. ACM SIGGRAPH, pp. 121–128 (1999)
12. Foster, N., Fedkiw, R.: Practical animation of liquids. In: Proc. ACM SIGGRAPH, pp. 23–
30 (2001)
13. Hoff, K., Culver, T.: Fast computation of generalized voronoi diagrams using graphics
hardware. In: Proc. ACM SIGGRAPH, pp. 277–286 (1999)
14. Trendall, C., Stewart, A.J.: General calculations using graphics hardware with applications
to interactive caustics. In: Proc. Eurographics Workshop on Rendering, pp. 287–298
(2000)
15. Harris, M., Coombe, G.: Physically-based visual simulation on graphics hardware. In:
Proc. SIGGRAPH/Eurographics Workshop on Graphics Hardware, pp. 109–118 (2002)
16. Goodnight, N., Woolley, C.: A multi-grid solver for boundary value problems using
programmable graphics hardware. In: Proc. SIGGRAPH/Eurographics Workshop on
Graphics Hardware (2003)
17. Bolz, J., Farmer, I., Grinspun, E.: Sparse matrix solvers on the GPU: conjugate gradients
and multi-grid. ACM Transactions on Graphics 22, 917–924 (2003)
18. Kruger, J., Westermann, R.: Linear algebra operators for GPU implementation of
numerical algorithms. ACM Transactions on Graphics 22, 908–916 (2003)
cuROB: A GPU-Based Test Suit
for Real-Parameter Optimization
Ke Ding and Ying Tan
Key Laboratory of Machine Perception and Intelligence (MOE), Peking University,

Department of Machine Intelligence, School of Electronics Engineering and Computer
Science, Peking University, Beijing, 100871, China
{keding,ytan}@pku.edu.cn
Abstract. Benchmarking is key for developing and comparing optimiza-

tion algorithms. In this paper, a GPU-based test suit for real-parameter
optimization, dubbed cuROB, is introduced. Test functions of diverse
properties are included within cuROB and implemented efficiently with
CUDA. Speedup of one order of magnitude can be achieved in compari-
son with CPU-based benchmark of CEC’14.
Keywords: Optimization Methods, Optimization Benchmark, GPU,

CUDA.
1 Introduction
Proposed algorithms are usually tested on benchmark for comparing both per-
formance and efficiency. However, as it can be a very tedious task to select and
implement test functions rigorously. Thanks to GPUs’ massive parallelism, a
GPU-based optimization function suit will be beneficial to test and compare
optimization algorithms.
Based on the well known CPU-based benchmarks presented in [1,2,3], we
proposed a CUDA-based real parameter optimization test suit, called cuROB,
targeting on GPUs. We think cuROB can be helpful for assessing GPU-based
optimization algorithms, and hopefully, conventional CPU-based algorithms can
benefit from cuROB’s fast execution.
Considering the fact that research on the single objective optimization algo-
rithms is the basis of the research on the more complex optimization algorithms
such as constrained optimization algorithms, multi-objective optimizations al-
gorithms and so forth, in this first release of cuROB a suit of single objective
real-parameter optimization function are defined and implemented.
The test functions are selected according to the following criteria: 1) the func-
tions should be scalable in dimension so that algorithms can be tested under
various complexity; 2) the expressions of the functions should be with good
parallelism, thus efficient implementation is possible on GPUs; 3) the functions
should be comprehensible such that algorithm behaviours can be analysed in

To whom the correspondence should be addressed.

cuROB: A GPU-Based Test Suit for Real-Parameter Optimization 67
the topological context; 4) last but most important, the test suit should cover
functions of various properties in order to get a systematic evaluation of the
optimization algorithms.
The source code and a sample can be download from code.google.com/
p/curob/.
1.1 Symbol Conventions and Definitions

Symbols and definitions used in the report are described in the following. By
default, all vectors refer to column vectors, and are depicted by lowercase letter
and typeset in bold.
– [·] indicates the nearest integer value
– · indicates the largest integer less than or equal to
– xi denotes i-th element of vector x
– f (·), g(·) and G(·) multi-variable functions
– fopt optimal (minimal) value of function f
– xopt optimal solution vector, such that f (xopt ) = fopt
– R normalized orthogonal matrix for rotation
– D dimension
– 1 = (1, . . . , 1)T all one vector
1.2 General Setup

The general setup of the test suit is presented as follows.
– Dimensions The test suit is scalable in terms of dimension. Within the
hardware limit, any dimension D ≥ 2 works. However, to construct a real
hybrid function, D should be at least 10.
– Search Space All functions are defined and can be evaluated over RD ,
while the actual search domain is given as [−100, 100]D .
– fopt All functions, by definition, have a minimal value of 0, a bias (fopt )
can be added to each function. The selection can be arbitrary, fopt for each
function in the test suit is listed in Tab. 1.
– xopt The optimum point of each function is located at original. xopt which
is randomly distributed in [−70, 70]D , is selected as the new optimum.
– Rotation Matrix To derive non-separable functions from separable ones,
the search space is rotated by a normalized orthogonal matrix R. For a given
function in one dimension, a different R is used. Variables are divided into
three (almost) equal-sized subcomponents randomly. The rotation matrix for
each subcomponent is generated from standard normally distributed entries
by Gram-Schmidt orthonormalization. Then, these matrices consist of the
R actually used.
1.3 CUDA Interface and Implementation

A simple description of the interface and implementation is given in the following.
For detail, see the source code and the accompanied readme file.
68 K. Ding and Y. Tan
Interface. Only benchmark.h need to be included to access the test func-

tions, and the CUDA file benchmark.cu need be compiled and linked. Before
the compiling start, two macro, DIM and MAX CONCURRENCY should be
modified accordingly. DIM defines the dimension of the test suit to used while
MAX CONCURRENCY controls the most function evaluations be invoked con-
currently. As memory needed to be pre-allocated, limited by the hardware, don’t
set MAX CONCURRENCY greater than actually used.
Host interface function initialize () accomplish all initialization tasks, so must
be called before any test function can be evaluated. Allocated resource is released
by host interface function dispose ().
Both double precision and single precision are supported through
func evaluate () and func evaluatef () respectively. Take note that device pointers
should be passed to these two functions. For the convenience of CPU code, C in-
terfaces are provided, with h func evaluate for double precision and
h func evaluatef for single precision. (In fact, they are just wrappers of the GPU
interfaces.)
Efficiency Concerns. When configuration of the suit, some should be taken

care for the sake of efficiency. It is better to evaluation a batch of vectors than
many smaller. Dimension is a fold of 32 (the warp size) can more efficient. For
example, dimension of 96 is much more efficient than 100, even though 100 is
little greater than 96.
1.4 Test Suite Summary

The test functions fall into four categories: unimodal functions, basic multi-
modal functions, hybrid functions and composition functions. The summary of
the suit is listed in Tab. 1. Detailed information of each function will given in
the following sections.
2 Speedup
Under different hardware, various speedups can be achieved. 30 functions are
the same as CEC’14 benchmark. We test the cuROB’s speedup with these 30
functions under the following settings: Windows 7 SP1 x64 running on Intel
i5-2310 CPU with NVIDIA 560 Ti, the CUDA version is 5.5. 50 evaluations
were performed concurrently and repeated 1000 runs. The evaluation data were
generated randomly from uniform distribution.
The speedups with respect to different dimension are listed by Tab. 2 (sin-
gle precision) and Tab. 3 (double precision). Notice that the corresponding di-
mensions of cuROB are 10, 32, 64 and 96 respectively and the numbers are as
in Tab. 1
Fig. 1 demonstrates the overall speedup for each dimension. On average,
cuROB is never slower than its CPU-base CEC’14 benchmark, and speedup of
one order of magnitude can be achieved when dimension is high. Single precision
is more efficient than double precision as far as execution time is concerned.
Table 1. Summary of cuROB’s Test Functions
No. Functions ID Description

0 Rotated Sphere SPHERE Optimum easy
1 Rotated Ellipsoid ELLIPSOID to track
Unimodal 2 Rotated Elliptic ELLIPTIC
Functions 3 Rotated Discus DISCUS
4 Rotated Bent Cigar CIGAR Optimum hard
5 Rotated Different Powers POWERS to track
6 Rotated Sharp Valley SHARPV
7 Rotated Step STEP
8 Rotated Weierstrass WEIERSTRASS With
9 Rotated Griewank GRIEWANK adepuate
10 Rastrigin RARSTRIGIN U global
11 Rotated Rastrigin RARSTRIGIN structure
12 Rotated Schaffer’s F7 SCHAFFERSF7
Basic
13 Rotated Expanded Griewank plus Rosenbrock GRIE ROSEN
Multi-modal
14 Rotated Rosenbrock ROSENBROCK
Functions
15 Modified Schwefel SCHWEFEL U
16 Rotated Modified Schwefel SCHWEFEL
With
17 Rotated Katsuura KATSUURA
weak
18 Rotated Lunacek bi-Rastrigin LUNACEK
global
19 Rotated Ackley ACKLEY
structure
20 Rotated HappyCat HAPPYCAT
21 Rotated HGBat HGBAT
22 Rotated Expanded Schaffer’s F6 SCHAFFERSF6
23 Hybrid Function 1 HYBRID1
24 Hybrid Function 2 HYBRID2 With different
Hybrid 25 Hybrid Function 3 HYBRID3 properties for
Functions 26 Hybrid Function 4 HYBRID4 different variables
27 Hybrid Function 5 HYBRID5 subcomponents
28 Hybrid Function 6 HYBRID6
29 Composition Function 1 COMPOSITION1
30 Composition Function 2 COMPOSITION2 Properties similar
31 Composition Function 3 COMPOSITION3 to particular
Composition 32 Composition Function 4 COMPOSITION4 sub-function
Functions 33 Composition Function 5 COMPOSITION5 when approaching
34 Composition Function 6 COMPOSITION6 the corresponding
35 Composition Function 7 COMPOSITION7 optimum
36 Composition Function 8 COMPOSITION8
Search Space: [−100, 100]D , fopt = 100
Table 2. Speedup (single Precision)
D NO.3 NO.4 NO.5 NO.8 NO.9 NO.10 NO.11 NO.13 NO.14 NO.15
10 0.59 0.20 0.18 12.23 0.49 0.28 0.31 0.32 0.14 0.77
32 3.82 2.42 2.00 47.19 3.54 1.67 3.83 5.09 2.06 3.54
64 4.67 2.72 2.29 50.17 3.56 0.93 3.06 2.88 2.20 3.39
94 13.40 10.10 8.50 84.31 11.13 1.82 9.98 9.66 8.75 6.73
10 0.80 3.25 0.36 0.20 0.26 0.45 0.63 0.44 2.80 0.52
32 5.57 10.04 3.46 1.22 1.42 6.44 3.95 3.43 11.47 3.36
64 5.45 13.19 3.27 2.10 2.27 3.81 4.62 3.07 14.17 3.34
96 14.38 23.68 11.32 8.26 8.49 11.60 13.67 10.64 30.11 10.71
10 0.65 0.72 0.70 0.55 0.71 3.49 3.50 0.84 1.28 0.70
32 2.73 3.09 3.63 3.10 4.10 12.39 12.51 5.25 5.19 3.33
64 3.86 4.01 3.21 2.67 3.38 12.68 12.63 3.80 5.27 3.13
96 12.04 11.32 8.15 6.27 8.49 23.67 23.64 9.50 11.79 7.93
Table 3. Speedup (Double Precision)
10 0.56 0.19 0.17 9.04 0.43 0.26 0.29 0.30 0.14 0.75
32 3.78 2.43 1.80 33.37 3.09 1.59 3.52 4.81 1.97 3.53
64 4.34 2.49 1.93 30.82 3.15 0.92 2.87 2.74 2.11 3.29
96 12.27 9.24 6.95 46.01 9.72 1.78 9.62 8.74 7.87 5.92
10 0.79 2.32 0.34 0.18 0.26 0.45 0.59 0.43 1.97 0.52
32 5.10 6.79 3.28 1.13 1.29 6.10 3.63 3.14 8.15 3.23
64 4.75 8.29 3.06 1.99 2.18 3.32 4.02 2.77 9.80 2.92
96 11.91 13.81 9.75 7.37 7.78 10.24 11.55 9.57 20.81 9.40
10 0.79 2.32 0.34 0.18 0.26 0.45 0.59 0.43 1.97 0.52
32 5.10 6.79 3.28 1.13 1.29 6.10 3.63 3.14 8.15 3.23
64 4.75 8.29 3.06 1.99 2.18 3.32 4.02 2.77 9.80 2.92
96 11.91 13.81 9.75 7.37 7.78 10.24 11.55 9.57 20.81 9.40
Fig. 1. Overall Speedup
3 Unimodal Functions
3.1 Shifted and Rotated Sphere Function

D
f1 (x) = z2i + fopt (1)
i=1
where z = R(x − xopt ).
Properties
– Unimodal
– Non-separable
– Highly symmetric, in particular rotationally invariant
3.2 Shifted and Rotated Ellipsoid Function

D
f4 (x) = i · z2i + fopt (2)
i=1
Properties
– Unimodal
– Non-separable
3.3 Shifted and Rotated High Conditioned Elliptic Function

D
i−1
f2 (x) = (106 ) D−1 z2i + fopt (3)
i=1
Properties
– Unimodal
– Non-separable
– Quadratic ill-conditioned
– Smooth local irregularities
3.4 Shifted and Rotated Discus Function

D
f5 (x) = 10 · 6
z21 + z2i + fopt (4)
i=2
Properties
– Unimodal
– Non-separable
– Smooth local irregularities
– With One sensitive direction
3.5 Shifted and Rotated Bent Cigar Function

D
f6 (x) = z21 + 10 · 6
z2i + fopt (5)
i=2
Properties
– Unimodal
– Non-separable
– Optimum located in a smooth but very narrow valley
3.6 Shifted and Rotated Different Powers Function

D

f4 (x) =
i−1
|zi |2+4 D−1 + fopt (6)
i=1
where z = R(0.01(x − xopt )).

Properties
– Unimodal
– Non-separable
– Sensitivities of the zi -variables are different
3.7 Shifted and Rotated Sharp Valley Function

D

f4 (x) = zi + 100 ·
2
z2i + fopt (7)
i=2
Properties
– Unimodal
– Non-separable
– Global optimum located in a sharp (non-differentiable) ridge
4 Basic Multi-modal Functions

4.1 Shifted and Rotated Step Function

D
f3 (x) = zi + 0.52 + fopt (8)
i=1
where z = R(x − xopt )
Properties
– Many Plateaus of different sizes
– Non-separable
4.2 Shifted and Rotated Weierstrass Function
k

D
max k
max
f9 (x) = a cos (2πb (zi + 0.5)) −D·

k k
ak cos (2πbk ·0.5)+fopt (9)
i=1 k=0 k=0
where a = 0.5, b = 3, kmax = 20, z = R(0.005 · (x − xopt )).
Properties
– Multi-modal
– Non-separable
– Continuous everywhere but only differentiable on a set of points
4.3 Shifted and Rotated Griewank Function
D
z2i
D
zi
f10 (x) = − cos( √ ) + 1 + fopt (10)
i=1
4000 i=1 i
where z = R(6 · (x − xopt )).
Properties
– Multi-modal
– Non-separable
– With many regularly distributed local optima
4.4 Shifted Rastrigin Function

D
2
f11 (x) = zi − 10 cos(2πzi ) + 10 · D + fopt (11)
i=1
where z = 0.0512 · (x − x opt
).
Properties
– Multi-modal
– Separable
4.5 Shifted and Rotated Rastrigin Function

D
2
f12 (x) = zi − 10 cos(2πzi ) + 10 + fopt (12)
i=1
where z = R(0.0512 · (x − x opt
)).
Properties
– Multi-modal
– Non-separable
4.6 Shifted Rotated Schaffer’s F7 Function
2
1 √
D−1
f17 (x) = (1 + sin (50 · wi )) · wi
2 0.2
+ fopt (13)
D − 1 i=1

where wi = z2i + z2i+1 , z = R(x − xopt ).
Properties
– Multi-modal
– Non-separable
4.7 Expanded Griewank plus Rosenbrock Function
Rosenbrock Function: g2 (x, y) = 100(x2 − y)2 + (x − 1)2

Griewank Function: g3 (x) = x2 /4000 − cos(x) + 1

D−1
f18 (x) = g3 (g2 (zi , zi+1 )) + g3 (g2 (zD , z1 )) + fopt (14)
i=1
where z = R(0.05 · (x − xopt )) + 1.
Properties
– Multi-modal
– Non-separable
–
4.8 Shifted and Rotated Rosenbrock Function

D−1

f7 (x) = 100 · (z2i − zi+1 )2 + (zi − 1)2 + fopt (15)
i=1
where z = R(0.02048 · (x − xopt )) + 1.
Properties
– Multi-modal
– Non-separable
– With a long, narrow, parabolic shaped flat valley from local optima to global
optima
4.9 Shifted Modified Schwefel Function

D
f13 (x) = 418.9829 × D − g1 (wi ), wi = zi + 420.9687462275036 (16)
i=1
⎧
⎪
⎪ wi · sin( |wi |) if |wi | ≤ 500
⎨
(wi −500)2
g1 (wi ) = (500 − mod(wi , 500)) · sin 500 − mod(wi , 500) − 10000D if wi > 500
⎪
⎩(mod(−w , 500) − 500) · sin 500 − mod(−w , 500) − (wi +500)2
⎪
if wi < −500
i i 10000D
(17)
where z = 10 · (x − xopt ).
Properties
– Multi-modal
– Separable
– Having many local optima with the second better local optima far from the
global optima
4.10 Shifted Rotated Modified Schwefel Function

D
f14 (x) = 418.9829 × D − g1 (wi ), wi = zi + 420.9687462275036 (18)
i=1
where z = R(10 · (x − xopt )) and g1 (·) is defined as Eq. 17.
Properties
– Multi-modal
– Non-separable
– Having many local optima with the second better local optima far from the
global optima
4.11 Shifted Rotated Katsuura Function
10
D 32
|2j · zi − [2j · zi ]| 10 10
f15 (x) = 2
(1 + i j
) D1.2 − 2 + fopt (19)
D i=1 j=1
2 D
where z = R(0.05 · (x − xopt )).
Properties
– Multi-modal
– Non-separable
– Continuous everywhere but differentiable nowhere
4.12 Shifted and Rotated Lunacek bi-Rastrigin Function

D

D
f12 (x) = min (zi − μ1 )2 , dD + s (zi − μ2 )2 ) + 10 · (D − cos(2π(zi − μ1 ))) + fopt (20)
i=1 i=1 i=1
where z = R(0.1 · (x − xopt ) + 2.5 ∗ 1), μ1 = 2.5, μ2 = −2.5, d = 1, s = 0.9.
Properties
– Multi-modal
– Non-separable
– With two funnel around μ1 1 and μ2 1
4.13 Shifted and Rotated Ackley Function

⎛
⎞

1 D
1 D
f8 (x) = −20 · exp ⎝−0.2 x2 ⎠ − exp cos(2πxi ) + 20 + e + fopt
D i=1 i D i=1
(21)
Properties
– Multi-modal
– Non-separable
– Having many local optima with the global optima located in a very small
basin
4.14 Shifted Rotated HappyCat Function

D
1 2
D D
f16 (x) = | z2i − D|0.25 + ( zj + zj )/D + 0.5 + fopt (22)
i=1
2 j=1 j=1
where z = R(0.05 · (x − xopt )) − 1.

Properties
– Multi-modal
– Non-separable
– Global optima located in curved narrow valley
4.15 Shifted Rotated HGBat Function
D D
1 2
D D
f17 (x) = |( z2i )2 − ( zj )2 |0.5 + ( zj + zj )/D + 0.5 + fopt (23)
i=1 j=1
2 j=1 j=1
where z = R(0.05 · (x − xopt )) − 1.

Properties
– Multi-modal
– Non-separable
– Global optima located in curved narrow valley
4.16 Expanded Schaffer’s F6 Function

sin2 ( x2 + y 2 ) − 0.5
Schaffer’s F6 Function: g4 (x, y) = + 0.5
(1 + 0.001 · (x2 + y 2 ))2

D−1
f19 (x) = g4 (zi , zi+1 ) + g4 (zD , z1 ) + fopt (24)
i=1
Properties
– Multi-modal
– Non-separable
5 Hybrid and Composition Functions

Hybrid functions are constructed according to [3]. For each hybrid function, the
variables are randomly divided into subcomponents and different basic functions
(unimodal and multi-modal) are used for different subcomponents.
Composition functions are constructed in the same manner as in [2,3]. The
constructed functions are multi-modal and non-separable and merge the proper-
ties of the sub-functions better and maintains continuity around the global/local
optima. The local optimum which has the smallest bias value is the global op-
timum. The optimum of the third basic function is set to the origin as a trip
in order to test the algorithms’ tendency to converge to the search center. Note
that, the landscape is not only changes along with the selection of basic function,
but the optima, σ and λ can effect it greatly.
The detailed specifications of hybrid and composition functions can be found
in the extended version of this paper1 , along with illustrations for all 2-D func-
tions except hybrid functions.
Acknowledgements. This work was supported by National Natural Science

Foundation of China (NSFC), Grant No. 61375119, 61170057 and 60875080.
References
1. Finck, S., Hansen, N., Ros, R., Auger, A.: Real-parameter black-box optimization
benchmarking 2010: Noiseless functions definitions. Technical Report 2009/20, Re-
search Center PPE (2010)
2. Liang, J.J., Qu, B.Y., Suganthan, P.N., Hernández-Dı́az, A.G.: Problem definitions
and evaluation criteria for the cec 2013 special session and competition on real-
parameter optimization. Technical Report 201212, Computational Intelligence Lab-
oratory, Zhengzhou University and Nanyang Technological University, Singapore
(2013)
3. Liang, J.J., Qu, B.Y., Suganthan, P.N.: Problem definitions and evaluation criteria
for the cec 2014 special session and competition on single objective real-parameter
numerical optimization. Technical Report 201311, Computational Intelligence Lab-
oratory, Zhengzhou University and Nanyang Technological University, Singapore
(2013)
1
Download from http://arxiv.org/abs/1407.7737
A Particle Swarm Optimization Based Pareto
Optimal Task Scheduling in Cloud Computing
A.S. Ajeena Beegom1 and M.S. Rajasree2

1
Dept. of Computer Science and Engineering,
College of Engineering, Trivandrum, India
ajeena@cet.ac.in
2
IIITM-K, Trivandrum, India
rajasree.ms@iiitmk.ac.in
Abstract. Task scheduling in Cloud computing is a challenging aspect

due to the conflicting requirements of end users of cloud and the Cloud
Service Provider (CSP). The challenge at the CSP’s end is to schedule
tasks submitted by the cloud users in an optimal way such that it should
meet the quality of service (QoS) requirements of the user at one end and
the running costs of the infrastructure to a minimum level at the other
end for better profit. The focus is on two objectives, makespan and cost,
to be optimized simultaneously using meta heuristic search techniques
for scheduling independent tasks. A new variant of continuous Particle
Swarm Optimization (PSO) algorithm, named Integer-PSO, is proposed
to solve the bi-objective task scheduling problem in cloud which out
performs the smallest position value (SPV) rule based PSO technique.
Keywords: Cloud Computing, Task Scheduling, Particle Swarm Opti-

mization, Integer-PSO.
1 Introduction
Recently, cloud computing has emerged as an attractive platform for

entrepreneurs as well as researchers in various domains. Cloud computing refers
to leasing computing resources over the Internet. Benefits of using such a set
up include reduced infrastructure cost, reduced overhead and pay only for the
components he has used for the given amount of time. There are many task
scheduling models available in literature for heterogeneous distributed systems
but these models aim at the improvement of specific performance metrics like
throughput and storage. For a cloud computing platform, apart from theses con-
siderations, user satisfaction in terms of QoS and CSP’s profit is to be considered
while scheduling tasks. In this work, we consider the scheduling of large set of
independent tasks of different size. The conflicting objectives of performance op-
timization considered are the overall execution time (makespan) of all the tasks
and the cost of service. The cost include computation cost, communication cost
and over all maintenance cost as well as power consumption cost.

80 A.S. Ajeena Beegom and M.S. Rajasree
Most of the existing task scheduling research in cloud computing addresses

any one factor, namely makespan as done in [5] or cost as seen in [11] and tries
to find an optimal schedule based on that factor alone. But single objective
optimization solutions will try to optimize one objective making another key
factor to worse. In the proposed work, we have used weighted sum approach for
pareto-optimality and uses Particle swarm algorithm to solve the same.
The rest of the paper is organized as follows. In Section 2, we present Re-
lated Work in the domain and Section 3 describes the Mathematical Model and
formally states the optimization problem. Section 4 details Particle Swarm Op-
timization Technique and the proposed Integer-PSO algorithm. Experimental
Results are presented in Section 5 and Section 6 gives the Conclusion.
2 Related Work
Most of the research in computational grids and cloud systems for scheduling
independent tasks to be executed in parallel tries to optimize a single objec-
tive function where the parameters are any one of makespan, profit earned, cost
of service, QoS, energy consumption and average response time. Meta heuristic
search techniques has been tried by many to solve the same. L.Zhu and J. Wu [15]
uses PSO technique combined with Simulated annealing to solve task schedul-
ing problem in a general scenario. M.F.Tasgetiren et al. [8] proposed Smallest
Position Value (SPV) rule based PSO algorithm to solve Single Machine Total
Weighted Tardiness Problem and has used a local search technique to intense
the search. Lei Zhang et al. [14] used this technique to solve task scheduling
problem in grid environment and has given a comparative study on the applica-
tion of PSO technique with Genetic algorithm for achieving minimal completion
time. PSO algorithm has also been used in solving task allocation / scheduling
problems in work flows in cloud [1], [13], [7].
Multi-objective optimization for resource allocation in cloud computing has
been addressed by Feng et al. [3] and uses PSO algorithm to solve the problem.
They have considered total execution time, resource reservation and QoS of each
task as their optimization objective and uses pareto domination mechanism to
find optimal solutions. Lizheng Guo [4] addresses task assignment problem in
cloud computing considering makespan and cost. They use smallest position
value based PSO algorithm for finding an optimal schedule.
Our proposed method can be used to schedule tasks in pubic cloud or in pri-
vate cloud for independent tasks. Our approach is unique because, to the best
our knowledge, pareto optimal task scheduling using particle swarm optimiza-
tion technique has not been addressed in the case of private or public cloud with
independent tasks to be scheduled, but has been proposed for work flow schedul-
ing in hybrid cloud[2]. Our work also proposes a new variant of PSO algorithm
namely Integer - PSO.
A Particle Swarm Optimization Based Pareto Optimal Task Scheduling 81
3 Mathematical Model
Assume an application consists of N independent tasks, n out of N are scheduled
at each time window, where the value of n is limited by the number of available
VMs m and k = N n similar epochs are needed to complete the execution of all
tasks. For each type of VM instance I = [very small, small, medium, large, extra
large, super], the associated cost of usage and the computing power are different.
Let P fj represent the processing power of j th VM instance type where j ranges
from 1 to |I| and Cj represents its cost for unit time. The task length of each
task T ASKi is precomputed and represented as Ti , the time needed to execute
each task in ’very small’ type VM. The optimization objectives for N tasks are :

k
n
M inimize M akespanf n = Ti ∗ P fj ∗ xij f or some jI (1)
p=1 i=1

k
n
M inimize Costf n = Cj ∗ Ti ∗ P fj ∗ xij f or some jI (2)
p=1 i=1
where xij is a decision variable, denoting T ASKi is scheduled on V Mj and n
tasks are there in an epoch. subject to the following constraints:
n≤m (3)

1 if T ASKi scheduledto V Mj
xij = (4)
0 otherwise
and

n
xij = m (5)
i=1
3.1 Pareto Optimality

A multi objective optimization problem consists of optimizing a vector of nobj
where the objective function F (x) = (f1 (x), f2 (x), . . . , fnobj (x)). The problem
here is to find an optimal task schedule considering both the objectives. i.e.,cost
and makespan. We have used weighted sum approach[9], [12] for solving the
bi-objective optimization problem, which in essence convert a multi-objective
optimization problem to a single objective one with weights representing prefer-
ences among objectives by the decision maker. This approach is easy to solve and
produce a single solution to the problem. Hence our bi-objective optimization
problem can now be represented using the formula:
M inimize θ ∗ Costf n + (1 − θ) ∗ M akespanf n (6)
where θ represents the relative weight or preference of one objective over the
other, in the range [0, 1]. When θ = 0, the optimization problem becomes that
of minimizing Makespanfn and when θ = 1, the problem becomes minimizing
Costfn.
4 Particle Swarm Optimization Technique
Finding an optimal schedule meeting the constraints of a bi-objective optimiza-

tion problem are well-known problems in N P hard category, hence one of the
heuristic techniques, Particle Swarm Optimization[6] is applied to obtain a fea-
sible solution in reasonable time. Initially, the PSO algorithm generates a set of
N solutions called particles, randomly in the D dimensional search space. Each
particle is represented by a D-dimensional vector Xi where i ranges from 1 to d
which stands for its location (xi1 , xi2 , ..., xid ) in space. Velocity of each particle
v is constrained by vmin and vmax and its position x is updated according to
the following equations:
n+1
vid = w ∗ vid
n
+ c1 ∗ rand1 ∗ (pbestnid − xnid ) + c2 ∗ rand2 ∗ (gbestnd − xnid ) (7)
xn+1
id
n+1
= xnid + vid (8)
where i = 1, 2, . . . , N ; n = 1, 2, . . . , itermax , the maximum iteration number,
w, the inertia weight; c1 and c2 are two positive constants called acceleration
coefficients and rand1 and rand2 are two uniformly distributed random numbers
in the interval [0, 1]. Each particle maintains its position and its velocity. It also
remembers the best fitness value it has achieved thus far during the search
(individual best fitness) and the candidate solution that achieved this fitness
(individual best position (pbest)). Also, the PSO algorithm maintains the best
fitness value achieved among all particles in the swarm (global best fitness) and
the candidate solution that achieved this fitness (global best position (gbest)).
Equations (7) and (8) enable the particles to search around its individual best
position pbest and update global best position gbest. This technique was initially
proposed for solving problems in the continuous domain through the velocity
updating rule. Since our problem work in the discrete domain, it has to be
modified to suit the discrete domain. The Smallest Position Value rule based PSO
(PSO-SPV) algorithm [8] is widely used for the same. This technique performs
poor when there exist high variance in the length of the tasks submitted by end-
users and when high variance exists in computational speed of resources[10]. We
Algorithm 1. Abstract of Particle Swarm Optimization Algorithm

1: P ← Initial Population
2: Evaluate (P)
3: Initialize pbest and gbest
4: while termination criterion not met do
5: Update Velocity(P) as indicated in equation (7)
6: Update Position (P) as indicated in equation (8)
7: Evaluate (P)
8: Find pbest and gbest
9: end while
10: Output gbest
too observed that the same technique is not able to converge to near optimal
solution with bi-objective optimization of task scheduling in cloud computing.
Hence a new method for generating discrete permutations is proposed, namely
integer-PSO. Here permutation encoding technique is used where every VM is
assigned a number from 1 to n and a solution sequence (5, 2, 1, 3, 4) means assign
Task 1 to VM 5, Task 2 to VM 2 and so on. Initial populations are randomly
generated. Each solution is evaluated to find its fitness based on equation (6) on
different values of θ.
4.1 Integer – PSO
An update in the position of the particle based on equations (7) and (8) should
result in new task assignment for a scheduling problem, but they produce floating
point values in the continuous domain. Many discrete versions of PSO rounds-off
the floating point position values and stores the discrete integer value for the
particle’s position. To preserve the stochastic nature of the continuous PSO, we
have modified equation (8) in our algorithm, as shown below:
Yidn = ceil((xnid + vid

n
) ∗ β) where β = 10y (9)
P osn+1
id = (Yidn ) mod m (10)

P osn+1 if P osn+1 >0
xcn+1
id = id id
(11)
m otherwise
Table 1. Integer-PSO Example
Task J 1 2 3 4 5
xkij 4 5 1 3 2
k+1
vij -0.6015 -0.2413 0.0327 -0.0352 -0.8544
Yijk+1 33985 47587 10327 29648 11456
P osk+1
ij 0 2 2 3 1
xck+1
ij 5 2 2 3 1
xk+1
ij 5 4 2 3 1
New variables Yidn and P osn+1

id are introduced to store the continuous value as
an integer of required accuracy and the temporary task assignment respectively.
This method may create more than one task assignment to some of the VMs
and may not assign any task to some other VMs. which need to be handled
separately. The procedure is as shown in Table 1, assuming m = 5 and k=4.
(a) Bi-objective optimization (Cost and Makespan) using

Integer - PSO
(b) Comparison on Integer - PSO (c) Improvement with respect to

and PSO-SPV algorithm number of tasks
(d) Improvement with respect to (e) Improvement with respect to type

number of VMs of VMs
Fig. 1. Performance Analysis

5 Results and Discussion

The algorithm has been simulated 10 times on the same data set with θ = 0.9,
θ = 0 and θ = 1 and the average value is taken to find an estimate on best
convergences, on a laptop PC with PIV processor of 3 GHz clock frequency and
8GB RAM. We have tested the algorithm on a task set of 99 tasks, setting the
population size (|P|) as 5 and other PSO parameters as w = 0.6, c1 = c2 = 0.2,
vmax = 4 and vmin = −4. A suitable value of θ is found through trial and error as
0.9. Assumptions regarding the capacity of each VM in terms of speed and cost
of usage per unit time and the task length are given as vectors. The algorithm
runs for 500 generations per task set and is repeated 11 times. The results are
shown in Figure 1(a).
For θ = 0.9, the proposed algorithm (Integer-PSO) finds optimal cost in 90
percentage of time, but SPV-PSO has not converged to optimal or near optimal
cost value on any value of θ, shown in Figure 1(b). When the algorithm is applied
to single objective optimization scenario, for 90 percentage of time, the Integer-
PSO algorithm converges to optimal value. A detailed performance analysis of
both the approaches were done with different number of tasks, different number
of VMs and different types of VMs and are shown in Figure 1(c) to 1(e).
6 Conclusion
Scheduling tasks in the cloud is a challenging one as the same involves many
factors such as cost and profit considerations, execution time, SLAs, Quality
of service parameters requested by the end user and committed by the CSP
and power considerations. Also the task arrival rate is highly unpredictable and
dynamic in nature. We have modelled the problem as a constraint bi-objective
optimization problem, where the objectives are makespan and cost and have used
Particle Swarm Optimization algorithm to solve the same, where the pareto op-
timality is achieved through weighted sum approach. A variant of PSO technique
is proposed (Integer-PSO) whose results are promising.
References
1. Szabo, C., Kroeger, T.: Evolving multi-objective strategies for task allocation of
scientific workflows on public clouds. In: Proc. of IEEE Congress on Evolutionary
Computation (CEC), pp. 1–8 (2012)
2. Farahabady, R.H., Lee, Y.C., Zomaya, A.Y.: Pareto optimal cloud bursting. Ac-
cepted for publication in IEEE Transactions on Parallel and Distributed Systems
(2013)
3. Feng, M., Wang, X., Zhang, Y., Li, J.: Multi-objective particle swarm optimization
for reseource allocation in cloud computing. In: Proc. of 2nd International Confer-
ence on Cloud Computing and Intelligent Systems (CCIS), vol. 3, pp. 1161–1165
(2012)
4. Guo, L., Shao, G., Zhao, S.: Multi-objective task assignment in cloud computing by
particle swarm optimization. In: Proc. of 8th International Conference on Wireless
Communications, Networking and Mobile Computing (WiCOM), pp. 1–4 (2012)
5. Jin, J., Luo, J., Song, A., Dong, F., Xiong, R.: Bar: An efficient data locality
driven task scheduling algorithm for cloud computing. In: Proc. of 11th IEEE/ACM
International Symposium on Cluster, Cloud and Grid Computing, pp. 295–304
(2011)
6. Kennedy, J., Eberhart, R.C.: Particle swarm optimization. In: Proc. of IEEE In-
ternational Conference on Neural Networks, pp. 1942–1948 (1995)
7. Rodrignez Sossa, M., Buyya, R.: Deadline based resource provisioning and schedul-
ing algorithm for scientific workflows on clouds. IEEE Transactions on Cloud Com-
puting (2014)
8. Liang, Y.-C., Tasgetiren, M.F., Sevkli, M., Gencylmaz, G.: Particle swarm opti-
mization algorithm for single machine total weighted tardiness problem. In: IEEE
Congress on Evolutionary Computation (CEC2004), vol. 2, pp. 1412–1419 (2004)
9. Marler, T., Arora, J.S.: The weighted sum method for multi-objective optimization:
new insights. Springer (2009)
10. Sidhu, M.S., Thulasiraman, P., Thulasiram, R.K.: A load-rebalnce pso heuristic
for task matching in heterogeneous computing systems. In: IEEE Symposium on
Swarm Intelligence (SIS), pp. 180–187 (2013)
11. Sadhasivam, G.S., Selvarani, S.: Improved cost-based algorithm for task scheduling
in cloud computing. In: Proc. of IEEE International Conference on Computational
Intelligence and Computing Research (ICCIC), pp. 1–5 (2010)
12. Ivan, P.: Stanimirovic, Milan Lj. Zlatanovic, and Marko D Petkovic: On the linear
weighted sum method for multi-objective optimization. FACTA UNIVERSITATIS
(NIS), Ser. Math. Inform, 49–63 (2011)
13. Wu, Z., Ni, Z., Liu, X.: A revised discrete particle swarm optimization for cloud
workflow scheduling. In: Proc. of International Conference on Computational In-
telligence and Security (CIS), pp. 184–188 (2010)
14. Zhang, L., Chen, Y., Sun, R., Jing, S., Yang, B.: A task scheduling algorithm based
on pso for grid computing. International Journal of Computational Intelligence
Research 4(1), 37–43 (2008)
15. Zhu, L., Wu, J.: Hybrid particle swarm optimization algorithm for flexible task
scheduling. In: Proc. of 3rd Internatinal Conference on Genetic and Evolutionary
Computing (WGEC 2009), pp. 603–606 (2009)
Development on Harmony Search Hyper-heuristic
Framework for Examination Timetabling Problem
Khairul Anwar1, Ahamad Tajudin Khader1,

Mohammed Azmi Al-Betar1, 2, and Mohammed A. Awadallah1
1
School of Computer Sciences, Universiti Sains Malaysia (USM),
11800 Pulau Pinang, Malaysia
2
Department of Information Technology, Al-Huson University College,
Al Balqa Applied University, P.O. Box 50, Al-Huson, Irbid-Jordan
{ka10_com097,mama10_com018}@student.usm.my,
{tajudin,mohbetar}@cs.usm.my
Abstract. In this paper, a Harmony Search-based Hyper-heuristic (HSHH) ap-

proach is proposed for tackling examination timetabling problems. In this ap-
proach, the harmony search algorithm will operate as a high level of abstraction
which intelligently evolves a sequence of low level heuristics. This sequence is
a combination of improvement heuristics which consist of neighborhood struc-
ture strategies. The proposed approach is tested using the examination timetabl-
ing tracks in Second International Timetabling Competition (ITC-2007)
benchmarks. Experimentally, the HSHH approach can achieve comparable re-
sults with the comparative methods in the literature.
Keywords: Examination Timetabling, Harmony Search, Hyper-heuristic.
1 Introduction
Examination timetabling is the process of scheduling a set of exams to a set of time-

slots and rooms, subject to hard and soft constraints. It becomes extremely difficult
when it involves a large number of events (could be hundreds or thousands) to be
scheduled in limited resources in accordance with a wide variety of constraints, which
need to be satisfied [1]. The hard constraints are mandatory to fulfill while the soft
constraints are desired but not absolutely necessary. The solution that satisfied all the
hard constraints is called feasible. Conventionally, the soft constraints play a major
role in measuring the quality of the solution. The main target is to find an examination
timetabling solution that satisfies all hard constraints and minimize the violation of
soft constraints as much as possible.
Several optimization techniques have been introduced to solve examination time-
tabling problems such as sequential techniques, constraints based techniques, local
search-based techniques, population-based techniques, and many others as surveyed
in [1]. But recently, the attention of the researches in the operation research and the
artificial intelligence fields has shifted to be concerned with a more general approach
88 K. Anwar et al.
as hyper-heuristic. Hyper-heuristics (HH) have a higher level heuristic to choose from

a set of heuristics that are applicable for the problem in hand. The main difference
between the hyper-heuristics and meta-heuristics is that hyper-heuristics explore the
heuristic search space while meta-heuristics explore the solution search space [2].
There are several hyper-heuristics to solve the examination timetabling problems.
Sabar et al.[3] investigated a new graph coloring constructive hyper-heuristic
(GCCHH) for solving examination timetabling problems using ITC-2007 dataset. The
results demonstrate that GCCHH produces good results and outperforms other ap-
proaches on some of the benchmark instances. In another study, Pillay and Banzhaf
[4] used genetic programming (GP) for the evolution on hyper-heuristics framework
to solve uncapacitated examination timetabling problems. From the research, it shows
that the genetic programming system was comparable to the other search algorithm,
and in some cases it can produce better quality timetables.
In our previous work [5], Harmony search algorithm (HSA) has been employed in
a hyper-heuristic framework as a high-level heuristic where some move heuristics (i.e.
move and swap) have been employed as low-level heuristics. HSA is a population-
based algorithm developed by Geem et al.[6]. HSA is a stochastic search mechanism,
simple in concept, and no derivation information is required in the initial search [6]. It
has been successfully applied to a wide range of optimization and scheduling prob-
lems such as course timetabling problem [7], nurse scheduling problem [8, 9], as well
as examination timetabling problem [10] and many others as reported in [11].
Evidently, the initial employment produced an impressive result for the initial in-
vestigation, with a chance to improve the results by modifying the proposed method.
In this research, the new low-level heuristics strategies and the pitch adjustment pro-
cedure are modified to enhance the proposed method. For purposes of evaluation,
the dataset of examination timetabling tracks established in the Second International
Timetabling Competition (ITC-2007) is used. The results produced by new HSHH
outperformed those produced by previous versions of HSHH and comparable to the
results of the ITC-2007 comparative methods.
This paper is organized as follows: Section 2 discusses the examination timetabling
problem (ETP) and benchmark dataset of ITC-2007 for examination track. The de-
scription of harmony search hyper-heuristic (HSHH) algorithm is presented in Section
3. Section 4 discusses the experimental setup and the computational results. Finally,
the conclusion and future works are provided in Section 5.
2 Examination Timetabling
In this research, we used ITC-2007 [12] examination version as a benchmark to

evaluate the proposed method. ITC-2007 provides a capacitated examination time-
tabling dataset which contains eight instances and four hidden ones. For this research
we experimented with the eight instances. This dataset consists of five hard con-
straints (i.e., H1-H5) and seven soft constraints (i.e., S1-S7) as shown in Table 1.
Development on Harmony Search Hyper-heuristic Framework 89
Table 1. ITC-2007 Hard and Soft constraints
Key Constraints
H1 No student sits for more than one examination at the same time.
H2 The capacity of individual rooms is not exceeded at any time throughout the exami-
nation session.
H3 Period Lengths are not violated.
H4 Satisfaction of period related hard constraints (e.g., exam B must be scheduled after
exam A).
H5 Satisfaction of room related hard constraints (e.g., exam A exclusively scheduled in
room X).
S1 Two exams in a row.
S2 Two exams in a day.
S3 Specified spread of examinations.
S4 Mixed duration of examinations within individual periods.
S5 Larger examinations appearing later in the timetable.
S6 Period related soft constraints – some period has an associated penalty.
S7 Room related soft constraints – some room has an associated penalty.
The objective function to summarize the ITC-2007 dataset is formalized in equa-

tion (1) and the details’ explanations is provided in [12].
(1)
where x is a complete timetabling solution; S refers to a set of students while refers

to the institutional penalty for each constraint except for period and room related soft
constraint (i.e., CP and CR). Table 2 shows the detail notation of the variables used in
equation (1).
Table 2. List of abbreviations given to the ITC-2007 soft constraint [12]
Math Symbol Description

“two exam in row” penalty for student s.
“two exam in day” penalty for student s.
“period spread” penalty for student s
“No mixed duration” penalty
“Front load” penalty
“Period” penalty
“Room” penalty
90 K. Anwar et al.
3 Harmony Search Hyper-heuristic for ITC-2007
In previous work of Harmony Search Hyper-Heuristic (HSHH) [5], the pitch adjust-
ment operator is deactivated in the improvisation step, and two neighborhood struc-
tures are utilized as low-level heuristics. In this study, the pitch adjustment operator is
added during the improvisation step, and seven different neighborhood structures have
been utilized as low level heuristics. They can be summarized as follows:
• h1: Move Exam. Select one exam at random and move to a new randomly se-
lected feasible timeslot.
• h2: Swap Timeslot. Select two exams at random and swap their timeslots.
• h3: Swap Exam. Select two timeslots randomly and exchange all exams between
them.
• h4: Swap Period. Select two periods and swap the exams between the periods.
• h5: Select two timeslots (e.g. t1 and t2) randomly and move some exams from the
timeslot t1 to t2 and vice versa.
• h6: This heuristic similar to h3 but it only swaps the conflicting exams in two
distinct timeslots. This heuristic is similar to kempe chain method in (Al-Betar et
al.,[10]).
• h7: do nothing.
Basically, HSHH has five main steps as follows:
Step 1: Initialization. The HSHH begins by setting the harmony search parameters:
harmony memory size (HMS), harmony memory consideration rate (HMCR), number
of iterations (NI) and Harmony Memory Length (HML) which represents the length of
heuristic vector. Furthermore, the Pitch Adjustment Rate (PAR) parameter also will be
set. Initially, the largest degree (LD) heuristic is used to construct the initial feasible
solution (xfeasible). If the solution is not feasible, then the repair procedure as used in [7]
will be triggered to maintain the feasibility of the solution.
Step 2: Initializing Harmony Memory. HSHH consists of two complemented search

spaces (heuristic search space and solution search space), each represented in a har-
mony memory: Heuristics Harmony Memory (HHM) and Solutions Harmony Memory
(SHM). HHM contains sets of heuristic vectors determined by HMS where every vec-
tor is a heuristics sequence. The length of the vector is determined by HML. Similarly,
SHM contains sets of solution vectors and the length is determined by the number of
exam, N.
In initializing HHM and SHM, the HSHH, firstly, generates the new heuristics vec-
tor (h’) randomly and apply this vector to the initial feasible solution (xfeasible) to pro-
duce the new solution (x’). The new solution will be evaluated using the objective
function as in equation (1). If the new solution (x’) is better than the initial solution
(xfeasible), then the new solution (x’) will be saved in the SHM and the new heuristic
vector (h’) will be saved in the HHM. This process will be repeated until HHM and
SHM are filled (see equations (2) and (3)). After completing the process, HSHH will
retain the worst solution (x worst) and the best solution (x best) in SHM.
Step 3: Improvise a new heuristic sequence. In this step, a new heuristics vector h’ =
( , ,… ) is generated from scratch, based on three HSA operators: memory
consideration, random consideration and pitch adjustment.
 h11 h21  hHML

1
  x11 x21  x1N
   2 
HHM =  h12 h22  hHML 
2
(2) SHM =  x1 x22  xN2
           
 HMS HMS   HMS 
 h1 h2HMS  hHML   x1 x2HMS  xNHMS 
(3)
Note that N refers to the number of examinations. In memory consideration opera-

tor, the new heuristic index of in the new heuristic vector (h’) is randomly selected
from the historical indexes (e.g. , , … . . ), stored in the heuristic harmony
memory with probability of HMCR, where HMCR ϵ [0, 1] .
For Random consideration operator, the new heuristic index is randomly assigned
from the set of heuristics ϵ {h1, h2, h3, h4, h5, h6, h7} with probability of (1-
HMCR) as in equation (4).
, ,….. . .
(4)
1, 2, 3, 4, 5, 6, 7 . . 1
In Pitch Adjustment operator, a simple adjustment is used. The new index of

will be added/subtracted by 1 with a probability of PAR where 0 1 as in
equation (5).
. .
(5)
. . 1
In case the decision of PAR is yes, the index of will be recalculated as follows:
1 (6)
Note that if the index is out of range, it will remain the same. Then the new harmony
of heuristic vector h’ will be applied to a solution (e.g.
= , ,….. ) where is randomly selected from the solu-
tion search space or SHM. The HSHH used random selection to select the solution
from the SHM to avoid the local optima. In this process, the heuristic in h’ will be
executed sequentially to the selected solution ( . The process will continue until
all the heuristics in h’ have been executed, and a new solution (x’) will be produced.
Pseudo-code for improvisation step is shown in the Algorithm 1.
Step 4: Update HHM and SHM. In hyper-heuristic environment, this step is called a
move acceptance step. HSHH will decide either to accept or neglect the new heuristic
vector h’. In this process, the new solution (x’) will be evaluated using the objective
92 K. Anwar et al.
function. The new solution must be complete and feasible. If the new solution is bet-
ter than the worst solution in solution harmony memory (SHM), the new h’ and x’
will be saved in the memory (h’ in HHM and x’ in SHM) and the worst heuristic vec-
tor and solution will be excluded from the memory (i.e., HHM and SHM).
Step 5: Check the stop criterion. Step 3 and step 4 in this approach are repeated until
the stop criterion (i.e., NI) is met.
Algorithm 1: Pseudo-code for selecting and generating heuristic vector during the
improvisation process in step 3.
h’= 0; //heuristic vector
for l = 0,…,HML do
if (U(0,1) ≤ HMCR) then
, ,….. ; //Memory consideration;
if (U(0,1) ≤ PAR) then
1 ; //Pitch adjustment;
else
ϵ {h1, h2, h3, h4, h5, h6, h7}; //Random consideration;
end if
end for
, ,…, ; //Select random solution from SHM;
x’ = apply h’ to xrand ;
In this section, Harmony Search Hyper-heuristic is evaluated using the real world
problem dataset (ITC-2007) for university examination timetabling problem. The
proposed method is coded in Microsoft Visual C++ 6 under Windows 7 on Intel pro-
cessor with 2G RAM. We chose to test the proposed method with each problem in-
stances in ITC-2007.
The characteristics of the ITC-2007 dataset are provided in Table 3. This table in-
cludes information such as number of students (Info1), actual number of students
(Info2), number of exams (Info3), number of timeslots (info4), and number of rooms
(Info5). We ran each experiment 10 times for each problem due to the stochastic na-
ture of the method [13]. The Harmony Search Hyper-Heuristic (HSHH) parameters
are set as HMS=10, HML=10, PAR=0.1 HMCR=0.95, and N1=100000, where these
parameter settings are used based on some experiments carried out previously.
Experimentally, the HSHH is able to find a feasible solution for seven out of eight
instances in ITC-2007 dataset. Table 4 provides the comparative results of the HSHH
and the other comparative methods that are working using the same dataset. The dif-
ferent comparative methods are provided as shown in Table 5. The numbers in table 4
referred to the penalty value of the soft constraint violations. The best results are hig-
hlighted in bold. The indicator ‘x% inf’ indicates that the percentage of such algo-
rithm could not find a feasible solution.
Table 3. The Characteristics of ITC-2007 Examination Timetabling Dataset
Dataset Info1 Info2 Info3 Info4 Info5

Exam1 7891 7833 607 54 7
Exam2 12743 12484 870 40 49
Exam3 16439 16365 934 36 48
Exam4 5045 4421 273 21 1
Exam5 9253 8719 1018 42 3
Exam6 7909 7909 242 16 8
Exam7 14676 13795 1096 80 15
Exam8 7718 7718 598 80 8
Table 4. Comparison with previous HSHH and other methods

Dataset HSHH 2 HSHH 1 M1 M2 M3 M4 M5 M6 M7 M8 M9
Exam1 9885 11823 8559 6235 6234 4775 4370 4633 6582 4368 12035
Exam2 393 976 830 2974 395 385 385 405 1517 390 3074
Exam3 19931 26770 11576 15832 13002 8996 9378 9064 11912 9830 15917
Exam4 100% inf 100% inf 21901 35106 17940 16204 15368 15663 19657 17251 23582
Exam5 4065 6772 3969 4873 3900 2929 2988 3042 17659 3022 6860
Exam6 29935 30980 28340 31756 27000 25740 26365 25880 26905 25995 32250
Exam7 8801 11762 8167 11562 6214 4087 4138 4037 6840 4067 17666
Exam8 12145 16286 12658 20994 8552 7777 7516 7461 11464 7519 16184
Table 5. The comparison methods for ITC-2007 Examination Timetabling Dataset
Key Method
HSHH 1 Harmony Search hyper-heuristic [5].
M1 Evolutionary Algorithm hyper-heuristic [14].
M2 Hybrid Approach hyper-heuristic[15].
M3 Graph Coloring Constructive hyper-heuristic[3].
M4 An improved multi-staged algorithmic[16].
M5 A Three phase constraint-based approach[17]
M6 An extended great deluge algorithm [18].
M7 Artificial Bee Colony algorithm [19] .
M8 Hybrid approach within great deluge algorithm[20].
M9 Developmental Approach [21].
As shown in Table 4, the performance of the new HSHH (i.e. HSHH 2) is much
better than the performance of the previous version of HSHH (i.e. HSHH 1). Figure 1
shows the comparison between the HSHH 1 and HSHH 2 in terms of convergence
behavior. Experimental results show that HSHH is able to produce good results and
one of these datasets (i.e. Exam2) has achieved comparable result as shown in Table
4. Furthermore, the proposed method has also been able to obtain better results com-
pared to the several other approaches. As compared with hybrid approach hyper-
heuristic (M2), HSHH are able to produce better results in five problem instances
(i.e., Exam2, Exam5, Exam6, Exam7 and Exam8) and six problem instances (i.e.,
Exam1, Exam2, Exam5, Exam6, Exam7 and Exam8) compared to the developmental
approach (M9).
94 K. Anwar et al.
Fig. 1. Comparison of Convergence behavior between HSHH I and HSHH II
This paper presented Harmony search-based hyper-heuristics (HSHH) for solving

examination timetabling problems using the ITC-2007 dataset. The harmony search is
utilized at the high-level to evolve a sequence of improvement low level heuristics.
In order to evaluate HSHH, problem instances of ITC-2007 dataset have been used.
The experimental result shows that HSHH is able to solve examination timetabling
problems. Although, the results produced by HSHH in this study have not reached the
best known results, they seem comparable or even better in some cases when com-
pared to the previous approaches using ITC-2007. Utilizing several low-level heuris-
tics in the HSHH framework is a very promising extension to the hyper-heuristic
domain in general. This is because each low-level heuristic can deal with a region of
the search and touching several regions in the search space might increase the chance
of improvements.
The main objective behind proposing the HSHH is to suggest an applicable frame-
work that is general enough to be re-implemented for other types of scheduling or
combinatorial optimization problems. Therefore, we plan to apply the proposed ap-
proach to solve the nurse rostering problem using INRC2010 dataset. For future re-
search, we plan to adapt learning mechanism within the HSHH algorithm in order to
improve the heuristic selection and to enhance the speed of convergence.
References
1. Qu, R., et al.: A Survey of Search Methodologies and Automated System Development for
Examination Timetabling. Journal of Scheduling 12(1), 55–89 (2009)
2. Burke, E.K., et al.: A survey of Hyper-heuristics. Computer Science Technical Report No.
NOTTCS-TR-SUB-0906241418-2747, School of Computer Science and Information
Technology, University of Nottingham (2009)
3. Sabar, N.R., et al.: A Graph Coloring Constructive Hyper-heuristic for Examination Time-
tabling Problems. Applied Intelligence, 1–11 (2011)
4. Pillay, N., Banzhaf, W.: A Genetic Programming Approach to the Generation of Hyper-
Heuristics for the Uncapacitated Examination Timetabling Problem. In: Neves, J., Santos,
M.F., Machado, J.M. (eds.) EPIA 2007. LNCS (LNAI), vol. 4874, pp. 223–234. Springer,
Heidelberg (2007)
5. Anwar, K., et al.: Harmony Search-based Hyper-heuristic for Examination Timetabling.

In: 9th International Colloquium on Signal Processing and its Applications (CSPA). IEEE
(2013)
6. Geem, Z.W., Kim, J.H., Loganathan, G.: A New Heuristic Optimization Algorithm: Har-
mony Search. Simulation 76(2), 60–68 (2001)
7. Al-Betar, M.A., Khader, A.T.: A Harmony Search Algorithm for University Course Time-
tabling. Annals of Operations Research, 1–29 (2012)
8. Awadallah, M.A., et al.: Nurse Scheduling Using Harmony Search. In: Sixth International
Conference on Bio-Inspired Computing: Theories and Applications (BIC-TA). IEEE
(2011)
9. Awadallah, M.A., et al.: Harmony Search with Novel Selection Methods in Memory Con-
sideration for Nurse Rostering Problem. Asia-Pacific Journal of Operational Research
(2013)
10. Al-Betar, M.A., Khader, A.T., Nadi, F.: Selection Mechanisms in Memory Consideration
for Examination Timetabling with Harmony Search. In: Proceedings of the 12th Annual
Conference on Genetic and Evolutionary Computation (GECCO 2010). ACM, New York
(2010)
11. Manjarres, D., et al.: A Survey on Applications of the Harmony Search Algorithm. Engi-
neering Applications of Artificial Intelligence 26(8), 1818–1831 (2013)
12. McCollum, B., et al.: The Second International Timetabling Competition: Examination
Timetabling Track. Technical Report QUB/IEEE/Tech/ITC2007/Exam/v4. 0/17, Queens
University, Belfast, UK (2007)
13. Al-Betar, M.A., Khader, A.T., Doush, I.A.: Memetic Techniques for Examination Time-
tabling. Annals of Operations Research, 1–28 (2013)
14. Pillay, N.: Evolving Hyper-heuristics for a Highly Constrained Examination Timetabling
Problem. In: Proceedings of the 8th International Conference on the Practice and Theory
of Automated Timetabling, PATAT 2010 (2010)
15. Burke, E.K., Qu, R., Soghier, A.: Adaptive Selection of Heuristics for Improving Con-
structed Exam Timetables. In: Proceedings of the 8th International Conference on the
Practice and Theory of Automated Timetabling, PATAT 2010 (2010)
16. Gogos, C., Alefragis, P., Housos, E.: An Improved Multi-staged Algorithmic Process for
the Solution of the Examination Timetabling Problem. Annals of Operations Re-
search 194(1), 203–221 (2012)
17. Müller, T.: ITC2007 Solver Description: a Hybrid Approach. Annals of Operations Re-
search 72(1), 429–446 (2009)
18. McCollum, B., et al.: An Extended Great Deluge Approach to the Examination Timetabl-
ing Problem. In: Proceedings of the 4th Multidisciplinary International Scheduling: Theory
and Applications (MISTA 2009), pp. 424–434 (2009)
19. Alzaqebah, M., Abdullah, S.: Artificial Bee Colony Search Algorithm for Examination
Timetabling Problems. International Journal of the Physical Sciences 6(17), 4264–4272
(2011)
20. Turabieh, H., Abdullah, S.: An Integrated Hybrid Approach to the Examination Timetabl-
ing Problem. Omega 39(6), 598–607 (2011)
21. Pillay, N.: A Developement Approach to the Examination Timetabling (2007),
http://www.cs.qub.ac.uk/itc2007
Predator-Prey Pigeon-Inspired Optimization
for UAV Three-Dimensional Path Planning
Bo Zhang1 and Haibin Duan1,2,*

1
Science and Technology on Aircraft Control Laboratory, School of Automation Science
and Electrical Engineering, Beihang University, Beijing, 100191, P. R. China
2
Provincial Key Laboratory for Information Processing Technology, Soochow University,
Suzhou, 215006, P. R. China
zhangbo0216@163.com, hbduan@buaa.edu.cn
Abstract. Pigeon-inspired optimization (PIO) is a new bio-inspired

optimization algorithm. This algorithm searches for global optimum through
two models: map and compass operator model is presented based on magnetic
field and sun, while landmark operator model is designed based on landmarks.
In this paper, a novel Predator-prey pigeon-inspired optimization (PPPIO) is
proposed to solve the three-dimensional path planning problem of unmanned
aerial vehicles (UAVs), which is a key aspect of UAV autonomy. To enhance
the global convergence of the PIO algorithm, the concept of predator-prey is
adopted to improve global best properties and enhance the convergence speed.
The comparative simulation results show that our proposed PPPIO algorithm is
more efficient than the basic PIO and particle swarm optimization (PSO) in
solving UAV three-dimensional path planning problems.
Keywords: pigeon-inspired optimization (PIO), unmanned aerial vehicle

(UAV), path planning, predator-prey.
1 Introduction
Three-dimensional path planner is an essential element of the unmanned aerial vehicle

(UAV) autonomous control module [1]. It allows the UAV to compute the best path
from a start point to an end point autonomously [2, 3]. Whereas commercial airlines
fly constant prescribed trajectories, UAVs in operational areas have to travel
constantly changing trajectories that depend on the particular terrain and conditions
prevailing at the time of their flight.
Pigeon-inspired optimization (PIO), which is a new swarm intelligence optimizer
based on the movement of pigeons, was firstly invented by Duan in 2014 [4]. Homing
pigeons can easily find their homes by using three homing tools: magnetic field, sun and
landmarks. In the optimization, map and compass model is presented based on magnetic
field and sun, while landmark operator model is presented based on landmarks.
In this paper, we propose a predator-prey pigeon-inspired optimization (PPPIO)
method, integrating the concept of predator-prey into PIO in order to improve its
*
Predator-Prey Pigeon-Inspired Optimization for UAV Three-Dimensional Path Planning 97
capability of finding satisfactory solutions and increasing the diversity of the

population. We also solve the UAV three-dimensional path planning problem by
PPPIO. Simulation results and comparisons verified the feasibility and effectiveness
of our proposed algorithm.
The rest of the paper is organized as follows: Section 2 provides the representation
and the cost function we developed to evaluate the quality of candidate trajectories.
Section 3 describes the principle of basic PIO algorithm. Section 4 shows the
implementation procedure of our proposed predator-prey PIO algorithm. Finally, we
compare the quality of the trajectories produced by the PIO, particle swarm
optimization (PSO) and the PPPIO in Section 5.
2 Problem Formulation
The first step of three-dimensional path planning is to discretize the world space into
a representation that will be meaningful to the path planning algorithm. In this work,
we use a formula to indicate the terrain environment. The mathematical function is of
the form [5]:
z ( x, y ) = sin( x / 5 + 1) + sin( y / 5) + cos( a x 2 + y 2 ) + sin(b x 2 + y 2 ) (1)
where z indicate the altitude of a certain point, and a, b are constants experimentally
defined. Our representation of cylindrical danger zones (or no-fly zones) to be in a
separate matrix where each row represents the coordinates (x i , y i ) and the radius ri
of the ith cylinder as shown in Eq. (2). Complex no-fly zone can be built by partially
juxtaposing multiple cylinders
 x1 y1 r1 
 
 x2 y2 r2  (2)
danger zones =
  
 
 xn yn rn 
The three-dimensional trajectories generated by the algorithm are composed of line
segments and (x i , y i , z i ) represents the coordinates of the ith way point. The
trajectories are flown at constant speed.
In the situation of UAV path planning, the optimal path is complex and includes
many different characteristics. To take into account these desired characteristics, a
cost function is used and the path planning algorithm becomes a for a path that will
minimize the cost function. We define our cost function as follows [6]:
Fcost = C length + C altitude + C danger zones + C power + C collision + C fuel (3)
In the cost function, the term associated with the length of a path is defined as
follows:
L
Clength = 1 − ( p1p2 ) (4)
Ltraj
C length ∈ [0,1] (5)
98 B. Zhang and H. Duan
where Lp1p2 is the length of the straight line connecting the starting point P1 and
the end point P 2 and Ltraj is the actual length of the trajectory.
The term associated with the altitude of the path is defined as follows:
Atraj − Z min
Caltitude = (6)
Z max − Z min
Caltitude ∈ [0,1] (7)
where Z max is the upper limit of the elevation in our search space, Z min is the
lower limit and Atraj is the average altitude of the actual trajectory. Z max and Z min
are respectively set to be slightly above the highest and lowest point of the terrain.
The term associated with the violation of the danger zones is defined as follows:
Linside d.z.
Cdanger zones = n
(8)
d
i =1
i
C danger zones ∈ [0,1] (9)

where n is the total number of danger zones, Linside d.z. is the total length of the
subsections of the trajectory which go through danger zones and d i is the diameter
of the danger zone i .
The term associated with a required power higher than the available power of the
UAV is defined as follows:
0, Lnot feasible = 0

Cpower =  Lnot feasible  (10)
 P +  L  , Lnot feasible > 0
  traj 
C power ∈ 0  [ P , P + 1] (11)
where Lnot feasible is the sum of the lengths of the line segments forming the trajectory
which require more power than the available power of the UAV, Ltraj is the total
length of the trajectory and P is the penalty constant. This constant must be higher
than the cost of the worst feasible trajectory which would have, based on our cost
function, a cost of 3. By adding this penalty P , we separate nonfeasible solutions
from the feasible ones.
The term associated with ground collisions is defined as follows:
0, Lunder terrain = 0

Ccollision =  Lunder terrain  (12)
 P +  L  , Lunder terrain > 0
  traj 
Ccollision ∈ 0  [ P , P + 1] (13)
where Lunder terrain is the total length of the subsections of the trajectory which travels
below the ground level and Ltraj is the total length of the trajectory.
The term associated with an insufficient quantity of fuel available is defined as
follows:
0, Ftraj ≤ Finit

Cfuel =  FP1P2  (14)
 P + 1 −  F  , Ftraj > Finit
  traj 
Cfuel ∈ 0  [ P , P + 1] (15)
where FP1P2 is the quantity of fuel required to fly the imaginary straight segment
connection the starting point P1 to the end point P 2 , Ftraj is the actual amount of
fuel needed to fly the trajectory, Finit is the initial quantity of fuel on board the UAV.
The search engine will be adopted to find a solution, which can minimize the cost
function during the optimization phase of our path planner algorithm. This can also be
explained as to find a trajectory that best satisfies all the qualities represented by this
cost function. Our cost function demonstrates a specific scenario where the optimal
path minimizes the distance travelled, the average altitude (to increase the stealthiness
of the UAV) and avoids danger zones, while respecting the UAV performance
characteristics. This cost function is highly complex and demonstrates the power of
our path planning algorithm. However, this cost function could easily be modified and
applied to a different scenario.
3 Principle of Basic PIO

PIO is a novel swam intelligence optimizer for solving global optimization problems.
It is based on natural pigeon behavior. Studies show that the species seem to have a
system in which signals from magnetite particles are carried from the nose to the
brain by the trigeminal nerve [4, 7]. Evidence that the sun is also involved in pigeon
navigation has been interpreted, either partly or entirely, in terms of the pigeon’s
ability to distinguish differences in altitude between the Sun at the home base and at
the point of release [8]. Recent researches on pigeons’ behaviors also show that the
pigeon can follow some landmarks, such as main roads, railways and rivers rather
than head for their destination directly. The migration of pigeons is summarized as
two mathematical models. One is map and compass operator, and the other is
landmark operator.
3.1 Map and Compass Operator

In PIO model, virtual pigeons are used. In the map and compass operator, the rules are
defined with the position X i and the velocity Vi of pigeon i , and the positions and
velocities in a D-dimension search space are updated in each iteration.

The new position X i and velocity Vi of pigeon i at the t-th iteration can be
calculated with the follows [3]:
Vi (t ) = Vi (t − 1) ⋅ e − Rt + rand ⋅ ( X g − X i (t − 1)) (16)
X i (t ) = X i (t − 1) + Vi (t ) (17)
where R is the map and compass factor, rand is a random number, and X g is the
current global best position, and which can be obtained by comparing all the positions
among all the pigeons.
3.2 Landmark Operator
In the landmark operator, half of the number of pigeons is decreased by N p in every

generation. However, the pigeons are still far from the destination, and they are
unfamiliar the landmarks. Let X c (t ) be the center of some pigeons’ position at the
t-th iteration, and suppose every pigeon can fly straight to the destination. The
position updating rule for pigeon i at t-th iteration can be given by:
N P (t − 1)
N P (t ) = (18)
2
X c (t ) =
 X i (t ) ⋅ fitness ( X i (t ) ) (19)
N P  fitness ( X i (t ) )
X i (t ) = X i (t − 1) + rand ⋅ ( X c (t ) − X i (t − 1)) (20)
where fitness is the quality of the pigeon individual. For the minimum optimization
1
problems, we can choose fitness ( X i (t ) ) = for maximum
f ( X i (t ) ) + ε
optimization problems, we can choose fitness ( X i (t ) ) = f ( X i (t ) ) .
4 PPPIO for Three-Dimensional Path Planning
4.1 Predator-Prey Concept

Predatory behavior is one of the most common phenomena in nature, and many
optimization algorithms are inspired by the predator-prey strategy from ecology [9].
In nature, predators hunt prey to guarantee their own survival, while the preys need to
be able to run away from predators. On the other hand, predators help to control the
prey population while creating pressure in the prey population. In this model, an
individual in predator population or prey population represents a solution, each prey
in the population can expand or get killed by predators based on its fitness value, and
a predator always tries to kill preys with least fitness in its neighborhood, which
represents removing bad solutions in the population. In this paper, the concept of
predator-prey is used to increase the diversity of the population, and the predators are
modeled based on the worst solutions which are demonstrated as follows:
Ppredator = Pworst + ρ (1 − t / tmax ) (21)
where Ppredator is the predator (a possible solution), Pworst is the worst solution in
the population, t is the current iteration, while tmax is the maximum number of
iterations and ρ is the hunting rate. To model the interactions between predator and
prey, the solutions to maintain a distance of the prey from the predator is showed as
follows:
 Pk+1 = Pk + ρ e − d , d>0
 −d
(22)
 Pk+1 = Pk − ρ e , d<0
where d is the distance between the solution and the predator, and k is the current
iteration.
4.2 Parallelization of the Map and Compass Operations and the Landmark
Operations
In the basic model of PIO algorithm, the landmark operation is used after several
iterations of map and compass operation. For example, when the number of
generations N c is larger than the maximum number of generations of the map and
compass operation N c max1 . The map and compass operator will stop and it the
landmark operation will be start. During my experiment, we found it’s easy to fall into
a local best solution before the number of generations got to N c max1 . Furthermore,
half of the number of pigeons is decreased by N p in every generation on the
landmark operator. The population of pigeons is decreased too rapidly according to
formula (18), which would reach to zero after a small amount of iterations. The
landmark operator would make only a small impact on the pigeons’ position by this
way. So we make a small modification on the basic PIO algorithm. The map and
compass operation and the compass operation are used parallelly at each iteration. A
parameter ω is used to define the impaction of the landmark increase with a
smoothly path. And a constant parameter c is used to define the number of pigeons
that are in the landmark operator. Our new formula of landmark operator is as
follows:
N P (t ) = c ⋅ N P max c ∈ (0,1) (23)
X c (t ) =
 X (t ) ⋅ fitness ( X (t ) )
i i (24)
N  fitness ( X (t ) )
P i
ω = s + (1 − s) ⋅ t/ N c max s ∈ (0,1) (25)
X i (t ) = X i (t − 1) + ω ⋅ rand ⋅ ( X c (t ) − X i (t − 1)) (26)

where s is a constant experimentally defined.
4.3 Proposed Predator-Prey PIO (PPPIO) Based Path Planner

In order to overcome the disadvantages of the classical PIO algorithm, such as the
tendency to converge to local best solutions, PPPIO, which integrates PIO with the
concept of predator-prey, was proposed in our work. After the mutation of each
generation, the predator-prey behavior is been conducted in order to choose better
solutions into next generation. In this way, our proposed algorithm takes the
advantage of the predator-prey concept to make the individuals of sub generations
distributed ergodically in the defined space and it can avoid from the premature of the
individuals, as well as to increase the speed of finding the optimal solution.
The implementation procedure of our proposed PIO approach to UAV path
planning can be described as follows:
Step 1: According to the environmental modeling in Section 2, initialize the
detailed information about the path planning task.
Step 2: Initialize the PIO parameters, such as solution space dimension D, the
population size N p , map and compass factor R, the number of iteration N c .
Step 3: Set each pigeon with a randomized velocity and path. Compare the fitness
of each pigeons, and find the current best path.
Step 4: Operate map and compass operator. Firstly, we update the velocity and path
of every pigeon by using Eqs. (16) and (17).
Step 5: Rank all pigeons according their fitness values. Some of pigeons whose
fitness are low will follow those pigeons with high fitness according to Eq. (23). We
then find the center of all pigeons according to Eq. (24), and this center is the
desirable destination. All pigeons will fly to the destination by adjusting their flying
direction according to Eq. (26). Next, store the best solution parameters and the best
cost value.
Step 6: Model the predators based on the worst solution as Eq. (15) demonstrates.
Then, use Eq. (16) to provide the other solutions to maintain a distance between the
predator and the prey.
Step 7: If Nc > N c max , stop the iteration, and output the results. If not, go to step 6.
5 Comparative Experimental Results
In order to evaluate the performance of our proposed PPPIO algorithm in this work,
series of experiments are conducted in Matlab2012a programing environment.
Coordinates of a starting point are set as (10, 16, 0), and the target point as (55, 100,
0). The initial parameters of PIO algorithm were set as: NP =150. The comparative
results of PPPIO with PIO and PSO are showed as follows:
0.6
PSO
0.55 PPPIO
PIO
0.5
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0 20 40 60 80 100 120 140 160 180 200
Fig. 1. Comparative evolutionary curves of PPPIO, PIO and PSO
100
90
80
70
60
50
40
30
20
10
0
0 10 20 30 40 50 60 70
Fig. 2. Comparative path planning results of PPPIO, PIO and PSO

4
2
0
100
80
60
40
70
60
20 50
40
30
20
0 10
0
Fig. 3. Comparative path planning results of PPPIO, PIO and PSO on 3D version
6 Conclusions
This paper proposed a novel PPPIO algorithm for solving the UAV three-dimensional
path planning problem in complex environments. The concept of predator-prey is
adopted to improve the performance of the basic PIO algorithm. Series of
comparative simulation results were given to show that our proposed PPPIO
algorithm is more efficient than basic PIO and PSO in solving UAV three-
dimensional path planning problems.
Acknowledgements. This work was partially supported by National Key Basic

Research Program of China(973 Project) under grant #2014CB046401, Natural
Science Foundation of China (NSFC) under grant # 61333004 and #61273054,
National Magnetic Confinement Fusion Research Program of China under grant #
2012GB102006, and Aeronautical Foundation of China under grant #20135851042.
References
1. Chen, H., Wang, X.M., Li, Y.: A Survey of Autonomous Control for UAV. In: International
Conference on Artificial Intelligence and Computational Intelligence, vol. 2, pp. 267–271
(2009)
2. Duan, H.B., Li, P.: Bio-inspired Computation in Unmanned Aerial Vehicles. Springer,
Heidelberg (2014)
3. Duan, H.B., Luo, Q.N., Ma, G.J., Shi, Y.H.: Hybrid Particle Swarm Optimization and
Genetic Algorithm for Multi-UAVs Formation Reconfiguration. IEEE Computational
Intelligence Magazine 8(3), 16–27 (2013)
4. Duan, H.B., Qiao, P.X.: Pigeon-Inspired Optimization: A New Swarm Intelligence
Optimizer for Air Robot Path Planning. International Journal of Intelligent Computing and
Cybernetics 7(1), 24–37 (2014)
5. Ioannis, K.N., Athina, N.B.: Coordinated UAV Path Planning Using Differential Evolution.
In: IEEE International Symposium on, Mediterrean Conference on Control and Automation,
vol. 70, pp. 77–111. Springer, Heidelberg (2005)
6. Vincent, R., Mohammed, T., Gilles, L.: Comparison of Parallel Genetic Algorithm and
Particle Swarm Optimization for Real-Time UAV Path Planning. IEEE Transactions on
Industrial Informatics 9(1), 132–141 (2013)
7. Mora, C.V., Davison, M., Wild, J.M., Michael, M.W.: Magnetoreception and Its Trigeminal
Mediation in the Homing Pigeon. Nature 432, 508–511 (2004)
8. Whiten, A.: Operant Study of Sun Altitude and Pigeon Navigation. Nature 237, 405–406
(1972)
9. Zhu, W.R., Duan, H.B.: Chaotic Predator-Prey Biogeography-Based Optimization
Approach for UCAV Path Planning. Aerospace Science and Technology 32(1), 153–161
(2014)
Research on Route Obstacle Avoidance Task Planning
Based on Differential Evolution Algorithm for AUV
Jian-Jun Li1,3, Ru-Bo Zhang1,2, and Yu Yang3

1
College of Computer Science and Technology,
Harbin Engineering University, Harbin 150001, China
2
College of Electromechanical & Information Engineering,
Dalian Nationalities University, Liaoning Dalian, 116600, China
3
School of Computer and Information Engineering ,
Harbin University of Commerce, Harbin 150028, China
Abstract. AUV mission planning route avoidance purpose is to be able to

successfully avoid the threat of a number of different levels of obstacles
between the start and end of the route , and plan the optimal route planning to
meet certain performance indicators. Through the differential evolution
algorithm analysis and description , the avoidance route mission planning
problem into a multi-dimensional function optimization problems, optimization
problems for AUV mission planning route avoidance functions , based on
differential evolution algorithm is proposed route obstacle avoidance task
planning methods and after a comprehensive analysis and simulation results
validate the differential evolution algorithm in high-dimensional function
optimization convergence and stability demonstrated good performance.
Keywords: Differential evolution algorithm, Autonomous Underwater Vehicle,

route avoidance, mission planning.
1 Introduction
Due to the complex undersea environment, (Autonomous Underwater Vehicle, AUV),

Also known as underwater robots, the need for obstructions on the route, torpedoes
and other potential threats to take evasive strategy. Josep et al proposed a low cost
computing underwater vehicle planning program, create a static or dynamic obstacle
avoidance optimization campaign mode [1]. Zouming Cheng et al use of electronic
navigation and positioning collision path planning , and application of artificial
intelligence ant colony optimization algorithm to construct collision model[2]. Cruz
made hydrodynamics based obstacle avoidance algorithm to determine the location of
obstacles and target drones by the harmonic function, thus avoiding local optimum[3].
China Shipbuilding Industry Corporation710Research Institute, Yan Gang, who
proposed an improved genetic algorithm to improve search speed and path
optimization AUV levels[4]. AUV route avoidance is a key component of AUV
mission planning system. Differential evolution algorithm (Differential Evolution
Research on Route Obstacle Avoidance Task Planning 107
Algorithm, DE) by the American scholar Storn and Price propose a heuristic
algorithm to solve optimization problems[5]. Differential evolution algorithm in
solving global optimization problems in complex environments and continuous
domain optimization problem, with outstanding advantages. Therefore, based on
differential evolution algorithm AUV route avoidance, the successful completion of
the task execution for AUV has important practical significance.
2 Differential Evolution Algorithm
Differential evolution algorithm remembers the evolution of individual groups and

groups in the optimal solution features internal information sharing, through
competition and cooperation between individuals within the group to achieve the
optimal solution. Assuming a population size of NP , m to the first generation of the
evolution of the population of X (m) ,The dimension of the solution space is
K .Initial population X (0) = { x10 , x20 ,  , xNP
0
} , Solutions of the i-th individual
xi0 =  xi0,1 , xi0,2 ,  , xi0,k  . Individual components of the formula:
xi0, j = x j ,min + rand ( x j ,max − x j ,min ) (1)
x j ,max upper bound for the solution space, x j ,min lower bound for the solution
space. Differential evolution algorithm including mutation , crossover and selection of
three operating [6-9].
2.1 Mutation
Differential evolution algorithm mutation is the last generation of linear combinations

of multiple individuals in the population , the variability of individual difference
vector generation. Process variation follows the formula:
vi = xr1 + F ⋅ ( xr 2 − xr 3 ), i = 1, 2,, NP (2)
From the previous generation arbitrary choice of three different populations of

individuals { xr1 , xr 2 , xr 3} , and r1 ≠ r2 ≠ r3 ,F constant factor between [0,2]
between, Also known as the scaling factor, Used to control the difference vector
( xr 2 − xr 3 ) . ( xr 2 − xr 3 ) smaller value of the difference vector, The smaller the
disturbance, That is closer to the optimal value group , the disturbance value is
automatically reduced.
108 J.-J. Li, R.-B. Zhang, and Y. Yang
2.2 Crossover
Differential evolution algorithm is a variation vector crossover target vectors vi

and xi random reorganization , thereby increasing the diversity of the population of
individuals. Process crossover following formula:
vi , j , randk ≤ CR or j = rand j ;

uij =  (3)
 xi, j , randk > CR or j ≠ rand j ;
i = 1,, NP, j = 1,, NP
randk Is a random variable [0,1],CRIs a constant [0,1]. CR The larger the value,
the greater the probability of crossover, CR=0 Cross probability 0.
2.3 Select Options
Differential evolution algorithm selection operation is to adapt to the new vector

values and objectives of the individual ui vector of individual fitness value xi
compare, When the value of ui is better than xi , replacing xi . The select
operation by the following formula :
ui , f (ui ) < f ( xi )

t
t +1
xi = t (4)
 xi , other
3 AUV Multiple Route Avoidance Model
AUV safe navigation area, Refers to the current position of the center AUV, Rφ
radius of the circular area , And there is no obstacle in this circular area , the AUV
can be achieved without collision safe navigation. If there is an obstacle , then the
AUV be single or multiple route avoidance[10-11].
AUV route avoidance of multiple models , which means that AUV underwater
work space with the ability to meet multiple obstacle avoidance.
obstacl
3
Sail
track
φ obstacle
AU 2
V
( xi , x j )
obstacle
1
Fig. 1. AUV safe navigation area
3.1 Single Obstacle Avoidance Route Model
AUV set maximum safe radius ra , Obstacle radius r is

+ +
between (cx , c y ) ∈ R , (a x , a y ) ∈ R between the starting point A and the target
+
point B between (bx , by ) ∈ R .If the following conditions are met.
(Δcx )2 + (Δcy )2 > (Δbx )2 + (Δby )2 + r + ra (5)
A starting point is the presence of AUV through navigational path and destination
point B without a single obstacle collision avoidance.
3.2 Route Multiple Obstacles Avoidance Model
AUV set maximum safe radius of ra , and set between the start and end of all
+
obstructions are located between. A starting point is between (a x , a y ) ∈ R ,B target
+
point is between (cx , c y ) ∈ R and an n the obstacle underwater space. Assuming
the k the obstacle b(k ) is located (bkx , bky ) ∈ R + , the radius of the obstacle is
rk . Then , if the following conditions.
(Δcx )2 + (Δcy )2 > (Δbkx )2 + (Δbky )2 + rmax + ra , ∀(bkx , bky ) ∈ O (6)

A starting point is the presence of AUV through navigational path and destination
point B without collision avoidance multiple obstacles.
4 Experimental Verification
4.1 Problem Description
AUV obstacle avoidance task route planning is planning to meet the optimal route
planning based on certain performance indicators mission objectives. The route
mission planning problem into a k dimensional function optimization problems.
y − y1
α = arcsin 2 (7)
AB
 x   cos α sin α   x '   x1 
 = ⋅  +   (8)
 y   − sin α cos α   y '   y1 
According to equation (7) and Equation (8) to the original coordinate system
conversion of the connection of the horizontal coordinate system for the new start and
end points.
α is a coordinate rotation angle. The new coordinate system x -axis is divided
'
'
into k segments, optimizing the corresponding y -coordinate.
After the conversion of the coordinates ( x ' , y ' ) connected in order to obtain a
path connecting the start and end points , Thereby converting the problem into a k
dimensional function route optimization problem.
4.2 Barriers Threat Level
AUV navigation path length Li , j ,The overall threat level barriers to the M n
obstacle
Mn
tk
 [( x − x )
Lij
λn , L =  dl (9)
ij 0
k =1 k
2
+ ( y − yk )2 ]2
The AUV navigation path into X segments, If the obstruction to navigation path
segment from the threat within a radius of obstacles, the barrier is calculated as the
threat level.
L5ij Mn
1 1 1 1 1
λn,L =
ij
x
α (l
k =1
k 4
+
l4
+ 4
l
+ 4
l
+
l4
) (10)
0.2,k 0.4,k 0.6,k 0.8,k 1.0,k
The length of the start and end points y z edge Lij , α k obstacle to obstacle
4
threat level, l0.2,k represents 1/5 the first pitch from the center of the k obstacle
edge Lij .
4.3 Program Flow

Steps are as follows :
Step1 Transformed coordinate system, the threat level obstacles to the rotating
coordinate system conversion, and the horizontal axis of rotation of the coordinate
system K aliquots.
Step2 Initialization K segment route , each route is calculated barriers threat level.
Step3 Iterative calculations.
Step4 For the population of feasible solutions consisting of K, perform mutation
operation.
Step5 Individual against individual variation generated with the original crossover
operation execution, generate new individuals.
Step6 Calculate the value of the cost function crossover operation to generate new
individuals, compared to individuals with newly generated target individuals choose
to perform the operation.
Step7 Iterations<maximum number of iterations, jump to Step3 iterative
calculation, otherwise exit the loop.
Step8 Coordinate inverse transform, output optimal route avoidance task planning
results.
4.4 Simulation
AUV starting point coordinates [10,10],End coordinates [55,100],Consideration

weights0.5. Set the initial parameters , population size NP=20,Optimization
dimension K=20,The maximum number of iterations Ncmax =200,Variability factor
F=0.5,Cross factor CR=0.9.
Table 1. AUV route barrier parameter

Obstacle radius Barriers threat Disorders Center
level
20 5 [45,50]
15 2 [12,40]
20 6 [32,48]
18 6 [36,26]
16 4 [22,40]
22 10 [30,45]
100
Obstacle 1 Obstacle 5
90 Obstacle 2
Obstacle 3
80 Obstacle 4
Obstacle 5
70 AUV avoidance route Obstacle 3
finishing point
starting point
60
Obstacle 4
Y
50
40
Obstacle 2
30
Obstacle 1
20
10
10 15 20 25 30 35 40 45 50 55
X
Fig. 2. AUV mission planning route avoidance
140
Differential evolution algorithm evolutionary curve
130
120
Generation of value
110
100
90
80
70
60
0 20 40 60 80 100 120 140 160 180 200
Number of iterations
Fig. 3. Differential evolution algorithm evolutionary curve

5 Conclusion
By differential evolution algorithm route planning AUV mission planning can

successfully avoid multiple obstacles, arrived in the end of the mission objectives.
Simulation results show the differential evolution algorithm solves the problem of
high-dimensional optimization , convergence speed and good performance.
Acknowledgments. This work was supported in part by the National Natural Science
Foundation of China ( 60975071,61100005 ) , Ministry of Education, Scientific
Research Project (13YJA790123).
References
1. Isern-Gonzalez, J., Hernandez-Sosa, D., et al.: Obstacle Avoidance in Underwater Glider
Path Planning. Physical Agents 6(1), 11–20 (2012)
2. Tsou, M.-C., Hsueh, C.-K.: The Study of Ship Collision Avoidance Route Planningby Ant
Colony Algorithm. Journal of Marine Science and Technology, 746–756 (2010)
3. Cruz, G.C.S., Encarnação, P.M.M.: Obstacle Avoidance for Unmanned Aerial Vehicles.
Journal of Intelligent & Robotic Systems, 203–217 (2012)
4. Yan, G., Wang, L., Zhou, J., Zha, Z.: Path Planning Based on Improved Genetic Algorithm
for AUV. Journal of Chongqing University of Technology, 115–120 (2010)
5. Storn, R., Price, K.: Differential Evolution A Simple and Efficient Heuristic for Global
Optimization over Continuous Spaces. Journal of Global Optimization, 341–359 (1997)
6. Price, K.: Differential Evolution A Fast and Simple Numerical Optimizer. In: Proceedings
of Biennial Conference of the North American Fuzzy Information Processing Society, pp.
524–527 (1996)
7. Islam, S.M., Das, S.: An Adaptive Differential Evolution Algorithm With Novel Mutation
and Crossover Strategies for Global Numerical Optimization. IEEE Transactions on
Systems, Man, and Cybernetics Part B: Cybernet ICS, 482–500 (2012)
8. Deng, Y., Beaujean, P.-P.J., An, E., Carlson, E.: Task Allocation and Path Planning for
Collaborative Autonomous Underwater Vehicles Operating through an Underwater
Acoustic Network. Journal of Robotics, 1–15 (2013)
9. Qu, H., Xing, K., Alexander, T.: An Improved Genetic Algorithm With Coevolution
Strategy for Global Path Planning of Multiple Mobile Robots. Neuro Computing 120,
509–517 (2013)
10. Ye, W., Wang, C., Yang, M., Wang, B.: Virtual Obstacles Based Path Planning for Mobile
Robots. Robot, 273–286 (2011)
11. Song, Q., Liu, L.: Mobile Robot Path Planning Based on Dynamic Fuzzy Artificial
Potential Field Method. International Journal of Hybrid Information Technology 5, 85–94
(2012)
An Improved Particle Swarm Optimization-Based
Coverage Control Method for Wireless Sensor Network
Huimin Du1,2 , Qingjian Ni1,2,3, , Qianqian Pan4 , Yiyun Yao1 , and Qing Lv1
1
College of Software Engineering, Southeast University, Nanjing, China
2
Provincial Key Laboratory for Computer Information Processing Technology,
Soochow University, Suzhou, China
3
School of Computer Science and Engineering, Southeast University, Nanjing, China
4
School of Information Science and Engineering, Southeast University, Nanjing, China
nqj@seu.edu.cn
Abstract. Coverage control plays a significant role in wireless sensor network

(WSN) design. To meet a layout with a certain cover rate, movable nodes are
maintained in deployment which accomplish self-organization through moving
and changing topological structure. This paper proposes an improved discrete par-
ticle swarm optimization algorithm aimed at coverage control method of WSN,
and the optimization is implemented under two processes: deployment planning
and movement control. The method interpreted in this paper can be easily used
solving such problems and the experiment result shows its efficiency, which will
inspire new insights in this field.
Keywords: Wireless Sensor Network, Coverage Control, Discrete Particle

Swarm Optimization.
1 Introduction
Wireless sensor network (WSN) in complex environment has typical characteristics like
large-scale, self-organization, limited energy for nodes and inconstant topology struc-
ture, etc. Every node in the network contains a small volume, cheap, energy-saving,
multifunction sensor and each sensor has the ability of signal acquiring, data handling
and communicating with its neighbors. These features have made WSN topology con-
trol a challenging issue.
The quality of topology control influences directly on the lifetime and performance
of networks, while a good topology scheme relies on a complete evaluation method-
ology. Composing those characters and system features, three following indicators are
taken into major considerations[1] to evaluate the WSN topology control:
– Coverage: Coverage is a measure of WSN service quality, which is mainly focused
on the coverage rate of initial nodes deployment and whether these nodes can ac-
quire signals of the region of interest(ROI), completely and accurately.
– Connectivity: Sensor networks are usually of large scale, thus connectivity is an
assurance that data information obtained by sensor can be delivered to sink nodes.


An Improved PSO-Based Coverage Control Method for WSN 115
– Network lifetime: Network lifetime is generally defined as the time duration from
the start to when the percentage of dead nodes comes to a threshold.
Coverage control, or deployment design, is the cornerstone of wireless sensor
networks. Node deployment can follow two trends: structured and randomized[2]. Struc-
tured method are suitable for small scope deployment where nodes positions are prede-
fined when planning, while randomized way are more pervasive. For supervisory region
with large scope which is hard to approach for humans, nodes are initialized (airdropped
usually) randomly and adjusted by topology control technology to achieve monitoring.
In such case, the mobility of nodes is rather crucial.
Aleksandra et al. stated a hexagonal repartition-based C 2 algorithm[3]. This algo-
rithm organizes the space to hexagonal grid, and chooses Cluster Heads (CHs) in the
center for each grid cell, using them to rearrange the nodes inside and adjacent to
the cells to improve the coverage ratio and connectivity. Zou et al. proposed a Vir-
tual Force-based deployment algorithm(VFA), dividing the nodes in the network into
clusters[4], where every cluster head node collects information of nodes inside the clus-
ter and computes their final positions and instructs the movement of nodes. Ma et al. put
forward an Adaptive Triangular Deployment Algorithm(ATRI) to deal with large-scope
situations[5]. This process adapts node deployments to regular triangles, and divide
node transmission range into six sector and thus nodes can be adjusted from view of its
neighbors from each sector.
The utilization of swarm intelligence has made the control processes more effective
and easier to implement. Liu[6] et al. introduced Easidesign algorithm for WSN cover-
age control based on Ant Colony Optimization(ACO), which combines greedy strategy
and additional pheromone evaporation methods to satisfy network connectivity of dif-
ferent sink positions. A Virtual force co-evolutionary PSO(VFCPSO) was proposed by
Wang et al.[7] In this algorithm a node is moved several times by the virtual force from
other nodes, and the virtual force vectors come from the distance information, their
moving direction and other factors. This can also reach a higher coverage ratio. An-
other situation where mixing stationary and mobile nodes is solved by Li ed al. using
a novel particle swarm genetic optimization(PSGA), combining PSO and Genetic Al-
gorithm(GA) to repair network holes in [8]. In this method, positions of mobile nodes
(or robots) are adjusted to improve the quality-of-service(QoS ). This method imports
mutation and selection operators to PSO and implements some extra update methods,
which are proved to be well-performed.
This paper proposed a novel discrete PSO strategy and applied it to WSN coverage
control, to improve QoS stated in the following part. The rest of this paper is arranged as
follows: Section 2 describes the abstraction and modeling of coverage problem, Section
3 explains the basic concepts of PSO, and a new discrete strategy with redefined opera-
tors is presented in this section. Experiments are conducted and analysed in Section 4,
and conclusion shows up in Section 5.
116 H. Du et al.
2 Problem Analyses
2.1 Problem Statement
In WSNs, every node has a certain length of sense radius R s and communication radius
Rc . Metrics of QoS include coverage rate, uniformity, time and distance[9], and we
mainly consider the coverage and distance problem in this paper.
Coverage. Measuring coverage rate is to detect the ratio of scope inside sense range to
the whole object range. Coverage scope is often interpreted as the amount of area. For
a node vi , its coverage scope COVi in the object region A equals to its sense range, and
the total amount of coverage range of the network is explained in formula (1):
COVA = ∪ki=1 COVi (1)

thus coverage rate can be represented as (2):
C = COVA /M (2)
where M is the area of object region.
Raster coverage is a meliorative strategy, where the ROI is meshing into a grid. The
grid points rather than the whole area are treated as the coverage object[10]. This strat-
egy is expanded as region-based point covering in [6]. In such a problem, some geo-
graphical points that can show the environment situation are chosen as covering points
(CPs), and the coverage object is to cover these CPs which are not necessarily grid dis-
tributed. Hence, coverage rate becomes the ratio of covered CPs. Typical applications
of such problems are Environmental Monitoring Systems and Targets Monitoring, etc.

For a node vi , its coverage scope is the number of CPs inside its sense range, coverage
rate is also formulated as (1)(2), where M is the total number of CPs. Research in this
paper are expended based on raster coverage.
Distance. The distance a node travels in the movement process is related to the energy
limitation. Therefore, optimization strategies are taken to minimize the distance of a
node and the total distance of a network. Distance a node takes is regard as the moving
range from its initial position to the objective.
2.2 Problem Modeling

Nodes are deployed randomly at initial stage and after the position optimization, a
higher coverage rate is obtained. The goal of this process is to maximize the cover-
age rate and minimize moving distances as well, achieving the minimum energy cost.
In deployment design, object region are divided into a grid, and every grid point
acts as a candidate position of nodes. Deployment process contains two procedures,
deployment planning and movement control, which are explained as followings:
– For a determined region and CP distribution (showing in Fig.1), design the objective
deployment layout for a certain amount of nodes.
– Adjust the positions of randomly-deployed nodes to meet the objective layout.
Formula (1) and (2) are detecting indicators in covering. As for moving process, whose
object sketch is shown in Fig.2, consider two sets, P = {p1 , p2 , ..., pN } and
Q = {q1 , q2 , ..., qN }, where P is the set of stochastic positions generated preliminar-
ily, and Q is the objective position set, N is the number of nodes. The purpose is to pair
the vertexes from different set completely in an nonredundant way, that is to make a
vertex qi of Q the moving target of vertex p j in P, and at the same time, minimizing
total moving distance. This can be measured by (3),

N
F= Distance(qi , p pair with qi ) (3)
i=1
where F is the objective function to be minimized.
Initial point
Objective point
200 CP 90
180 80
160
70
140
60
120
50
100
80 40
60
30
40
20
20
0 50 100 150 200 10 20 30 40 50 60 70 80 90 100
Fig. 1. Distribution of CPs Fig. 2. Objective Sketch for Moving
3 Improved Particle Swarm Optimization for WSN Coverage

Control
The section gives a brief summary of basic Particle Swarm Optimization and puts for-
ward an improved method for discrete PSO.
3.1 Basic Principle of PSO

Particle Swarm Optimization (PSO) is a meta-heuristic algorithm that simulates the
action of bird folks, proposed by J. Kennedy and R.C. Eberhart, and has been wildly
used in solving NP-hard problems. Basic PSO method is:
vid (t + 1) = vid (t) + c1 · r1 · (pid (t) − xid (t)) + c2 · r2 · (pgd (t) − xid (t)) (4)
xid (t + 1) = xid (t) + vid (t) (5)

118 H. Du et al.
Assume that in a searching space of D-dimension, a population consists of m parti-

cles. Position of particle i in the population can be represented by a D-dimension vector:
Xi = {xi1 , xi2 , · · · , xiD }, its present velocity is expressed as Vi = {vi1 , vi2 , · · · , viD }, and
its particle best position is recorded as Pi = {pi1 , pi2 , · · · , piD }, where i = 1, 2, · · · , m.
During optimization, a fitness function is built to judge the quality of a particle position.
3.2 Typical Discrete Strategy

Three typical discrete particle swarm optimization(DPSO) will be introduced: Binary
Strategy, Integer Strategy[11] and DPSO aimed at TSP.
Binary PSO: In Binary PSO, each particle is made of binary decisions, valid values
are like 0(False) and 1(True), and the probability of those values are modeled as[12]:
P(Xid = 1) = f (Xid (t − 1), Vid (t − 1), pid , pgd ) (6)
It’s proved that the probability for a dimension to choose 1/True of a particle is a mul-
tivariate function, depending on its previous position Xid and velocity(trend), which
are radically decided by the particle itself and the environment, i.e., pid and pgd . Use
sigmoidal function (7)
1
s(Vid ) = (7)
1 + exp(−Vid )
can educe the corresponding threshold determined by Vid in the probability function P.
And ⎧
⎪
⎪
⎨1, i f rand() < s(Vid )
Xid (t) = ⎪
⎪ (8)
⎩0, otherwise
gives the value conditions of Xid .

Integer PSO: Integer PSO comes out when solutions are constrained in only integer
space but not necessarily binary. This can be seemed as a constrained problem in con-
tinuous space, where optimal values are approximated to integers. Advantage of this
method is that, though solution space are truncated, algorithm performs still well[13].
DPSO for TSP: Clerc[14], Wang[15] et al., Shi[16] et al. all proposed novel DPSOs
when solving traveling salesman problem(TSP). Basic thoughts of DPSO for TSP is like:
particle Xi (t) = {xi1 , xi2 , · · · , xiD } stands for a traveling routine. Start from city xi1 , go
through xi2 , xi3 ...,reach xiD . In this kind of DPSO, velocity are defined as a swap table.
That’s to say, (πai , πbi ) means to swap city πai and πbi , vi j = ((πai1 , πbi1 ), ..., (πain, πbin )),
for example. For the velocity above, |vi j | = n, vector length equals to the number of swap
items. Therefore, velocity here contains no repeated item.
3.3 Proposed DPSO for WSN

Based on those existed DPSOs, this paper proposed an improved discrete strategy aimed
at WSN coverage problem.
For PSO basic updating formulas (4)(5), there are elements like: position information
(such as X, P), velocity information (V), weighting coefficients (ω, c ∗ rand()). In this
discrete algorithm, the information are defined as follows:
Position Information: A multi-dimension vector that contains no repeated element,

it stands for the spatial situation of a particle and is a sequential list.
Velocity Information: A multi-dimension vector in which elements may repeat, and
is the difference of two positions.
Weighting Coefficient: A float number in [0.0, 1.0] that shows the weight of a certain
element among all the elements.
Updating method for velocity and position are redefined as:
Vi(t+1) = W(ω) ⊗ Vi(t) ◦ W(c1 r1 ) ⊗ (Pid Xi(t) ) ◦ W(c2 r2 ) ⊗ (Pgd Xi(t) ) (9)
X (t+1) = X (t) ⊕ V (t) = {x1(t) + ...x(t) (t) (t)

N } ⊕ {v1 , ..., vN } (10)
where W is a normalizing function to produce weighting coefficient. Assume there are
k factors in all (k = 3 in (9)) for a formula, then W is built as:
f actori
W( f actori ) = (11)

k
f actori
i=0
Operators are defined as follows:

⊗ Multiplication operator for coefficient and velocity: The result is a vector contain-
ing null values. Retain a certain number (round of coe f f icient ×|vector|) of dimensions
using some strategy (randomly, for example) and set other dimensions to null value and
composes a resultant vector. Indexes of retained items should be unique among all the
vectors to be added.
◦ Addition operator for two velocities: Merge to a new vector according to the item
indexes. For example, {vi1 , vi2 , null, vi4 } ◦ {null, null, vi3 , null} = {vi1 , vi2 , vi3 , vi4 }. Ac-
cording to (11) and multiple method above, all the sub-vectors will be merged to exact
one complete vector.
Subtraction operator for two positions: This operation results in a velocity vector.
Define operate mode for each dimension as follow:
⎧
⎪
⎪
⎪ pd , i f rand() ≥ α
⎪
⎪
⎨
p d xd = ⎪
⎪ xd + γ(pd − xd ), i f α > rand() > β (12)
⎪
⎪
⎪
⎩ xd , otherwise
where 0 < γ < 1, β ≤ α < 1, rand() ∈ [0, 1], and (β, α) is the probability interval of
interference factor, which is optional.
⊕Addition operator for a velocity and a position: This will produce a new position.
Here, assume that velocity and position vectors are of the same dimension. Operation
for each dimension are defined as:
⎧
⎪
⎪
⎪ vd , i f rand() ≥ α

⎪
⎪
⎨
xd ⊕ vd = ⎪
⎪ xd + γ (vd − xd ), i f α > rand() > β (13)
⎪
⎪
⎪
⎩ xd , otherwise
120 H. Du et al.
T HEN do xi ← xd where xi == xd ⊕ vd

where 0 < γ < 1, β ≤ α < 1, rand() ∈ [0, 1], and (β , α ) is the probability interval
of interference factor which is optional. This operation has two steps: calculate xd ⊕ vd
and swap the result with xd inside the vector.
4 Experiment Results and Analyses
4.1 Deployment Planning
While implementing PSO algorithm, every particle is a candidate solution maintaining

the coordinates of all the nodes, i.e., a deployment layout. Thus, for N nodes in the two-
dimensional space, every particle is a 2N-dimensional vector searching those discrete
vertexes in the space, i.e., the grid points. For particle i, Xi = {xi1 , xi2 , ..., xi2N }.
CPs are chosen in the space as Fig.1, and N nodes are placed randomly. Appropri-
ate decision of N can be different, but it should usually be more than needed to face
unexpected conditions. Algorithm 1 and 2 deal with this situation. Use PSO process
(algorithm 1) for elementary optimization. LocalBest version of PSO using ring topol-
ogy [11] is adopted in this process. In particular, this is a constraint problem that the
position of a node should stay in a certain scale and should have neighbor nodes near
around which it can deliver message to. For a node ni , define NCi as the set of neighbor
nodes which are inside the communication range of ni .
NCi = {n j |distance(n j, ni ) ≤ Rc } ∅ (14)

Then the position update formula (5) can be strengthened as update-method 2.
Update-Method 2
1 tempPi(t+1) = Pi(t) + Vi(t+1) = {xi(t+1) , yi(t+1) , zi(t+1) , ...}

2 if xi(t+1) > threshold
3 then xi(t+1) =threshold
4 judge yi(t+1) , zi(t+1) ...
5 obtain NCi(t+1)
6 if NCi(t+1) null
7 then Pi(t+1) = tempPi(t+1)
8 else Pi(t+1) = Pi(t)
Result of algorithm 1 is shown in Fig.3. This algorithm is aimed at single-cover con-

dition, and some CPs in this region are covered more than once. When multi-coverage is
not required, nodes are said to be of no coverage benefit where CPs inside its coverage
range are all covered by other nodes already. These nodes will be chosen as dormant
Algorithm 1: Pso Process
1 Initialize D randomly particles, P(0) (0)

1 , · · · , PD
2 for i=1 to D
3 Set Pbests(0)
i =Pi
(0)
4 Judge quality of Pi and set Lbests(0)

5 for i=1 to D
6 Update Pi value by (4) and Update-Method2
7 Judge quality of Pi
8 Update Pbests and Lbests
9 if a terminal condition is met
10 then go to step 5
11 else go back to step 12
12 Stop and output the best solution
nodes and adapted to sleep mode. Algorithm 2 using a traverse accomplished this job,
reducing the number of working nodes to a smaller quantity.
Object deployment layout is like Fig.4 after the optimization, which reached a high
coverage rate with a smaller number of nodes for single-cover case, and the connectivity
of this network can then be achieved overtly. Flexibility is a significant advantage of this
deployment method, since some extra nodes will be awaiting inside the region, a new
layout will comes up quickly without external aid once environment changes occur or
working nodes get problem. In such case the coverage rate C ≥ 98%.
Node
Sense Range
Node
CP
Sense Range
CP
220
220
200
200
180
180
160
160
140
140
120
120
100
100
80
80
60
60
40 40
20 20
0 50 100 150 200 250 0 50 100 150 200 250
Fig. 3. Elementary Coverage Result Fig. 4. Adapt Dormant Nodes

122 H. Du et al.
Algorithm 2: Traverse Method

1 Nodes are represented as {x1 , x2 , ..., xN }
2 for i = 1 to N
3 j=i
4 Mark all the unmarked CPs in the sense range of x j ,
the amount is denoted as n
5 if n = 0
6 then pick x j into dormant set
7 Let j = ( j + 1)%N, if j i
8 then go back to step 4
9 else go to step 10
10 Get plani
11 plan = min{plan1 , ..., planN }
4.2 Movement Control
Geographical positions of nodes are stochastically initialized and will be changed to

specific ones. Since movement consumes energy and reduces life cycle, a shortest mov-
ing plan is expected. Assume that every two vertexes in the region can be connected
with one straight path (no obstacle between).
Improved DPSO proposed in this paper is the solver. P is the initial position set and
Q is the object set, which are taken as sequences. Purpose of this algorithm is to realign

P to P = {p1 , p2 , ..., pN }, as to let pi pair with qi , making up the moving plan. This plan
are illustrated in Fig.5.
Initial point
Object point
90 Move path
80
70
60
50
40
30
20
10 20 30 40 50 60 70 80 90 100
Fig. 5. Moving Plan Sketch Generated by DPSO

Parameters in formulas (9)(12)(13) are set as following: ω = 0.6, c1 = 2.2, c2 =

1.4, α = 0.6, β = 0.4, γ = 0.6, α = 0.7, β = 0.7, γ = 0.6.
Besides PSO, Ant Colony Optimization (ACO) also has dominant performance solv-
ing discrete problems[17]. Comparisons are made between ACO with MMAS strategy
and DPSO proposed in this paper. Fig.6 shows when area scope is 200 × 200, perfor-
mance difference with different node amount. Solid line indicates DPSO performance
while dotted line is for ACO. Fig.7 represents the coverage quality for different object
area scope with a certain amount of nodes (N = 45). Figures stand for the average result
of decades runs of the algorithms.
As can be seen, discrete strategy put forward in this paper performs dominantly
within a certain range. A lower total distance in the figures claims a better performance.
While the node amount is within a threshold(around 100), a lower energy is obtained,
and so is the result when the ROI scale is less than 450. This signifies that when imple-
mented in a relatively small scale (of node amount and scale), DPSO proposed in this
paper can provide a faster and easier way to a solution. What should be noticed is about
the parameters adopted above. Adjustment to those parameters may be needed when
situations changes, such as different sensing range and irregular shape of ROI. In such
situations, actual concrete data of algorithm may also be different.
1600 1200
1100
1400
1000
1200
900
Total Distance
Total Distance
1000 800
700
800
600
600
500
DPSO
400 DPSO
ACO 400
ACO
200 300
20 30 40 50 60 70 80 90 100 110 120 100 150 200 250 300 350 400 450 500
Node Amount Area Scale
Fig. 6. Performance Comparison with Fig. 7. Performance Comparison with

Different Node Amount Different Area Scope
5 Conclusion
Coverage control for WSN includes many aspects. Besides coverage rate and connec-
tivity, multi-dimensional environments in real world with obstacles and change still
need more considerations. This paper mainly discusses application of PSO to two-
dimension coverage problem and improved traditional DPSO based on characteristics
of WSN coverage control. Results illustrate that in movement control, the improved
DPSO which is easy to handle, is of high performance which is no less than traditional
discrete problem solver. Successive research in multi-dimension covering problem is to
be conducted with further applications of improved DPSO. Multi-objective problems
on network uniformity and life-cycling are also crucial points following up.
124 H. Du et al.
Acknowledgement. This paper is supported by Provincial Key Laboratory for Com-

puter Information Processing Technology, Soochow University (KJS1223), Suzhou,
China and NSFC (Grant No.61170164).
References
1. Ghosh, A., Das, S.: Coverage and Connectivity Issues In Wireless Sensor Networks: A Sur-
vey. Pervasive and Mobile Computing 4, 303–334 (2008)
2. Younis, M., Akkaya, K.: Strategies and Techniques For Node Placement in Wireless Sensor
Networks: A survey. Ad Hoc Networks 6, 621–655 (2008)
3. Mateska, A., Gavrilovska, L.: Wsn coverage and connectivity improvement utilizing sensors
mobility. European Wireless, pp. 686–693 (2011)
4. Zou, Y., Chakrabarty, K.: Sensor Deployment and Target Localization Based on Virtual
Forces. In: International Conference on Computer Communications (INFOCOM), NC, USA,
pp. 1293–1303 (2003)
5. Ma, M., Yang, Y.: Adaptive Triangular Deployment Algorithm For Unattended Mobile Sen-
sor Networks. IEEE Transactions on Computers 56, 946–958 (2007)
6. Wei, L., Li, C.: Ant Based Approach to The Optimal Deployment in Wireless Sensor Net-
works. Journal on Communications 30, 25–33 (2009)
7. Wang, X., Wang, S., Ma, J.J.: An Improved Co-Evolutionary Particle Swarm Optimization
for Wireless Sensor Networks with Dynamic Deployment. Sensors 7, 354–370 (2007)
8. Li, J., Li, K., Zhu, W.: Improving Sensing Coverage of Wireless Sensor Networks by Em-
ploying Mmobile Robots. In: Proceedings of the International Conference on Robotics and
Biomimetics (ROBIO), pp. 899–903 (2007)
9. Heo, N., Varshney, P.K.: A distributed self spreading algorithm for mobile wireless sensor
networks. In: IEEE Conference on Wireless Communications and Networking, March 16-20,
vol. 3, pp. 1597–1602 (2003)
10. Chakrabarty, K., Iyengar, S.S., Qi, H., Cho, E.: Grid coverage for Surveillance and Target
Location in Distributed Sensor Networks. IEEE Transactions on Computers 51, 1148–1153
(2002)
11. del Valle, Y., Venayagamoorthy, G.K., Mohagheghi, S., Hernandez, J.-C., Harley, R.G.: Par-
ticle Swarm Optimization: Basic Concepts, Variants and Applications in Power Systems.
IEEE Transactions on Evolutionary Computation 12(2), 171–196 (2008)
12. Kennedy, J., Eberhart, R.: A Discrete Binary Version of The Particle Swarm Algorithm. In:
Computational Cybernetics and Simulation (ICSMC), vol. 5, pp. 4104–4108 (1997)
13. Parsopoulos, K., Vrahatis, M.: Recent Approaches to Global Optimization Problems
Through Particle Swarm Optimization. Natural Computing 1, 235–306 (2002)
14. Clerc, M.: Discrete Particle Swarm Optimization Illustrated by The Traveling Salesman
Problem. In: Onwubolu, G.C., Babu, B.V. (eds.) New Optimization Techniques in Engineer-
ing. STUDFUZZ, vol. 141, pp. 219–239. Springer, Heidelberg (2004)
15. Wang, K.P., Huang, L., Zhou, C.G., Pang, W.: Particle Swarm Optimization for Travel-
ing Salesman Problem. In: International Conference on Machine Learning and Cybernetics,
vol. 3, pp. 1583–1585 (2003)
16. Shi, X.H., Lianga, Y.C., Leeb, H.P., Lub, C., Wanga, Q.X.: Particle Swarm Optimization-
Based Algorithms for TSP And Generalized TSP. Information Processing Letters 103, 169–
176 (2007)
17. Dorigo, M., Stutzle, T.: Ant Colony Optimization: Overview and Recent Advances. Interna-
tional Series in Operations Research and Management Science, Handbook of Metaheuristics,
IRIDIA/2009-013, pp. 227–263 (2009)
An Improved Energy-Aware Cluster Heads
Selection Method for Wireless Sensor Networks
Based on K-means and Binary Particle Swarm
Optimization
Qianqian Pan1,2 , Qingjian Ni2,3,4 , Huimin Du3 , Yiyun Yao3 , and Qing Lv3
1
School of Information Science and Engineering,
Southeast University, Nanjing, China
2
Laboratory of Military Network Technology, PLA University
of Science and Technology, Nanjing, China
3
College of Software Engineering, Southeast University,
Nanjing, China
4
School of Computer Science and Engineering,
Southeast University, Nanjing, China
nqj@seu.edu.cn
Abstract. The limited and non-replenishable energy supply is the main

character of Wireless Sensor Networks (WSNs). Hence, maximizing the
lifetime of WSNs becomes a critical issue in sensor networks. Clustering
is one of the most effective means to extend the lifetime of the whole
network. In this paper, an energy-aware cluster heads selection method,
based on binary particle swarm optimization (BPSO) and K-means, is
presented to prolong the network lifetime. We apply the BPSO and make
it suitable for this issue. The selection criteria of the objective cost func-
tion are based on minimizing the intra-cluster distance as well as the
distance between cluster heads and base station, and optimizing the en-
ergy consumption of the whole network. In addition, the sensor nodes
are divided into several clusters based on K-means algorithm at the be-
ginning, which can reduce the complexity of the whole algorithm. The
performance of our technique is compared with the well-known cluster-
based sensor network protocols, LEACH-C and PSO-C respectively. The
simulation results demonstrate that our proposed work can achieve bet-
ter network lifetime over its comparatives.
Keywords: Wireless Sensor Networks, Binary Particle Swarm

Optimization, Energy-aware Cluster, K-means Algorithm.
1 Introduction
Wireless sensor network is a kind of wireless network composed of a large number

of sensor nodes [1][2]. These tiny nodes are usually scattered in a sensor field


126 Q. Pan et al.
and each of these scattered sensors is equipped with a capability to collect data
and route data back to the sink, base station (BS). Recent advancement in
wireless communication and electronics enable the development of the WSNs.
The network has a wide range of applications such as health, military and home.
However, sensor node, usually powered by batteries, is limited in energy supply.
Therefore, energy efficiency should be considered as the critical design objective.
Prolonging lifetime of the WSNs becomes a key issue.
Clustering is one of the most popular methods to prolong the lifetime of the
WSNs. One of the well-known clustering protocols is called LEACH [3], which
uses randomized rotation of cluster heads to evenly distribute the energy load
among the sensors in the network. LEACH-C (LEACH Centralized) [4] is the
extension to LEACH. In LEACH-C, BS finds the optimal cluster heads among
sensor nodes whose energy are above average, using the simulated annealing
algorithm [5].
Particle swarm optimization (PSO) [6][7] is a popular optimization technique,
simulating the social behavior of a flock of birds flying to the food. PSO algo-
rithm is applied to find cluster heads and produces better results [8]. A popular
clustering algorithm based on PSO is PSO-C [9], which selects cluster heads
considering both energy available to nodes and distances between the nodes and
their cluster heads.
In this paper, we develop an energy-aware cluster heads selection method
based on binary particle swarm optimization (BPSO) [10] and K-means [11]
(BPSO-K). Our proposed method selects the high-energy nodes as the clus-
ter heads and evenly distributes the energy load among nodes in the network.
The main idea of our protocol is selecting cluster heads that can minimize the
intra-cluster distance as well as the distance between cluster heads and BS, and
optimize the energy consumption of the whole network. K-means algorithm is
utilized at the beginning to divide the nodes into several initial clusters. The
rest of this paper is organized as follows: the network and energy models are
described in section II. The detailed description of our proposed energy-aware
cluster heads selection method for WSNs based on K-means and BPSO is out-
lined in section III. In section IV, we discuss the simulation study of the proposed
protocol. Finally, the concluding remarks appear in section V.
2 The System Model

2.1 Network Model
We assume the network model similar to those used in paper [4] and [9], with
the following properties.
1. Sensor nodes are scattered in a field at random and all nodes are static.
2. All sensor nodes are energy constrained, generally powered by batteries.
3. A BS is fixed inside or outside of the sensor filed.
4. Each sensor node has capabilities of processing data and sending data to the
BS.
5. Each node can compute its own location and energy level, and send them to
the BS.
An Improved Cluster Heads Selection Method for WSNs Based on BPSO-K 127
2.2 Energy Model

The energy model of our protocol is based on the classical model used in paper [3].
The radio hardware energy dissipation of this model is that the transmitter dis-
sipates energy to run the radio electronics and power amplifier, and the receiver
dissipates energy to run the radio electronics. Both free space and multipath
fading channel model are used according to the distance between transmitter
and receiver. Free space model with d2 is used, if the distance is less than a
threshold d0 . Otherwise, the multipath model with d4 is applied. Thus, in order
to transmit an l-bit message over a distance d, the radio energy extended is given
in the equation (1).

lEelec + l
fs d2 , if d < d0
ETx (l, d) = . (1)
lEelec + l
mp d4 , if d ≥ d0
And to receive an l-bit message, the energy expended by the radio is given as
equation (2).
ERx (l) = lEelec . (2)
where Eelec denotes the electronics energy, depending on the energy dissipated
per bit to run the transmitter or the receiver. The amplifier energy
fs d2 and
mp d4 depend on the distance between transmitter and receiver.
3 Method Description
3.1 Binary Particle Swarm Optimization
PSO [6][12] is a simple, effective, and computationally efficient optimization
algorithm for continuous optimization [14][15]. Binary PSO (BPSO) [16] is an
extension of PSO based on the binary coding scheme, proposed by Kennedy
and Eberhart. BPSO consists of a swarm of S particles. An individual possible
solution of a problem is presented by a D-dimensional particle. A particle i has a
coordinates xid and a velocity vid in the dth dimension, 1 ≤ i ≤ S and 1 ≤ d ≤ D.
The velocity is defined as changes of probabilities that decide bits of coordinate
in one state or the other. Thus, each dimension of a particle moves to a state
restricted to 0 or 1 depending on velocity. Each bit vid of velocity represents
the probability of bit xid taking value 1. Velocity of a particle is determined by
equation (3).
vid (t + 1) = vid (t) + c1 r1 (pid − xid ) + c2 r2 (pgd − xid ). (3)
where c1 and c2 are positive numbers, r1 and r2 are two random numbers between
0 and 1 with uniform distribution, and pid , pgd denote particle’s and global best
position respectively.
Since vid is a probability, it must be constrained in the interval of [0,1]. Sigmoid
function is used to normalization velocity vid based on equation (4).
1
s(vid ) = . (4)
1 + e−vid
128 Q. Pan et al.
where s(vid ) denotes the velocity after normalization.

The position of a particle is defined as equation (5).

1, if rand() ≤ s(vid )
xid = . (5)
0, otherwise
3.2 Proposed Method of Initialization Using K-means

In the process of establishing clusters, the nodes of location proximity are easier
to be assigned to a cluster, mainly because of lower energy for transmission
among location closer nodes. Hence, we propose a method of initialization to
divide sensor nodes into several initial clusters [13] according to location of nodes.
K-means [17] as a classical clustering algorithm can get different K groups
which are described by their centroid of nodes. Besides, the number of clusters
is determined by user. The algorithm of initialization is given as algorithm 1 in
detail.
Algorithm 1. Initialization using K-means

1: function K-means(Array, K, N )
2: for j = 1 → K do
3: C[j] ← random(1, N )
4: end for
5: repeat
6: for j = 0 → K do
7: Q[j] = ∅
8: end for
9: for i = 1 → N do
10: MinDist ← ∞
11: for j = 1 → K do
12: if Distance(Array[i], C[j]) < MinDist then
13: MinDist ← Distance(Array[i], C[j]);
14: MinNode ← j
15: end if
16: end for
17: Q[MinNode] ← Q[MinNode] + i
18: end for
19: C ← Updata(Q)
20: until C[j] not changed for j = 1 → K
21: return Q
22: end function
We divide sensor nodes into K initial clusters based on K-means algorithm

before the operation of the network and ensure that K ≤ M , where M denotes
the predetermined number of cluster heads. The proposed method of initializa-
tion will reduce the complexity of the whole algorithm with slight influence of
lifetime.
3.3 Proposed Cluster Heads Selection Method Based on BPSO
Our proposed method is a centralized clustering algorithm to form clusters by

dispersing the cluster heads throughout the network. The selection process of
our method is working on the BS. The operation is divided into rounds. Each
round begins with a set-up phase when the clusters are formed, followed by a
steady-state phase when data are transmitted to cluster heads and to the BS
[3]. During the set-up phase, each node sends information about current location
and energy level to the BS. Based on these statistics, BS computes the average
energy of the network. To evenly distribute energy among the whole network,
only the nodes with an energy level above the average can be possible cluster
heads for current round [18][19]. BS selects M cluster heads using BPSO, where
M is the optimizing number of cluster heads for the network.
In the cluster heads selection method based on BPSO [14], a D-dimensional
particle represents a selection of cluster heads, where D denotes the number of
the possible cluster heads with sufficient energy in current round. The position
xid represents the state of the possible cluster head d, where 1 ≤ d ≤ D. If xid
is restrained to 0, that means the possible cluster head d is not selected as a
cluster head in this particle. Otherwise, the value of xid is 1, in this particle,
possible cluster head d is chosen as head node. The protocol should ensure each
D-bit particle of the flock has exactly Mj bits restrained to 1 and the values of
other bits are 0, where Mj (j = 1, 2, . . . , K) represents the optimizing number of
K
cluster heads for initial cluster j and also ensure j=1 Mj = M . To do this, we
can adjust the value of xi with normalized velocity s(vi ). If the number of bits
whose value is 1 in a particle is more than Mj , we select the nodes with larger
normalized velocity as cluster heads. On the other hand, when the number is less
than Mj , we obtain other cluster heads from remaining possible cluster heads
according to larger s(vi ).
BS determines Mj cluster heads that suit the cost function best. The main
objective of cost function is to optimize the combined influences of distance and
energy [20]. Thus, we define the function depending on the following factors.
(1)The longest average distance between nodes and their cluster heads, defined
by equation (6).
⎧C ⎫
⎪
⎪ ⎪
d(CMjik , CHji ) ⎪
ji
⎪
⎨ ⎪
⎬
k=1
distj1 = max . (6)
i=1,2,··· ,Mj ⎪
⎪ Cji ⎪
⎪
⎪
⎩ ⎪
⎭
where Mj and Cji denote the number of initial clusters and nodes in cluster i
of the initial cluster j respectively, d(CMjik , CHji ) is distance between nodes
CMjik and its cluster head CHji .
(2)The longest distance between cluster heads and BS, given as equation (7).
distj2 = max {d(CHji , BS)}. (7)

i=1,2,··· ,Mj
130 Q. Pan et al.
(3)Energy consumption of whole network, represented as Ejsum . We can get

it based on the system model.
All these three factors should be considered in an integrated manner. However,
data for each of them has difference in dimension and magnitude between others.
To eliminate such effects, we introduce a normalized function to these three
factors, given as equation (8).
2
y= tanh(x). (8)
π
where y is the normalized data of x.
Combining three factors mentioned above, the cost function, represented as
fj , is specified in the equation (9).
fj = αdisttj1 + βdisttj2 + γEtjsum . (9)
where disttj1 , disttj2 and Etjsum are the normalized data of distj1 , distj2 and
Ejsum , α, β, γ are positive factors determining the priority weighting of disttj1 ,
disttj2 and Etjsum , with α + β + γ = 1. In this paper, we assume disttj1 and
disttj2 have identical impact on the cost function and Etjsum has a bit higher
influence due to energy often being considered as a main factor of WSNs. Hence,
we set α = 0.3, β = 0.3, γ = 0.4.
For a wireless sensor network with N nodes and M predetermined cluster
heads, the clusters formed in each round is given as algorithm 2.
Algorithm 2. Cluster heads selection method using BPSO-K

1: Q ← K-means(Arrary, K, N )
2: repeat
3: for j = 1 → K do
4: repeat
5: Select Mj cluster heads among candidates of Qj
6: for i = 1 → Mj do
7: P [j, i] ← ∅
8: end for
9: for i = 1 → N umOf N ode[j] do
10: ClosestHead ← the closest cluster head from node i
11: P [j, ClosestHead] ← P [j, ClosestHead] + i
12: end for
13: Calculate cost function
14: Update(LocalBest)
15: Update(GlobalBest)
16: Update(v, x)
17: until reach the set number of iteration
18: end for
19: until the network is dead
After selecting cluster heads of the network and each nodes is decided which
cluster it belongs to, the cluster heads act as local control centers to coordinate
the data transmissions in their cluster and send the fused data to the BS.
4 Simulations and Analysis
The proposed cluster heads selection method is simulated to evaluate its per-
formance. We define that a wireless sensor node is dead when it runs out of
energy and the network is dead at the moment the first node dies. We ran the
simulation for 100 nodes in a 200m × 200m network area with both equal and
unequal initial energy of nodes. Paraments used in energy model are similar to
paper [4], Eelec = 0.5nJ/bit,
fs = 10pJ/bit/m2,
mp = 0.0013pJ/bit/m4. The
number of clusters is set to be 5 percent [9] of the nodes M = 5. The initial
clusters divided by K-means is set as K = 5. For the paraments of BPSO, we
use S = 30 particles and c1 = c2 = 2. In addition, BS is set at location (0, 200)
and the data message size is fixed at l = 4000bit. The performance of our pro-
posed method is compared with the well-known cluster-based sensor network
protocols, LEACH-C and PSO-C.
Fig.1 illustrates the system lifetime, defined by the time of the first node died.
It also shows the performance of our proposed method compared with LEACH-
C and PSO-C with equal initial energy of nodes. We set the total nodes have
0.5J of initial energy. Fig.2 shows the performance with unequal initial energy of
nodes set from 0.3J to 0.7J randomly. The results shown in Fig.1 and Fig.2 are
selected randomly from our experiences, which are simulated 20 times for each.
We can find that when the first dead node occurs, the LEACH-C and PSO-C do
not run exceeding 50 rounds. Whereas the BPSO-K has run about 500 rounds
until the first node die.
105 105
BPSO−K BPSO−K
LEACH−C LEACH−C
PSO−C PSO−C
100 100
95 95
Number of nodes alive
90 90
85 85
80 80
75 75
70 70
0 100 200 300 400 500 600 0 100 200 300 400 500 600
Rounds Rounds
Fig. 1. Number of nodes alive with equal Fig. 2. Number of nodes alive with un-
energy equal energy
Clearly our proposed protocol can prolong the lifetime of network significantly
compared to LEACH-C and PSO-C. It fairly assigns energy consumption to
each node in the field by selecting cluster heads periodically based on BPSO-K,
which is helpful to avoid some sensor nodes scattered in the field dying too early.
Besides, the cost function consist of both distance and energy also plays a critical
role in prolonging the lifetime of the WSNs.
132 Q. Pan et al.
Fig.3 shows the clusters divided by K-means when K = 5. Wireless sensor

nodes are assigned to different clusters based on their location. The nodes of
location proximity are assigned to same cluster.
Fig.4 and 5 show the comparison of BPSO and BPSO-K with equal and un-
equal initial energy nodes. The results shown in Fig.4 and Fig.5 are selected
randomly from our experiences, which are simulated 20 times for each. The sim-
ulation results demonstrate K-means can reduce the complexion of the method
while having slight impact on the lifetime of the network.
200
180
160
140
120
100
y
80
60
40
20
0
0 20 40 60 80 100 120 140 160 180 200
x
Fig. 3. Clusters divided by K-means
100 BPSO−K 100 BPSO−K

BPSO BPSO
90 90
80 80
70 70
60 60
50 50
40 40
30 30
20 20
10 10
0 0
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000 1100
Rounds Rounds
Fig. 4. Number of nodes alive with equal Fig. 5. Number of nodes alive with un-
energy equal energy
5 Conclusion
In this paper, we presented an improved energy-aware cluster heads selection
method for wireless sensor networks based on K-means and BPSO. We defined
a new cost function that takes into account the maximum of the intra-cluster
distance as well as the distance between cluster heads and BS, and the min-
imum of energy consumption. Results from the simulations indicate that the
proposed protocol using BPSO and K-means algorithm gives a higher network
lifetime compared to LEACH-C and PSO-C. Furthermore, the proposed proto-
col produces better clustering by evenly allocating the cluster heads throughout
the network area. The extension of this work would be a further discussion of
the parameters setting to the BPSO-K and to prolong the lifetime of networks
consist of mobile nodes.
Acknowledgment. This paper is supported by Laboratory of Military Network

Technology, PLA University of Science and Technology (LMNT2012-1), Nanjing,
China and NSFC (Grant No.61170164).
References
1. Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., Cayirci, E.: Wireless sensor net-
works: a survey. Computer Networks 38(4), 393–422 (2002)
2. Akkaya, K., Younis, M.: A survey on routing protocols for wireless sensor networks.
Ad Hoc Networks 3(3), 325–349 (2005)
3. Heinzelman, W.R., Chandrakasan, A., Balakrishnan, H.: Energy-Efficient Commu-
nication Protocol forWireless Microsensor Networks. In: 33rd Hawaii International
Conference on System Sciences. IEEE Press, Hawaii (2000)
4. Wendi, B., Heinzelman, A.P.: Chandrakasan, Hari Balakrishnan: An Application-
Specific Protocol Architecture for Wireless Microsensor Networks. IEEE Transac-
tions on Wireless Communications 1(4), 666–670 (2002)
5. Murata, T., Ishibuchi, H.: Performance Evaluation of Genetic Algorithms for Flow-
shop Scheduling Problems. In: 1st IEEE Conference on Computational Intelligence,
pp. 812–817. IEEE Press, Orlando (1994)
6. del Valle, Y., Venayagamoorthy, G.K., Mohagheghi, S., Hernandez, J.-C., Harley,
R.G.: Particle Swarm Optimization:Basic Concepts,Variants and Applications in
Power Systems. IEEE Transactions on Evolutionary Computation 12(2), 171–196
(2008)
7. Kulkarni, R.V., Venayagamoorthy, G.K.: Particle Swarm Optimization in Wireless-
Sensor Networks: A Brief Survey. IEEE Transactions on Systems, Man, and Cy-
bernetics Part C: Applications and Reviews 41(2), 262–267 (2011)
8. Bala Krishna, M., Doja, M.N.: Swarm intelligence-based topology maintenance
protocol for wireless sensor networks. IET Wireless Sensor Systems 1(4), 181–190
(2011)
9. Abdul Latiff, N.M., Tsimenidis, C.C., Sharif, B.S.: Energy-aware Clustering for
Wireless Sensor Networks Using Particle Swarm Optimization. In: 18th Annual
IEEE International Symposium on Personal, Indoor and Mobile Radio Communi-
cations, pp. 1–5. IEEE Press, Athens (2007)
10. Luh, G.-C., Lin, C.-Y., Lin, Y.-S.: A Binary Particle Swarm Optimization for Con-
tinuum Structural Topology Optimization. Applied Soft Computing 11(2), 2833–
2844 (2011)
11. Jain, A.K.: Data Clustering: 50 years beyond K-means. Pattern Recognition Let-
ters 31(8), 651–666 (2010)
12. Shi, Y., Eberhart, R.: A Modified Particle Swarm Optimizer. In: The 1998 IEEE
International Conference on Computational Intelligence, pp. 577–584. IEEE Press,
Anchorage (1998)
134 Q. Pan et al.
13. Wagstaff, K., Cardie, C.: Constrained K-means Clustering with Background
Knowledge. In: The Eighteenth International Conference on Machine Learning,
pp. 577–584. Morgan Kaufmann Publishers Inc., San Francisco (2001)
14. Chuang, L.-Y., Tsai, S.-W., Yang, C.-H.: Improved Binary Particle Swarm Opti-
mization Using Catfish Effect for Feature Selection. Expert Systems with Applica-
tions 38(10), 12699–12707 (2011)
15. Fernández-Martı́nez, J.L., Garcı́a-Gonzalo, E.: Stochastic Stability Analysis of the
Linear Continuous and Discrete PSO Models. IEEE Transactions on Evolutionary
Computation 15(3), 405–423 (2011)
16. Kennedy, J., Eberhart, R.C.: A Discrete Binary Version of the Particle Swarm
Algorithm. In: 1997 IEEE International Conference on Systems, Man, and Cyber-
netics, pp. 4104–4108. IEEE Press, Orlando (1997)
17. Na, S., Xumin, L., Yong, G.: Research on k-means Clustering Algorithm: An Im-
proved k-means Clustering Algorithm. In: 3rd International Symposium on Intel-
ligent Information Technology and Security Informatics, pp. 63–67. IEEE Press,
Jinggangshan (2010)
18. Shi, S., Liu, X., Gu, X.: An Energy-Efficiency Optimized LEACH-C for Wireless
Sensor Networks. In: 7th International ICST Conference on Communications and
Networking, pp. 487–492. Kun Ming (2012)
19. Karaboga, D., Okdem, S., Ozturk, C.: Cluster Based Wireless Sensor Network
Routing Using Artificial Bee Colony Algorithm. Wireless Networks 18(7), 847–860
(2012)
20. Singh, B., Lobiyal, D.K.: A Novel Energy-aware Cluster Head Selection Based on
Particle Swarm Optimization for Wireless Sensor Networks. Human-centric Com-
puting and Information Sciences 2(1) (2012)
Comparison of Multi-population PBIL and Adaptive
Learning Rate PBIL in Designing Power System
Controller
Komla A. Folly
Department of Electrical Engineering, University of Cape Town,

Private bag., Rondebosch 7701, Cape Town, South Africa
Komla.Folly@uct.ac.za
Abstract. Population-Based Incremental Learning (PBIL) is a combination of

Genetic Algorithm with competitive learning derived from Artificial Neural
Network. It has recently received increasing attention due to its effectiveness,
easy implementation and robustness. Despite these strengths, it has been
reported recently that PBIL suffers from issues of loss of diversity in the
population. To deal with the issue of premature convergence, we propose in this
paper a parallel PBIL based on multi-population. In parallel PBIL, two
populations are used where both probability vectors (PVs) are initialized to 0.5.
The approach is used to design a power system controller for damping low-
frequency oscillations. To show the effectiveness of the approach, simulations
results are compared with the results obtained using standard PBIL and another
diversity increasing PBIL called herein as PBIL with Adapting learning rate
(APBIL). It is shown that Parallel PBIL approach performs better than the
standard PBIL and is as effective as APBIL.
Keywords: Adaptive learning rate, Low frequency oscillations, Population-

based incremental learning, Parallel PBIL.
1 Introduction
Recently, a novel type of Evolutionary Algorithm called Population-Based

Incremental Learning (PBIL) [1]-[2] has received increasing attention [3]-[6].
Population-Based Incremental Learning (PBIL) is a combination of Genetic
Algorithm with competitive learning derived from Artificial Neural Network. Like
other Evolutionary Algorithms such as GAs [7]-[9], Differential Evolution (DE) [10]-
[11], and variants such as Particle Swarm Optimization (PSO) [12]-[13], PBIL works
with a population of individuals rather than a single individual (e.g., point) [1], [2].
Over successive generations, the population “evolves” toward an optimal solution.
PBIL is simpler than Genetic Algorithms GAs, and yet more effective than GAs. In
PBIL, the crossover operator of GAs is abstracted away and the role of population is
136 K.A. Folly
redefined [1]. PBIL works with a probability vector (PV) which controls the random
bit strings generated by PBIL and is used to create other individuals through learning.
Learning in PBIL consists of using the current probability vector (PV) to create N
individuals. The best individual is used to update the probability vector, increasing the
probability of producing solutions similar to the current best individuals [2], [3]. It has
been shown that PBIL outperforms standard GAs approaches on a variety of
optimization problems including commonly used benchmark problems [1], [2]. PBIL
has also been applied for controller design in power systems for small-signal stability
improvement. In [4], PBIL based power system stabilizers (PSSs) were compared
with GA based PSSs and were found to give better results GA based PSSs. In [5]-[6],
it was shown that PBIL based PSS performed as effectively as BGA based PSS.
However, there are still some issues related to PBIL [14]. It has been reported in [15]-
[17] that PBIL suffers from diversity loss making the algorithm to converge to local
optima. To cope with this problem, a PBIL with adaptive learning rate strategy was
proposed in [16]-[17]. In this paper, a new approach that can improve population
diversity in PBIL is presented. The idea of using parallel PBIL (PPBIL) based on
multi-population to improve population diversity is explored [18]-[19]. The proposed
approach is applied to a power system controller design in a multi-machine power
system. The effectiveness of the proposed approach is demonstrated by comparing it
to the Adaptive PBIL (APBIL) introduced in [16]-[17] and the standard PBIL
(SPBIL). Simulation results show that the parallel PBIL based on multi-population
performs better than the standard PBIL and is as effective as APBIL.
2 Overview of the Standard PBIL
Population–based incremental learning (PBIL) is a technique that combines aspects of

Genetic Algorithms and simple competitive learning derived from Artificial Neural
Networks [1], [2]. PBIL belongs to the family of Estimation of Distribution
Algorithms (EDAs), which use the probability (or prototype) vector to generate
sample solutions. Unlike GAs which performance depends on crossover operator,
PBIL performance depends on the learning process of the probability vector. The
probability vector guides the search, which produces the next sample point from
which learning takes place. The learning rate determines the speed at which the
probability vector is shifted to resemble the best (fittest) solution vector [3]. Initially,
the values of the probability vector are set to 0.5 to ensure that the probability of
generating 0 or 1 is equal. As the search progresses, these values are moved away
from 0.5, towards either 0.0 or 1.0.
Like in GA, mutation is also used in PBIL presented in this paper to maintain
diversity. In this paper, the mutation is performed on the probability vector; that is, a
forgetting factor is used to relax the probability vector toward a neutral value of 0.5
[3], [4]. The pseudocode for the standard PBIL is shown in Fig. 1, [1]-[4].
Comparison of Multi-population PBIL and Adaptive Learning Rate PBIL 137
If the learning rate is fixed during the run, it cannot provide the flexibility needed to
achieve a trade-off between exploration and exploitation. To achieve a trade-off
between exploration and exploitation, PBIL with adaptive learning rate strategy
presented in [16]-[17] could be used. However, the approach proposed here is to use
multi-population PBIL instead of a single population to achieve the same objective as
discussed in section 4
Begin
g:= 0;
//initialize probability vector
for i:=1 to l, do PVi0 = 0.5;
endfor;
while not termination condition do
generate sample S(g) from (PV(g) , pop.)
Evaluate samples S(g)
Select best solution B(g)
// update probability vector PV(g) toward best
solution according to (1)
//mutate PV(g)
Generate a set of new samples using the new
probability vector
g=g+1
end while // e.g., g>Gmax
Fig. 1. Pseudocode for standard PBIL
3 Overview of Multi-Population PBIL
For the multi-population or Parallel PBIL (PPBIL), two populations are used with two
probability vectors (PV1 and PV2). Each probability vector is initialized to 0.5 and
sampled to generate solutions independently from each other. The PVs are updated
independently according to the best solution generated by each. Initially, each
probability vector has equal sample solutions. That is, the total population is divided
into two populations and a PV is assigned to each population. As the run progresses,
the population of the probability vector (PV) that performs better is allowed to
increase its share of samples. The sample sizes of the probability vectors are slightly
adapted within the range [popmin popmax] = [0.4*pop 0.6*pop] according to their
relative performances. The probability that outperforms the other is increased by a
constant value Δ = LR*pop, where LR is the learning rate (which was selected as 0.1
in this paper). Fig. 2 shows the pseudocode of PPBIL.
138 K.A. Folly
Begin
g:= 0;
//initialize probability vector
for i:=1 to l, do PVi10 = PVi20 = 0.5;
endfor;
// initialize the sizes of the probability vectors
such that: pop1= pop 2= pop/2
while not termination condition do
generate sample S1(g) from (PV1(g) , pop1.)
generate sample S2(g) from (PV2(g) , pop2.)
Evaluate samples (S1(g), S2(g))
Select best solutions B1(g)and B2(g)
// update probability vectors PV1(g) and PV2(g)
toward bests solution B1(g)and B2(g) according to (1)
If f(B1(g))> f(B2(g) )
then pop1= min [(pop1 + Δ) popmax]
If f(B1(g))< f(B2(g) )
then pop1= max [(pop1 -Δ) popmin]
pop2= pop-pop1
//mutate PV1(g) and PV2(g)
g=g+1
end while // e.g., g>Gmax
Fig. 2. Pseudocode for parallel PBIL
Fig. 3. Power system model

4 Problem Description and Formulation
4.1 Problem Description

The controller to be designed is also known as Power System Stabilizer (PSS) and is
needed to damp low frequency oscillations ranging from 0.1 Hz to 2.5 Hz which
occur in overly stressed power systems or when power is transmitted over weak
transmission lines [20]-[21]. These oscillations are highly undesirable because they
can lead to fatigue of machine shafts and limit the ability of the system to transfer the
maximum power. It is therefore important that low-frequency oscillations are damped
quickly if the security of the system is to be maintained. The power system model
used in this paper is the IEEE 2-area system, 4-machine power system as shown in
Fig. 3. Each machine is represented by the detailed six order differential equations.
The machines are equipped with simple exciter systems. For more information on this
system, the reader is referred to [18], [20]. To design the controller, several operating
conditions have been considered. However, for simplicity only three operating
conditions are shown in Table 1. This Table also shows the eigenvalues and damping
ratios in brackets of the three operating conditions.
The system exhibits two local modes one in area 1 and the other in area 2 and one
inter-area mode. For the purpose of this study, only the inter-area modes are shown in
Table 1 since they are the most difficult to control.
Case 1 is the light load condition, where about 200 MW of real power is
transferred from area 1 to area 2. The system is stable for this case as can be seen by
the negative value of the real part of the eigenvalue. Case 2 is the nominal condition,
under this operating condition, there is a transfer of 400 MW power from area 1 to
area 2. The system is unstable for this case, since the real part of the eigenvalue is
positive. Case three is the heavy load condition where about 500 MW of power is
transferred from area 1 to area 2. This case is also unstable.
4.2 Problem Formulation and Objective Functions

The purpose of the design is to optimize the parameters of the generator excitation
controls (i.e., PSSs) simultaneously and in a coordinated and decentralized manner
such that adequate damping is provided to the system over a wide range of operating
conditions, while keeping the structure of the PSS as simple as possible. The structure
of the widely used conventional PSS was adopted here. The PBILs are applied to
optimize the parameters of a fixed structure (Δω input) PSS of the form:
 T s  1 + T1 s  1 + T3 s  (1)
K ( s) = K p  w   
 1 + Tw s  1 + T2 s  1 + T4 s  .
140 K.A. Folly
where, Kp is the gain, T1-T4 represent suitable time constants. Tw is the washout time
constant needed to prevent steady-state offset of the voltage. The value of Tw is not
critical for the PSS and has been set to 5sec. Therefore, five parameters are required
for the optimization.
Since most of oscillation modes considered in this paper are unstable and dominate
the time response of the system, it is expected that by maximizing the minimum
damping ratio, a set of system models could be simultaneously stabilized over a wide
range of operating conditions [10]-[12]. The following objective function was used to
design the PSSs.
J = max min (ζ i , j ) (2)

 .
where i = 1, 2 … n, and j = 1, 2, … m
− σ i, j
and ζ i, j = is the damping ratio of the i-th eigenvalue in the j-th
σ i, j 2 + ω i, j 2
operating condition. σij is the real part of the eigenvalue and the ωij is the frequency. n
denotes the total number eigenvalues and m denotes the number of operating
conditions.
Table 1. Selected open-loop operating conditions including eigenvalues and damping ratios
Case Pe [MW] Eigenvalue (ζ)

1 200 -0.35±3.92i (0.0889)
2 400 0.0096±3.84 i (-0.0025)
3 500 0.148±3.09i (-0.0478)
4.3 Application of Standard PBIL to Controller Design

The configuration of the standard PBIL is as follows:
Length of chromosome: 15 bits
Trial solutions (population): 10
Generations: 400
Learning rate (LR): 0.1
Mutation (Forgetting factor-FF ): 0.005
4.4 Application of APBIL to Controller Design

The configuration of the APBIL is as follows:
Generations: 400
Initial Learning rate (LR0) = 0.0005
Final Learning rate (LRmax): 0.2

Mutation (Forgetting factor-FF): 0.005
4.5 Application of PPBIL to Controller Design

The configuration of the APBIL is as follows:
Initial population for PV1: 5
Initial population for PV2: 5
Generations: 400
Final Learning rate (LR): 0.1
Mutation (Forgetting factor-FF): 0.005
For all the controllers, the parameter domain is as follows:
0≤Kp≤30
0≤ T1,T3≤1
0.010≤ T2, T4 ≤ 0.3
5.1 Convergence Rate

Figs. 4-6 show the convergence rate of SPBIL, APBIL and PPBIL, respectively. It
can be seen that APBIL and PPBIL converge to higher fitness values of 0.514 and
0.502, respectively, compared to a value of 0.484 for SPBIL. From the simulation
results, it can be seen that APBIL has more diversity in the population at the middle
of the run between generation 100 and 200 than PPBIL and SPBIL. This can be
attributed to the small value of learning rate, at these generations. Small learning rate
increases the exploration of the algorithm and thereby introduces more diversity in the
population. Unlike SPBIL which diversity is much more concentrated at the
beginning of the run between generation 1 and generation 150, the diversity in PPBIL
is somehow spread across all generations. At generations 300 to 400 for example, the
SPBIL and APBIL have converged (i.e., almost no diversity). On the other hand,
PPBIL still has some diversity. Therefore, it can still explore the search space
although at a limited pace. Table 2 shows the comparison between the best, mean
and worst, fitness values. It can be seen that on average SPBIL and PPBIL have
practically the same fitness. The mean for PPBIL and SPBIL which is approximately
0.434 is higher than the mean of APBIL which is 0.383. The main reason for this is
that APBIL has much more spread, with the worst fitness value at 0.09 compared to
0.124 for PPBIL and 0.166 for SPBIL.
142 K.A. Folly
In terms of the distance between the best and the worst fitness values, APBIL has
the highest distance (0.424), followed by PPBIL (0.378) and then SPBIL (0.318). This
suggests that both APBIL and PPBIL have more diversity in their populations than
SPBIL. Table 3 shows the number of functions evaluations for each algorithm before
the best fitness was found. It can be seen that that SPBIL has the lowest function
evaluations (3810) and APBIL has the highest function evaluations (15950). PPBIL is
somehow in the middle (10610). In terms of the speed in finding the best fitness
value, SPBIL is better and APBIL is the worst. However, the best value found by
SPBIL is lower than that found by APBIL and PPBIL. This suggests that although
SPBIL converges faster, it converges to local optima, which may not be appropriate.
5.2 Eigenvalue Analysis

Table 4 shows the eigenvalues and damping ratios in brackets of the closed-loop
systems with the three controllers. It can be seen that PPBIL and APBIL give better
performances (i.e., better damping ratios) than SPBIL. However, PPBIL provides
slightly a better damping than APBIL.
0.5
0.45
0.4
Best Fitness
0.35
0.3
0.25
0.2
0 50 100 150 200 250 300 350 400

Generation
Fig. 4. SPBIL convergence rate
0.55
0.5
0.45
0.4
Best Fitness
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0 50 100 150 200 250 300 350 400
Generation
Fig. 5. APBIL convergence rate

0.55
0.5
0.45
0.4
Best Fitness
0.35
0.3
0.25
0.2
0.15
0.1
0 50 100 150 200 250 300 350 400
Generation
Fig. 6. PPBIL convergence rate
Table 2. Fitness Values
Fitness SPBIL APBIL PPBIL

Best 0.484 0.514 0.502
Mean 0.434 0.383 0.434
Worst 0.166 0.090 0.124
Table 3. Number of Function Evaluation
Controllers Evaluations
SPBIL 3810
PPBIL 10610
APBIL 15950
Table 4. Closed-system eigenvalues and damping ratio
Case SPBIL APBIL PPBIL

1 -1.13 ± j2.04 (0. 48) -1.54 ± j2.57 (0.51) -1.53 ± j2.53 (0.52)
2 -0.778 ± j1.78 (0.44) -1.26 ± j2.11 (0. 51) -1.26 ± j2.08 (0. 52)
4 -0.582 ± j1.25 (0.42) -1.15 ± j1.52 (0.61) -1.14 ± j1.54 (0.60)

144 K.A. Folly
6 Conclusions
By using Parallel PBIL based on multi-population we have been able to increase the
diversity in the population. This is important to prevent premature convergence that is
inherent to the standard PBIL. The effectiveness of the proposed approach is
demonstrated by comparing it to the Adaptive PBIL (APBIL) and the standard PBIL
(SPBIL). Simulation results show that the performance of PPBIL in increasing the
population diversity is as effective as that of APBIL. Both the PPBIL-PSS and the
APBIL-PSS performed better than the standard SPBIL-PSS in terms of improving the
damping of the system.
Acknowledgement. This work is based on the research supported in part by the

National Research Foundation of South Africa, UID 83977 and UID 85503.
References
1. Baluja, S.: Population-Based Incremental Learning: A Method for Integrating Genetic
Search Based Function Optimization and Competitive Learning. Technical Report, CMU-
CS-94-163, Carnegie Mellon University (1994)
2. Baluja, S., Caruana, R.: Removing the Genetics from the Standard Genetic Algorithm.
Technical Report CMU-CS-95-141, Carnegie Mellon University (1995)
3. Greene, J.R.: Population-Based Incremental Learning as a Simple, Versatile Tool for
Engineering Optimization. In: EvCA 1996, Moscow (1996)
4. Folly, K.A.: Design of Power System Stabilizer: A Comparison Between Genetic
Algorithms (GAs) and Population-Based Incremental Learning (PBIL). In: Proc. of the
IEEE PES 2006 General Meeting, Montreal, Canada (2006)
5. Sheetekela, S., Folly, K.: Power System Controller Design: A Comparison Between
Breeder Genetic Algorithm (BGA) and Population-Based Incremental Learning (PBIL).
In: Proc. of the Int. Joint Conference on Neural Networks, IJCNN (2010)
6. Folly, K.A., Sheetekela, S.P.: Optimal design of Power System Controller using Breeder
Genetic Algorithm. In: Bio-Inspired Computational Algorithms and Their Applications,
Intech, pp. 303–316 (2012)
7. Goldberg, D.E.: Genetic Algorithms in Search, Optimization & Machine Learning.
Addison-Wesley (1989)
8. Davis, L.: Handbook of Genetic Algorithms. International Thomson Computer Press
(1996)
9. Yao, J., Kharma, N., Grogono, P.: Bi-objective Multipopulation Genetic Algorithm for
Multimodal Function Optimization. IEEE Trans. On Evol. Comput. 14(1), 80–102 (2010)
10. Kennedy, J.F., Kennedy, J., Eberhart, R.C., Shi, Y.: Swarm Intelligence. Morgan
Kaufmann (2001)
11. Mulumba, T., Folly, K.A.: Design and Comparison of Multi-machine Power System
Stabilizer base on Evolution Algorithms. In: In Proc. of the 46th International Universities’
Power Engineering Conference (UPEC), Soest – Germany, September 5-8 (2011)
12. Abido, A.A.: Particle swarm Optimization for Multimachine Power System Stabilizer
Design. IEEE Trans. on Power Syst. 3(3), 1346–1351 (2001)
13. Venayagamoorthy, G.K.: Improving the Performance of Particle Swarm Optimization

using Adaptive Critics Designs. In: IEEE Proceedings on Swarm Intelligence Symposium,
pp. 393–396 (2005)
14. Gosling, T., Jin, N., Tsang, E.: Population-Based Incremental Learning Versus Genetic
Algorithms: Iterated Prisoners Dilemma. Technical Report CSM-40, University of Essex,
England (2004)
15. Rastegar, R., Hariri, A., Mazoochi, M.: The Population-Based Incremental Learning
Algorithm Converges to Local Optima. Neurocomputing 69(13-15), 1772–1775 (2006)
16. Folly, K.A., Venayagamoorthy, G.K.: Effect of learning rate on the performance of the
Population-Based Incremental Learning algorithm. In: Proc. of the International Joint
Conf. on Neural Network (IJCNN), Atlanta Georgia, USA (2009)
17. Folly, K.A.: An Improved Population-Based Incremental Learning Algorithm.
International Journal of Swarm Intelligence Research (IJSIR) 4(1), 35–61 (2013)
18. Folly, K., Venayagamoorthy, G.: Power System Stabilizer Design using Multi-Population
PBIL. In: Proc. of the 2013 IEEE Symposium Series on Computational Intelligence (2013)
19. Yang, S., Yao, X.: Experimental Study on Population-Based Incremental Learning
Algorithms for Dynamic Optimization Problems. Soft Computing 9(11), 815–834 (2005)
20. Kundur, P.: Power System Stability and Control, McGraw-Hill, Inc. (1994)
21. Gibbard, M.J.: Application of Power System Stabilizer for Enhancement of Overall
System Stability. IEEE Trans. on Power Systems 4(2), 614–626 (1989)
Vibration Adaptive Anomaly Detection of Hydropower
Unit in Variable Condition Based on Moving Least
Square Response Surface
Xueli An and Luoping Pan
China Institute of Water Resources and Hydropower Research, 100038 Beijing, China
an_xueli@163.com
Abstract. It is difficult to effectively analyze and identify the conditions of

hydropower unit, due to its complex operation conditions, frequent start-stop
conditions, continual working status switch, less fault samples, single static
alarm threshold. Lots of test research shows that active power and working head
are key factors which affect the operation conditions of hydropower unit. The
health standard condition of unit is determined. An adaptive real-time anomaly
detection model of hydropower unit vibration parameters is proposed based on
moving least square response surface. In the proposed model, active power and
working head are comprehensively considered. This model can adapt variable
conditions of hydropower unit. The model is used to real time detect the
anomaly of hydropower unit vibration parameters. The results show that this
model can effectively evaluate the performance of unit vibration, can more
accurately detect the abnormal of unit vibration.
Keywords: hydropower unit, vibration parameter, adaptive anomaly detection,

moving least square response surface.
1 Introduction
Hydropower plant is the most favorable power source of power system. It is used to
undertake the tasks of peaking, filling valley, frequency modulation, phase
modulation, and emergency reserve. Hydropower plant can improve the efficiency of
thermal power plants and nuclear power plants, increase the reliability of power grid.
It has a significant role in ensuring the safety of the power grid operation and
improving the economy of power system [1] and [2]. Due to the complexity of
hydropower units’ operating conditions, frequent start-stop and working conditions
conversion, making the unit easy to malfunction. To ensure units’ safe and stable
operation, it is needed to mine their condition monitoring data. The data mining can
better get the real operating condition of units and early warn the possible
abnormalities.
The research of online monitoring and fault diagnosis of hydropower units mainly
focuses on the development and integration of condition monitoring system and fault
diagnosis methods. The current research achievements don’t meet site requirements
Vibration Adaptive Anomaly Detection of Hydropower Unit 147
[1]. The studies for effectively analyzing the mass monitoring data aren't many. The
studies for building the anomaly detection model based on online monitoring data are
fewer [1, 3]. This makes the vast majority of hydropower plants can only use
preventive maintenance strategy, namely scheduled overhaul strategy. This strategy
will inevitably lead to the problem for inadequate maintenance or excess
maintenance.
Anomaly detection aims to find the relationship between the abnormities of units’
condition parameters and their potential failure, to reveal hidden information of
abnormal parameters. The field personnel can timely take appropriate measures
according to units’ abnormal condition information, to restraint the further
deterioration of abnormities [4], [5] and [6]. Using this method, faults can be nipped
in the bud. The failure rate will be reduced. The security, stability of units operation
and economy of maintenance will be increased.
In this paper, based on a long time condition monitoring data of hydropower units,
the moving least square response surface is used to build adaptive anomaly detection
model. This model considers the factors of active power and working head, which
affect the operation condition of hydropower units. The proposed model provides a
new way to online assessment of units’ running condition.
2 Moving Least Square Response Surface
In a local domain Ωx of the fitting area Ω, supposing the function values of a

calculated function f(x) of N sampling points xI (I=1, 2, …, N) in the local domain Ωx
are known, yI=f(xI). In the local domain, the calculated function f(x) can be
approximated as g(x)≈f(x), the fitting function g(x) [7] can be expressed as:
m (1)
g ( x) =  β i ( x) ρ i ( x) = ρ T ( x) β ( x) .
i =1
where β ( x) = (β1 ( x), β 2 ( x),  , β m ( x) )T is calculated coefficient,

ρ ( x) = (ρ1 ( x), ρ 2 ( x),  , ρ m ( x) ) is m-dimensional k-order complete polynomial,
T
m is the number of basis function.

The fitting accuracy of g(x) is varietal with the different order of basis function. As
a secondary basis function in two-dimensional space ρ(x) is
ρ ( x) = (1 x y x 2 xy y 2 ) . (2)
where m=6.
In moving least-squares fitting, the coefficient β(x) is determined based on the
weighted least squares. This makes the weighted sum of squares of each sampling
point’s errors minimized for the approximate function g(x) in the neighborhood Ωx of
the point x.
148 X. An and L. Pan
N
Ψ =  wI ( x)[ g ( x) − f ( xI )]2
I =1 (3)
.
2
N
 m

=  wI ( x)  β i ( x) ρ i ( xI ) − f ( xI ) 
I =1  i=1 
∂Ψ N
m 
= 2 wI ( x)  β i ( x) ρ i ( xI ) − f ( xI )ρ i ( xI ) = 0 . (4)
∂β i ( x) I =1  i =1 
where i = 1, 2, ..., m, N is number of sampling points within the neighborhood Ωx of

the point x. The wI(x)=wI(x-xI) is weighting function at the sampling point xI. The
wI(x) is greater than zero in a limited area ΩI around the sample point xI, beyond the
ΩI are zero. The ΩI is support domain of weight function, also called impact domain
of sampling points xI. The accuracy of moving least square approximation depends
largely on the weight function, which often using Gaussian weighting function in the
application, detailed introduction refers to the [7].
3 Adaptive Anomaly Detection Model of Hydropower Unit

Vibration
The real-time monitoring and timely warn of pumped storage unit is important for
power grid’s stable operation. For the moment pumped storage power units have
implemented monitoring system to online collect key parts’ monitoring signals. The
measured values of monitoring signals are simply compared with a preset threshold to
achieve the alarm, which guide the operation and maintenance of the unit. This
method of a static alarm threshold is a single judge. It ignores the unit’s performance
differences in different working conditions. When an alarm occurs, the equipment
performance of unit may have largely deviated from the design conditions. A situation
may occur. That is the unit’s equipment may have seriously deteriorated, but the
alarm level of condition monitoring system has not yet been reached. It can be seen
that static alarm threshold lacks the thorough study of monitoring signal’s hidden
abnormities’ or faults’ information. And it lacks the warning capacity for early
potential failures. It is far from insufficient to fully reflect the operational condition of
the unit. Meanwhile, with the constantly expanding of hydropower plant capacity and
the gradual improvement of monitoring auxiliary systems, the information quantity of
unit’s control and monitoring data increase continuously. The operation personnel are
often difficult to understand unit’s situation based on such a large amount of data [8].
They can’t timely find unit’s abnormity and fault. This can happen that there are huge
amounts of data, but lack of information to guide decision.
In the same or similar operating conditions, when the equipments of pumped

storage units are normal, their monitoring parameters should be random fluctuations
in the mean nearby. After a long time’s running, the unit will gradually deviate
normal operation condition, enter the non-normal operation condition. As time went
on, unit’s deterioration will gradually accelerate, from quantitative change to
qualitative change, may lead to serious consequences. Therefore, unit’s abnormal
operating condition should be paid great attention and closely observed. The reasons
for an abnormality occurs should be analyzed. The effective solutions should be
searched. Unit equipments’ abnormality can be early found by effectively mining the
useful implicit information of online monitoring data. This can guide the operator to
adjust and control unit, ensure efficient and stable operation of the unit, minimize
accidents probability.
A large number of field data analysis show that the main factors affecting
hydropower unit’s operating condition are active power and working head. In this
paper, considering the effect of active power and working head, the health standard
three-dimensional surface model c=f(P, H) is built, where c is the unit’s condition
parameter, P is active power, H is working head. Based on the three-dimensional
surface model, the adaptive anomaly detection of hydropower unit is made, concrete
steps are as follows:
(1) Analyzing the conditions monitoring data of pumped storage power plant units
in different operating condition, determining unit’s standard health condition.
Selecting the characteristic parameters which can reflect unit’s operating condition.
(2) Inputting the unit’s characteristic parameters in the health condition into LS-SVM
to train, building a three-dimensional surface model c=f(P, H), and validating the model.
(3) Substituting the real-time condition monitoring data of active power and
working head into the trained LS-SVM model to calculate the health standard value
c(t) of condition parameters in the current operating conditions. Comparing the
current real value r(t) and health standard value c(t), calculate the vibration deviation
of the unit in current condition:
r (t ) − c(t )
d v (t ) = × 100% . (5)
c (t )
where t is run time of pumped storage units. If dv(t) exceeds a preset threshold, an
alert of abnormal vibration can be made. This can promptly find abnormal conditions
of the units.
4 Case Study
The real condition monitoring data of a pumped storage power plant unit in
September 22, 2008 ~ December 15, 2011 are studied to validate the effectiveness of
adaptive anomaly detection model of hydropower unit’s vibration parameters in
varying conditions. The model is based on moving least squares response surface.
Due to the complexity of pumped storage units’ operating conditions, frequent starts
and stops and working conditions switch, the validity of the proposed model in
changing conditions can be better reflected by using pumped storage units’

monitoring data. The x-direction horizontal vibration of upper bracket is selected as
the study object.
(a) The real data of operating head.
(b) The real data of power.
(c) The real data of upper bracket horizontal vibration in direction-X.

Fig. 1. The condition monitoring data of pumped storage power station unit
Figure 1 shows unit’s real data of working head, active power and upper bracket
horizontal vibration in direction-X in from May 16, 2011 to May 30, 2011. It can be
seen from Figure 1 that unit’s active power focused on 250MW for pumping
conditions; power is concentrated in 150MW, 200MW and 250MW for generating
operation. The working head has strong volatility. The conversion of pumping and
generating conditions is frequent. So the changes of upper bracket horizontal
vibration in direction-X are complex. The effective information which reflects the
true condition can’t be obtained only from this Figure. Research shows that active
power, working head have an important impact on the unit’s vibration parameters. If
setting a single static alarm threshold for units, the performance difference, hidden
information of abnormality and fault will be greatly neglected. And the unit’s real
condition can’t be truly reflected. Therefore, it is need to build a three-dimensional
surface model to detect the abnormality of vibration parameters. This model should be
adaptive the changes of working conditions for pumped storage units.
Firstly, determining unit’s standard health condition, selecting the characteristic
parameter which can reflect unit’s operating condition.
The online monitoring data (unit has good condition and without fault) of unit
initial operation are adopted to build vibration standard model of unit in healthy
condition. The 800 sets online monitoring data from September 22, 2008 to
September 18, 2009 are selected. The peak-peak value of x-direction horizontal
vibration of upper bracket is selected as the characteristic parameter.
Then, inputting unit’s healthy condition parameter into the moving least squares
response surface to train, building a three-dimensional surface model c=f(P, H), and
validating the model.
To real-timely get a true operating condition of hydropower units, it is need to
build a health condition model. Considering the important influence of power and
working head on hydropower unit’s vibration characteristics, and moving least square
response surface has good fitting performance for scattered data, a vibration-power-
working head three-dimensional surface model v=f(P, H) of hydropower unit is built.
This model is based on moving least squares response surface. Through this model,
the mapping relationship in health condition among power (P), working head (H) and
vibration parameter (v) can be obtained. For the 800 sets data from September 22,
2008 to September 18, 2009, 600 sets data are selected to establish a health standard
model, the remaining 200 sets data as the test samples to validated the model. In order
to make the moving least squares response surface model has good performance, the
selected 800 sets health standard data should cover possible changes in working head
and active power. The 200 test samples’ active power and working head are inputted
this model, the results can be seen that the health standard values of upper bracket
horizontal vibration in direction-X based on moving least squares response surface
model is consistent with the measured values. The average relative error is 3.36 %.
Finally, substituting the unit’s online monitoring data of power and working head
into the trained three-dimensional surface model (moving least squares response
surface), calculating the health standard value c(t) of the condition parameter in
current condition. Using formula (5) to online assess unit’s real-time operating
condition, achieving early warn to unit’s vibration anomalies.
Fig. 2. The three-dimension surface of vibration-power-working head for hydropower unit
Substituting the power and working head of unit’s condition monitoring data that
after two years (May 12, 2011 ~ December 15, 2011) into unit’s health model
v(t)=f(P(t), H(t)), calculating the health standard values v(t) of condition parameter in
the current condition, and comparing the v(t) and the real values r(t). The comparison
can be seen in Figure 3. The formula (5) is using to calculate the vibration deviation
of the unit in current condition. The results are shown in Figure 4. It can be seen from
Figure 4 that after two years of operation, the deterioration of pumped storage unit’s
component occurs, the unit is gradually deviating its healthy operation condition. If
dv(t) exceeds a preset threshold, an alert of abnormal vibration is made. This can
promptly detect the unit’s abnormal condition.
Fig. 3. The fault early warning results of pumped storage unit

Fig. 4. The fault early warning results of pumped storage unit.
In summary, hydropower unit’s vibration anomaly can be found early by using the
presented three-dimensional surface anomaly detection model. When making unit’s
maintenance plan, the unit parts whose condition is abnormal, can be checked
purposefully. This can effectively avoid the possibility of forced outages, truly realize
the prevention of unit’s fault. So field operator of hydropower plant can real-timely
and comprehensivly obtain the health condition of unit’s key components. The
meaningless alarms will be reduced. The repair time will be shorten and the operation
time will be increased.
5 Conclusions
Pumped storage unit has complex operating conditions. It has many condition
monitoring points, less fault samples. It is difficult to effectively diagnose it’s falut.
Setting static alarm thresholds will ignore unit’s differences of dynamic performance
in varying operating conditions. So an adaptive real-time anomaly detection model of
hydropower unit vibration parameters is proposed based on three-dimensional
surfaces. Firstly, the unit’s conditions monitoring data in different operating condition
are analyzed. This reveals the key factors of active power and working head, which
affect unit’s performance. The unit’s health standard condition is determined. Then,
the characteristics parameters which can reflect of unit’s operating condition are
selected. The selected parameters of health condition are inputted into LS-SVM to
train. Finally, unit’s current condition monitoring data are inputted into the three-
dimensional surface model to online assess unit’s condition ment, achieve early
warning of abnormal vibration. The example shows that the proposed model can
effectively demonstrate the hidden information of condition monitoring data, real-
timely track unit’s operating condition, early warn unit’s potential failures. The model
has are good application prospects.
Acknowledgments. This work was supported by the National Natural Science

Foundation of China (grant number 51309258) and the Special Foundation for
Excellent Young Scientists of China Institute of Water Re-sources and Hydropower
Research (grant number 1421).
References
1. An, X.L., Pan, L.P., Zhang, F.: Condition degradation assessment and nonlinear prediction
of hydropower unit. Power System Technology 37(5), 1378–1383 (2013)
2. An, X.L.: Vibration characteristics and fault diagnosis for hydraulic generator units. Thesis
for the degree of doctor of engineering, Huazhong University of Science & Technology,
Wuhan (2009)
3. Lu, W.G., Dai, Y.P., Gao, F.: A hydroelectric-generator unit faults early warning method
based on distribution estimation. Proceedings of the CSEE 25(4), 94–98 (2005)
4. Renders, J.M., Goosens, A., Viron, F.D.: A prototype neural network to perform early
warning in nuclear power plant. Fuzzy Sets and Systems 74, 139–151 (1995)
5. Karlsson, C., Larsson, B., Dahlquist, A.: Experiences from designing early warning system
to detect abnormal behaviour in medium-sized gas turbines. In: 23rd International Congress
on Condition Monitoring and Diagnostic Engineering Management, pp. 117–120. Sunrise
Publishing, Hikone (2010)
6. Jardine, A., Lin, D., Banjevic, D.: A review on machinery diagnostics and prognostics
implementing condition-based maintenance. Mechanical Systems and Signal
Processing 20(7), 1483–1510 (2006)
7. Zhang, Y., Li, G.Y., Zhong, Z.H.: Design optimization on lightweight of full vehicle based
on moving least square response surface method. Chinese Journal of Mechanical
Engineering 44(11), 192–196 (2008)
8. Yan, J.F., Yu, Z.H., Tian, F.: Dynamic security assessment and early warning system of
power system. Proceedings of the CSEE 28(34), 87–93 (2008)
Capacity and Power Optimization
for Collaborative Beamforming
with Two Relay Clusters
Bingbing Lu1,2 , Ju Liu1,2 , Chao Wang1 , Hongji Xu1,2 , and Qing Wang1
1
School of Information Science and Engineering,
Shandong University, Jinan, 250100, China
juliu@sdu.edu.cn
2
National Mobile Communications Research Laboratory,
Southeast University, Nanjing, 210096, China
hongjixu@sdu.edu.cn
Abstract. In this paper, we study two approaches to optimize the prob-

lems between capacity and power in a three-hop multi-relay network. In
the first approach, two beamforming weight vectors are designed to max-
imize the capacity under the power constraints of relay clusters. While in
the second approach, we minimize the power of total relay nodes as well
as meet the minimal capacity demand. In both of design schemes, we turn
into two beamformer vectors to only one though a series of mathematical
manipulation. Then apply genetic algorithm (GA) to obtain the optimal
weight value of the nonconvex problems. Simulation results show that
our proposed approaches significantly outperform the previous methods
conducted.
Keywords: cooperative communication, three-hop, multi-relay, genetic

algorithm.
1 Introduction
Exploiting relay nodes to improve information capacity and link reliability has
attracted increasing interest recently [1–4]. Many swarm intelligence algorithms
[5–8] which can achieve optimal results also quickly development. Recently, the
dual-hop relay systems have attracted attentions in the research academia. As
the serious signal fading and path loss problems in some specific situation, we
consider a three-hop relay system which consists of a transmitter, a receiver
and two clusters of relay nodes. The relays at both terminals will form like a
multi-input multi-output (MIMO) beamforming system [9, 10]. Some methods
are proposed to optimize this problem like in [11], the cooperative relay weight
coefficients are optimized by maximizing the destination SNR under the sum-
power constraints at the relay clusters.
In this paper, we develop two distributed beamforming approaches in a three-
hop AF cooperative communication system. In the first approach, we aim to

156 B. Lu et al.
maximize the information capacity subject to the separate total power con-
straint. In the second approach, we minimize the total transmit power which
maintains the mutual information above a predefined threshold. As illustrated,
we turning multi-variables into single one then solve by Genetic Algorithm (GA).
Compared with the defect that be easy trapped into local optimal solution in [12],
our proposed approaches can obtain the global optimal solution in statistically.
The remainder of the paper is organized as follows. In Section II, we present
the system model and the optimization problem. The approaches of two relay
weights optimization are presented in section III. Simulation results are provided
in Section IV. Finally section V concludes the paper.
2 System Model and Optimization Problem
2.1 System Model
Consider a relay network in Fig. 1. We suppose that there is no direct link be-
tween the transmitter and receiver. Each node is equipped with a single antenna,
and is subject to the half-duplex constraint. The Rayleigh flat fading channel
coefficients between all the nodes are identical independent distributed.
R1 L F1
h1 g1
R2 F2
h2 g2
S x x D
hM x x gK
x x
RM FK
Fig. 1. System model [11]
The transmission information is divided √ into three parts. In the first part,
the source node broadcasts the signal P0 s to the first relay √ cluster, where
2 M
E{|s| } = 1. The signal x sent by {Rm }m=1 is given as x = P0 Whs + WnR ,
where W = diag([w1 , w2 , · · · wM ]) and nR with zero mean and variance of σ 2 ,
K
similar as the noise nF at {Fk }k=1 and nD at the destination.
K
In the second part, {Fk }k=1 retransmits the received signal y multiplied by
the beamforming matrix P, which can be expressed as y = PLx + PnF , where
P = diag([p1 , p2 , · · · pK ]). Finally, the signal d received at the destination can be
written as

d = gH y + nD = P0 pH GLHws+pH GLWnR +pH GnF +nD (1)
H H
where G = diag (g), H = diag (h), p = [p1 , p2 , ..., pK ] and w = [w1 , w2 , ..., wM ] .
Capacity and Power Optimization for Collaborative Beamforming 157
2.2 Optimization Problem
We want to find the optimal solution from the capacity maximization and re-
lay power minimization. By Shannon’s second theorem, the mutual information
between the source and destination is given by I = 12 log2 (1 + SNR). The capac-
ity maximization is equivalent to maximized the receiving SNR at D since I is
increasing along with SNR.
The two problems will be optimized with two relay beamformers w and
p, which are the capacity maximization under the relay power constraints
and the relay power minimization as well as meet the capacity demand.
M
In the both problems, the transmit power PR of relay cluster {Rm }m=1

M
can be calculated as PR = E |xm |2 = wH DR w, where DR =
m=1
P0 diag E{|h1 |2 }, E{|h2 |2 } · · · E{|hM |2 } +σ 2 I. In the same way, the total trans-

K
mit power PF can be obtained as PF = E |yk |2 = pH DF p, where
k=1

M
[DF ]k,k = |lk,m |2 [DR ]m,m |wm |2 + σ 2 , k = 1, 2 · · · , K.
m=1
3 Relay Weights Optimization Based on GA
GA is a metaheuristic optimization algorithm that use survival of the fittest

evolutionary scheme to refine a set of solution candidates iteratively [13]. The
computing process of GA includes mutation, crossover, selection and inheritance.
The algorithm begins with creating an initial population of random individuals
which represented as binary strings. Then we compute the objective values of the
solution candidates to select the fittest individuals under some constraints. The
selected individuals are reproduced by crossover and mutation so as to create the
new individuals. The GA with the features of randomness and heuristic search
nature has made the algorithm a suitable tool for finding the global optimal
solution.
In the following two problems of capacity maximization and relay power min-
imization, the GA is a basic approach to solve the final optimization problems.
3.1 Capacity Maximization
In this subsection, our goal is to maximal the information capacity subject to the
separated power constraint of the two relay clusters. Because of the relationship
between I and SNR, the optimal I can be calculated by the optimal solution of
SNR, then this optimization problem is equivalent to
P0 pH GLHw(GLHw)H p
max
w,p σ 2 pH GLw(GLw)H + ggH p + 1 (2)
s.t. w DR w
H
PRmax , pH DF p PFmax
158 B. Lu et al.
Our goal is to obtain the optimal vector w and p so that the SNR is maxi-
mized. We can see that, this problem is not convex, and the two beamforming
vectors w and p depend on each other in problem (2), which make the prob-
lem more difficult to solve. We find that the initial problem can be regard as
a question with variable p when w is a constant selected within the feasible
region. And then for any available w, the maximum achievable SNR [10] can be
expressed as
−1/2 −1/2
SN Rmax (w) = PFmax P0 MH DF XDF M (3)
−1/2 −1/2
where M = GLHw, X = (σ 2 I + PFmax DF QDF )−1 , Q = σ 2 GLw(GLw)
H
+ σ 2 ggH . Therefore the optimization problem (2) is simplified into a question

with only one variable w with the constraint, which can be written as
−1/2 −1/2
max PFmax P0 wH (GLH)H DF XDF GLHw
w (4)
s.t. wH DR w PRmax
Genetic Algorithm
Initial w
Calculate p by (3)
for g=1; g<=G; g=g+1 //G: generation limit
for i=1; i<= N; i=i+1 //N: Number of individuals
if wi don’t meet constraints
then utilize penalty function to generate new wi
else
H
q(i) = wwHi BwAwi
//q(i):fitness faction of wi
i i +c
if q(i) > ε //ε: a predefined constant
then reproduction wi
else
give up wi
end if
end if
wi crossover with wi+j of probability pc
wi mutation of probability pm
end
end
It can be shown that, the problem (4) is not convex. GA is often applied as an
approach to obtain a statistics global optimal solution for this nonconvex prob-
lems. The details process of the algorithm are expressed in Genetic Algorithm.
The maximum I can be calculated by the obtained SNR. Because of this lengthy
process, the computational complexity of GA is relatively high. Compared with
using GA to solve the problem in (2) directly, the initial population of solving (4)
is produced only by w. Thus the coding length of initial population individuals will
reduce to half, which leads to a decrease of computation time through abundant
crossover, mutation and duplication in every generational population of GA.
3.2 Power Minimization
In this subsection, our goal is to minimize the total transmit power while keeping
the capacity at the destination above a certain preconcerted threshold. Similar
as the first problem, the optimization of I can be converted to the problem about
SNR, which meet the threshold γ. The problem can be expressed as
min wH DR w + pH DF p
w,p
P0 pH GLHw(GLHw)H p (5)
s.t. γ
σ 2 pH GLw(GLw)H + ggH p + 1
Similar to the previous subsection, wH DR w in the objective function is fixed

for a certain w. Under the condition of every fixed w, this question can be
simplified as the problem about p. It can be known [10] that this optimization
problem will get the global optimal solution when p is chosen as
1/2
γσ 2 −1/2
p = g(w) = DF u (6)
−1/2 −1/2
uH DF (R−γQ)DF u
−1/2 −1/2 H
where u = P{DF (R−γQ)DF } and the matrix R = P0 GLHw(GLHw) .
We put (6) into (5), which turns (5) into a problem with only one variable w.
Mathematically, the simplified optimization problem can be expressed as
min wH DR w + gH (w)DF g(w)

w
wH Ãw (7)
s.t. ≥γ
wH B̃w+c̃
H H
where Ã =P0 (g(w)H GLH) (g(w)H GLH), B̃ = (g(w)H GL) (g(w)H GL) and
the notation c̃ = σ 2 g(w)H ggH g(w) + σ 2 .
We solve the problem in (7) to obtain the optimal solution using GA. For the
implement of GA, the initial population is produced according to the vector w,
and the beamforming weight p can be calculated by w. The fitness function and
the constraints in GA are the same as the objective function and the restrict
function expressed in (7), respectively. We can obtain the optimal value of vector
w, and calculate the corresponding optimal p and the global minimum of PR +
PF through GA. By constrast, our proposed method has a lower complexity of
computing than solving the problem in (5) with GA directly.
4 Simulation Result
In this section, simulations are designed to assess the performance of the pro-
posed algorithms. Over all the simulation process, all the nodes are with the
same noise power level and the transmit power P0 of the source node is set to be
160 B. Lu et al.
2.5
Information capacity I (bit/s/Hz)

2
1.5
Proposed approach M=K=10

Method of [11] M=K=10
Method of [11] M=K=6
Fixed power allocation M=K=10
1
0 2 4 6 8 10 12 14 16
PRmax (dBW )
Fig. 2. Information capacity against PRmax with 6 (dash line) and 10 (solid line) relay
nodes, respectively
2.6
2.4
2.2
1.8
1.6
Proposed Method M=K=10

1.4
Proposed Method M=K=6
Method in [11] M=K=10
1.2 Method in [11] M=K=6
1
0 2 4 6 8 10 12 14 16 18 20
PFmax (dBW )
Fig. 3. Information capacity against PFmax with 6 (dash line) and 10 (solid line) relay
nodes, respectively
20
min total relay powerPR + PF (dBW)

15 Proposed approach M=K=8
10
−5
0 0.5 1 1.5 2 2.5
Fig. 4. Minimum total relay power PR + PF versus capacity threshold for M=K=4, 6,
8, respectively
10dBW. In the simulation process, only beamforming vector w is the random

variable, and p is calculated by w. The objective function and constraints in GA
are presented in (4) and (7).
Fig. 2 shows the maximum achievable information capacity versus the maxi-
mum allowable total transmit power PR max . In Fig. 3, we have shown the max-
imum achievable mutual information versus the power PF max . We can see that
the proposed methods outperform the method presented in [11] and greatly ex-
ceed the fixed power allocation solution. As can be seen in these figures, the
maximum information capacity tend to be saturated near some reaching thresh-
old with the increasing of PR max , while the maximum capacity almost keep linear
increase with the rising of PF max . Therefore, PF max has more contribution than
PR max after they reaching a certain power.
In Fig. 4, we have plotted the minimum total power of relay nodes PR + PF
versus the capacity threshold. The total transmit power decrease with the rise
of relay numbers due to the diversity gain. It can be observed that the proposed
method suffers certain power reduction compared to the method in [12]. This is
because the iterative computation between w and p in [12] is easy to fall into
local optimal solution, while on the contrary, our proposed method based on GA
can avoid this problem effectively.
5 Conclusion
In this paper, we consider the problem of distributed beamforming in a coop-
erative communication network which consists of a transmitter, a receiver and
two relay clusters equipped at the transmitter and receiver side, respectively. The
beamforming weight vectors are designed in two different approaches. In the first
approach, we aim to obtain the maximum achieved information capacity subject
to the power constraints of two relay clusters. In the second approach, we design
the beamformer through minimizing the total transmit power of all the relay
nodes subject to a constraint which guarantees the mutual information above a
predefined threshold. In both of the two approaches, the two random variables
can be reduced to only one. For this reason, the computational complexity will
be reduced due to the halving of the initial population coding length. Simu-
lation results show that the proposed methods can achieve great improvement
compared to the existing solutions.
Acknowledgment. This work was supported by National Natural Science

Foundation of China under Grant(61371188), Research Fund for the Doctoral
Program of Higher Education under Grant(20130131110029), Natural Science
Foundation of Shandong under Grant(ZR2011FM027), Open Research Fund of
National Mobile Communications Research Laboratory under Grant (2012D10),
China Postdoctoral Science Foundation funded project (2011M501092), Inde-
pendent Innovation Foundation of Shandong University (2012ZD035), Special
Fund for Postdoctoral Innovative Projects of Shandong Province (201103003),
Scientific Research Foundation for the Excellent Young and Middle-aged Scien-
tists of Shandong Province (BS2012DX024).
162 B. Lu et al.
References
1. Laneman, J.N., Tse, D.N.C., Wornell, G.W.: Cooperative diversity in wireless net-
works: Efficient protocols and outage behavior. IEEE Transactions on Information
Theory 50(12), 3062–3080 (2004)
2. Lim, G., Cimini, L.J.: Energy-efficient cooperative beamforming in clustered wire-
less networks. IEEE Transactions on Wireless Communications 12(3), 1376–1385
(2013)
3. Madan, R., Mehta, N.B., Molisch, A.F., Zhang, J.: Energy-efficient cooperative
relaying over fading channels with simple relay selection. IEEE Transactions on
Wireless Communications 7(8), 3013–3025 (2008)
4. Yang, Y., Li, Q., Ma, W.-K., Ge, J., Ching, P.C.: Cooperative secure beamform-
ing for AF relay networks with multiple eavesdroppers. IEEE Signal Processing
Letters 20(1), 35–38 (2013)
5. Ge, M., Wang, Q.-G., Chiu, M.-S., Lee, T.-H., Hang, C.-C., Teo, K.-H.: An effec-
tive technique for batch process optimization with application to crystallization.
Chemical Engineering Research and Design 78(1), 99–106 (2000)
6. Precup, R.-E., David, R.-C., Petriu, E.M., Preitl, S., Radac, M.-B.: Novel adaptive
gravitational search algorithm for fuzzy controlled servo systems. IEEE Transac-
tions on Industrial Informatics 8(4), 791–800 (2012)
7. Saha, S.K., Ghoshal, S.P., Kar, R., Mandal, D.: Cat swarm optimization algorithm
for optimal linear phase FIR filter design. ISA Transactions 52(6), 781–794 (2013)
8. Yazdani, D., Nasiri, B., Azizi, R., Sepas-Moghaddam, A., Meybodi, M.R.: Op-
timization in dynamic environments utilizing a novel method based on particle
swarm optimization. International Journal of Artificial Intelligence 11(A13), 170–
192 (2013)
9. Chalise, B.K., Vandendorpe, L.: MIMO relay design for multipoint-to-multipoint
communications with imperfect channel state information. IEEE Transactions on
Signal Processing 57(7), 2785–2796 (2009)
10. Nassab, V.H., Shahbazpanahi, S., Grami, A., Luo, Z.-Q.: Distributed beamforming
for relay networks based on second-order statistics of the channel state information.
IEEE Transactions on Signal Processing 56(9), 4306–4316 (2008)
11. Chen, L., Wong, K.-K., Chen, H., Liu, J., Zheng, G.: Optimizing transmitter-
receiver collaborative-relay beamforming with perfect CSI. IEEE Communication
Letters 15(3), 314–316 (2011)
12. Wang, C., Liu, J., Dong, Z., Xu, H., Ma, S.: Multi-hop Collaborative Relay Beam-
forming. In: 78th IEEE Vehicular Technology Conference, pp. 1–5. IEEE Press,
Las Vegas (2013)
13. Tang, K.S., Man, K.F., Kwong, S., He, Q.: Genetic algorithms and their applica-
tions. IEEE Signal Processing Magazine 13(6), 22–37 (1996)
Solving Power Economic Dispatch Problem
Subject to DG Uncertainty via Bare-Bones PSO
Yue Jiang, Qi Kang, Lei Wang, and Qidi Wu
Department of Control Science and Engineering,

Tongji University, Shanghai 201804, China
qkang@tongji.edu.cn
Abstract. Distributed generation (DG) is becoming increasingly important in

power system. However the of DG will lead risks in power system due to its
failure or uncontrollable power outputs which is usually relied on renewable
energy. In this work, we solve the economic dispatch (ED) problem by
considering controllable and uncontrollable DG in power system. This paper
applies the bare-bones particle swarm optimization (BBPSO) method to solve
the ED problem. The performance of BBPSO method is evaluated via IEEE
118-bus test system, and would be compared with other methods in terms of
convergence performance and solution quality. The results may verify the
effectiveness and promising application of the proposed method in solving the
ED problem when we are considering both controllable and uncontrollable DG
in power system.
Keywords: particle swarm optimization, bare-bones PSO, economic dispatch,

uncontrollable.
1 Introduction
With the increasing use of distributed generation (DG), the power system will face
many risks of system disruption. Because DG mainly relies on renewable energy such
as solar and wind power which is unstable, thus the real power output is changing
with weather. There are many types of DG, so the changing rules are different. For
instance, the real power output of wind turbines is related to wind speed, which
Weibull distribution is generally considered as the optimal probability density
function. Therefore, this paper adopts Weibull distribution [1] as a probability
distribution of wind speed over time.
Economic dispatch (ED) is an important problem in power system. ED is used in
real-time energy management power system to control the production of thermal
power stations. Its objective is to minimize the total cost of operating the generators,
subject to load and operational constraints [10]. This paper proposes the notion which
is when some DG units have a conspicuous rand feature, we can optimize the power
output of available generators in order to ensure the safe, effectiveness and efficiency
operation of the power system. Its objective is to minimize the total cost of operating
the generators and optimize the network voltage profile.
164 Y. Jiang et al.
Traditionally, ED problem can be addressed by a series of mathematical

programming methods, such as lambda-iteration method [2], gradient method [3], and
dynamical programming [4]. Nevertheless, such deterministic numerical methods
cannot effectively solve nonsmooth and nonconvex cost function, moreover, it has
larger scale dimension which is too difficult to calculate. In order to effectively solve
such problems in power systems, many researchers used a variety of computational
intelligence to handle it. For example the genetic algorithm [6], differential evolution
algorithm [7], neural network [8], Tabu search algorithm [9] and so on. But the
methods mentioned above have many shortcomings, including premature and poor
convergence performance.
Usually speaking, people consider ED problem as a static optimization problem
discussing in a normal power system. In this paper, however, we want to obtain a
real-time power output of controllable generators in a dynamic environment. PSO
[10]-[14] has been employed to handle the ED problem for many years. We propose
the BBPSO [15]-[19] to solve ED problems.
2 Mathematical Model
The objective in this paper is to minimize the total short-term cost of operating the
generators and network loss. It is given by:
Minimize F ∑ F P P (1)
where F P is the cost function of the jth generator, P is the real output of the
jth generator, n is the total # of the jth generators in the power system, P is the
network loss in the power system.
F P is related to the real power injected into the power system, which is modeled
by the function.
(2)
where a , b and c are cost coefficient of the jth generator.

This paper adopts the Weibull distribution as probability distribution of wind speed
over time .The probability density is given by.
f v exp (3)
where, v is wind speed, k and c are shape parameter and scale parameter of
Weibull distribution respectively. k and c can be calculated by the average wind speed
µ and standard deviation .
.
k (4)
c (5)
Solving Power Economic Dispatch Problem Subject to DG Uncertainty 165
where Γ is Gamma function.

The relationship between the power output and wind is as followed[21].
P (6)
Where v is cut-in wind speed, v is rated wind speed, P is rated power, v is

cut-out wind speed. k , k can be calculated as follows:
(7)
(8)
According to statistics, wind speeds stay between cut-in speed and the rated wind
speed most of the time. In this paper DG is to simplify as PV node in the power flow
caculation.
In this work, we consider the following constraints.
∑
j 1, … , n
j 1, … , n (9)
Where PD is the load demand of the power system; The operating range of all on-
line is restricted by ramp rate limits. If power generation increases, then P P
UR . If power generation decreases, then P P DR . P is the previous power
output power, UR the up-ramp limit of the jth generator, and DR is down-ramp
limit of the jth generator.
3 Bare-Bones PSO Algorithm
PSO commonly used decimal encoding. It was motivated by the social behavior of
fish schooling and bird flocks. There are N particles in the PSO algorithm. The search
space assumes dimension D. Each particle has three vectors. Its current position
x x , x , … , x D , its velocity v v , v , … , v D and its personal best position
pBest.
However, F. van den Bergh proved the update formula which cannot guarantee
convergence to the global optimal solution [22]. Therefore, Kennedy proposed the
BBPSO[15].
166 Y. Jiang et al.
The particles’ position in BBPSO is updated by the following equations:

B
pBest est rand 0.5
(10)
pBest else
Where, if rand>0.5, current position x pBest , otherwise the current position
B B
will be derived from a Gaussian distribution with the mean and
standard deviation pBest gBest .
By the formula (10) to update x , it avoids adjusting parameter in the PSO
algorithm and easy to implement. And it has the strong global convergence.
4 Solving ED Problems via BBPSO
Step 1. Input data of the power system, and set algorithm parameters. The power
system data includes load demand, minimum P and maximum P of each
generator, minimum U and maximum U of each node varies.
Step 2. Set k=1. Initialize the current position of each particle. Each particle’s
position is represented as matrix X. Set its personal best position equal to be its
current position:
Step 3. By Power flow calculation, we can obtain the fitness value for each
particle.
Step 4. Update pbest and gbest.
Step 5. Update the position according to (12). If the position cannot fulfill
constraints in (9).
Step 6. Examine the termination condition. If it is not met, Set k=k+1 and return to
Step 3. Otherwise, end and output results.
5 Simulations and Results
In order to verify the effectiveness of the BBPSO, the power system we used for
simulation is the IEEE 118-bus system[25] in this work. This system consists of 118
buses and 54 generator nodes. 1,4,6,8,10,12,15,18,19,24 and 25 node are DGs. The
rang of the real power output of GD is 0-50Mw. Node 69 is a slack one in power flow
calculation. The entire load of the power system is 4242MW. It assumed that node 1
accessing a wind turbine that has a conspicuous rand feature. The wind speed will be
derived from a weibull distribution with the k 2.3466 and c 8.0928. Rated
power of wind turbine is 1Mw, cut-in wind speed is 3m / s, rated wind speed is 12m/s.
The problem now is how to optimize the real power output of all controllable online
generators and to satisfy the load demand.
According to (3)-(6), we can get the real power output of node 1 at interval of 10
minutes. Fig.1 shows the power output in each period.
Fig. 1. Power change of node 1
The parameters of GA were set in[10], Pc=0.8 and Pm=0.1 are the crossover
probability and mutation probability respectively. The parameters of RDPSO were set
in[10], and are the thermal coefficient and the drift coefficient respectively, and the
decrease from 0.9 to 0.3 on the course of the search. The parameters of PSO were
set in[20], the range of inertia weight ω is decrease from 1.2 to 0.8 and the leaning
factors were set as c1=c2=2.
Table 1 lists the total cost by each method mentioned when the power output of
node 1 is 4.49Mw (in the first change). The mean cost and the standard deviation got
by 100 runs of these methods. Fig.2 is convergence properties of the tested
optimization methods. Both of them indicate that BBPSO has a better performance
and robustness than other methods .
Table 1. The results for 100 runs
Methods Min.Cost Mean.Cost Std.Dev Mean.Time/min

GA 147856 161504 13648 10.0
PSO 150489 151815 1326 9.0
RDPSO 152352 153093 714 8.0
BBPSO 148023 148275 252 5.0
Fig. 2. Convergence properties of methods Fig. 3. The output power solution

168 Y. Jiang et al.
Fig.3 is the solution of the power output of 53 generators when the power output of
the node 1 keep changing in the interval of 10 minutes. It shows the power output
solution of 52 generators in 10 times, when the wind speed of the node 1 is derived
from a weibull distribution with the k 2.3466 and c 8.0928.
6 Conclusion
In this paper, we propose a BBPSO-based method to solve ED problems. The

performance of the proposed method is evaluated in an IEEE 118-bus power system,
and compared with other optimization methods in terms of convergence performance
and solution quality. The results show that BBPSO method performs better than other
compared methods in solving the ED problem. Our future work will focus on the
dynamic economic dispatch problem.
Acknowledgement. This work was supported in part by the Natural Science

Foundation of China (71371142, 61005090, and 61034004), the Fundamental
Research Funds for the Central Universities, and the Research Fund of State Key Lab.
of Management and Control for Complex systems.
References
1. Balouktsis, A., Chassapis, D., Karapantsios, T.D.: A nomogram method for estimating the
energy produced by wind turbine generators. Solar Energy 72(3), 251–259 (2002)
2. Chowdhury, B.H., Rahman, S.: A review of recent advances in economic dispatch.
Institute of Electrical and Electronics Engineers (1990)
3. Dhar, R., Mukherjee, P.: Reduced-gradient Method for Economic Dispatch. Proc. Inst.
Elec., Eng. 120(5), 608–610 (1973)
4. Abarghooee, R., Aghaei, J.: Stochastic Dynamic Economic Emission Dispatch
Considering Wind Power. In: IEEE Engineering and Automation Conference, vol. 1,
pp. 158–161 (2011)
5. Loia, V., Vaccaro, A.: Decentralized Economic Dispatch in Smart Grids by Self-
organizing Dynamic Agents. IEEE Trans. Systems, Man and Cybernetics: Systems 44(4),
397–408 (2014)
6. Chiang, C.: Improved Genetic Algorithm for Power Economic Dispatch of Units with
Value-point Effects and Multiple Fuels. IEEE Trans. Power Syst. 20(4), 1690–1699 (2005)
7. Elsayed, S., Sarker, R., Essam, D.: An Improved Self-adaptive Differential Evolution
Algorithm for Optimization Problems Industrial Informatics. IEEE Trans. on Industrial
Informatics 9(1), 89–99 (2013)
8. Yang, C., Deconimck, G., Gui, W.: An Optimal Power-dispatching Control System for
Electrochemical Process of Zinc Based on Back-propagation and Hopfield Neural
Network. IEEE Trans. Electron 50(5), 953–961 (2003)
9. Lin, W., Cheng, F., Tsay, M.: An Improved Tabu Search for Economic Dispatch with
Multiple Minima. IEEE Trans. Magn. 38, 1037–1040 (2002)
10. Sun, J., Palade, V., Wu, X.: Solving the Power Economic Dispatch Problem with
Generation Constraints by Random Drift Particle Swarm Optimization. IEEE Trans. on
Industrial Informatics 10(1), 222–232 (2014)
11. Faria, P., Soares, J.: Modified Particle Swarm Optimization Applied to Integrated Demand
Response and DG Resources Scheduling. IEEE Transaction on Smart Grid 4(1), 606–616
(2013)
12. Chen, S., Manalu, G.: Fuzzy Forecasting Based on Two-factors Second-order Fuzzy-trend
Logical Relationship Groups and Particle Swarm Optimization Techniques. IEEE
Transaction on Cybernetics 43(3), 1102–1117 (2013)
13. Niknam, T., Doagou-Mojarrad, H.: Multiobjective Economic/Emission Dispatch by
Multiobjective θ-particle Swarm Optimization, Generation, Transmission & Distribution,
6(5), 363–377 (2012)
14. Fu, Y., Ding, M.: Route Planning for Unmanned Aerial Vehicle (UAV) on the Sea Using
Hybrid Differential Evolution and Quantum-behaved Particle Swarm Optimization. IEEE
Trans. Systems, Man, and Cybernetics: Systems 43(6), 1451–1465 (2013)
15. Kennedy, J.: Bare Bone Particle Swarm. In: Proceedings of the IEEE Swarm Intelligence
Symposium, pp. 80–87 (2003)
16. Blackwell, T.: A Study of Collapse in Bare Bones Particle Swarm Optimization. IEEE
Transactions on Evolutionary Computation 16(3), 354–372 (2012)
17. Chen, C.: Cooperative Bare Bone Particle Swarm Optimization. Information Science and
Control Engineering, 1–6 (2012)
18. Krohling, R., Mendel, E.: Bare Bones particle Swarm Optimization with Gaussian or
Cauchy Jumps. In: IEEE Congress on Evolutionary Computation, pp. 3285–3291 (2009)
19. Wang, P., Shi, L.: A Hybrid Simplex Search and Modified Bare-bones Particle Swarm
Optimization. Chinese Journal of Electronics 22(1), 104–108 (2013)
20. Kang, Q., Zhou, M.: Swarm Intelligence Approaches to Optimal Power Flow Problem
with Distributed Generator Failures in Power Networks. IEEE Transactions on Automation
Science and Engineering 10(2), 343–353 (2013)
21. Shapic, E., Balzer, G.: Power Fluctuation from a Large Wind Farm. In: International
Conference on Future Power Systems, November 16-18 (2005)
22. van den Bergh, F.: An Analysis of Particle Swarm Optimizers, Pretoria, South Africa:
Depatment of Computer Science, University of Pretoria (2002)
23. Pan, F., Hu, X., Eberhart, R.: An Analysis of Bare Bones Particle Swarm. In: IEEE Swarm
Intelligence Symposium, pp. 1–5 (2008)
24. Kang, Q., Lan, T., Yan, Y.: Group Search Optimizer Based Optimal Location and
Capacity of Distributed Generations. Neurocomputing 78(1), 55–63 (2012)
25. Power Systems Test Case Archive, http://www.ee.washington.edu/
research/pstca/pf118/pg_tca118bus.htm
Extracting Mathematical Components Directly
from PDF Documents for Mathematical Expression
Recognition and Retrieval*
Botao Yu1,2, Xuedong Tian1,2,**, and Wenjie Luo1,2

1
College of Mathematics and Computer, Hebei University, Baoding, Hebei, China
2
Hebei key laboratory of Machine Learning and Computational Intelligence, Baoding, China
txdinfo@yahoo.com
Abstract. PDF document gains its popularity in information storage and ex-
change. With more and more documents, especially the scientific documents,
available in PDF format, extracting mathematical expressions in PDF docu-
ments becomes an important issue in the field of mathematical expression
recognition and retrieval. In this paper, we proposed a method of extracting ma-
thematical components directly from PDF documents rather than cooperating
indirectly with corresponding images converted from PDF files. Compared with
traditional image-based method, the proposed method makes full use of the in-
ternal information of PDF documents such as font size, baseline, glyph bound-
ing box and so on to extract the mathematical characters and their geometric in-
formation. The experimental result shows the method could meet the needs of
the following processing of mathematical expressions such as formula structural
analysis, reconstruction and retrieval, and has a higher efficiency than tradition-
al image-based ways.
Keywords: PDF, Mathematical expression component, Mathematical

expression recognition and retrieval, Font size, Baseline, Glyph bounding box.
1 Introduction
PDF (Portable Document Format) [1] documents present their contents and layouts in
a manner independent of application software, hardware, and operating systems,
which provides users with a consistent experience in sides of the displaying and print-
ing pattern. With an increasing number of documents presented with PDF, more and
more attentions are paid to this format of document for making good use of this re-
source.
Current researches on PDF documents involve extracting components from PDF
documents or converting PDF documents into other formats such as XML and
*
This work is supported by the National Natural Science Foundation of China (Grant No.
61375075) and the Natural Science Foundation of Hebei Province (Grant No.
F2012201020).
**
Extracting Mathematical Components Directly from PDF Documents 171
HTML. Chao and Fan [2] developed a method of extracting layout and content of
PDF documents. The document was separated into text, image and vector graphics
according to the object type. After that, words were formed to lines, then segments
and images and vector graphics were saved. Marinai [3] developed a software tool to
extract administrative metadata from PDF documents which could assist for building
personal digital libraries. Déjean and Meunier [4] designed a system for converting
PDF documents into structured XML format. In this system, streams that contain text,
bitmap and vector images were extracted and converted respectively, and the ex-
tracted components were expressed in XML format. Rahman and Alam [5] proposed
a method of converting PDF documents into HTML. By applying document image
analysis techniques to retrieval logical and layout information of the document, the
document was output in HTML format.
As an important component in PDF documents, mathematical expressions are also
needed to be extracted for further recognizing and searching processes. Different from
the images recognized and analyzed by mathematical formula recognition system, a
PDF document that is not generated from scanned images has already contained the
information of character code, baseline and font size. Therefore, the traditional opera-
tions applied to formula images for improving image quality and obtaining symbol
codes such as image preprocessing (including binarization, denosing, skew detection
and correction, etc.) and symbol recognition are not required. Nevertheless, the PDF
documents do not provide the syntax and semantic information of symbols, it is ne-
cessary to locate the precise geometrical information of symbols and obtain their logic
relationships for the following mathematical expression extraction, reconstruction and
retrieval operations. Yang and Fateman [6] stated the significance of accessing ma-
thematical expressions on line in digital documents like Postscript and PDF docu-
ments. And a method of extracting formulas from Postscript documents is proposed.
First, a modified version of a program called Prescript was used to output information
about strings and bounding boxes about typeset expressions. Then broken string
fragments were assembled into words and items were determined as part of mathe-
matical expressions by the characteristics of fonts and words (e.g. sin) commonly
used in mathematical expressions and mathematical characters. Finally, the built-up
mathematical expression in Lisp data structure was generated from stored data by
applying clumping heuristics based on an existing Math/OCR program. In literature
[7], Chan and Yeung summarized the existing work of mathematical expression rec-
ognition on symbol recognition and structural analysis. Lin et al. [8] combined rule-
based and learning-based methods to detect both isolated and embedded mathematical
expressions in PDF documents. For isolated formulas, they first used the character
features to remove text lines which didn't seem to contain expressions. Then, the con-
fidence level of classifying a line as a formula line was calculated by exploiting the
geometric layout features. For embedded formulas, they focused much on the charac-
ter features combined with additional layout features. Then the confidence level of a
character being a math symbol was calculated. If the confidence level was higher than
a threshold, the corresponding line was detected as an isolated formula or an embed-
ded formula. In literature [9], they further discussed the identification of embedded
mathematical formulas in PDF documents. First, text lines were segmented into words
which are classified into formula type and ordinary text type with an SVM classifier.
Then, formulas were extracted by merging formula type words as formulas. Baker et
172 B. Yu, X. Tian, and W. Luo
al. [10, 11] proposed a method of extracting mathematical expression by accessing the
PDF document and a rasterized version of the PDF document. First, characters and
their related information of font size, baseline and bounding box were extracted from
the original PDF documents. In order to solve the problem of bounding box overlap-
ping and obtain the exact character bounding boxes, they rendered PDF documents
into images. After searching bounding boxes of the glyphs in the image, all the
bounding boxes were registered with characters obtained from the original PDF doc-
uments. Then the expression parse tree was established with characters and related
geometric information. This method paid attention to internal information in PDF
documents such as font size, baseline, font name and font bounding box, which
helped to locate the characters and got the minimal character bounding box from the
corresponding images converted from original PDF document indirectly.
In this paper, we propose a method of extracting mathematical components directly
from PDF documents for mathematical expression recognition and retrieval. Different
from the method proposed in literature [10, 11], we obtain all information about glyph
bounding box, baseline and font size, by directly accessing the original PDF docu-
ments, which makes full use of the internal information in the PDF documents and is
also efficient. The glyph bounding boxes, together with baselines and character codes
are used for the following processing.
A PDF document has complex structure. It belongs to a text and binary integrated
format with compressed data, which leads to low readability of the original code of
the documents. PDF documents can be generated by many tools and each tool has its
own standard based on the PDF reference which has 7 editions so far. Font types used
in PDF document could also frequently vary. When it comes to mathematical sym-
bols, some symbols are generated by path construction operators (e.g. the long frac-
tion symbol) or made of a character and a shape defined by the path construction
operator (e.g. √ and a horizontal line make up the radical symbol). All these facts
make it complex to extract components from a PDF document directly.
Our proposed method focuses on PDF documents with type1 font only and doesn't
constrain tools that generate PDF documents.
2 Extracting Mathematical Components from PDF Documents

Directly
The workflow of our method mainly contains 3 steps described as following and
shown in Fig. 1.
1. Parsing of PDF documents. Parse the PDF documents to get the information of the
fonts in Resource dictionary and the content of the Content stream of the Page ob-
jects.
2. Components extraction. Extract font and character information from the content
parsed in step 2.
3. Expression output. Compose the mathematical expression with components in step
2 by using existing technique.
Fig. 1. Workflow of the proposed method
2.1 Parsing of PDF Documents

For a better understanding of PDF and considering our specified research purpose, we
decide to develop a special parsing algorithm of PDF documents, although there are
many open source SDKs or open source software such as iTextSharp, iText, PDFBox
and so on that can handle PDF documents. The parsing process mainly contains the
following steps:
1. Read the global objects such as Xref Table and root object to get all the Page ob-
jects.
2. Get basic font information in Resource dictionary of the Page object.
3. Decode the content stream of the Page object.
Once the PDF document is parsed and the contents of content stream and font infor-
mation in resource dictionary are obtained, we move to the extraction process.
2.2 Extraction of Glyph Bounding Box in Font File

By accessing the font dictionary and the font descriptor obtained in 2.1, we get the
font type, font name, font bounding box, character encoding and width of the glyph in
the font. Font bounding box and width are expressed in glyph coordinate system and
are measured in units in which 1000 units corresponds to 1 unit in text space. When
the font bounding box is transformed by the transformation method which will be
described in 2.3, we find that the font bounding box is actually larger than the area the
character pixels take and cannot be treated as the parameter to calculate the precise
area the character occupy, as shown in Fig. 2.
174 B. Yu, X. Tian, and W.
W Luo
Fig. 2. An example of tran

nsformed font bounding boxes of the corresponding characters
In Fig.2, it is obvious th
hat character a, b and c that share the same font have the
font bounding boxes with the same height which can not reflect the real areas tthey
occupy. The bounding box x for the character - is also much bigger than expected.
These boxes are grosser than the exact coordinates of each character's bounding bbox
and cannot be used as geom metric information for mathematical expression recognitiion.
To solve the problem, we canc take advantage of PDF documents in consistence tthat
different PDF customer ap pplications can display the PDF documents with same ap-
pearance, i.e. characters at the same position with the same appearance to get m more
precise geometrical data froom the parameters in PDF documents that control the ppat-
tern of the layouts. Now wee concentrate on the font file in PDF documents.
ntaining Type 1 font program. It's in binary format and de-
Font file is a stream con
fines logic to render the ch
haracter. Positional values in font file are also expressedd in
glyph coordinate system. Type1
T font has two types, the Compact Font Format [[12]
and Adobe Type 1 Font Forrmat [13]. Each type has its own format, but uses the saame
method to render the charaacter. Type1 font uses commands to draw lines and Bezzier
curves to describe the appeaarance of the characters, as shown in Fig. 3.
Fig. 3. An
n example of character b described in font file
In Fig. 3, there are manyy key position points on or off the outline of the characcter.
These points control the dissplay behaviors of the lines and Bezier curves that makee up
the outline, which could bee used to calculate the exact bounding box, also called the
glyph bounding box.
Fig. 4(a) shows one kind d of binary data commands in font file that draws the ccha-
racter b and Fig. 4(b) showss the same commands but in decoded format which is m more
readable.
(a) (b)
Fig. 4. Commands in font fille rendering character b. (a) Binary format; (b) Decoded formaat
The process of our prograam that parses the decoded commands to get the glyyph
bounding box is composed of the following steps:
1. Read one single comman nd and denote it as CurrentCmd.
2. If CurrentCmd is endchaar, go to step 5; otherwise, go to step 3.
3. Obtain all the key pointss that control the boundary of the shape in CurrentCmd and
calculate the minimal boox that holds all the points. Denote the box as CurrentBoox.
4. If CurrentBox is the firstt box got, denote it as BOX; otherwise calculate the minnim-
al box that holds CurreentBox and BOX, and update BOX with the result. Goo to
step 1.
5. Output BOX. End.
By decoding the binary daata and parsing the commands, the process of renderring
characters is reproduced, frrom which we obtain the glyph bounding boxes. As w with
the font bounding box, the glyph bounding box is also expressed in glyph coordinnate
system. The glyph boundin ng box in font file is actually a unit bounding box, whhich
means it equals to the exactt area that the corresponding character whose font size is 1
takes. After the transformaation in 2.3, the glyph bounding box could be used as the
precise bounding box of th he corresponding character with a certain font size. So far
we have gotten the glyph bounding
b boxes in font file, the next is to extract geomeetric
shapes and characters from content stream and transform the glyph bounding boxess.
2.3 hematical Elements

Extraction of Math
Mathematical elements aree classified into 4 types, single character, multi-charactters,
character with geometric sh hapes, and geometric shapes. In this process, all elemeents
are extracted with their relaated attributes. We design a structure called ElementInfo
fo to
represent a mathematical ellement.
struct ElementInfo{
string glyphName;;/*element name, e.g. plus for + */
double baseLine;
gBox actualGBB;/*transformed glyph
TextSpaceBounding
bounding box*/
};
W Luo
Fig
g. 5. An example of the extracted data
Fig. 6. An example
e of transformed glyph bounding boxes
In order to get the math hematical elements, we write a stack based program to ex-
tract characters, geometric shapes and the related attributes such as font size, sstart
point and baseline from decoded
d content stream obtained in 2.1. Each charactter's
position information can be b calculated with the text showing commands in conttent
stream. During the processs, the glyph bounding boxes are transformed from glyyph
coordinate system to text coordinate
c system in 3 steps. First, the value of the glyyph
bounding box is divided by 1000. Then, multiply the value of the bounding box by the
font size. Finally, move th he origin of the glyph bounding box to the correspondding
character's start point. For example,
e the glyph bounding box of character x in font file
is [0 500 0 600] which tak kes the form [llx urx lly ury] specifying the lower-left x, up-
per-right x, lower-left y andd upper-right y coordinates of the bounding box. In conttent
stream, the start point of x is (100, 600) in text space and its font size is 6. After the
first step, the value of the glyph
g bounding box is [0 0.5 0 0.6]. Then, it is changedd to
[0 3.0 0 3.6]. Finally, it beccomes [100 103.0 600 603.6], which is the precise bouund-
ing box of x. All the positioonal information is expressed in the coordinate that sets the
lower-left corner of the pag ge as the origin. An example of the result obtained by our
program for the expression is shown in Fig. 5. The transformed glyph bounding booxes
of the expression are shown n in Fig. 6.
From Fig. 5 and Fig. 6 we
w can see that all characters share the same baseline exccept
character 2 and all boundin ng boxes are equal to the real sizes of the characters, whhich
tell the exact position relatio
onship between each character.
(a) (b) (c)

Fig. 7. Special mathematical elements. (a) Racial; (b) Long vertical bar; (c) Long fraction line.
Now the most of mathematical elements and their precise geometric information are
ready for mathematical expression recognition. But some special situations still exist
which would result in error results without analyzing and processing.
• Some mathematical elements are composed of characters and geometric shapes,

e.g. the radical as shown in Fig. 7(a). The geometric shapes are added to the cha-
racters and the glyph bounding boxes are recalculated. For the radical in Fig. 7(a),
the character √ intersects with the horizontal line, so the character and the line
are treated as a whole and the glyph bounding box is recalculated to enclose the
two items.
• For those who are composed of multi-characters, e.g. the vertical bar | as shown in
Fig. 7(b), which may contain more than one shorter vertical bar to form a longer
one, all the components are integrated as a new mathematical element with the
same character name but a larger glyph bounding box and new baseline. In Fig.
7(b), the shorter vertical bars overlap with each other. The combining process is
similar to the situation of the radical.
• The elements who are composed of geometric shapes, e.g. the longer fraction line
as shown in Fig. 7(c) will be label as certain mathematical elements due to the con-
text and their shapes.
2.4 Output of Expressions

After all the characters and their related attributes are extracted, the method of struc-
tural analysis of mathematical expressions proposed in [14] is used to analyze the
structure to build the syntax tree for the expected expression. Once the syntax tree is
build, the expression could be output in LATEX format. We will not talk about it
much here.
In this paper, MikTex and LATEX editors are used to generate PDF documents with
Type1 fonts of varying font sizes and mathematical expressions in different forms
such as matrix, integral, radical and so on. We use our proposed method to extract
mathematical components directly from PDF documents that are similar to the layout
as shown in Fig. 8(a).
W Luo
(a) (b)
Fig. 8. Layout of a PDF file and its experimental results. (a) Layout; (b) Results.
Fig.8 (b) is the result sh

howing exact geometric information of mathematical eele-
ments in Fig.8 (a) by appllying our method. Experimental results show that all the
information of the characters in the generated mathematical expressions, includding
character name, glyph bou unding box, position, and baseline could be extracted ccor-
rectly. The mathematical co omponents could be used for further mathematical exprres-
sion recognition and retriev
val.
4 Conclusions
In this paper, a method off extracting mathematical components directly from P PDF
documents for mathematicaal expression recognition and retrieval is proposed, whhich
makes full use of the interrnal information of the PDF documents and doesn't drraw
support from any other meethods like format conversion. The extracted componeents
could be used to mathematiical expression extraction, reconstruction and retrieval.
Although the proposed method
m could extract the components correctly, it is oonly
for Type1 fonts. Our furtheer work is to transferred this method to other type fonts and
improve its robustness.
Acknowledgements. This work

w is supported by the National Natural Science Founnda-
tion of China (Grant No. 61375075)
6 and the Natural Science Foundation of Heebei
Province (Grant No. F2012201020).
References
1. Adobe Systems Incorporated, PDF Reference, 6th edn. (November 2006)
2. Chao, H., Fan, J.: Layout and Content Extraction for PDF Documents. In: Marinai, S.,
Dengel, A.R. (eds.) DAS 2004. LNCS, vol. 3163, pp. 213–224. Springer, Heidelberg
(2004)
3. Marinai, S.: Metadata Extraction from PDF Papers for Digital Library Ingest. In: 10th In-
ternational Conference on Document Analysis and Recognition, pp. 251–255. IEEE Press,
New York (2009)
4. Déjean, H., Meunier, J.-L.: A System for Converting PDF Documents into Structured
XML Format. In: Bunke, H., Spitz, A.L. (eds.) DAS 2006. LNCS, vol. 3872, pp. 129–140.
Springer, Heidelberg (2006)
5. Rahman, F., Alam, H.: Conversion of PDF Documents into HTML: a Case Study of Doc-
ument Image Analysis. In: Conference Record of the Thirty-Seventh Asilomar Conference
on Signals, Systems and Computers, vol. 1, pp. 87–91. IEEE Press, New York (2004)
6. Yang, M., Fateman, R.: Extracting Mathematical Expressions from Postscript Documents.
In: Proceedings of the 2004 International Symposium on Symbolic and Algebraic Compu-
tation, pp. 305–311. ACM (2004)
7. Chan, K.-F., Yeung, D.-Y.: Mathematical Expression Recognition: a Survey. J. Interna-
tional Journal on Document Analysis and Recognition. 3(1), 3–15 (2000)
8. Lin, X.Y., Gao, L.C., Tang, Z., Lin, X.F., Hu, X.: Mathematical Formula Identification in
PDF Documents. In: 2011 International Conference on Document Analysis and Recogni-
tion, pp. 1419–1423. IEEE Press, New York (2011)
9. Lin, X.Y., Gao, L.C., Tang, Z., Hu, X., Lin, X.F.: Identification of Embedded Mathemati-
cal Formulas in PDF Documents Using SVM. In: IS&T/SPIE Electronic Imaging,
pp. 82970D–82970D. International Society for Optics and Photonics (2012)
10. Baker, J.B., Sexton, A.P., Sorge, V.: Extracting Precise Data on the Mathematics Content
of PDF Documents. Towards Digital Mathematics Library, Birmingham , pp. 75–79
(2008)
11. Baker, J.B., Sexton, A.P., Sorge, V.: A Linear Grammar Approach to Mathematical For-
mula Recognition from PDF. In: Carette, J., Dixon, L., Coen, C.S., Watt, S.M. (eds.)
MKM 2009, Held as Part of CICM 2009. LNCS, vol. 5625, pp. 201–216. Springer,
Heidelberg (2009)
12. Adobe Systems Incorporated, The Compact Font Format Specification, Version 1.0, 4
(December 2003)
13. Adobe Systems Incorporated, Adobe Type 1 Font Format, Version 1.1 (February 1993)
14. Tian, X.D., Li, N., Xu, L.J.: Research on Structural Analysis of Mathematical Expressions
in Printed Documents. J. Computer Engineering 32(23), 202–204 (2006)
An Efficient OLAP Query Algorithm Based
on Dimension Hierarchical Encoding Storage and Shark
Shengqiang Yao and Jieyue He*
School of Computer Science and Engineering,

MOE Key Laboratory of Computer Network and Information Integration,
Southeast University, Nanjing 210096, China
shengq_yao@163.com,jieyuehe@seu.edu.cn
Abstract. The on-line analytical processing (OLAP) queries always include

multi-table joins and aggregation operations in their SQL clauses. As a result,
how to reduce multi-table joins and effectively aggregate the query data with
“big data” is the key issue for query processing. Therefore, the novel OLAP
query algorithm is proposed in this paper based on the dimension hierarchical
encoding (DHE) storage strategy with the In-Memory computing in Shark.
With DHE and Shark, a star join with hierarchy level is mapped to a multidi-
mensional range query on the fact table and the large-scale data by transforma-
tions and actions are computed on resilient distributed datasets (RDDs). The
experimental results show that, compared with the data analysis operations in
Hive, complex multi-table joins and I/O overhead are reduced by DHE and
Shark. The query performance is greatly improved than that of the ordinary star
schema.
Keywords: Big data, Data warehouse, Dimension hierarchical encoding, Shark.
1 Introduction
Data warehousing and on-line analytical processing (OLAP) [1] are essential ele-
ments of decision support system. However, the rapidly growing size of the data sets
for business intelligence makes traditional warehousing solutions unsuitable. The
main idea underneath this evolution is that those data sets need to be stored in the
cloud and be accessed by a set of services. Following this consideration, there have
been several proposals to store and process extremely large data sets.
Hadoop [2] is a popular open-source map-reduce implementation inspired by
Google’s MapReduce [3]. It is being widely used in web search, log analysis and
other large-scale data processing filed. Hive [4] is a data warehousing solution built
on top of Hadoop. It supports SQL-like declarative language (HiveQL) and is widely
used. HiveQL is compiled into map-reduce jobs executed on Hadoop. However, ex-
pensive data materialization for fault tolerance and costlier execution strategies [5, 6]
makes Hive slow. Shark [7, 8] is a new data warehouse system capable of deep data
*
An Efficient OLAP Query Algorithm 181
analysis using the Resilient Distributed Datasets (RDDs) memory abstraction. Though
Shark shows quite better performance than hive, complex OLAP queries still take lots
of system resources and affect the query efficiency, like pre-shuffle and intermediate
data outputs, especially when the data sets become larger.
OLAP over big data repositories has recently received a great deal of attention
from the research communities [10]. For example, there is a research focus on the
performance of aggregation operations by using dimension-oriented storage in HBase
[11]. A decomposed snowflake schema was proposed in [12] to get better perfor-
mance in their parallel database. The work in [13] shows that HBase can’t original
support the complex OLAP queries. The performance of “data loading time” and
“grep select time” in HBase is poorer than Hive. Other studies, like [14], they use a
distributed cube model in the cloud platform. However, cube needs to be pre-
computed and extra space to store. It is also quite difficult to decide which queries to
pre-compute. In traditional warehouse system, a multidimensional hierarchical me-
thod [15, 16] is used to reduce table joins and improve the efficiency of queries.
Therefore, an efficient OLAP query method is proposed in this paper by using dimen-
sion hierarchical encoding (DHE) storage and shark. The code for representing the
hierarchies of each dimension replaces the foreign keys in the fact table. As a result,
the complex table joins reduced and the efficiency of OLAP queries improved. Our
experimental results on star schema benchmark (SSB) [9] show that with DHE sto-
rage and shark, the OLAP query method has better performance.
Briefly then, the outline of this paper is as follows. In Section 2, the method of
OLAP query based on DHE and shark are described in detail. In Section 3, our algo-
rithm is applied to SSB and the results are analyzed. In Section 4, the conclusions are
given.
2 OLAP Query Based on DHE and Shark
2.1 Dimension Hierarchical Encoding

In data warehouse systems, a central fact table contains the measures, and the dimen-
sion tables connect to it via foreign keys. In most situations, the size of dimension
table is far less than the fact table and the number of hierarchy level members is
small. So we need only a little effort to tackle with the DHE when an OLAP query
comes without losing any original semantics.
Assuming that is the j-th level attribute of dimension table and is the
binary code of . can be denoted as formula (1):
… … . (1)
The value of is 0 or 1 and k is the binary code length of . Its value can be cal-
culated by formula (2):
k Bit log (m . (2)
m is the number of different members in .
182 S. Yao and J. He
Assuming that is the dimension hierarchy encoding of dimension table . To

obtain , DHE combines each hierarchy level of dimension in a level-
descending order (from the top level to low level). can be calculated by formula
(3):
… | … | .
(3)
“ ” is the left shift operator and “|” is the bitwise-Or operator.
For example, as shown in table 1, the “Customer” dimension has 3 different hie-
rarchy levels {Region, Nation, City}. “Region” is the top level attribute and has total
5 different members, which is originally stored as string type.
According to formula (1) and (2), 3 bits are used to replace the 5 different members
in Region with {001,010,011,100,101}. “Nation” and “City” can be respectively
represented by 4bit and 8bit. By formula (3), we can obtain the DHE of dimension
table “Customer”. Table 1 displays the part of the DHE of “Customer”.
Table 1. Part of the DHE of Customer
Customer(Dimension) Region(Level) Nation(Level) City(Level)

CustKey BCustomer Region BRegion Nation BNation City BCity
1 0010001…00001 AFRICA 001 ALGERIA 0001 ALGE 00000001
0
… 010…010100001 … … ARGENTINA 0010 …… ……
300 … ASIA 011 … … BRAZ 00010101
… … … … BRAZIL 1111 0 …
…
In order to speed up the OLAP query with “big data” and reduce storage space, we
use DHE to replace the dimension foreign keys in fact table and this phase can be
done by ETL tools. Fig.1 shows the example of the storage strategy with DHE in star
schema.
BTime BSupplier
datekey suppkey
Year Region
Month BTime Nation
Week Bcustomer City
… Bsupplier ...
Bpart
Quantity
BCustomer … BPart
custkey Tax partkey
Region MFGR
Nation Categ
City Brand
… …
Fig. 1. Star schema with DHE

2.2 Overview of Spark and Shark

Spark [7] is a fast and general engine for large-scale data processing which outper-
forms Hadoop. Spark introduces an abstraction called RDDs (resilient distributed
datasets). An RDD is a read-only collection of objects partitioned across a set of ma-
chines. RDDs achieve the abstraction to manipulate distributed data sets in local oper-
ations. RDDs can be calculated in two ways: Transformations (e.g. map, filter) and
Actions (e.g. count, collect). Transformations are lazy operations that define a new
RDD, while actions launch a computation to return a value.
Spark is composed of one Master node and several Worker nodes. The program
developed by users called driver program. The workers are long-lived processes that
can store RDD partitions in RAM across operations. The driver defines one or more
RDDs and invokes actions on them.
Shark [8] is a data warehousing implementation on Spark. Shark is compatible with
Hive and uses Hive query compiler. However, it transforms the operators into the
operation on RDDs rather than MapReduce jobs. The Hadoop Distributed File System
(HDFS) data obtained by shark will be computed by spark.
2.3 The OLAP Query Algorithm Based on DHE and Shark

Usually OLAP users are not interested in single measures but in some form of sum-
marized data. An important concept of OLAP data models is the notion of dimension
hierarchies. Hierarchies provide an appropriate method of describing the level of ag-
gregation for a dimension. Thus, typical OLAP queries in star schema contain restric-
tions on multiple dimension tables that are then used as restrictions on the very large
fact table:
Select F.attributes,D.attributes,Agg(Measures)
From Fact,Dimension
Where <Join Constraint> and <Dimension Restriction>
and <Fact Restriction>
Group By Gb-attributes
Order By Ob-attributes
Attributes in “<Dimension Restriction>”, “Gb-attributes” and “Ob-attributes” mostly
contain hierarchy level attributes and a few non-hierarchy-level attributes.
By DHE, a star join with hierarchy level is mapped to a multidimensional range
query on the fact table. In Shark, the large-scale data is computed by transformations
and actions on RDDs. Map Sets < , > are preprocessed according to formula
(1), (2) and is got according to Formula (3). As the example shown in Fig.2, the
hierarchy level restriction in dimension table “Customer” is Region=“AFRICA” and
Nation=“ALGERIA”. According to formula (3), the first 3 bits in
represent “Region” and the middle 4 bits represent “Nation”. So we create the mask
code "111111100000000" . "001" and "0001"
are obtained from formula (1) and (2). Thus, we create the filter code
"001000100000000" for this restriction. Bitwise-And operator is then used between
and to filter the tuples to new RDDs.
Thus, the query output can be obtained by the main steps as follows:
─ Step1: Analyze the level attributes in “<Dimension Restriction>”. Obtain the hie-
rarchy level code from Map Set < , > and their corresponding offset
in .
─ Step2: create the mask and the filter key from Step 1.
─ Step3: Sequential Scan the Fact table with and . Create new RDDs
for those filtered tuples in Fact table as the example shown in Fig.2.
─ Step4: If there exists non-hierarchy-level attribute in <Dimension Restriction>,
Group By and Order By clause, join the corresponding Dimension table with the
RDDs created in Step 3 and then generate new RDDs too.
─ Step5: GroupByPreShuffle (part-aggregate the data at map-side) is executed on the
RDD got from Step4 according to in and the other attributes. As a result,
the amount of data processed at reduce-side will decrease.
─ Step6: According to Gb-attributes, value will be distributed to different reducer. Then
merge the values and compute the sum at the reduce side, sort and extract data.
─ Step7: Submit the final results from worker nodes to master node.
<Customer Restriction> Example:

Condition: Region=”AFRICA” Nation=”ALGERIA”
Region offset Mask: 111 0000 00000000

Nation offset Mask: 000 1111 00000000
AFRICA: 001 0000 00000000 Part1
For each tuple in partition
ALGERIA: 000 0001 00000000 {
If (Bcustomer & Bmask == Bfitler)
Part2 Filter the tuple to new
RDDs
}
Bmask: 111 1111 00000000
Bfilter: 001 0001 00000000 Part3
Create the mask Bmask and filter key Bfilter Transformations on RDDs
Fig. 2. Create the filter key for <Customer Restriction> and transform actions on RDDs
For example, the following complex multi-table joins query in SSB:

Select sum(revenue),year,brand
From datetbl join lineorder
on (datetbl.datekey=lineorder.orderdate) join part
on (part.partkey=lineorder.partkey) join supplier
on (supplier.suppkey=lineorder.suppkey)
Where category='MFGR#12' and region='AMERICA'
Group by year,brand
Order by year,brand
It filters data set according to region attribute in dimension table “supplier” and cate-
gory attribute in “part”. The summarized “revenue” is group by and order by year in
dimension table “datetbl” and brand in “part”. As shown in Fig.3, new RDDs reflect
an operator’s transformation on the RDD that resulted from the previous operator’s
transformation. Based on DHE and Shark, We do not need to write intermediate re-
sults to HDFS in a temporary file, and only simply write results to local disk. The
complex multi-table joins and I/O overhead are also reduced by DHE.
RDD RDD
TableScan: Lineorder
HDFS FilterCondi-
tion:category=’MFGR#12’,region=’AMERICA’ SELECT
RDD RDD
Group By Reduce
Final RDD
Attributes={year,brand} Output Extract
(Sort)
Fig. 3. Query plan based on DHE and Shark

SSB [9] is used in the experiment which is based on the TPC-H benchmark to meas-
ure the performance of data warehousing applications. SSB provides both functional
coverage (different common types of Star Schema queries) and selectivity coverage
(varying fractions of the fact table that must be accessed to answer the queries). There
are four major query types in SSB. The number of dimension tables increases from
query type 1 (Q1) to query type 4 (Q4). For example, Q1 is shown in the below code:
Select sum(lo_extendedprice*lo_discount) as revenue
From lineorder,date
Where lo_orderdate=d_datekey
and d_year=[YEAR]
and lo_discount between [DISCOUNT]-1
and [DISCOUNT]+1 and lo_quantity<[QUANTITY]
The above query is meant to quantify the amount of revenue increase that would have
resulted from eliminating certain company-wide discounts in a given percentage
range for products shipped in a given year. This is a “what if” query to find possible
revenue increases.
The experiments utilized IBM high performance computing platform. The platform
has total 279 computing nodes and 3500 CPU cores. We used 1 master node and 3
worker nodes. The available RAM of each worker node is set to 2GB.
We use three data sizes of 5G, 10G and 20G with DHE and pre-load them into
HDFS. The records number in fact table is respectively 30 million, 60 million and 120
million. For each query type, we complete the query with different values several
times (e.g. discount=0.5, discount=0.6) and use the average time as the result.
Q1 to Q4 on Hive-SSB and Shark-SSB are based on the original SSB. Q1 to Q4 on
Shark-DHE are based on our OLAP query algorithm and the DHE star schema which
is generated from the original SSB.
Fig.4, Fig.5 and Fig.6 illustrate that the data analysis performance built on Shark is
much better than Hive. Moreover, Our OLAP algorithm based on DHE and Shark
further improves the query time than the original star schema.
10 50
Hive-SSB Hive-SSB
time(minute)
time(minute)
8 40
Shark-SSB Shark-SSB
6 30
Shark-DHE Shark-DHE
4 20
2 10
0 0
Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4
Fig. 4. Query performance on 5G data Fig. 5. Query performance on 20G data
8
25 Q1
Hive-SSB
6
time(minute)
20 Q2
time(minute)
Shark-SSB
15 4 Q3
Shark-DHE
10 Q4
2
5
0 0
Q1 Q2 Q3 Q4 5G 10G 20G
Fig. 6. Query performance on 10G data Fig. 7. Performance trends on Shark-DHE
50
40 Hive-SSB
time(minute)
30 Shark-SSB
20 Shark-DHE
10
0
Q1 Q2 Q3 Q4
Fig. 8. Q1 to Q4 trends on 20G data
Fig.7 shows that the OLAP query time increases along with the size of the data.
The trend of query time with 20G data is demonstrated in Fig.8. The trend in our me-
thod is quite smooth because most join operations can be removed by using DHE and
map the query to a range query on the fact table. Q1 involves only one dimension and
does not contain group by operations, so the improvement of the query time is not
obvious between Shark-SSB and Shark-DHE.
4 Conclusion
With the era of “big data” and expanding amount of data size, the processing technol-
ogy which computing distributed data set to satisfy the complex OLAP queries
has become a research hot spot. In this paper, the novel OLAP query algorithm is
proposed based on the dimension hierarchical encoding storage strategy with the In-
Memory computing in Shark. This method reduces the complex multi-table joins in
star schema. The results of experiment reveal that our OLAP algorithm based on DHE
and Shark greatly improves the query time.
Acknowledgements. This work was supported by the Natural Science Foundation of

Jiangsu Province under Grant No.BK2012742.
References
1. Chaudhuri, S., Dayal, U.: Data warehousing and OLAP for decision support. ACM Sig-
mod Record 26(2), 507–508 (1997)
2. Apache Hadoop, http://wiki.apache.org/hadoop
3. Dean, J., Ghemawat, S.: MapReduce: simplified data processing on large clusters. Com-
munications of the ACM 51(1), 107–113 (2008)
4. Thusoo, A., Sarma, J., Jain, N.: Hive: a warehousing solution over a map-reduce frame-
work. Proceedings of the VLDB Endowment 2(2), 1626–1629 (2009)
5. Pavlo, A., Paulson, E., Rasin, A.: A comparison of approaches to large-scale data analysis.
In: Proceedings of the 2009 ACM SIGMOD International Conference on Management of
Data, pp. 165–178. ACM (2009)
6. Stonebraker, M., Abadi, D., DeWitt, D.: MapReduce and parallel DBMSs: friends or foes.
Communications of the AC 53(1), 64–71 (2010)
7. Zaharia, M., Chowdhury, M., Das, T.: Resilient distributed datasets: A fault-tolerant ab-
straction for in-memory cluster computing. In: Proceedings of the 9th USENIX conference
on Networked Systems Design and Implementation, p. 2. USENIX Association (2012)
8. Pavlo, A., Paulson, E., Rasin, A.: A comparison of approaches to large-scale data analysis.
In: Proceedings of the 2009 ACM SIGMOD International Conference on Management of
Data, pp. 165–178. ACM (2009)
9. O’Neil, P., O’Neil, E., Chen, X.: The star schema benchmark (SSB). Pat. (2007)
10. Jing-hua, Z., Ai-mei, S., Ai-bo, S.: OLAP Aggregation Based on Dimension-oriented Sto-
rage. In: IEEE 26th International Parallel and Distributed Processing Symposium Work-
shops & PhD Forum (IPDPSW), pp. 1932–1936. IEEE (2012)
11. Apache HBase, http://www.hbase.apache.org
12. Wang, H., Qin, X., Zhang, Y., Wang, S., Wang, Z.: LinearDB: A relational approach to
make data warehouse scale like mapReduce. In: Yu, J.X., Kim, M.H., Unland, R. (eds.)
DASFAA 2011, Part II. LNCS, vol. 6588, pp. 306–320. Springer, Heidelberg (2011)
13. Shi, Y., Meng, X., Zhao, J.: Benchmarking cloud-based data management systems.
In: Proceedings of the Second International Workshop on Cloud Data Management,
pp. 47–54. ACM (2010)
14. Brezany, P., Zhang, Y., Janciak, I.: An Elastic OLAP Cloud Platform. In: IEEE Ninth
International Conference on Dependable, Autonomic and Secure Computing (DASC),
pp. 356–363. IEEE (2011)
15. Fa, H., Sheng, Y., Zhen, X.: A Novel Aggregation Algorithm for Online Analytical
Processing Queries Evaluation Based on Dimension Hierarchical Encoding. J. Computer
Research and Development. 4, 608–614 (2004)
16. Markl, V., Ramsak, F., Bayer, R.: Improving OLAP performance by multidimensional hie-
rarchical clustering. In: International Database Engineering and Applications Symposium,
pp. 165–177. IEEE (1999)
The Enhancement and Application of Collaborative
Filtering in e-Learning System
Bo Song and Jie Gao
College of Software, Shenyang Normal University, Shenyang City,

Liaoning Province, China, 110034
songbo63@aliyun.com, gaojiexy@126.com
Abstract. Collaborative Filtering recommendation algorithm is one of the most

popular approaches for determining recommendations at present and it can be
used to solve Information Overload issue in e-Learning system. However the
Cold Start problem is always one of the most critical issues that affect the
performance of Collaborative Filtering recommender system. In this paper an
enhanced composite recommendation algorithm based on content
recommendation tags extracting and CF is proposed to make the CF
recommender system work more effectively. The final experiment results show
that the new enhanced recommendation algorithm has some advantages on
accuracy compared with several existing solutions to the issue of Cold Start and
make sure that it is a feasible and effective recommendation algorithm.
Keywords: Recommendation Algorithm, Cold Start, Collaborative Filtering,

e-Learning.
1 Introduction
With the development of Internet technology, huge information resource is presented

to us constantly, thus the issue of information overload formed. It becomes more
difficult for users to find out what they need or they are interested in. Facing
information overload issue, recommender system is one of the solutions.
Collaborative Filtering is one of the most popular approaches for determining
recommendations at present and the CF recommendation algorithm is proposed
originally in 1992 by Goldberg, Nichols and Oki [1]. It is an intelligent and
personalized information service system and describes the user's lone-term
information need by user modeling, based on which it can customize the personalized
information with the specific recommendation strategy [2]. The implementation of CF
recommender system must rely on users’ explicit rating, such as Amazon, Drugstore;
these electronic commerce platforms possess a large number of users who have rated
their purchased goods. The CF recommendation algorithm can be divided into two
parts. One is user-based CF and the other one is item-based CF.
The e-Learning system stays on the situation that they may confront with Cold
Start issue while using Collaborative Filtering. When new users register the system or
new items are added into the system, the CF recommender system may recommend
The Enhancement and Application of Collaborative Filtering in e-Learning System 189
no items for the new users or the new items may not be recommended for existing
users because of there is no relevant rating information in the user-item rating matrix
known as user modeling. This is the problem of Cold Start of CF recommendation
algorithm. Aiming at above problems, in this paper an enhanced composite
recommendation algorithm based on content information tags extracting and
Collaborative Filtering will be proposed to solve the new users’ Cold Start problem to
make the recommender system works more effectively.
2 Collaborative Filtering Recommendation Algorithm

Personalized recommendation is an important part in user behavior analysis. In simple
term, it is a process of searching the resource which the user might interest in. At
present, Collaborative Filtering is a relatively mature and popular recommendation
algorithm. In CF, an item is considered as a black box-we don’t look at its content and
user interactions (such as rating, saving, purchasing) with the item are used to
recommend an item of interest to the user. The main idea of Collaborative Filtering is
to exploit information about the past behavior or opinions of an existing user
community for predicting which items the current user will most probably like or be
interested in [3]. The recommender process is simply as follows: first given a user-
item rating dataset and a current user as input and calculate the similarity between
users; second use the similarity matrix made up of the above similarities to find
similar users that had similar preferences to those of the current user, that is called
nearest neighbors; third for every item that the current user has not yet seen, the
prediction ratings will be calculated based on the ratings for the items made by the
similar users; finally sort the prediction ratings and the top-N items will be
recommended for the current user. Such a top-N list should not contain items that the
current user has already rated.
The matrix of given user-item ratings is known as user modeling which is obtained
by collecting and arranging users interactions. If the number of items is “n” and the
number of users is “m” in the system, the user modeling is a rating matrix of “m*n”.
We use U = {User 1 , User 2 ,..., User m } to denote the set of users,
I = { Item 1 , Item 2 ,..., Item m } for the set of items, and R as a m * n matrix of
ratings r( i , j ) , with i ∈ {1 ... m } and j ∈ {1 ... n } . The possible rating values are defined
on a numerical scale from 1 to 5 in this paper and the higher the rating value, the more
strongly the user likes the item.
 
  a ⋅b
sim( a , b ) =   (1)
a *b
To find nearest neighbors, the formulae applied to calculate the similarity between
users or items must be used. There are several similarity calculation methods, such as
the cosine similarity measure, adjusted cosine similarity measure, Pearson correlation
coefficient measure. Cosine similarity is established as the standard metric in item-
based Collaborative Filtering; as it has been shown that it produces the most accurate
190 B. Song and J. Gao
results [3]. The formula (1) shows the cosine similarity measure which calculates the
rating similarity between Item a and Item b in item-based Collaborative Filtering. The
range of the ratio in formula (1) represents similarity which is from 0 to 1. The larger
the ratio is, the higher the similarity between two items is. However, the basic cosine
measure does not take the differences in the average rating behavior of the users into
account [3]. The adjusted cosine measure solves the problem and subtracts the user
average. And the values for the adjusted cosine measure correspondingly range from
−1 (strong negative correlation) to +1(strong positive correlation) as is shown in
formula (2). Here ru ,a is the rating given by user u to the item a , ru ,b is the rating
given by user u to the item b and ru represents the average of user u ratings.
sim(a, b) =
 u∈U
(ru ,a − ru )(ru ,b − ru )
(2)
u∈U
(ru ,a − ru ) 2  u∈U
(ru ,b − ru ) 2
As in the adjusted cosine measure, the Pearson correlation coefficient measure

shows in the formula (3). The symbol ra corresponds to the average rating of
user a and rb corresponds to the average rating of user b and takes the values from −1
to +1. Pearson correlation coefficient is used to calculate the similarity between users
in user-based Collaborative Filtering.
sim(a, b) =
 i∈I
(ra ,i − ra )(rb ,i − rb )
 i∈I
(ra ,i − ra ) 2  i∈I
(rb ,i − rb ) 2 (3)
After the nearest neighbors were found, the recommender engine will work
according to the prediction rating of non-rated items based on the current user. The
prediction rating formula is shown in formula (4).
predictedRating =
 (sim(a, b) * itemrating ) (4)
 sim(a, b)
Where, the prediction rating is the ratio of similar users’ rating weighted sum and
similarity sum. Then add up all similar users’ prediction rating and sort them
according to the values of sum and the top N items will be the recommendation items
for the current user. The larger the prediction rating sum of an item rated by all
similar users is, the more effective recommendation results the current user will get.
Correspondingly the scale of the sum will not be more than “5”.
3 The Enhancement of Collaborative Filtering

The success of Collaborative Filtering relies on the availability of a sufficiently large
set of quality preference ratings provided by users, but it method may suffer from the
new user problem, in which there is no rating record on new users in user modeling
[4]. This is the issue of new users’ Cold Start. Usually there are several solutions to
Cold Start problem such as random recommendation and Mean Value
recommendation. In recent years, combining content information with CF
recommendation proved to be an effective solution to Cold Start problem and has
become a hot research [5].
3.1 Content-Based Tags Extracting Solution to Cold Start

The recommendation process of the new enhance composite recommendation
algorithm can be divided into four steps: firstly content information of new users and
all existing items will be analyzed and the tags that represent their feature will also be
extracted; secondly find the similar matrix between the tags matrix of new users and
the tag matrix of all items with the method of linear algebra, then recommend the
items for the new users, this is the first time recommendation in the new algorithm;
Thirdly the new users can choose the items and rate them, then the interaction can fill
in the user modeling and the new users will acquire the final more accurate
recommendation results.
When new users register the e-Learning system, it is necessary for them to fill in
some individual information, especially education background, preference and forte.
Usually the items in the e-Learning system are added in some text depiction about
them, for instance lesson name, major classification, lesson introduction. We can use
the above information to build tags matrix. Tags may contain a single term or multiple
terms and it is very important to convert terms into numerical value.
The typical steps involves in building tags matrix are tokenization, normalization,
eliminating stop words and quantitative filling. At this stage, we will introduce the use
of term frequency and document frequency to compute the weight associated with
each term. Tokenization is to parse the text to generate terms and sophisticated
analyzers can also extract phrases from the text. The TokenStream class and Analyzer
class of Apache Lucene jar file can help to extract tags from the text [6]. Apache
Lucene is a Java-based open source search engine developed by Doug Cutting [7].
And we can create a Map collection class buildTermFrequencyMap to build a
HashMap to store the terms and their frequencies. It is with all the text information in
the system as input and uses Loops and conditional statements to put the terms and
frequencies into the termFrequencyMap Map. Then getTopNTermFrequencies
method will find the top N terms appeared most frequently in each text information
file by sorting the frequencies.
The second step is normalization to convert the terms to lowercase. To make the
matrix more accurately, we will also use the StandardAnalyzer class of Apache
Lucene to eliminate stop words such as “a, an, the” that appear in the text too often
[7]. The stop words list can use a String type array to preserve. Eliminating stop
words can not only enhance the importance of meaningful words, but also abate noise
apparently at the same time. The Java static code block shows the above methods.
static {
List<String> allStopWords = new ArrayList<String>();
allStopWords.addAll(Arrays.asList(StandardAnalyzer.STOP_WORDS));
allStopWords.addAll(Arrays.asList(ADDITIONAL_STOP_WORDS));
MERGED_STOP_WORDS = allStopWords.toArray(new String[0]); }

public CustomAnalyzer() {
super(CustomAnalyzer.MERGED_STOP_WORDS); }
ADDITIONAL_STOP_WORDS is the String type array that contains stops words. It

is added into the allStopWords ArrayList.
When the first top N terms were picked out and then they can consist of coordinate
space of tags. So every new user Useri can be described by term vectors:
Useri = {U 1 ,U 2 ,...,U n } . Among them, U j express a tag which is a column
vector. For each new user’ registration text information, take 80% term frequency as
the cut-off point, If the term frequency is higher than 80%, we can set the term vector
element value as 1, or else is 0. For each item, we can also use the above method to
build the corresponding tags matrix. So, all the text information of new users and
existing items can be indicated by a 0-or-1 matrix of tags by quantitative filling.
After defining the tags matrix, the tags matrixes of new users and items have the
same dimensions. Then we can use the method of linear algebra to find similarity
matrixes. Methods are as follows: Assume A, B as the matrixes of order n, if there
exist invertible matrices P of order n, such that P −1 * A * P = B was
−1
established, we call matrix A is similar to B, denoted by A~B. P is the inverse
matrix of P. Here also involves matrix multiplication algorithm. The steps for finding
similarity matrix are as follows. Firstly calculate the Eigen values of the matrix:
A − λE = 0 , here E is a unit matrix, from upper left to lower right corner of the
main diagonal, the elements are all 1, outside are 0. The matrix A can be seen as tags
matrix of random new user and matrix B can be tags matrix of existing item. Then for
every Eigen value, calculate the solution space of matrix equation ( A − λE ) X = 0 .
We can exploit the feature vector provided by basic solutions as column vector to
comprise the matrix P. The situation shows that the invertible matrix P exists and
matrix A is similar to B. If the above mentioned matrix equation has no solution
space, there will not be invertible matrix P. On this situation, matrix A is not similar
to B; there is no similarity between the new user and the item. Using this algorithm,
the tags matrix of new user will be compared with every tags matrix of existing item,
if there is tags matrix of item is similar to the tags matrix of the new user, the
corresponding item will be recommended to the new user. The number of matrix B
−1
that can enable the equality P * A * P = B succeeds is certainly more than one in
the system. So the first recommendation list for the new user can form. Similarity
matrix is an equivalence relation and has reflexive, symmetry and transitivity three
properties. The items that tags matrix are similar to the tags matrix of the new user
can recommend to him or her. We call it the first-time recommendation because the
recommendation this time is not enough accurate and we can further make efforts to
let the recommender system work more effectively. Therefore, the items of first time
recommendation will be chosen by the new user. The new user then can rate the items
by ones preference and education background. After this step, the new user’
interactions are added into the User Modeling. After the new user’ interactions are
added the User Modeling, we can use Collaborative Filtering Recommendation
algorithm to obtain more accurate recommendation results. A series of process,
calculating similarity between the new user and existing users, predicting ratings for
the new user, sorting the predicting ratings, will work as usual. The recommendation
results of first-time recommendation process will be as seed candidates to create the
final more accurate recommendation results. Therefore we can name the new solution
to Cold Start problem as two-times recommendation. Using the new content-based
tags extracting composite recommendation algorithm, Collaborative Filtering
recommendation algorithm can provide better service for new users and improve the
performance of recommender system.
3.2 Experimental Results and Analysis
In order to verify the effectiveness of the algorithm, this paper will use the public data
set MoviLens provided by Grouplens working group to test. MoviLens is a research
recommender system developed by the researchers of Grouplens working group in
USA University of Minnesota based Web. It accepts user evaluation of the films and
provides the corresponding movie recommendation list. At present, the system has
more than 43000 users and the user ratings of the items are more than 1600 [8].
In this paper, the ML data set provided by the Movielens will be used, which is
composed of 943 user evaluations of the 10000 items with 1-to-5 ratings. The data set
has a total of 1682 items and each user makes evaluation on the 20 items at least [8].
In this paper we will select 10% of the data set as the experimental data randomly.
Then we will choose 10% users data of the experimental data as new users’
information. Their rating items are more than 100 so that enough information can be
used. In this experiment, the new users’ ratings will be deleted and stored in another
place. Using the three different recommender algorithms of Random
Recommendation, Mean Value Recommendation and the new recommendation
algorithm proposed in this paper, compare the predict ratings with real ratings of the
new users. This experiment process will be carried out 9 times totally. The accuracy
of recommendation algorithm usually uses MAE (Mean Absolute Error) and RMSE
(Root Mean Squared Error) to measure. They express that what degree the new users
will like or dislike the items recommended by recommendation algorithm.
p i ,i − ri , j
(5)
MAE = i∈N
Where pi , j is on behalf of the predicted values. In fact, MAE is the average value
of the sum of differences between predicted values and real ratings values. The lower
the MAE value, the better the recommendation accuracy. Fig.1 shows the MAE curve
of three recommendation algorithms that solve Cold Start problem. The
recommendation results of Random Recommend are random and the MAE curve is a
random fluctuating curve that the value is more than 0.6. As the ratings from new user
are more, the accuracy of Random Recommend is much bad. The MAE of Mean
Value Recommend is almost a fixed value. The MAE of the new recommendation
algorithm is a low value firstly. When ratings from new user are added, the MAE
becomes much lower and the accuracy of new recommendation is best of all.
1.60
1.40
1.20
1.00
Random Recommend
MAE
0.80 Mean Value Recommend

0.60 New Recommender Algorithm
0.40
0.20
0.00
0 10 20 30 40 50 60 70 80 90
Number of Ratings from a New User
Fig. 1. The MAE Curve of Three Solutions to Cold Start
1.50
1.40
1.30
1.20 Random Recommend
RMSE
1.10 Mean Value Recommend

1.00 New Recommender Algorithm
0.90
0.80
0.70
0 10 20 30 40 50 60 70 80 90
Number of Ratings from a New User
Fig. 2. The RMSE Curve of Three Solutions to Cold Start
RMES in formula (6) is another metric to measure accuracy of a recommender

system. MAE is widely used because of its simple calculation, easy to understand.
But MAE also has some limitations, because the low ratings that are difficult to
predict accurately usually make great contribution to the MAE metric. RMSE firstly
square each absolute error, so it will have heavier punishment for the larger absolute
error relatively [9]. In formula (6), N is the total number of experimental items and
E p is the entire experimental data set. The lower the RMSE value, the better the
recommendation accuracy, like MAE. Fig.2 shows the curve of three recommendation
algorithms that solve Cold Start problem. The curve changes of RMSE are almost
unanimously with MAE, just the RMSE value is more precise. The two metrics of
accuracy prove that the new recommendation algorithm in this paper is a feasible and
effective recommendation algorithm to solve the problem of Cold Start.
 (r − pu ,i )
2
u ,i (6)
( u ,i )∈E p
RMSE =
N
4 Conclusions
In this paper an enhanced composite algorithm based on content information and tags
extracting is proposed to solve the issue of Cold Start to make Collaborative Filtering
recommender system works more effectively. The final experiment results show that
the new enhanced recommendation algorithm has some advantages on accuracy
compared with several existing solutions to the problem of Cold Start and make sure
that it is a feasible and effective recommendation algorithm. Therefore, we can adopt
the new recommendation algorithm in e-Learning system to recommend appropriate
lessons for learners. When new learners register e-Learning system, they can also
acquire recommendation lessons which meet their demand. The new enhanced
recommendation algorithm can be applied to the e-Learning system used in small and
medium enterprises and if the scale of e-Learning system is too large, tags extracting
workload will be correspondingly become too volume.
Acknowledgment. This work was supported by the Science and Technology Project
of Education Department of Liaoning Province, China (Research of e-Learning
System of small and medium-size Enterprises Based on SaaS, No.L2013417).
References
1. Goldberg, D., Nichols, D., Oki, B.M.: Using collaborative filtering to weave an
information tapestry. Communications of the ACM 35(12), 145–147 (1992)
2. Lei, R.: The Key Technology Research of Recommender System. East China Normal
University (2012)
3. Dietmar, J., Markus, Z., Alexander, F., Gerhard, F.: Recommender System an
Introduction, pp.13–14, 19. Cambridge University Press (2011)
4. Liu, Q., Gao, Y., Peng, Z.: A novel collaborative filtering algorithm based on social
network. In: Tan, Y., Shi, Y., Ji, Z. (eds.) ICSI 2012, Part II. LNCS, vol. 7332, pp. 164–
174. Springer, Heidelberg (2012)
5. Dongting, S., Tao, H., Fuhai, Z.: Summary for Research on the Cold Start Problem in
Recommender Systems. Computer and Modernization 5, 59–62 (2012)
6. Haralambos, M., Dmitry, B.: Algorithms of the Intelligent Web, pp.100–101. Publishing
House of Electronics Industry (2011)
7. Satnam, A.: Collective Intelligence in Action, pp. 349–350. Manning Publications (2009)
8. Yanhong, G.: Hybrid Recommendation Algorithm of Collaborative Filtering Cold Start
Problem of New Items. Computer Engineering 34(23), 11–13 (2008)
9. Yuxiao, Z.: Summary for Evaluating Indicator of Recommender System. Journal of
University of Electronic Science and Technology of China 41(2), 163–172 (2012)
A Method to Construct a Chinese-Swedish Dictionary
via English Based on Comparable Corpora
Fang Li, Guangda Shi, and Yawei Lv
Intellingence Engineering Lab, Beijing University of Chemical Technology,

Beijing 100029, China
lifang@mail.buct.edu.cn, shiguangdabuct@163.com,
lvyawei7@gmail.com
Abstract. Taking advantage of existing bilingual dictionaries and a third language

can construct a new bilingual dictionary. However, the access of some existing bi-
lingual dictionaries is difficult, which makes it impossible to construct a new bi-
lingual dictionary. For solving the problem, the paper proposes a method that we
can construct the bilingual dictionaries of the source language and the target lan-
guage with a third language by using the corresponding comparable corpora, re-
spectively. The proposed method is applicable to the languages which are lack of
comparable corpora and the available dictionaries, such as Chinese and Swedish.
This paper constructs a Chinese-Swedish bilingual dictionary by using English as
a third language to prove the validity of the proposed method. The result of expe-
riment shows that the proposed method has a good performance in the construc-
tion of the Chinese-Swedish bilingual dictionary.
Keywords: bilingual dictionary, intermediary third language, comparable corpora.
1 Introduction
With the rapid development of natural language processing such as machine transla-
tion and cross-language information retrieval, a bilingual dictionary plays a more and
more important role due to the growing demand for the bilingual dictionary. In the
process of constructing bilingual dictionary, especially for the languages which are
not widely available, the existing bilingual dictionaries via a third language (usually
English) are utilized as the usual method. But there is a problem in this process, which
is difficult or even impossible to get the existing bilingual dictionaries for less com-
mon languages such as Swedish. For the problem, this paper proposes a new method
of constructing a dictionary with the help of the comparable corpora.
The goal of the experiment is to construct Chinese-Swedish bilingual dictionary by
using Chinese-English and English-Swedish comparable corpora to construct Chi-
nese-English and English-Swedish dictionary, eventually we get a Chinese-Swedish
bilingual dictionary.
The remainder of this paper is structured as follows: Section 2 describes the
related work of the bilingual dictionary construction. Section 3 proposes a new
method in detail to solve the problem of the lack of dictionary when constructing
A Method to Construct a Chinese-Swedish Dictionary 197
Chinese-Swedish dictionary by using English as a third language. Section 4 provides

our experiment and analysis. The last section is a summary of this paper and pros-
pects.
2 Related Work of Bilingual Dictionary Construction

Tanaka and Umemura[1] built a Japanese-French bilingual dictionary by using Eng-
lish as third language intermediary. They made use of the existing Japanese-English
bilingual dictionary and the English-French bilingual dictionary to construct Japa-
nese-French bilingual dictionary. At the beginning, they searched the English transla-
tion candidates of Japanese words by looking up the Japanese-English bilingual
dictionary, and found the French translation candidates of the words in English-
French bilingual dictionary. Similarly, the English translation candidates of French
translation candidates were available by looking up the French-English bilingual dic-
tionary. After that, they got the French translation candidates of Japanese words by
comparing the English translation candidates of Japanese words with the English
translation candidates of French words. Bond and other scholars[2] took advantage of
the Japanese-English bilingual dictionary and the English-Malay bilingual dictionary
to construct Japanese-Malay bilingual dictionary. They used a method called semantic
classification to sort final candidate translations.
The construction of another bilingual dictionary using comparable corpora is based
on the following hypothesis: if two words have similar meaning, the contexts of the
two words should also be similar. That’s to say, the words in source language and its
translations in target language have the similar contexts. Fung (1998)[3] extracted the
contexts of the words in the comparable corpora, and then the similarity between two
words is calculated by using co-occurrence vector of the words, and constructed a
bilingual dictionary. Zhang Yongchen and other scholars captured the datas on the
web and built a Chinese-English bilingual dictionary by using Vector Space Model.
Haghighi[4] used the Matching Canonical Correlation Analysis to construct a Chi-
nese-English bilingual dictionary. Rapp[5] considered the sequence and relationship
of words based on Vector Space Model when constructing a English-German bilin-
gual dictionary. Daille and Morin[6] used the variant methods of the Vector Space
Model to construct a French-English bilingual dictionary. About the construction of
bilingual dictionary, a lot of scholars did a plenty of researches, involved a lot of lan-
guages. However, the research about the dictionary construction of Chinese-Swedish
is seldom seen.
3 Chinese-Swedish Dictionary Construction via English
This section detailedly describes how to extract the words of the source language and
the target language to construct a bilingual dictionary when we do not have the dic-
tionaries of the source language and the target language with a third language.
As you know, Chinese-Swedish comparable corpora are different to obtain. Our
method is to construct a Chinese-English dictionary and an English-Swedish dictio-
nary using the corresponding corpus respectively.
198 F. Li, G. Shi, and Y. Lv
3.1 The Construction of Chinese-English Dictionary and English-Swedish

Dictionary
In the process of constructing the Chinese-English dictionary, we make use of a me-
thod called the standard method, the technological process can be seen in Figure 1.
This detailed process of the method is as follows:
1. Preprocessing. Word segmentation and filtering stop words should be done in this
process. As to the preprocessing of English words, stemming and filtering stop
words should be used[7].
2. Extracting the contexts. We select 3 as our window size and extract the contexts of
the words in the range of window size.
3. The construction of the context space vector. We make use of the bag-of-words to
create the context vectors and use the TF-IDF (term frequency-inverse document
frequency) of each word to measure the importance of each word. They are calcu-
lated as follows:
(1)
(2)
log
1
(3)
In the formulas above, wi represents a word. nwi represents the frequency of the
word wi in the documents. N represents the number of the same words in the docu-
ments. K represents the number of the documents. k represents the number of docu-
ments which contain the word wi.
4. Calculating similarity. We select the Cosine Similarity[8] method to calculate the

similarity between Chinese-English word vectors. The highest similarity words in
the target language are selected as the translation candidates for words in the
source language.
∑ (4)
,
∑ ∑
WS and Wt represent a source word and a target word respectively.

Analogously, the process of constructing English-Swedish bilingual dictionary is
similar to the process of constructing Chinese-English bilingual dictionary.
Fig. 1. The method of the Chinese-English bilingual dictionary
3.2 The Construction of Chinese-Swedish Dictionary
The process of constructing Chinese-Swedish can be seen in Figure 2.

We construct Chinese-Swedish bilingual dictionary by extracting words in the
Chinese-English bilingual dictionary and the English-Swedish bilingual dictionary
that we get in section 3.1.
We look up English translations for Chinese words, and then look for Swedish
translations of these English translations. Then, for each Swedish word, we look up
all of its English translations. After that, we count the number of shared English trans-
lations. If they have the common English translations, they are treated as the transla-
tion pairs.
We use a similarity score S for a Chinese word c and a Swedish word s is given in
Equation (5), where E(w) is the set of English translations of w.
2 | | (5)
,
| | | |
Fig. 2. The method of the Chinese-Swedish bilingual dictionary
4 Experiments
This section mainly describes the evaluation criteria of the related work and experi-
mental results and analysis.
The paper uses the Chinese-English comparable corpora that are the XinHua news
about finance in the Gigawords and the English-Swedish comparable corpora that are
derived from Wikisource.
4.1 Evaluation Criterion

We use the precision and the Mean Reciprocal Rank (MRR) as the evaluation criteria.
In the process of the bilingual word of construction, the precision of the evaluation
criteria is often used and it is the average accuracy of the top n translation candidates.
The Mean Reciprocal Rank (MRR) is that the mean reciprocal of the rank of the
correct translation. It is a measure of the ranking of the correct translation in the
translation candidates.
The Mean Reciprocal Rank (MRR) is a more lenient evaluation criteria than
the precision. The precision and the Mean Reciprocal Rank(MRR) are calculated as
follows:
(6)
1 1
(7)
Counttop5 represents the total numbers of the top 5 translation candidates, ranki
represents the ranking of the correct translation, N represents the numbers of the word
pairs. Different from the precision, the Mean Reciprocal Rank (MRR) is not consider-
ing the number n of the word translation candidates; therefore, it can better measure
the performance of the bilingual dictionary construction.

Table 1 shows how the link is realized and the similarity scores in section 3. The simi-
larity score shows how many English words are shared by the two dictionaries. The
higher the score, the higher the probability of successful linking. As Table 1 shows,
we can see that, if the number of shared English translated words is more, then we get
a higher possibility of accurate matching of Chinese and Swedish. However, the accu-
racy reduces when the number of the shared English translated words decreases.
Sometimes we have to inappropriately sort out the matched pairs such as 首都
(kapital). The reason is that a lot of intermediate words are polysemous such as capi-
tal.
Table 1. Example of linking through English translations
Score ChineseEnglish SwedishEnglish Chinese-Swedish

2 支票(check, cheque) Check(check, cheque) 支票(check) Yes
2 现金(cash, money) Kassa(cash, money) 现金(kassa) Yes
1 资本(capital) Kapital(capital) 资本(kapital) Yes
1 费用(cost, expense) Kostnad (cost) 费用(kostnad) Yes
1 首都(capital) Kapital (capital) 首都(kapital) No
Table 2 shows the experiment results of the method to solve the problem of the
lack of dictionary when constructing Chinese-Swedish dictionary by using English as
a third-party language. We evaluate the precision of the top 5 translation candidates.
As Table 2 shows, we totally extract 17368 translation pairs from the Chinese-
English comparable corpora, and the precision is 54.38%, the Mean Reciprocal Rank
(MRR) is 58.13%, moreover, there are 94 translation pairs that are not quite accurate,
but they are accepted. We get 14273 translation pairs in total, and the precision is
51.26%, the Mean Reciprocal Rank (MRR) is 56.34%, moreover, there are 137 trans-
lation pairs that can be accepted. Finally, we have a Chinese-Swedish dictionary that
contain 7369 translation pairs, its precision is 46.17%, the Mean Reciprocal Rank
(MRR) is 58.26% and there are 63 translation pairs accepted.
Table 2. The Experiment Results
Translated Precision MRR Accepted

Chinese-English 17368 54.38% 58.13% 94
English-Swedish 14273 51.26% 56.34% 137
Chinese-Swedish 7369 46.17% 51.26% 63
As a conclusion, we proposes a new method in detail to solve the problem of the lack
of dictionary when constructing Chinese-Swedish dictionary by using English as a
third language, and this method is applicable to other language resources. We make
use of the public available resources Gigawords and Wikisource for the construction
of a new language pair. The process is divided into three parts, the first part is con-
structing Chinese-English bilingual dictionary by using the Chinese-English compa-
rable corpora, and the second part is constructing English-Swedish dictionary by
using the English-Swedish comparable corpora, finally we can get the Chinese -
Swedish dictionary based on the two above dictionary. The accuracy obtained is
，
46.17% which proves the effectiveness of the proposed method when lack of the
dictionary between the source language and the third language or between the target
language and third language.
The current study about the Chinese-Swedish bilingual dictionary construction is
relatively few, during the construction of Chinese-Swedish bilingual dictionary
process, the proposed method proves its feasibility and effectiveness, and this method
is also applicable to other language resources, which has an important contribution to
the study of related work.
However, there is still a lot of work to do. As future work, firstly, we plan to com-
pare different definitions of context in the process of constructing the context vector,
such sentence-based context and syntax-based context. Secondly, we plan to conduct
experiments on other comparable corpora such as Wikipedia and different language
pairs. Finally, we plan to extend our method to deal with some compound and rare
words.
Acknowledgments. This research is supported by the Chinese-English comparable

corpora of Gigawords and the English-Swedish comparable corpora of Wikisource.
We would also like to thank everyone who offers help.
References
1. Tanaka, K., Umemura, K.: Construction of a bilingual dictionary intermediated by a third
language. In: 15th COLING International Conference on Computational Linguistics,
pp. 297–303. ACL, Stroudsburg (1994)
2. Bond, F., Sulong, R., Yamazaki, T., Ogura, K.: Design and construction of a machine-
tractable Japanese-Malay dictionary. In: 8th MT Summit, pp. 53–58. Santiago de Compos-
tela (2001)
3. Fung, P.: Compiling bilingual lexicon entries from a non-parallel English-Chinese corpus.
In: 3rd Annual Workshop on Very Large Corpora, Boston, pp. 173–183 (1995)
4. Haghighi, A., Liang, P., Berg-Kirkpatrick, T.: Learning Bilingual Lexicons from Monolin-
gual Corpora. In: the Association for Computational Linguistics on Computational Linguis-
tics, pp. 771–779. ACL, Stroudsburg (2008)
5. Rapp, R.: Automatic identification of word translations from unrelated English and German
corpora. In: 37th Annual Meeting of the Association for Computational Linguistics on
Computational Linguistics. Association for Computational Linguistics, pp. 519–526. ACL,
Stroudsburg (1999)
6. Daille, B., Morin, E.: French-English Terminology Extraction from Comparable Corpora.
In: Dale, R., Wong, K.-F., Su, J., Kwong, O.Y. (eds.) IJCNLP 2005. LNCS (LNAI),
7. Fung, P.: A Statistical View on Bilingual Lexicon Extraction: From Parallel Corpora to
Non-parallel Corpora. In: Farwell, D., Gerber, L., Hovy, E. (eds.) AMTA 1998. LNCS
(LNAI), vol. 1529, pp. 1–17. Springer, Heidelberg (1998)
8. Kaji, H., Erdenebat, D.: Automatic Construction of a Japanese-Chinese Dictionary via Eng-
lish. In: the International Conference on Language Resources and Evaluation, Marrakech,
pp. 699–706 (2008)
The Design and Implementation of the Random HTML
Tags and Attributes-Based XSS Defence System
Heng Lin, Yiwen Yan, Hongfei Cai, and Wei Zhang
School of Software, Beijing Institute of Technology, Beijing 100081 China

anonymous.joker.lin@gmail.com,
278567461@qq.com
Abstract. At present, cross site scripting (XSS) is still one of the biggest threat
for Internet security. But the defensive approach is still feature matching
mostly; that is, to check for a matching and filter in all information submitted.
However, filtering technology has many disadvantages as heavy-workload,
complex-operation, high-risk and so on. For this reason, our system use the
randomization techniques of HTML tags and attributes innovatively, based on
the prefix of HTML tags and attributes, to determine the tags and attributes are
Web designers expect to generate or other users insert in, and then we follow
the results to carry out different policies, only tags and attributes that Web
designers expected to generate can be rendered and implemented. By this way,
we can defend against XSS attacks completely. The test results show that the
system is able to solve a variety of problems in filtering technology. It uses
simple and convenient operation and safe and secure effect to free developers
from heavy filtering work. System has a good compatibility and portability
across platforms, it also can connect with all web-based applications
seamlessly. In all, system defend against XSS better and meet the need of
today's XSS attacks defence.
Keywords: random, tag prefix, cross-site scripting, defence system.
1 Project Background
Cross Site Script Execution (usually abbreviated as XSS) is a kind of attack that
attacker using the lack of filtering on the user's input, manufacturing malicious input
which can affect other users, so as to achieve the purpose of attack, stealing user data,
using user identity to make some things or using virus to attack the visitors.
In recent years, XSS vulnerability has been ranked in the top three Web security
.According to the data published by OWASP(Open Web Application Security
Project) in 2010 and 2013,in the top ten Web security vulnerabilities, XSS is both
ranked the top three[1].
In November 2012, Anheng Institute for information security found XSS
vulnerability exist in the Web applications of Baidu, Tencent, Sina and other Internet
companies[2]; In April 2013,a fact was exposed by WooYun that Taobao had Cross
Site Vulnerabilities [3];In June 2013, WooYun reported again that the main Web site
The Design and Implementation of the Random HTML Tags 205
of Tencent and Baidu had XSS vulnerability [4-5]. XSS vulnerability has been a
serious threat to the information security of majority of Internet users.
To prevent from XSS attack, someone present model-based testing evolved as one
of the methodologies which offer several theoretical and practical approaches in
testing the system under test (SUT) that combine several input generation strategies
like mutation testing, using of concrete and symbolic execution etc. by putting the
emphasis on specification of the model of an application [6]; someone detecting
Cross-Site Scripting Vulnerability using Concolic Testing[7]. Existing cross-site
scripting attack protection mainly contains four aspects: the input validation, escape,
filtering, and character set specifies, there are Http-Only, input check, output check,
three kinds of defense in total. But the reality is that each method has its own
defensive shortcomings, it’s also very hard to handle the text properly.
So we develop XSS-Defender, a system that allows the server to identify untrusted
content and reliably convey this information to the client, and that allows the client to
enforce a security policy on the untrusted content. Analogously, XSS-Defender
randomizes HTML tags and attributes to identify and defeat injected malicious web
content. These randomized tags and attributes serve two purposes. First, client proxy
distinguish a tag or attribute is legal or not through whether the tag or attribute has a
random prefix client proxy generated so that identify untrusted content. Second, they
prevent the untrusted content from distorting the document tree.
We make the following contributions:

• We innovatively develop a system randomizing HTML tags and attributes to
defend against XSS attacks.
• We lead a simple way for current web application to defend against popular XSS
attack efficiently.
2 Systematic Realization
2.1 Processing of Client Proxy for the User Data

When a user opens a client proxy, the client proxy automatically connect to the server
proxy, to get the public key that server proxy stored in the cache, after that, the public
key will use as a basis of the RSA encryption algorithm to transfer randomized prefix.
When a user of Web site requests a Web page or submit data to the server, first of all,
the user send data to the client proxy that is deployed on the user's computer. Via the
client proxy, system forwards the request to the server proxy that is deployed on the
server. Meanwhile, the client proxy use Python random number generator to generate
a random string as the identification of random prefix, after a non-equivalent
encryption, it will be sent to server proxy with user's requests together. The random
string is effective only at this time of page request and getting response to server, it
will be destroyed at once after using it one time, in order to ensure the security of the
system.
206 H. Lin et al.
2.2 Processing of Server Proxy for the User Data

As shown in Fig. 1, when the server proxy starts, it generates a pair of public and
private keys used for RSA encryption. When the server proxy receives requests or
data, it detects the requested type first, if the request is a request of RSA public key,
proxy will send the generated public key to client proxy. If the request is other
networks' interactive request, proxy will judge whether it has a random prefix the
client proxy generates or not. If not, the request or data will be discarded; if so, it will
forward requests or submitted data to the Web server, then the one-time random
prefix the client proxy generates will be stored in the cache.
Fig. 1. Processing flow chart of server proxy for the user data
2.3 Processing of Web Server for the User Data
In the system, the Web server pre-stores a confidential 16-bit string which can be
modified at any time as a tag prefix to prove that the tags or attributes come from the
server. In a static page stored on the server, each HTML tag and attribute will have
the confidential prefix, and in the code of Web page dynamically generated by the
server, each HTML tags and attributes generated within the plan of Web site
developer (legal) will have the confidential prefix, tags and attributes not in the
developers' plan (illegal) does not have this prefix. Confidential prefix only transfer
across the local network connection between the server proxy and server which are
deployed on the same machine, in addition to the site maintainers, others all could not
get the confidential prefix.
When receiving a request, the Web server dynamically generates or read the user
requests' website from the cache. In this page, all legal tags and attributes have the
confidential prefixes, meanwhile illegal tags and attribute does not have the
confidential prefixes. In this way, the server generates a page that is able to
distinguish whether each tab comes from the server or not. Then, the server sends this
page to the server proxy through a local connection inside the computer.
2.4 Processing of Server Proxy for the Page Sent to Users
After receiving the Web server's response to the request, server proxy will replace all
confidential prefixes in the code of Web page with confidential prefix the client proxy
generates. Then, the server proxy sends modified code of Web page to the client
proxy. As the prefix for the Web site's users cannot be revealed to other users, so that
the replaced page still can distinguish this tag is legal or not by the tag or attribute
whether the client has proxy generated random prefix. This way, the client proxy can
classify and deal with them according to the different tag prefix.
2.5 Processing of Client Proxy for the Page Sent to Users
When the client proxy receives the treated page sent by server proxy, first of all, it
extracts all tag and attribute information in the code of Web pages. Then, it can
classify according to the different tag and attribute prefix. (Process as Fig. 2).
Fig. 2. Processing flow chart of client proxy for each tag in the code of Web pages
208 H. Lin et al.
Client proxy distinguish a tag or attribute is legal or not through whether the tag or
attribute has a random prefix client proxy generated. According to the result of
judgments, client proxy deletes the illegal tags; but deletes the random prefix of legal
tags, transforms tags into a format that can be rendered and implemented by browsers.
After the sorting, the client proxy will send the last generated and transformed page to
the user's browser. In this way, legal tags and attributes can be rendered and
implemented on the user's browser, while illegal tags and attributes cannot be
rendered and implemented on the user's browser. To avoid the browser executing
malicious code submitted by the attackers efficiently, and thus system can prevent
XSS occurred.
3 The Performance of Testing and Experiment
Through a large number of comprehensive testing of the system, the testing is able to
detect the operations whether efficient or stable. To achieve protection against XSS
attacks, we get the performance of the each module and the overall operational status
of the system.
Test content is divided into system functional testing and system performance
testing in two parts.
3.1 System Functional Testing
System Resources
Examine the CPU and memory usage. Turn on the system, using the Windows Task
Manager to see the CPU and memory usage. See the memory usage of server and the
clients which run XSS-Defender and detect the whether exist the presence of a
memory leak.
Response Speed Test
Statistics the user's access speed, page load time when using XSS-Defender system
and using XSS-Defender system. Then we can conclude the comparisons of running
the system which affect the ability of the server's response and effect.
Server Performance Impact
Specific test operation to record the XSS-Defender server for each user request
response time is calculated for each user request processing response speed; recording
XSS-Defender client browser page request to the server's response time is calculated
browser page request response speed.
3.2 The Results of System Testing
Functional Test for XSS Protection

We designed four different methods to attack by XSS.
Table 1. Four pieces of script code
We established a blog site. Then we inject four script codes into the blog message
board.
Based on the experimental results, the system makes the harmful input invalid, so
it defends against the several of XSS attacks efficiently.
Function Test that Submission does not Affect the Contents of the User's Normal
Display
Inject Code:
<script>alert(“XSS”);</script>
The browser of client shows:
Fig. 3. The result of injecting XSS code
The injected code is displayed as text and not executed properly in the site content:
i think <script> is the best label
The browser of client shows:
Fig. 4. Submit a normal script

210 H. Lin et al.
Users fill out a script tag containing the ‘script’ contents and the browser is
normally displayed.
Based on the above experimental results, the system can tell the difference between
legal and illegal entry.
The Function Test of Adding the Prefix on the Service Side of the Page File
Source page on the server part:
Fig. 5. Source code
The client sends a request to obtain the corresponding part of the html file:
Fig. 6. The code received
All tags add prefix ‘sad 32’.
Table 2. The function test of adding the prefix on the service side of the page file
Test of Log Function

Try to request the server, the log on the server side shows as follows:
Fig. 7. Test of log function
Logs completely records every request and its generated random prefix.
Function Test of Cross-Platform

The use of Python development in Linux, Mac OS X, Windows under all operating
normally, the client implementation for the client agent, it can be supported by all
browsers.
Table 3. Function test of adding prefix on the server side
Performance Test on the Client Side

When the client program is not running, the s peed of opening a web page:
Fig. 8. The speed of opening a web page
When client program is not running.

After running the client program ：
Fig. 9. The speed of opening a web page
When client program is running

The speed is almost unchanged.
When running the client, open the Task Manager, we can see that:
Fig. 10. The resources occupancy after running the client system
Based on the above experimental results, the system occupy very small CPU and
memory.
References
1. OWASP: OWASP Top- 2013 10 rcl The Ten Most Critical Web Application Security Risks
(2013)
2. eNet, http://www.enet.com.cn/article/2012/
1112/A20121112190987.shtml
3. WooYun, http://www.wooyun.org/bugs/wooyun-2010-022080
6. Top 25 most dangerous software errors, http://cwe.mitre.org/top25/
.CWE/SANS
7. Bozic, J., Wotawa, F.: XSS Pattern for Attack Modeling in Testing. In: 8th International
Workshop on Automation of Software Test (AST), pp. 71–74. IEEE (2013)
DWT and GA-PSO Based Novel Watermarking
for Videos Using Audio Watermark
Puja Agrawal1 and Aleefia Khurshid2

1
Department of Electronics and Communication Engineering,
Ramdeobaba College of Engineering and Mangement, Nagpur, India
2
Department of Electronics Engineering,
Ramdeobaba College of Engineering and Mangement, Nagpur, India
{agrawalps,khurshidaa}@rknec.edu
Abstract. This paper presents a digital video watermarking scheme that can
embed invisible and robust watermark information into the video streams of
MPEG-1, MPEG-2, H.264/AVC, MPEG-4 standards. Watermark embedding
process is in Discrete Wavelet Domain. Trade off between transparency and
robustness is considered as optimization problem and is solved by Genetic
Algorithm - Particle Swarm Optimization (GA-PSO) based hybrid optimization
technique. An audio signal is converted into 9 bit planes by using bit plane
slicing and then embedded into the frames of video signals as watermark. The
performance evaluation results based on the Peak Signal to Noise Ratio (PSNR)
and Normalized Correlation (NC) confirm that the proposed video processing
method shows reliable improvements for various sequences compared to
existing ones for geometrical attacks like rotation and cropping.
Keywords: Watermarking, Robustness, Transparency, GA-PSO, Bit Plane

Slicing.
1 Introduction
Protection of multimedia data has become one of the major challenges due to the
rapid growth of unauthorized access and copy of digital media objects like images,
audio and video. Digital Watermarking is a process where some valuable
information is embedded into the host media like images, video and audio etc. The
secret message embedded as watermark can be almost anything, for example: a serial
number, plain text, image, random signal, an organization’s trademark, or a copyright
message for copy control and authentication. Potential applications of digital
watermarking includes, copy control, transaction tracking, authentication, and legacy
system.
In general, digital watermarking involves two major operations: (i) Watermark
embedding, and (ii) Watermark extraction. The two most important properties viz.
robustness and transparency are required for preserving the security of videos from
unauthorized access. The ability to detect the watermark content after application of
common signal processing distortions like filtering, lossy compression, color
DWT and GA-PSO Based Novel Watermarking for Videos Using Audio Watermark 213
correction, noise, contrast distortions, geometric distortions is known as robustness.

Imperceptibility/Transparency means that the presence of watermark is not noticed by
the human eyes. Watermarking techniques can be classified according to the nature of
host data (text, image, audio or video), or according to the working (spatial or
frequency) domain.
1.1 Literature Review

Most of the proposed video watermarking scheme based on the techniques of the
image watermarking and applied to raw video or the compressed video. As some
issues in video watermarking are not present in image watermarking, such as video
object and redundancy of the large amount video data, researchers have made use of
those characteristics to develop different schemes [1].
Hui-Yu Huang et.al. have presented an approach consists of a pseudo-3-D DCT.
The watermark message represents an index for selection of a particular quantizer
from a set of possible quantizers. The selected quantizer is applied to the host data to
encode the watermark message [1]. This is invisible, robust for Raw Videos, but
complex process for implementation.
Jing Zhang et.al have described that a grayscale watermark pattern is first modified
to accommodate the H.264/AVC computational constraints, and then embedded into
video data in the compressed domain. With the proposed method, the video
watermarking scheme can achieve high robustness and good visual quality without
increasing the overall bit-rate [2].
Gwenael Doerr et.al have given the in depth overview of video watermarking and
have pointed out that video watermarking is not just a simple extension of still image
watermarking[3].
Ersin ELBASI [4]¸ has proposed a novel video watermarking system based on the
Hidden Markov Model (HMM). This novel watermarking scheme splits the video
sequences into a Group of Pictures (GOP) with HMM. Portions of the binary
watermark are embedded into each GOP with a wavelet domain watermarking
algorithm.
Sanjoy Deb Roy et.al. have presented a hardware implementation of a digital
watermarking system that can insert invisible, semi fragile watermark information
into compressed video streams in real time [5].
Sourav Bhattacharya et.al. have presented a survey on different video
watermarking techniques and comparative analysis with reference to H.264/AVC [6].
Salva A. K. Mostafa et. al. have presented a video watermarking scheme based on
principal component analysis and wavelet transform [7].
Chuen-Ching Wang et. al. have presented a simple but effective digital
watermarking scheme utilizing a context adaptive variable length coding method for
wireless communication systems [8].
In this paper, we have presented an efficient video watermarking technique using
discrete wavelet transform and GA-PSO based hybrid optimization to protect the
copyright of digital videos. In the watermark, first of all the input video is segmented
into shots. The frames of all the shots are decomposed into four sub-bands as HH,
214 P. Agrawal and A. Khurshid
HL, LH and LL. Watermark is embedded into the high middle frequency HL and LH
sub-bands, where acceptable performance of imperceptibility and robustness could be
achieved. An audio signal is chosen as watermark which is unusual. Before
embedding into the sub-bands, the watermark audio signature is processed into a 9 bit
plane slices. Then it is embedded into HL sub-band, and LH sub-band. Here, every
audio bit is embedded into the chosen sub-bands with the aid of our proposed
embedding process. Subsequently, the watermark audio bits are extracted with the
help of our proposed extraction process. Powerful GA-PSO optimization guarantees
the performance. Since, the watermarking is performed in the wavelet domain; the
attained watermark image is of good quality. The efficiency of our proposed
watermarking technique is proved by good PSNR and NC values obtained for the
watermarked videos in the experimental results.
2 Watermark Embedding and Extraction
This process is the most important in this scheme where all the different parts of
watermarks are embedded into different scenes of the video. Video Preprocessing and
Embedding is done by changing position of some DWT coefficients with the
following condition:
2.1 Selection of Best Embeddable Locations
Human visual system has a very strong error correction mechanism. An image contains
lot of redundancies. Small changes made to an image remain undetected by the human
eyes. On the other hand, it has been observed that if an effort is made to increase the
invisibility of a watermark, then robustness of the scheme suffers and vice a versa. A
compromise therefore has to be made in order to get an optimum system.
Wavelet based transforms gained popularity recently since the property of multi-
resolution analysis that it provides [12]. The higher level sub bands are more significant
than the lower level sub bands. They contain most of the energy coefficients, so
embedding in higher level sub bands is providing more robustness. On the other hand
lower level sub bands have minor energy coefficients so watermark in these sub bands
are defenseless to attacks. The sub band LL is not suitable for embedding a watermark
since it is a low frequency band that contains important information about an image and
easily causes image distortions. Embedding a watermark in the diagonal sub band HH is
also not suitable since the sub band can easily be eliminated, for example by lossy
compression as it has minor energy coefficient. So the middle frequency sub bands LH
and HL are the best choice for embedding [12].
There has been a considerable amount of research proposals on the applications of
DWT in digital image and video watermarking systems by virtue of its excellent and
exceptional properties mentioned above, but the scope of optimization in this area is
tremendously less. An optimized DWT for digital image watermarking is capable of
producing perceptual transparency and robustness among the watermarked and the
extracted images [13].
Extending the above concept to videos with suitable modifications, we have

proposed the DWT and GA-PSO hybrid optimization based scheme for video
watermarking.
2.2 Embedding
Input: Original video sequence: Vo [a,b] , watermark audio Aw [a,b]
Output: Watermarked video V w [a,b]
Initially, the shot segmentation technique is applied to original input video sequence
Vo [a,b] is segmented into number of non-overlapping shots D[a,b] . For embedding
purpose, we identify number of frames E[a,b] in all the segmented shots D[a,b]. Then
watermark audio signal Aw[a,b] is converted into 9-bit plane W[a,b], by the use of bit
plane slicing. The video frames have R, G and B components. The blue channel is
selected for embedding because this channel is more resistant to changes compared to
red and green channels and the human eye is less sensitive to the blue channel, a
perceptually invisible watermark embedded in the blue channel can contain more
energy than a perceptually invisible watermark embedded in the luminance channel of
a color image [12]. The blue components EB [a,b] of all the separated frames are
extracted. Then each bit of 9-bit plane sliced audio watermark W [a,b] is applied into
the blue components of each frame; Discrete Wavelet Transform is applied to blue
component EB[a,b]. Discrete Wavelet Transform converts each frame into four sub-
bands such as HH, HL, LH and LL to attain the transformed T [a,b] frames. Then we
select the middle frequency sub-bands (HL, LH) from the transformed frames to
embed the watermark audio Aw[a,b] into the appropriate sub-bands. In order to choose
the embedding locations in the sub-bands, we find the similarity matrix for the video
signal. The similarity matrix for HL sub-band is denoted by Up(x,y) and the LH sub-
band similarity matrix is denoted by lower part Lp(x,y).Then we calculate the mean
value TLH(m) and the maximum value TLH(M) of the chosen embedding part TLH.
Watermark bits are embedded in the video frames according to following
conditions:
Condition 1:For embedding the Watermark Pixel 1

If TLH(a) > 1 then Lp(x,y) << [TLH(a)]
else Lp(x,y)<< TLH (a) + TLH(M)
end if
Condition 2: For embedding the watermark Pixel 0

If TLH(a)>0 then
Lp(x,y) << Abs[TLH(a)]
else
Lp(x,y) << TLH(a) –TLH(M)
end if
Likewise the watermark bits can also be embedded into the HL band. Then the
modified sub-bands are mapped into its original position and inverse wavelet
transform is applied to attain the watermarked video sequence Vw [a,b].
2.3 Extraction Process
Input: Watermarked video sequence V w[a, b]

& the size of the audio watermark Aw[a, b].
Output: Recovered watermark audio Arw[a’, b’]
The watermarked video sequence Vw[a, b] is segmented into number of non-

overlapping shots D'[a, b] by the use of shot segmentation technique. Then we
identify number of frames E'[a, b] in each segmented shots. Blue components of the
partitioned frames are extracted. Afterwards Discrete Wavelet Transform is applied to
the each partitioned frame EB'[a, b]. The DWT is splits each frame into four sub-
bands such as HH, HL, LH and LL and then to attain the transformed T’[a, b] frames.
The middle frequency LH and HL sub-bands are selected from the transformed
frames. The watermark audio bits are extracted from LH and HL sub-bands. If the
embedded bit value is greater than the mean pixel value, then the extracted pixel value
is one. If it is lesser, then the extracted pixel is zero.
Matrix with the size of the audio watermark is prepared and the extracted bits are
placed to attain the watermark audio. The extracted audio watermark Arw [a',b'] is
obtained by the use of reverse process of vector finding operation.
3 GA-PSO Based Hybrid Optimization
In order to achieve both imperceptibility and robustness of the watermarked media,

we use the Genetic algorithm (GA) and Particle swarm Optimization (PSO) based
hybrid optimization. GA is applied for generating the chromosome and PSO for
selecting the optimal location for embedding the watermark media into host media.
GA-PSO based hybrid optimization techniques are applied in embedding as well as
extraction process.
The function of the randomly generated set of genes is the generation of
chromosomes. Population size plays an important role in presenting the solution to the
problem at hand. The beginning population set up is done by producing a population
set P that comprises of set of chromosome vectors having half size of the HL or LH
sub-band. Subsequently, we initialized it with “1” according to the size of the
watermark in that vector in a random manner, and the remaining places are filled
down with “0” values. Then, the beginning set of chromosomes is brought forth at
random with minimum number.
The watermark embedding process is iterated till the optimal locations are obtained
for each chromosome in the population set. Embedding and extraction process is
carried out using these procedures which were defined in the section 2. Fitness
computation formula is depicted below,
Fitness = PSNR + NC . (1)
PSNR: The Peak-Signal-To-Noise Ratio (PSNR) is used to measure deviation of the

watermarked and attacked frames from the original video frames.
NC: The normalized coefficient (NC) gives a measure of the robustness of
watermarking and its peak value is 1. For calculating the PSNR and NC we have used
standard formula as mentioned in [1].
Selection of optimized chromosomes is done based on the values for fitness. Select
the optimized chromosome = Np / 2.Where, Np is Number of Parent chromosomes.
Remaining chromosomes enter the next iteration in the search of finding the optimal
solutions according to fitness function.
Based on the selected optimal chromosomes, we have the value of optimal
solution. This solution is fed to the crossover operation. These set of fitness value
corresponding to chromosomes provide the new offspring by the use of crossover
operation. Every two individuals are chosen from the better set of chromosome to
produce two new offspring by single crossover point.
In mutation operation, the output of crossover operation is used as input. The
process of this function is to modify one gene value that is randomly selected and then
that chromosome is fed to the fitness computation operation. Here mutation operation
is replaced with a velocity computation.
The velocity computation operation is the part of the Particle swarm optimization.
During each iteration, each particle accelerates in the direction of its own personal
best solution found so far as well as in the direction of the global best position
discovered so far by any of the particle in the swarm. This means that if a particle
discovers a promising new solution, all the other particles will move closer to it,
exploring the region more thoroughly in the process. The velocity computation is
done with the standard formula as described in [15].
Newly obtained set of chromosomes velocity computation operation can be
evaluated for best fitness using fitness function. If the optimal solution for embedding
the watermark media into original media is obtained then and then this process will be
terminated, otherwise that solution will move to fitness computation operation again
and the selection and velocity computation operators are performed iteratively. The
PSO process will be iteratively performed until the desired termination is satisfied.
We used many different videos of varying standards, framerate, framesize, payload

for experimentation. Which includes Akiyo, Coastguard, Claire, Carphone, Shutttle,
container, football and silent. Apart from these we used bradman.mpg,
barryrichards.mpg, and chrisold.mpg and many others.
The original 30th frame and its corresponding audio watermark are shown in Fig. 1.
Watermarked frame appears visually identical to the original. The performance of
algorithm can be measured in terms of its imperceptibility and robustness against the
possible attacks. Watermarked frame is subjected to a variety of attacks such as Salt

and Pepper Noise, Median Filtering, Gaussian Noise, and geometric attacks etc. In
case of geometric attacks, the scheme is tested against 900, 1800, 2700 frame rotation,
and 25 % frame cropping. To evaluate the performance of any watermarking system,
Peak Signal to Noise Ratio (PSNR) is used as a general measure of the visual quality.
And the NC values are used as a measure of robustness. The NC values for extracted
watermark for test videos by our proposed system are greater than .9. The most
important aspect is the consistency of the results for different attacks. For robustness
of compressions, our proposed system can effectively resist the MPEG-1, MPEG-2,
MPEG-4 and H.264 compressions. Proposed system causes very slight distortion and
simultaneously provides high visual quality. The embedded watermark is an audio
signal which is unusual. An audio signal Example.wave, 529kb size is used and
converted into suitable dimensions using wavread and signalslices functions in
Matlab and transformed into a suitable watermark. Human Auditory system is more
sensitive than the human visual system. Any modification in the extracted audio
watermark will be noticed more easily as compared image or other watermarks.
Evidently audio watermark bits are spread over LH and HL sub bands and it would be
difficult to extract and reconstruct proper watermark and achieve high NC values. Our
proposed system is strongly resistant to geometrical attacks like rotation and cropping
in comparison to [1], [2] and [7].
Fig. 1. Experimental Results for Akiyo, Carphone of frame size 720X 480, 80 frames
5 Conclusion
A DWT and GA-PSO based novel video watermarking technique has been proposed
using an audio signal as watermark. The performance of our purposed watermarking
scheme is evaluated with common image processing attacks such as salt and pepper
noises, rotation, cropping, Experimental results demonstrate this watermarking
technique is robust against various attacks including the geometrical attacks. This
proposed method is an extension to (HWT- Haar Wavelet Transform) HWT-GA-PSO
based Image watermarking method [10].
References
1. Hui-Yu, H., Cheng-Han, Y., Wen-Hsing, H.: A Video Watermarking Technique Based on
Pseudo 3-D DCT and quantization Index Modulation. IEEE Transactions on Information
Forensics and Security 5(4), pp. 625–637 (2010)
2. Jing, Z., Anthony, T., Ho, S., Gang, Q., Pina, M.: Robust Video Watermarking of
H.264/AVC. IEEE Transactions on Circuits and Systems-II: Express Briefs 54(2), 205–
209 (2007)
3. Gwenael, D., Jean-Luc, D.: Guide Tour of Video Watermarking Signal Processing: Image
Communication Elsevier. Signal Processing Image Communication 18, 263–282 (2003)
4. Elbasi, E.: Robust Multimedia Watermarking: Hidden Markov Model Approach for Video
Sequences. Turk J. Elec. Eng. & Comp. Sci. 18(2), 159–170 (2010), doi:10.3906/elk-0906-
85
5. Sonjoy, D.R., Xin, L., Yonatan, S., Alexander, F., Orly, Y.-P.: Hardware Implementation
of a Digital Watermarking System for Video Authentication. IEEE Transactions on
Circuits And Systems For Video Technology 23(2), 289–301 (2013)
6. Saurav, B., Chattopadhyay, T., Arpan, P.: A Survey on Different Video Watermarking
Techniques and Comparative Analysis with Reference to H. 264/AVC. IEEE (2006)
7. Mostafa, S.A.K., Tolba, A.S., Abdelkader, F.M., Elhindy, H.M.: Video Watermarking
Based on Principal Component Analysis and Wavelet Transform. International Journal of
Computer Science and Network Security 9(8), 45–52 (2009)
8. Chuen-Ching, W., Yao-Tang, C., Yu-Chang, H.: Post-Compression Consideration in
Video Watermarking for Wireless Communication. World Academy of Science,
Engineering and Technology, 199–204 (2011)
9. Noorkami, M., Marsereau, R.M.: Digital Video Watermarking in P-Frames With
Controlled Video Bit Rate Increase. IEEE Transactions on Information Forensics and
Security 3(4), 441–455 (2008)
10. Puja, A., Khurshid, A.: Novel Invisible Watermarking for Various Images using HWT-
GA-PSO based Hybrid Optimization. International Journal of Advanced Research in
Computer Science and Software Engineering 3(8), 1093–1101 (2013) ISSN: 2277 128X
11. Martin, Z.: Master Thesis on Video Watermarking, Department of Computer Science
Education, Charles University Prague (2007)
12. Shekhawat., R.S., Rao, S., Shrivastava, V.K.: A Robust Watermarking technique based on
Biorthogonal Wavelet Transform. IEEE (2012) 978-1-4673-0455-9/12
13. Surekha, P., Sumathi, S.: Application of GA and PSO to the Analysis of Digital Image
Watermarking Process. International Journal of Computer Science & Emerging
Technologies 1(44), 350–362 (2010) (E-ISSN: 2044-6004)
14. Ramesh, S.M., Shnamugam, A.: An Efficient Robust Watermarking Algorithm in Filter
Techniques for Embedding Digital Signature into Medical Images Using Discrete Wavelet
Transform. European Journal of Scientific Research 60(1), 33–44 (2011) ISSN 1450-216X
15. Wang, Z., Sun, X., Zhang, D.: A Novel Watermarking Scheme Based on PSO Algorithm.
In: Li, K., Fei, M., Irwin, G.W., Ma, S. (eds.) LSMS 2007. LNCS, vol. 4688, pp. 307–314.
Application and Comparison of Three Intelligent
Algorithms in 2D Otsu Segmentation Algorithm
Lianlian Cao1,2, Sheng Ding1,2, Xiaowei Fu1,2, and Li Chen1,2

1
College of Computer Science and Technology,
Wuhan University of Science and Technology, China
2
Hubei Province Key Laboratory of Intelligent Information Processing
and Real-time Industrial System, China
Abstract. 2D Otsu thresholding algorithm has been proposed based on Otsu

algorithm, it is more effective in image segmentation. However, the
computational burden of finding optimal threshold vector is very large for 2D
Otsu method. In this paper, three kinds of intelligent algorithm are applied to
improve and compare the efficiency of search. Experimental results show that
these methods can not only obtain the ideal segmentation results but also
greatly reduce the launch time. Moreover, it is proved that the quantum particle
swarm optimization (QPSO) algorithm has the highest efficiency.
Keywords: 2D Otsu, image segmentation, intelligent algorithm, QPSO.
1 Introduction
In the image processing field image segmentation is a very important part, it is the
basis of image analysis and understanding as well. The shareholding method is an
effective method of image segmentation, and the most representative method is Otsu
[1]. 2D Otsu [2, 3] method was presented on the basis of Otsu method, and it is based
on image pixel and the two-dimensional histogram of pixel domain average. This
method can get better segmentation results. However, it increases the computational
complexity and limits the application of the algorithm. In order to overcome these
disadvantages, the intelligent algorithms [4] are proposed to improve search
efficiency. These three algorithms are respectively particle swarm optimization (PSO)
algorithm [5], quantum particle swarm optimization (QPSO) algorithm [6, 7] and
genetic algorithm (GA) [8].
The experiment results show that the threshold search efficiency of the three
intelligent algorithms is greatly increased, and the QPSO algorithm searching
efficiency is highest in these intelligent algorithms.
2 Three Intelligent Algorithms and the Main Parameter Setting
2.1 Particle Swarm Optimization Algorithm

PSO is a stochastic search method that was developed in 1995 based on the
sociological behavior of bird flocking. This algorithm is easy to implement and has
222 L. Cao et al.
been successfully applied to solve a wide range of optimization problems. Now, the
PSO technique is used to solve the problem of threshold based segmentation.
Main Parameter Setting. Set the learning factor c1 and c2 , c1 = c2 = 2 ; the inertia
weight w , w = wmax − wmax − wmin × iter , where the wmax = 0.9, wmin = 0.4 , Maxiter
Maxiter
is the maximum number of iterations.
2.2 Quantum Particle Swarm Optimization Algorithm

QPSO is combining the classical PSO algorithm and the quantum theory, and it is
based on the concept of quantum theory. The main iterative formula for particles
1 m 1 m 1 m 1 m
mbest (t) = 
m i =1
pi (t) = [  pi1 (t),  pi 2 (t),...,  piD (t)].
m i =1 m i =1 m i =1
(1)
pid (t ) = (r1pid + r2 pbd ) (r1 + r2 ). (2)
1
X id (t + 1) = pid (t) ± β | mbest (t) − X id (t) | ln( ). (3)
u
where mbest is the average best position in group; pid as a random point
between pid and pbd ; r1 and r2 for the interval [0, 1] random number; t for the current
iteration number; D is the dimension of particles, u is a random number in [0, l];
β as the contraction coefficient of expansion of the algorithm.
t
Main parameter Setting. β = m − ( m − n ) , where m = 1, n = 0.5 .
Maxiter
2.3 Genetic Algorithm

GA is a kind of adaptive global optimization probability search algorithm, which can
speed up the overall algorithm to complement real-time processing. Because the
essence of 2D Otsu threshold is seeking an optimal solution process, so the available
genetic algorithm has the speediness of its optimization, in order to achieve the
purpose of improving the efficiency.
Main Parameter Setting. Crossover probability, pc = 0.8 ; Mutation probability,

pm = 0.02 .
Application and Comparison of Three Intelligent Algorithms 223
3 Image Segmentation Based on Intelligent Algorithm
3.1 2D Otsu Segmentation Methods
Given an image f represented by L gray levels, size of M × N , each pixel in the

image of the value corresponds to a greyscale.
Fig. 1. Planar projection of two-dimensional gray histogram
The two-dimensional vector (s, t ) in Fig.1 is the threshold, and it divides two-
dimensional histogram into four parts (I, II, III and IV). The regions I and III contain
the distributions of object and background classes, respectively.
Let S 0 and S1 represent the object and the background respectively. The between-
class discrete matrix is defined as
1
σ B ( s , t ) =  P ( S k )[( μ k − μT )( μ k − μT )T ]. (4)
k =0
where w0 and w1 are the probabilities of class occurrence; μ0 and μ1 are the mean
value vectors of S0 and S1 ; μT is the total mean level vector of the 2D histogram.
s t L −1 L −1
w0 ( s, t ) = P ( S0 ) =  pij , w1 ( s, t ) = P ( S1 ) =  pij . (5)
i =0 j = 0 i = s +1 j = t +1
s t ipij s t jpij L −1 L −1 ipij L −1 L −1 jpij

μ 0 (s, t ) = ( ,  )T , μ 1(s, t ) = (   ω ,  )T . (6)
i =0 j =0 ω0 i =0 j = 0 ω0 i = s +1 j =t +1 1 i = s +1 j =t +1 ω1
L −1 L −1 L −1 L −1
μT = ( ipij ,  jpij )T . (7)
i =0 j =0 i = 0 j =0
224 L. Cao et al.
The trace of discrete matrix could be expressed as

1
tr (σ B ) =  (ωk [( μki − μTi )2 + ( μ kj − μTj )2 ]). (8)
k =0
3.2 QPSO Concrete Realization of Image Segmentation
In this paper, the trace of discrete matrix function tr (σ B ) is the fitness function; find
the maximum value, namely
f (s, t ) = max tr (σ B ). (9)
QPSO Algorithm Process:

1) Initialize position xi of each particle.
2) Calculate the fitness of each particle (according to the formula (9)).
3) Update individual extremum: let the current fitness value of the i-th particle is
compared with the particle individual extreme pid . If the former is more optimal, then
update the pid , otherwise the pid unchanged.
4) Update the global extremum: choose the optimal value from all pid (choose
the maximum value of the algorithm) as the global extremum pbd .
5) Optimization process, according to the formula (1), (2) and (3) update all the
particles in the QPSO algorithm.
6) Check whether meet the termination conditions, if satisfied, exit; Otherwise,
k = k + 1 ( k is the number of iterations) return to step 2, until meet the termination
conditions.
4 Experimental and Result Analysis

All the algorithms are implemented on a personal computer with CPU of 2.6GHz
using MATLAB R2012a programming language. In the experiments, 4 images (Lena,
Cameraman, Peppers and Rice) used in this study and they are shown in Fig. 1. In
order to demonstrate the effectiveness of the algorithms, below, will do some specific
analysis of these algorithms and their results.
4.1 Segmentation Results

In this paper, the total number of particles is set to 10, and the largest number of
iterations is set to 100. The experimental results show that the algorithm achieves the
same segmentation results with the 2D Otsu algorithm. Results are as shown below:
Fig. 2. Test Images. (a) Lena; (b) Cameraman; (c) Peppers; (d) Rice
Fig. 3. Segmentation by 2D Otsu
Fig. 4. Segmentation by PSO+ 2D Otsu
Fig. 5. Segmentation by QPSO + 2D Otsu
Fig. 6. Segmentation by GA + 2D Otsu

226 L. Cao et al.
All the figures show the experimental results of these algorithms. Because PSO,
QPSO and GA algorithms can achieve the same effect as the 2D Otsu algorithm, so
we can obtained that the PSO, QPSO and GA algorithms are as good as the 2D Otsu.
4.2 Calculation Results

Firstly, 2D Otsu method is used to calculate the maximum variance threshold, get the
optimal threshold value of image a is (125,116), the corresponding maximum variance
is 2975.5; the image b is (89,103) and corresponding maximum variance is 6209.5;
the image c is (125,134) and corresponding maximum variance is 3912.3; the image d
is (133,138) and corresponding maximum variance is 2433.7. The contrast of the
thresholds and searching time of the various methods is shown in Table 1.
Table 1. Performance comparison
2D PSO+2D
Image GA+2D Otsu QPSO+2D Otsu
Otsu Otsu
Lena(125,116) 0.8452s 0.4359s 0.4164s 0.3220s
Cameraman(89,103) 0.8408s 0.4446s 0.4137s 0.3351s
Peppers(125,134) 0.8314s 0.4490s 0.4226s 0.3325s
Rice(133,138) 0.8388s 0.4638s 0.4154s 0.3644s
Table 1 lists the average elapsed time of 40 experiments. For each figure, each
algorithm tested 40 times. During the experiment, the time efficiency of these
algorithms is repeated comparative. The experimental results is as shown in Table 1, it
can come to a conclusion: the introduction of the three intelligent algorithms has
increased 2D Otsu operation efficiency and reduced the search time.
The following is the efficiency comparison among the three kinds of intelligent
algorithm. It is already to know, the efficiency of these algorithms is higher than the
2D Otsu algorithm. However, in the three algorithms whose efficiency is the highest.
Similarly, it can find the answer from Table 1. In respect of the consumption time of
three algorithms, QPSO algorithm has the shortest search time. So it is proved that the
efficiency of QPSO algorithm is optimal.
5 Conclusions
2D Otsu method is an effective method of image segmentation, and it considers the
image gray level information and the space between the pixel neighborhood
information. Usually, the method can get better segmentation result than one-
dimensional Otsu method, but the consumption of time is greatly increased. In order
to solve this problem, the GA, PSO and QPSO algorithms are used to search the
optimal two-dimensional threshold vector. The experiment results show that the use
of the three intelligent algorithms can reduce the search time, thereby it increase
search efficiency. What is more, among these algorithms, the search time of QPSO is
shortest, so we can draw a conclusion that the QPSO algorithm is optimal.
Acknowledgments. This work was supported by Open Project Program of Hubei

Province Key Laboratory of Intelligent Information Processing and Real-time
Industrial System (znss2013A008) and National Natural Science Foundation of China
(No.61201423, 61375017)
References
1. Sthitpattanapongsa, P., Srinark, T.: A two-stage Otsu’s thresholding based method on a 2D
histogram. In: 2011 IEEE International Conference on Intelligent Computer
Communication and Processing (ICCP), pp. 345–348. IEEE (2011)
2. Lu, C., Zhu, P.: The Segmentation Algorithm of Improvement a Two-dimensional Otsu and
application research. In: 2nd International Conference on software Technology and
Engineering (ICSTE) V1-76–V1-79 (2010)
3. Wang, X., Chen, S.: An improved image segmentation algorithm based on two-dimensional
Otsu method. Inf. Sci. Lett 1, 77–83 (2012)
4. Kennedy, J., Eberhart, R.: Swarm Intelligence. Morgan Kaufmann Publishers, San
Francisco (2001)
5. Tang, H., Wu, C., Han, L., Wang, X.: Image Segmentation Based on Improved PSO. In:
The Proceedings of the International Conference on Computer and Communication
Technologies in Agriculture Engineering (CCTAE 2010), pp. 191–194 (2010)
6. Yang, S., Wang, M., Jiao, L.: A quantum particle swarm optimization. In: Congress on
Evolutionary Computation, CEC 2004, vol. 1, pp. 320–324. IEEE (2004)
7. Chao, Z., Jun, S.: Hybrid-Search Quantum-Behaved Particle Swarm Optimization
Algorithm. In: 2011 Tenth International Symposium on Distributed Computing and
Applications to Business, Engineering and Science (DCABES), pp. 319–323. IEEE (2011)
8. Sheta, A., Braik, M.S., Aljahdali, S.: Genetic Algorithms: A tool for image segmentation.
In: 2012 International Conference on Multimedia Computing and Systems (ICMCS),
pp. 84–90. IEEE (2012)
A Shape Target Detection and Tracking
Algorithm Based on the Target Measurement
Intensity Filter
Weifeng Liu, Chenglin Wen, and Shuyu Ding
Hangzhou Dianzi University, Hangzhou, Zhejiang 310018, China

{liuwf,wencl}@hdu.edu.cn
Abstract. The probability hypothesis density (PHD) is the expectation

intensity in a point in state space. The intensity integral in any region
of the state space is the expected number of targets contained in that
region. In this paper, we propose a target measurement intensity (TMI)
filter. Compared with the existing methods, the proposed approach is
simpler. Since the conventional PHD filter can not directly deal with the
shape target detection and tracking, we give the detection and tracking
algorithm based on the TMI filter by modeling the parameter dynamics
and measurement function of the shape target.
1 Introduction
The probability hypothesis density (PHD) was first proposed by Stein and Win-
ter [1]. Literally, it is a hypothesis density and does not exist in practice. Its
physical meaning is the expected number of targets in a point in state space.
Therefore, its integral in certain region in state space proposes the number of
targets in that region. Mahler showed that the first order moment of multitarget
random finite set (RFS)[2], which is an extension of the first order moment of
random point process [3], is equal to the PHD almost everywhere. He also pro-
posed the PHD recursive filter as an alternative of RFS Bayesian equation. One
can estimate target state from the PHD filter. The PHD filter is a joint decision
and estimation algorithm. It can be seen as an implicit association-estimation al-
gorithm for the association step is substituted by an estimation step. It can deal
with the uncertain number of targets such as the surviving targets, the sponta-
neous birth of new targets, and the spawned targets. Similar to the traditional
approaches, the PHD filter is used in the point target tracking.
The PHD is equal to the expected number of measurement originating from
a point x in state space. This is built on the following viewpoint: under the
assumption of target being a point, a target produces at most one measurement.
Therefore, the number of targets statistically equals the number of measure-
ments. The PHD thus can be seen as the target measurement intensity (TMI).
In RFS framework, Mahler got the PHD filter through probability generating
functionals (PGF). Erdinc et al alternatively derived the PHD filter by using
the physical-space approach - a bin model [4], where the PHD is interpreted

A Shape Target Detection and Tracking Algorithm 229
as the bin-occupancy probability in the traditional probability framework. In

another ref.[5], Streit derived a multitarget intensity filter from a Bayesian first
principles approach using a Poisson point process approximation at one step.
Streit’s intensity filter is very similar to the PHD filter except the estimation of
the target birth and measurement clutter processes. In this paper, an alternative
derivation of the PHD filter is first proposed by using the TMI.
A point target produces one measurement at most in the traditional re-
searches. In contrast, a shape target might give multiple measurements in each
scan. The TMI can also be used in shape target tracking, not only a point target.
we here refer it to be as the TMI filter and extend the point target tracking to
the shape target tracking. Nevertheless, the detection and survival probability
also have different means. For example, how to define a shape target is detected
when the shape is partly covered. How to define the probability of detection in
this case. Besides, the key to shape target tracking is to model dynamics and
measurement function. All these implies that the original PHD filter cannot be
directly used to the shape target tracking. Note that in this paper we confine
the shape target to be with a parameter model, i.e., the shape target can be
described by using a parameter model.
2 Background and Problem Description
Two methods were proposed to derive the PHD filter. The first is the Mahler’s
PGF method. The second is the physical space method given by Erdinc et al.
Roy Streit proposed the intensity filter using the Poisson point process (PPP)
method. In single sensor case, the intensity filter and the PHD filter have the
same form. In this section, we review the PHD filter and the three methods.
Mahler’s PHD filter consists of the following two recursive steps of predicted
and update step [2],[6].
3 The Derivation of the Target Measurement Intensity

Filter
3.1 Non-parameter State Mixture Models
The finite mixture models (FMM) are used to describe observations coming from
various random sources and the models have a finite number of distributions
f (y|θ) = π1 f1 (y|θ1 ) + · · · + πm fm (y|θm ) (1)
Where y {yy 1 , · · · , y n } are observations, π1 , · · · , πm are the mixing weights,

θ1 , · · · , θm are the parameters for distributions f1 (·|·), · · · , fm (·|·), called com-
ponents here. we defined the following state mixture models:
fk (zz k |πk , Xk ) = πk,0 fk,0 (zk,i |xk,0 ) + πk,1 fk,1 (zk,i |xk,1 ) + · · · + πk,mk fk,mk (zk,i |xk,mk )
(2)
230 W. Liu, C. Wen, and S. Ding
For consistence, the former observations y are replaced by zk {zz k,1 , · · · , z k,nk },
k is the sampling time, , πk {πk,0 , πk,1 , · · · , πk,mk }, Xk {x x0 , x 1 , · · · , x k,mk },
where x k,0 is the clutter state, x k,1 , · · · , x k,mk are the states of targets, fk,0
is the clutter density, πk,0 is the clutter mixing weight, mk can be also inter-
preted as the number of targets. {πk,j }m k
j=1 are the mixing weights of the targets
and {fk,j (zz k,i |x mk
xk,1 )}j=1 are the corresponding measurement densities of targets.
Streit et al introduced the state mixture models in the probabilistic multiple hy-
pothesis tracking (PMHT) [12]-[13]. The state mixture models describes how the
measurements are generated, but the derivation of the measurement likelihood
function is still a difficult one. Under the independent k assumption, the measure-
ment likelihood can be given by product Lz (x) = ni=1 fk (zz k,i |πk , Xk ). But this
expression is an implicit form for the state Xk need to be first estimated. Then,
a problem is: can we obtain the states of targets while avoid the calculation of
the likelihood function? This needs one to consider certain characteristic func-
tion of the target state x in the target state space. Thus the states of targets
can be estimated from the characteristic function. For RFS, the characteristic
function is corresponding to its first order, i.e., the PHD Dk (x x ). In this paper we
consider the target measurement intensity (TMI) ek (x) in the state space. The
TMI describes the distribution of the number of the target measurements in the
state space. The further research shows that the TMI are statistically equal to
the PHD under certain assumptions.
Assume that the target states are all in the same state space. Consider the
state models of a point x in state space
fk (zk,i |xx ) = P (D0 |x x , D0 ) + P (D1 |x

x )fk (φ|x x )fk [zz k,i |x
x , D1 ]
= (1 − PD (x x ))πk,φ fk [φ|x, ek,0 (x x )] +
PD (x x )ck [zk,i |x
x ){πk,c (x x, ek,i (x
x ) = 0] + πk,t (xx)gk [zz k,i |x
x, ek,i (xx) = 1]} (3)
Where D1 , D0 are the events that the target is detected and is not detected,
respectively. φ is the event of no measurements, ck (·|·) is the clutter distribution,
x) is the indicating variables
g(·|·) is the target measurement distribution, ek,i (x
defined in {0, 1}. ek,i (xx) = 1 implies that the ith measurement produced by
state x . We extend the above state mixture models to the total state space S as
follows.
fk (zk,i |S) = fk (zz k,i |S, D0 )P (D0 ) + fk (zz k,i |S, D1 )P (D1 )
= [fk (zz k,i |x

x , D0 )P (D0 |x
x) + fk (zz k,i |x
x , D1 )P (D1 |x
x )]dx
x
x∈s
= {[1 − PD (x
x )]fk [φ|x
x , ek,0 (x
x )]}dx +
x∈s
{πk,c (x
x )ck [zz k,i |x
x, ek,i (x
x ) = 0] + PD (x
x )πk,t (x
x ))gk [zz k,ix , ek,i (x
x ) = 1]}dx(4)
x∈s
x ) and πk,t (x
Where πk,c (x x ) are the weights of clutter measurements and target
measurements. We define the mixing weight to be the probability of target ex-
isting in the point x

x ) = pk [ek,i (x
πk,c (x x ) = 0|x
x]
x ) = pk [ek,i (x
πk,t (x x ) = 1|x
x]
Then the target measurements intensity in the state x is defined by:

x ) = ek,0 (x
ek (x x ) + · · · + ek,nk (x
x ) + ek,1 (x x) (5)
The TMI consists of two types of terms: the intensity ek,0 (x x) of no measurements
and individual measurement intensities {ek,i (x
x )}ni=1
k
. In the following subsection,
we focus on deriving the recursive equations of the TMI and the mixing weights.
3.2 Derivation of the Recursive Equation for the TMI Filter

The recursive equations of the TMI filter involve two step, i.e., the predicted
step and the update step.
The Predicted TMI. The assumption in the predicted step is proposed as

follows.
A.1: Each target moves and generates individually and independently of all the
other targets.
Proposition 1. Under the assumption A1. Assume that at time k the intial
x ). Then, the predicated TMI is given by:
TMI is ek (x
x ) = eγk+1 (x
ek+1|k (x x ) + eβk+1|k (x
x ) + esk+1|k (x x) (6)
Where
esk+1|k (x
x) = PS (x x |ω)ek (ω
x )fk+1|k (x ω )dω
ω (7)
ω ∈S
eβk+1|k (x
x) = PS (x x |ω
x)βk+1|k (x ω )ek (ω
ω )dω
ω (8)
ω ∈S
Where eγk+1 (x x ), eβk+1|k (x

x ), esk+1|k (x x) are respectively the TMI for the sponta-
neous births, the surviving target and the spawned target. PS (x x ) is the survival
probability. fk+1|k (x x |ω) and βk+1|k (x x|ω
ω ) are respectively the Markov models of
the surviving target and the spawned target, which are the same as eq.(1). It
can be derived based on the weighted sums of the TMI. This is similar as the
PHD filter.
3.3 The Update TMI

In this step, we propose the following assumptions
A.2 The number of target measurements and the number of clutter measurement
follow Poisson with intensities Mk+1 and λk+1 , respectively.
A.3 Clutter measurements are independent of target states.
x ). Then, under the as-

Proposition 2. Assume that predicted TMI is ek+1|k (x
sumptions A.2 and A.3, the update TMI is given by:
x ) = ek+1,0 (x
ek+1 (x x ) + · · · + ek+1,nk+1 (x
x) + ek+1,1 (x x) (9)
x ) is the TMI of no measurements, ek+1,i≥1 (x
Where ek+1,0 (x x ) is the TMI of the
ith measurement. And they can be calculated by the following equation:
x) = (1 − PD (x
ek+1,0 (x x ))ek+1|k (x
x) (10)
PD (x x )gk+1 (zk+1,i |x
x)ek+1|k (xx)
x) =
ek+1,i≥1 (x ! (11)
λck+1 (zz k+1,i ) + PD (x x )gk+1 (zk+1,i |x
x )ek+1|k (x
x )dxx
4 The Shape Detection and Tracking
K=1
Y
K=3
r1 K=2
r3
(vx,1,vy,1) (vx,3,vy,3)
(px,1,py,1) (vx,2,vy,2)
r2
(px,3,py,3)
(px,2,py,2)
O X
Fig. 1. A movement of a circular shape in x-y plane which is described by a five-

dimensional parameter vector
4.1 The Target Dynamics and Measurement Function

A shape can be modeled in the parameter space and be described by a parameter
vector. We therefore propose the dynamics and measurement function in the
parameter space. The movement of the shape target can be modeled by the
dynamics of the parameter vector. This is the same as the usual state function.
For a nonlinear function and a linear system function, we can describe it by
using the following functions:
X k = f (Xk−1 , ω k ) nonlinear parameter function (12)

X k = Ak−1 Xk−1 + Bk−1ω k linear parameter function (13)
We example two targets with a circular shape and a linear shape, respectively.
Fig.1 shows a circular target moving in the planar. We select the center point
and the radius of the circle as the parameter vector X = (px,k , ṗx,k , py,k , ṗy,k , rk ),
where (px,k , py,k ) is the center coordinate of the circular shape, (ṗx,k , ṗy,k ) is the
velocity of the center point, rk is the circle radius.
In Figure 2, a movement of a linear target is given. It involves two types of
movements which include rotation around the center of the linear shape and
CV movement of the center point. A seven-dimensional parameter vector is pro-
posed as Xk = (px,k , ṗx,k , py,k , ṗy,k , θk , θ̇k , lk ), where (px,k , py,k ) is the the center
coordinate of the linear shape, (ṗx,k , ṗy,k ) is the velocity of the center point, θk
and θ̇k are respectively the angle and the angular velocity of the linear shape, lk
is the length of the target.
Y
K=1 K=2 K=3
l1 T2 l3
(px,1,py,1) T1 T3
(vx1,vy1) (vx,3,vy,3)
(px,2,py,2) (vx,2,vy,2)
l2
(px,3,py,3)
O X
Fig. 2. A movement of a linear shape in x-y plane which is described by a seven-

dimensional parameter vector
5 Simulations
In this section, a simulation of detection and tracking of three circular targets
is proposed to verify the proposed TMI filter. Three circular targets with CV
movement are proposed in this simulation. The parameter vector is given in
50 50
45
The number of target measruements
0 40
35
−50 30
y(m)
25
−100 20
15
−150 10
True number of target measurements
5 Estimated number of target measurements
−200 0
−50 0 50 100 150 200 0 10 20 30 40 50 60
x(m) t(s)
(a) The clutter measurements and (b) The measurement intensities

target measurements
Fig. 3. The measurements and intensities

20 20
True shape True tracks
Estimated shape Estimated tracks
0 0
−20 −20
−40 −40
y(m)
y(m)
−60 −60
−80 −80
−100 −100
−120 −120
−140 −140
−20 0 20 40 60 80 100 120 140 160 −20 0 20 40 60 80 100 120 140
x(m) x(m)
(a) The true shapes and the esti- (b) The true tracks and the esti-
mated shapes mated tracks
Fig. 4. The Estimated shapes and tracks
eq.(26). The initial states for the three targets are respectively:
X01 = (−10m, 3.2m/s, −10m, −2.2m/s, 5m), X02 = (10m, 2.2m/s, 10m, −2.2m/s,
5m) and X03 = (−10m, 1.8m/s, −50m, −1.9m/s, 5m). Covariance matrixes:
P0i = diag([10, 2, 2.5, 10, 2.5, 1]), Qik = diag([0.01, 0.01, 0.2]) and
Rki = diag([0.04, 0.04]), i = 1, 2, 3. Clutter density ρ(x x) = 1.0 × 10−3 m−2 . The
surveillance region is [−50, 200] × [−200, 50]m . The measurements are produced
2
by eqs.(27)-(29), where parameters {ψk (i)} samples in [0, 2π] per π/6 and thus
λk = 13.
The Gaussian mixture based TMI filter, which is analogous to the Gaussian
mixture PHD filter, is proposed here. Sub figure (a) of Fig.3 shows the clutter
measurements and shape measurements every 4 seconds. Obviously, these two
types of measurements are mixed. Thus, our first step is to detect the target. It
can be seen from (b) of Fig.3 that the measurement intensities and number of
targets indicate the existing of targets and their values are approximated to the
true value. In estimation of shape parameter, it can be seen from sub figure (a)
of Fig.4 that the circle radiuses are close to the true values. Sub figure (b) of
Fig.4 shows that the estimations of position are near the true tracks.
6 Conclusion
This paper proposed a target measurement intensity (TMI) filter based on the
mixture distributions. We provided the predicted TMI and the update TMI.
Under some assumptions, the TMI filter is equal to the original PHD filter. The
PHD filter focus on point target tracking. A potential advantage of the TMI
filter is that it can be extended the target tracking with parameter shape. Thus
we can use it in the extended target tracking. Correspondingly, our next work is
to extend the TMI to the target with parameter shape. The key is to model the
parameter dynamics and the parameter measurement function. Based on these
functions, we extend the TMI filter to the shape target tracking. Finally, we
propose two experiments involving three circular targets and three linear targets
to verify the proposed TMI filter.
Nevertheless, under the multiple measurements condition, estimation of the

number of targets is still an intractable problem. We propose a simple formula-
tion under the Poisson assumption of the target measurements. Besides, in this
paper we confine our object to the targets with parameter shapes and they are
in the same parameter space. Future works are still needed for the general shape
target.
Acknowledgements. This work was supported in part by the NSFC (61175030,

61273170, 61333011, and 61271144)
References
1. Winter, C.L., Stein, M.C.: IES/BTI system overview. In: Proceedings of 8th Na-
tional Symposium on Sensor Fusion, Dallas TX, vol. I, pp. 15–17, 27–46 (1995)
2. Mahler, R.: Multitarget Bayes Filtering via First-Order Multitarget Moments.
IEEE Transactions on Aerospace and Electronic System 39(4), 1152–1178 (2003)
3. Daley, D.J., Vere, J.D.: An Introduction to the Theory of Point Processes, Ele-
mentary Theory and Methods, 2nd edn., vol. I, pp. 123–131. Springer, New York
(2003)
4. Erdinc, O., Willet, P., Bar-Shalom, Y.: A Physical-Space Approach for the Proba-
bility Hypothesis Density and Cardinalized Probability Density Filters, Signal and
Data Processing of Small Targets. In: Proc. of SPIE, vol. 6236, pp. 1–12 (2006)
5. Streit, R.L., Stone, L.D.: Bayes Derivation of Multitarget Intensity Filters. In:
The 11th International Conference on Information Fusion, Colgon, Germany, pp.
1686–1693 (July 2008)
6. Mahler, R.: An Introduction to Multisource-Multitarget Statistics and Its Appli-
cations. Technical Monograph, Lockheed Martin: 1–20 (Mar 15, 2000)
7. Vo, B.N., Ma, W.K.: The Gaussian Mixture Probability Hypothesis Density Filter.
IEEE Transactions on Signal Processing 54(11), 4091–4104 (2006)
8. Vo, B.T., Vo, B.N., Cantoni, A.: Bayesian Filtering with Random Finite Set Ob-
servations. IEEE Transactions on Signal Processing 56(4), 1313–1326 (2008)
9. Challa, S., Vo, B.N., Wang, X.Z.: Bayesian Approaches to Track Existence - IPDA
and Random Sets. In: The 11th International Conference on Information Fusion,
Annapolis, Maryland, USA, pp. 1228–1235 (2002)
10. Ulmke, M., Erdinc, O., Willett, P.: Gaussian Mixture Cardinalized PHD Filter
for Ground Moving Target Tracking. In: The 10th International Conference on
Information Fusion, Quebec, Canada, pp. 1–8 (July 2007)
11. Liu, W.F., Han, C.Z., Lian, F.: An alternative derivation of a Bayes tracking filter
based on finite mixture models. In: Proceedings of the 12th International Confer-
ence on Information Fusion, Seattle, USA, pp. 842–849 (2009)
12. Streit, R.L., Luginbuhl, T.E.: Maximum likelihood method for probabilistic multi-
hypothesis tracking. In: Proceedings of SPIE International Symposium, Signal and
Data Proceeding of Small Targets, Bellingham, WA, USA, pp. 394–405 (1994)
13. Streit, R.L., Luginbuhl, T.E.: A Probabilistic multi-hypothesis tracking algorithm
without enumeration and pruning. In: Proceedings of the Sixth Joint Service Data
Fusion Symposium, Laurel, MD, pp. 1015–1024 (1993)
Multi-cell Contour Estimate
Based on Ant Pheromone Intensity Field
Qinglan Chen1, Benlian Xu2, Yayun Ren2, Mingli Lu2, and Peiyi Zhu2
1
School of Mechanical Engineering,
Changshu Institute of Technology, 215500 Changshu, China
2
School of Electrical & Automatic Engineering, Changshu Institute of Technology,
215500 Changshu, China
chenql@cslg.cn, xu_benlian@cslg.cn, zpy2000@126.com
Abstract. In this paper, we propose an ant pheromone based approach to

accurately extract the contours of multiple small cells in low contrast bio-
medical images. With the local information of intensity variation of each pixel,
the initial distribution of ant colony is generated as ants’ starting positions.
Following the heuristic information, such as the pixel grayscale variance, the
ant inertial heading and the image intensity, ant’s searching behavior is
modeled appropriately to make each of ants move along the edge of interested
object as possible. Due to modeling an accurate depositing mechanism of
pheromone, the corresponding ring pheromone field is formed and used to
extract interested cells’ contours after simple morphological operations.
Experiment results show that our algorithm could give an accurate contour
estimate of each cell for several different image sequences.
Keywords: Image Processing, Ant Colony, Contour Estimate.
1 Introduction
As an important branch of cell motion analysis, the estimate of cell contour could
directly or indirectly encompass rich contents about each individual cell, and the
related research is challenging and emerging due to poor image quality, small size and
discontinuities or sharp changes in intensity, etc.. In most computer applications,
image contour extraction constitutes a crucial initial step before performing the task
of object recognition and representation. The conventional approaches are
computationally expensive because each set of operations is conducted for each pixel.
Thus, many researchers resort to other promising techniques. An ACO-based
approach, a nature-inspired optimization algorithm, has the potential of solving these
intractable problems because of its parallelized and intelligent searching mechanisms,
which makes the algorithm easily adaptable for processing multiple objects
simultaneously. In terms of the combination between the ACO-based approaches and
image segmentation, the related work can be divided into two main strands of
research. The first strand focuses on the fusion of ACO and other edge detection and
contour extraction algorithms [1, 2], mainly because of the strong and effective
Multi-cell Contour Estimate Based on Ant Pheromone Intensity Field 237
optimization capabilities of ACO. Oliver et al.[3] formulate the shape correspondence

as a Quadratic Assignment Problem (QAP), incorporating proximity information into
the point matching objective function, and propose the first ACO algorithm directly
aimed at solving the point and contour correspondence problems. Lai et al. [4] present
a novel system for active contour tracking of moving objects in video sequences,
incorporating the use of edge flows in the ACO algorithm to improve the efficiency.
Li [5] presents a novel image contour extraction by ant colony algorithm and B-snake
model. The method can enhance the flexibility of B-snake to describe complex shape.
The second strand is those approaches [6-9] which construct the edge map based on
the pheromone matrix. In these methods, each entry of the pheromone represents the
edge information at each pixel of the image. Anna et al. [10] propose a ACO-based
edge detection method which takes advantage of the improvements introduced in ant
colony system, a extension of AS. Aminu et al. [11] use the discrete wavelet
transform (DWT) as a preprocessing step with ACO to enhance image edge detection.
Carla et al. [12] employ the ACO-based algorithm preceded by anisotropic diffusion
to segment the optic disc in color images, and good performance is achieved as the
optic disc was detected in most of all the images, even in the images with great
variability.
2 Ants for Multi-cell Contour Estimate
2.1 Ants’ Initial Distribution

In the original ant system, the initial ant colony is usually uniformly distributed in a
searching space, however, an alternative layout is proposed to assign a given number
of ants to pixels of an image where objects (cells) probably occur. For this purpose,
the grayscale distribution of current image is utilized to measure the relative intensity
variation of each pixel within its local region. As illustrated in Fig.1, an 8-
neighbouring pixel configuration, a neighboring region of a given pixel (i, j ) with
intensity I ( i , j ) , is defined, and its local region grayscale variance is computed as
1
Δσ ( i , j ) = 
| N (i , j ) | ( i ′, j ′)∈N( i , j )
( I ( i ′, j ′) − I ( N ( i , j ) )) 2 , (1)
where N (i , j ) denotes the neighboring pixels of pixel (i, j ) , I ( N (i , j ) ) denotes the

average gray intensity of N (i , j ) , and | N ( i , j ) | is the number of neighboring pixels of
pixel (i, j ) . As implicated in Eq. (1), the grayscale variance has a smaller value in the
area of background and interior of cell, whereas a larger value is probably taken
between two sides of edge, as well as in the vicinity of each edge pixel.
238 Q. Chen et al.
(i −1, j −1) (i −1, j) (i −1, j +1)
(i, j −1)
(i, j ) (i, j +1)
Boundar y mar ker

(i +1, j −1) (i +1, j) (i +1, j +1)
Fig. 1. 8-neighbouring pixel configuration
2.2 Ants’ Working Mechanism
To emulate clumping behavior in ant colony (i.e., aggregation), ant working

environment and ant decision constraint should be defined appropriately. The ant
operating environment corresponds to the current cell image and is denoted by a tuple
P, N , where P is a finite set of image pixels, and N ⊆ P × P is a closest
propagation neighboring relation among pixels. For any pixel index g ∈ P , an 8-
neighbouring configuration is represented as N ( g ) = {s ∈ P : gNs} with ( N ( g ) ≤ 8 ).
With this definition, each ant can move directly towards one of its neighbors at a time,
and any pixel can be visited simultaneously by several ants.
Cell contour extraction is challenging due to incomplete or discrete grayscale
variance curves and irregular or partial ant colony distribution in the neighboring
region of each edge pixel. Therefore, a novel ant decision is designed and modeled as
 arg max ( (τ (Ci′, j ′) (tˆ))λ (Δσ (i ′, j ′ ) )ς (W (Δθ (i ′, j ′) ))γ ) if q ′ < q0

 (i ′, j ′)∈N( i , j )
(i ′, j ′) =  (i ′, j ′)∉Ωv , (2)
  
 (i , j ) otherwise
where λ , ς , and γ are the adjustment parameters of contour pheromone τ (Ci′, j ′) (tˆ) ,
heuristic grayscale variance Δσ (i ′, j ′) , and weight factor of ant heading change
W (Δθ ( i ′, j ′) ) , respectively; Ω v is the set of pixels visited by the current ant; q0 is a
threshold which takes the value between 0 and 1.
The pixel (i, j ) will be visited according to the probability distribution given by:
 (τ (Ci , j ) (tˆ))λ (Δσ (i , j ) )ς (W (Δθ (i , j ) ))γ

 if N ( i , j ) ⊄ Ω v

P(Ci , j ) →(i , j ) (tˆ) =   (τ (Cm,n ) (tˆ))λ (Δσ ( m, n) )ς (W (Δθ( m,n ) ))γ
( m , n )∈N ( i , j )
. (3)
 ( m , n )∉Ωv

 0 otherwise
It is implicated that the above model not only propels an ant towards edge pixels of
cell, but also forces an ant keeping on moving along the edge of cell instead of staying
in the vicinity of its starting pixel. According to the configuration in Fig.1, the
destination candidates for each ant decision are up to eight, and the angle between two
neighboring candidates is 450 . Therefore, we assume that if the current heading of an
ant is known, the weight factor of ant heading change in the following decision is
defined as: W (±00 ) = 1/ 3 , W (±450 ) = 1/ 3 , W (±900 ) = 1/10 , W (±1350 ) = 1/ 16 ,
and W (±1800 ) = 1/ 20 .
Once an ant has made m decisions, it will deposit an amount of pheromone on
corresponding visited pixels (up to m pixels at a time) with two levels

(
1/ max {Δσ(m′) , μmin } ) if std {1/ Δσ(m′) }m′=1 <δ0 , dmmax >
m
m

−tˆ /T 2 m
c ⋅ (1 − e ) /
r C (tˆ) = 
2
m′=1 6 , (4)
c ⋅ (1 − e−tˆ /T 2 )
0 otherwise
where c2 , δ 0 , T 2 , and μmin are constants, std {}

⋅ denotes the standard deviation of a
given set, and Δσ ( m′) denotes the grayscale variance value of pixel corresponding to
the m′ -th decision of ant.
We observe that not all ants deposit the same amount of pheromone according to
the above two constraints, but they do offer a hint of where continuous cell edges are
located when an ant makes a decision. Specifically, the first constraint
std {1 / Δσ ( m′) }
m
<δ 0 tries to strengthen tour pixels within the same level of
m′=1
grayscale invariance, while the second constraint d mmax > m / 6 encourages ant to
move as far as possible along the edges.
2.3 The Formation of Pheromone Field

In this work, we model three pheromone working mechanisms to jointly produce
pheromone field. First, the pheromone deposited by different ant individuals are
aggregated and merged on the corresponding pixel, which results in pheromone peaks
in the interested areas. Second, pheromone performs evaporation over time, and this
simulates the memory ability of ant individuals as we observe in nature. Finally,
pheromone propagation is considered to build a connection between neighboring
pixels, and it builds a bridge for access by neighboring agents.
Pheromone aggregation is defined as a combination of evaporation, external input,
and propagation. For any pixel (i, j ) , the evolution of pheromone amount follows
τ (Ci , j ) (tˆ + 1) = E ⋅τ (Ci , j ) (tˆ) + r(i , j ) (tˆ) + q( i , j ) (tˆ) , (5)
where r( i , j ) (tˆ) denotes the pheromone external input to pixel (i, j ) at the tˆ -th
iteration, and q( i , j ) (tˆ) models the propagation input to pixel (i, j ) . Note that the
above model applies to both location and contour pheromone fields.
240 Q. Chen et al.
It is observed that, unlike the traditional ant system (AS), the pheromone
propagation q( i , j ) (tˆ) is introduced to coincide with the pixel intensity continuity in
an image, and its evolution form is defined as
q( i , j ) (tˆ) =  | N
D
( r(i ', j ') (tˆ − 1) + q(i ', j ') (tˆ − 1) ) , (6)
( i ′, j ′ )∈N ( i , j ) ( i ', j ') |
where | N (i ', j ') | defines the cardinality of N (i ', j ') , D denotes the propagation
D
coefficient with 0 < D < 1 , and characterizes the averaged propagation
| N ( i ', j ') |
proportion of total received pheromone intensity on pixel (i ′, j ′) at the tˆ − 1 -th
iteration to its neighboring pixels.
As defined in Eq. (6), the propagation pheromone field at next iteration is the
propagating results of the field of itself and the external input pheromone field both at
current iteration. Furthermore, we assume that the pheromone amount on each pixel is
an un-weighted sum of pheromones of its neighbors, thus the Eq. (6) could be divided
into two parts and rewritten as
D D
q( i , j ) (tˆ) =  | N |
r(i ', j ') (tˆ − 1) + 
| N
q(i ', j ') (tˆ − 1)
( i ′, j ′ )∈N ( i , j ) ( i ', j ') ( i ′, j ′ )∈N ( i , j ) ( i ', j ') | . (7)
=D ⋅ r(i , j ) (t − 1) + D ⋅ q(i , j ) (t − 1)
ˆ ˆ
Upon the contour pheromone field τ C (tˆ) is formed, we first treat it as an input
image, and then three steps of morphological operations, including bridging
unconnected pixels, filling image regions and holes, and removing interior pixels, are
done to generate the contour of each cell.
3 Experiments
The performance of our proposed cell contour estimate algorithm is evaluated using
two challenging low-contrast multiple cell image sequences, which considers various
cases including cell dynamic difference, cell shape variation, and varying number of
cells. In terms of the initial ant colony distribution, a predefined threshold is used to
allocate ants to corresponding pixels. A larger value of threshold could generate fewer
ants and save computational burden at the expense of the loss of more useful
information, whereas a smaller value improves tracking accuracy with more ants
generated and more computational cost required. To obtain the desirable tracking
results, we set the threshold to be 0.1 in both image sequences.
Figs. 2 and 3 give the multi-cell contour estimates for these two sequences, and it
can be observed that our method could give an accurate contour estimate of each
existing cell of an image.
Frame 5 Frame 8 Frame 9 Frame 19 Frame 20
Frame 5 Frame 8 Frame 9 Frame 19 Frame 20
Fig. 2. Cell contour estimates of image sequence 1
Frame 21 Frame 26 Frame 27 Frame 30 Frame35
Frame 21 Frame 26 Frame 27 Frame 30 Frame35
Fig. 3. Cell contour estimates of image sequence 2
Also, we note that, despite the weakness in intensity and the discontinuity of
potential edges in the normalized grayscale variance field Δσ ( i , j ) (Fig.4(a)), our
algorithm could obtain the cell contour pheromone field in the form of a series of
continuous and close loops, as shown in Fig.4(b). Through three steps of morphological
operations, all cell-related contours are extracted in the end, as illustrated in Fig.4 (c).
Since cell contour estimate is dependent directly on the contour pheromone field,
and the working mechanism of pheromone is appropriately adjusted to form multiple
close, smooth and continuous belt loops. As shown in Fig.5 (a), if the propagation
coefficient D increases, it means that the effect of propagation increases as well, and
the continuity of each contour is guaranteed but the size of contour is larger than the
true one. For a smaller value of D , it will result in contour discontinuity and debris
due to lack of link bridge between neighboring pixels in terms of pheromone.
Similarly, for a smaller value of E , which means that more current pheromones
evaporate and only few are used for the following iteration, an undesirable estimate of
each contour is achieved as a net structure. However, with less pheromone
evaporation, i.e., a larger value of E , more pheromone are utilized as a guide for ant
to search for possible segment of contour, as illustrated in Fig.5 (b).
242 Q. Chen et al.
(a) (b) (c) (d)
Fig. 4. Cell contour estimate
E = 0.8, D = 0.8 E = 0.8, D = 0.5 E = 0.8, D = 0.01

(a)
E = 0.1, D = 0.1 E = 0.5, D = 0.1 E = 0.9, D = 0.1

(b)
Fig.5. Sensitivity analysis of various combination of E and D on contour estimate.
4 Conclusions
Cell motion analysis has become a major research direction for understanding the full
potential of time-lapse microscopy in biological research or drug discovery. In this
paper, we propose an ant pheromone based approach to accurately extract the
contours of multiple small cells. Experiment results show that 1) our algorithm could
give an accurate estimate of contour of each cell in various scenarios; 2) the tracking
accuracy depends on how the pheromone field models, which is affected mainly by
parameters E and D . As part of future work, we would like to expand the scope of
cell contour estimate to multi-parameter joint estimate, which could give us a broad
quantitative view of cell cycle progression.
Acknowledgments. This work is supported by national natural science foundation of

China (61273312) and natural science foundation of higher education colleges in
Jiangsu (14KJB510001).
References
1. Li, L., Ren, Y.M., Gong, X.P.: Medical Image Segmentation Based on Modified Ant
Colony Algorithm with GVF Snake Model. In: 2008th IEEE International Seminar on
Future BioMedical Information Engineering, pp. 11–14. IEEE Press, Washington (2008)
2. Ruberto, C.D., Morgera, A.: ACO contour matching a dominant point approach. In: 2011
4th IEEE International Congress on Image and Signal Processing, pp. 1391–1395. IEEE
Press, Shanghai (2011)
3. Kaick, O.V., Hamarneh, G., Zhang, H., Wighton, P.: Contour Correspondence via Ant
Colony Optimization. In: 2007 15th Pacific Conference on Computer Graphics and
Applications, pp. 271–280. IEEE Press, Washington (2007)
4. Chang, Y.H., Lai, D.X., Zhong, Z.H.: Active Contour Tracking of Moving Objects Using
Edge Flows and Ant Colony Optimization in Video Sequences. In: Wada, T., Huang, F.,
Lin, S. (eds.) PSIVT 2009. LNCS, vol. 5414, pp. 1104–1116. Springer, Heidelberg (2009)
5. Li, J.: Image Contour Extraction Based on Ant Colony Algorithm and B-snake. In: Huang,
D.-S., Zhao, Z., Bevilacqua, V., Figueroa, J.C. (eds.) ICIC 2010. LNCS, vol. 6215,
pp. 197–204. Springer, Heidelberg (2010)
6. Tao, W., Jin, H., Liu, L.: Object segmentation using ant colony optimization algorithm and
fuzzy entropy. J. Pattern Recognition Letters 28(7), 788–796 (2007)
7. Agrawal, P., Kaur, S., Kaur, H., Dhiman, A.: Analysis and Synthesis of an Ant Colony
Optimization Technique for Image Edge Detection. In: ICCS 2012th International
Conference on Computing Sciences, pp. 127–131. IEEE Press, Washington (2012)
8. Haase, R., Böhme, H.J., Perrin, R., Zöphel, K., Abolmaali, N.: Self-reproduction versus
transition rules in ant colonies for medical volume segmentation. In: Dorigo, M., Birattari,
M., Blum, C., Christensen, A.L., Engelbrecht, A.P., Groß, R., Stützle, T. (eds.) ANTS
2012. LNCS, vol. 7461, pp. 316–323. Springer, Heidelberg (2012)
9. Huang, P., Cao, H.Z., Luo, S.Q.: An artificial ant colonies approach to medical image
segmentation. J. Comput Methods Programs Biomed 92, 267–273 (2008)
10. Baterina, A.V., Oppus, C.: Image Edge Detection Using Ant Colony Optimization. J.
WSEAS Transactions on Signal Processing 6, 58–67 (2010)
11. Muhammad, A., Bala, I., Salman, M.S., Eleyan, A.: Discrete Wavelet Transform-based
Ant Colony Optimization for Edge Detection. In: 2013th International Conference on
Technological Advances in Electrical, Electronics and Computer Engineering (TAEECE),
pp. 280–283. IEEE Press, Konya (2013)
12. Pereira, C., Goncalves, L., Ferreira, M.: Optic disc detection in color fundus images using
ant colony optimization. J. Med Biol Eng Comput 51, 295–303 (2013)
A Novel Ant System with Multiple Tasks for Spatially
Adjacent Cell State Estimate
Mingli Lu, Benlian Xu, Peiyi Zhu, and Jian Shi
School of Electrical & Automatic Engineering, Changshu Institute of Technology,

215500 Changshu, China
luxiaowenwp@sohu.com, xu_benlian@cslg.cn,
zpy2000@126.com, yievans2010@hotmail.com
Abstract. Multi-cell tracking is an important problem in studies of dynamic cell

cycle behaviors. This paper models a novel multi-tasking ant system that jointly
estimates the number of cells and their individual states in cell image
sequences. Our ant system adopts an interactive mode with cooperation and
competition. In simulations of real cell image sequences, the multi-tasking ant
system integrated with interactive mode yielded better tracking results .
Furthermore, the results suggest that our algorithm can automatically and
accurately track numerous cells in various scenarios, and is competitive with
state-of-the-art multi-cell tracking methods.
Keywords: Ant System, Cell Tracking, Object Motion Analysis.
1 Introduction
During cell image sequencing, two or more cells will very likely contact or present
occlusions. In such cases, the image is not easily associated with spatially adjacent
images because the joint observation cannot be easily segmented. Because the
components of corresponding cell states are now coupled, the tracking of spatially
adjacent cells or cell occlusions becomes a challenging task, rendered more
complicated by low SNR in the image data. For efficiency and accuracy, the
development of automated tracking methods for spatially adjacent cell is of great
importance.
Many efforts have been made over the past decades. Dufour et al. [1] presented a
fully automated technique for segmenting and tracking cells in 3-D+time microscopy
data. This method uses coupled active surfaces with or without edges, together with a
volume conservation constraint and several optimizations to handle touching and
dividing cells, and cells entering the field of view during the sequence. In the method
of Nguyen et al. [2], multiple cell collisions cells are automatically tracked by
modeling the appearance and motion of each collision state, and testing collision
hypotheses of possible state transitions. Although some of the above algorithms have
resolved special challenges in spatially adjacent objects, they are problem-dependent
and not applicable to generic cell tracking problems.
A Novel Ant System with Multiple Tasks for Spatially Adjacent Cell State Estimate 245
In this paper, we employ a novel ant system with multiple tasks that jointly
estimates the number of cells and their individual states in sequences of cell images.
Depending on the initial distribution of the ant colony, the colony is roughly divided
into several groups, each assigned the task of finding a potential cell through the
defined ant working mode, namely interactive mode with cooperation and
competition.
2 Methods
In this section, we introduce a novel ant system with multiple tasks (AS-MT) for
estimating multi-cell parameters. It is noted that our algorithm builds solutions in a
parallel way on N + 1 pheromone fields (where N is the initial number of divided ant
groups),whereas the conventional ACO algorithm computes solutions in an
incremental way on only one pheromone field.
2.1 Initial Distribution of Ant Colony
Since the background in most cell image sequences exhibits slowly varying
background signals, such a problem can be solved by simplistic, static-background
models. In our work, the approximate median method, a kind of recursive technique,
is employed for the purpose of fast background subtraction [3], in which each pixel in
the background model is compared to the corresponding pixel in the current frame,
and finally to be incremented by one if the new pixel is larger than the background
pixel or decremented by one if smaller. With the iteration evolves, a pixel in the
background model converges to a value where half of the incoming pixels are greater
than and half are less than its value, and this value is known as the median.
The approximate median foreground detection compares the current frame to the
background model and further identifies the foreground pixels I1 ( i ) and binary image
pixels I 2 (i ) as
 I (i ), if I (i ) − B(i ) > Th
I1 (i ) =  (1)
 0, otherwise
 1, if I (i ) − B(i ) > Th
I 2 (i ) =  (2)
 0, otherwise
where I ( i ) denote the current frame pixels, B(i) are background pixels estimated by
 B(i ) + 1 if I (i ) > B (i )
recursive technique , B(i ) =  , and Th is a predefined
 B(i ) − 1 if I (i ) < B (i )
threshold. If the absolute difference between current frame pixel and background
pixel is larger than the predefined threshold Th , the foreground pixel I1 ( i ) = I (i) ,
246 M. Lu et al.
otherwise I1 ( i ) = 0 . Binary image pixel I 2 (i ) follows the same rule as Eq. (2). We
further assume that an ant is distributed at the location of pixel i if I 2 ( i ) = 1 in the
binary image, and thus a given number of birth ants are generated in the current
frame. Using the K-means clustering method[4], these ants are further divided
into N subgroups.
2.2 Ant System with Multiple Tasks

In this section, working mode in ant system with multiple tasks (AS-MT) is proposed
and investigated in details. Each ant can move directly towards one of it neighbors at
each time, and any pixel can be visited simultaneously by several ants with the
guidance of our proposed pheromone update mechanism.
Interactive Mode with Cooperation and Competition (IMCC). In our defined
interactive mode with cooperation and competition (IMCC), ants with different tasks
are modeled to work together with appropriate cooperation and repulsion. The
τ sj (t )
repulsion term is characterized by indicated the ratio of pheromone level of
τ j (t )
task s to total pheromone level at the t -th iteration, and the larger the pheromone
amount of task s , the more important this pheromone field will play in ant decision.
while the cooperation is represented by the total pheromone τ j (t ) . Therefore, the
model of ant decision is a function of the pheromone amount τ j (t ) , heuristic
information function η j and the pheromone amount of current corresponding
task τ sj (t ) , which is formulated as
γ
 α  τ s (t ) 
 τ j (t )  η βj  j 
  
τ j (t ) 
 γ
, if j ∈ H (i )
Pi , j (t ) = 
s,k
α τ sj (t )  (3)
  τ j (t )  η j 
β

 j∈H ( i )  τ j (t ) 

0, otherwise
where H (i ) denotes the set of neighbors of pixel i , τ j (t ) is the total sum of
pheromone amount left by all ants with different tasks on pixel j , and parameters
α , β and γ regulate the relative importance of corresponding terms. It is noted
that, during the process of searching solutions, each ant of a given task is assumed to
sense some information of its neighboring pixels such as the total pheromone
amount τ j (t ) , the ratio of pheromone level of a given task to total pheromone level
τ sj (t )
, and the heuristic function η j . In the definition of IMCC, if both the relative
τ j (t )
proportion of pheromone s and the total pheromone amount keep in high level at
pixel j , the ant of task s will select the corresponding pixel j as its next position
with a lower probability than the relative proportion term is considered only, since
both the ant cooperation and competition between different tasks are in effect
simultaneously in a trade-off mode.
Heuristic Information. If an ant moves from pixel i to pixel j , the corresponding
heuristic value can be defined as
T M
 min( wi ( j ), w i ( j )))υ
1
− u (1−
ηj = e
T i =1 j =1
(4)
Where μ and υ are the adjustment coefficients designed for achieving better
likelihood difference comparison between the candidate blob and cell sample blobs,
η j lies in the range of 0 and 1, w i ( j ) denotes the value of the j -th element of w i in
cell sample pool, wi ( j ) denotes the histogram at pixel j , M is the total number of
elements in histogram w , and T is the number of cell samples in template pool.
2.3 Merge and Prune Processes

Considering the different ant pheromone fields, if more than one ant groups tend to
search for the same cell, the corresponding pheromone fields are probably partially
overlapped and the absolute distance between pheromone peaks is relatively small.
However, for those spatially distant ant groups, they naturally search for different
objects, and the peak distances between ant pheromone fields are easily discriminated
and well separated. Therefore, the overlapping ratio Ooverlap based on pheromone
peak is calculated between pheromone peak blobs and treated as a criterion. In our
experiments, the merging procedure is performed between two pheromone peak blobs
if the overlapping ratio Ooverlap > σ , where σ is threshold and set to σ = 0.3 in our
studied cell image data.
In order to remove the false alarms caused by noise and clutters, prune procedure is
employed. Suppose that we have the prior information on cell size, if the number of
ants of group is less than threshold, the prune processes is carried out and the
irrelevant object is removed. Finally, data association based on the easily-
implementing nearest neighboring method between frames is done to establish
individual trajectories of interested cells.
To visualize our proposed algorithm in a full view, we summarize the procedure in
Table 1.
248 M. Lu et al.
Table 1. Pseudo-code of our proposed algorithm (not considering data association)
Input: Image frame by frame

Generate initial distribution of ant groups by the approximate median method;
The initial ant distribution is roughly divided into N groups using K-means method;
t = 1, q sj (0) = 0,τ sj (0) = c ;
While t < tmax
For task s = 1: N
For ant k = 1: K
Ant k moves from pixel i to pixel j with a probability:
γ
 α τ s (t ) 
 τ j (t )  η βj  j 
  
 τ j (t ) 
 γ
, if j ∈ H (i )
Pi , j (t ) = 
s,k
α τ sj (t ) 
  τ j (t )  η j 
β

 j∈H ( i ) τ j (t ) 

0, otherwise
Deposit corresponding pheromone amount according to Δrjk , s (t ) = Δτ 0 ;
end
Propagated input to pixel j
P P
q1,s j (t ) = 
j ′∈H ( j ) H ( j′)
rjs′ (t ) =
H ( j ′)
  Δr
j ′∈H ( j ) k
s ,k
j′ ;
Pheromone update on each pixel at task s τ sj (t + 1) = ρτ sj (t ) + rjs (t ) + q sj (t ) ;

P
Propagated input evolution q sj (t ) = q1,s j (t ) + 
j ′∈H ( j ) H ( j ′)
q sj ′ (t ) ;
end
Total pheromone τ j (t + 1) = τ
s
s
j (t + 1) ;
end
If the blob overlap ratio between two pheromone field peaks is greater than a given
threshold then
The merge process is performed.
end
If the number of ant group is less than the given threshold then
the prune processes is carried out.
end
Output: Cell state
3 Experiments
In this section, we will test the tracking performance of our proposed algorithm in
terms of cells dividing, different dynamics and varying number in cell image
sequences. To evaluate the tracking accuracy between frames, we adopt three

measure criterions, namely, label switching rate (LSR), lost tracks ratio (LTR) and
false tracks ratio (FTR). The label switching rate is the number of label switching
events normalized over total number of ground truth tracks crossing event which
happen when two objects get very close each other(and they are sometimes merged
into one object) and after they are separated, one object would be treated as a new
object and its label is changed. The lost tracks ratio is the number of tracks lost over
total number of ground truth tracks. The false tracks ratio is the number of false object
that are tracked over total number of ground truth tracks. All experiments were
performed in MATLAB (R2012a) on a 1.7 GHz processor computer with 4G random
access memory.
Fig.1 presents an example of tracking results of selected images with our proposed
algorithm.
Frame 1 Frame 15 Frame 17 Frame 19 Frame23 Frame27
(a) The resulting ant distribution in each frame

Frame 1 Frame 15 Frame 17 Frame 19 Frame23 Frame27
(b) Tracking results of original RGB image sequences
125
120
115
110
140
105
120
100 100
80
95
60
110 115 120 125 130 135 140

40
20
20 40 60 80 100 120 140
(c) Initial ant colony distribution and the resulting of ant pheromone field in frame 23
Fig. 1. Tracking results of multi close moving cells( ρ = 0.8, P = 0.6, α = 2.5, β = 1, γ = 1.1 )
According to the tracking results presented in Fig.1(b), our proposed algorithm

could tackle the following challenging cases: cell 3 partly enter the field of view in
frame 1, then moves left, partially leaves the field of view in frames 15, and fully
leaves the field of view in frame 17. New cells 6 and cell 5 enter the field of view in
frame 15, cells 6 leaves the field of view in frame 19. Cell 4 divides into two cells
250 M. Lu et al.
(cell 4 and cell 7) in frame 23. All cells are kept on being tracked with our algorithm
in the following frames. It can be observed that initial ant distribution of three
spatially adjacent cells is adhered in frame 23, with the cooperation and compete of
our proposed algorithm all spatially adjacent cells are successfully separated and
tracked. After 50 times of iteration, the adhesion of pheromone field is well separated.
All these are illustrated in Fig. 1(c). In addition, Figs. 2 and 3 plot the position and
instant velocity estimates of each cell. It can be seen that cell 1 undergoes fast
dynamics, and cell 3 also moves quickly both in x and y direction.
140 140
cell 1
cell 2
120
120 cell 3
cell 4
100 cell 5
100 cell 6
x-coordinate[pixel]
y-coordinate[pixel]
cell 7
80
80
60
cell 1
cell 2 60
40 cell 3
cell 4
cell 5 40
20
cell 6
cell 7
0 20
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time step Time step
Fig. 2. The position estimate of each cell in x and y directions
15 4
cell 1 cell 1
cell 2 2 cell 2
cell 3 cell 3
10
cell 4 0 cell 4
X direction Velocity[pixel/sec]
Y direction Velocity[pixel/sec]
cell 5 cell 5
cell 6 -2 cell 6
5 cell 7 cell 7
-4
-6
0
-8
-10
-5
-12
-10 -14
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time step Time step
Fig. 3. The instant velocity estimate of each cell in x and y directions
To get insight into tracking performance of our proposed algorithm, we have

thoroughly compared our proposed algorithm based on mode IMCC with other three
recently developed multi-cell tracking algorithms, i.e., the particle filter (PF) [5] , the
multi-Bernoulli filter [6] and Gaussians Mixture Probabilistic Hypothesis Density
(GM-PHD) filter [7]. To perform an objective and fair comparison, both the PF and
GM-PHD filter use the same detection data obtained by a hybrid cell detection
algorithm [8] due to the fact that these two belong to “detect-before-track” methods;
meanwhile, for the multi-Bernoulli filter, the likelihood function takes the same form
as the heuristic information used in our ant system, denoted by Eq. (4), since both fall
in the category of “track-before-detect” techniques. We record all label switching
reports, tracks lost reports and false track reports in each frame over 50 Monte-Carlo
simulations, and the averaged LSR, LTR and FTR are computed as illustrated in
Table 2. According to the statistic results in Table 2, the averaged LSR, LTR and FTR
are only 1.57%, 3.17% and 1.57%, respectively, using our algorithm. The comparison
results demonstrate that our algorithm performs better than the other methods in the
case cells closely.
Table 2. Comparison results for tracking performance of various methods
Method LSR(%) LTR(%) FTR(%)

PF 15.87 9.52 19.05
Multi-Bernoulli filter 12.39 11.11 17.46
GM-PHD 9.52 6.34 3.17
Our method 1.57 3.17 1.57
Without loss of generality, we present the averaged position errors using the
manual tracking result as the ground truth. The comparison of cell 1 position error
estimates per frame by various methods is shown in Fig.4 and the same conclusions
are drawn as the above.
5 5
Our method Our method
4.5 Multi-Bernoulli filter 4.5 Multi-Bernoulli filter
Position error estimates in x-coordinate[pixel]
Position error estimates in y-coordinate[pixel]

PF PF
4 4
GM-PHD GM-PHD
3.5 3.5
3 3
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
0 0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 1 3 5 7 9 11 13 15 17 19 21 23 25 27
Time step Time step
Fig. 4. The comparison of cell 1 position error estimates by various methods
4 Conclusions
The problem of properly tracking spatially adjacent objects is one of the most difficult
issues in automated cell tracking. In this paper, a novel ant system with multiple tasks
is modeled for jointly estimating the number of cells and individual states in cell
image sequences. According to statistic results, our ant system with multiple tasks
algorithm demonstrates a robust tracking performance in terms of the measures of
LSR, LTR and FTR when comparing with other three recently developed multi-cell
tracking algorithms.
Acknowledgments. This work is supported by national natural science foundation of

China (No.61273312) and the natural science fundamental research program of higher
education colleges in Jiangsu province(No. 14KJB510001).
References
1. Dufour, A., Shinin, V., Tajbakhsh, S., Guillen-Aghion, N., Olivo-Marin, J.C., Zimmer, C.:
Segmenting and tracking fluorescent cells in dynamic 3-D microscopy with coupled active
surfaces. IEEE Transactions on Image Processing 14, 1396–1410 (2005)
252 M. Lu et al.
2. Nguyen, N.H., Keller, S., Norris, E., Huynh, T.T., Clemens, M.G., Shin, M.C.: Tracking
Colliding Cells In Vivo Microscopy. IEEE Transactions on Biomedical Engineering 58,
2391–2400 (2011)
3. Bandi, S.R., Varadharajan, A., Masthan, M.: Performance evaluation of various foreground
extraction algorithms for object detection in visual surveillance. Comput. Eng. Res. 2,
1339–1443 (2012)
4. Hartigan, J.A., Wong, M.A.: Algorithm AS 136: A K-Means Clustering Algorithm. Journal
of the Royal Statistical Society. Series C (Applied Statistics) 28, 100–108 (1979)
5. Smal, I., Draegestein, K., Galjart, N., Niessen, W., Meijering, E.: Particle Filtering for
Multiple Object Tracking in Dynamic Fluorescence Microscopy Images: Application to
Microtubule Growth Analysis. IEEE Transactions on Medical Imaging 27, 789–804 (2008)
6. Hoseinnezhad, R., Vo, B.-N., Vo, B.-T., Suter, D.: Visual tracking of numerous targets via
multi-Bernoulli filtering of image data. Pattern Recognition 45, 3625–3635 (2012)
7. Juang, R.R., Levchenko, A., Burlina, P.: Tracking cell motion using GM-PHD. In: IEEE
International Symposium on Biomedical Imaging: From Nano to Macro, ISBI 2009, pp.
1154–1157 (2009)
8. Lu, M., Xu, B., Sheng, A.: Cell automatic tracking technique with particle filter. In: Tan, Y.,
Shi, Y., Ji, Z. (eds.) ICSI 2012, Part II. LNCS, vol. 7332, pp. 589–595. Springer, Heidelberg
(2012)
A Cluster Based Method for Cell Segmentation
Fei Wang, Benlian Xu, and Mingli Lu
School of Electrical & Automatic Engineering,

Changshu Institute of Technology,215500 Changshu, China
wangleea@aliyun.com
Abstract. In the field of cell biology, cell segmentation is an essential task in

biomedical application. For this purpose, a cluster based method for cell
segmentation is proposed. Firstly, an ant colony clustering algorithm is used to
make pre-segmentation from which cell candidates are identified, then some
noise spots are filtered with area feature, after that, a novel cluster algorithm is
proposed to divide adhering cells into individuals. Finally, good results of
segmentation can be achieved. Experimental result show that the method
remains both the advantage of image segment of ant colony cluster and the
ability of further process of pre-segmentation, which improves the performance
of cell segmentation.
Keywords: Ant colony algorithm, clustering algorithm, cell segmentation.
1 Introduction
Many scientific biological applications as well as experiments for drug development

require the observation of cell responses to a variety of stimuli. Some of the responses
that need to be quantified are cell migration, cell proliferation, and cell differentiation.
The corresponding conclusions require the observation of cells over extended periods
of time. An effective way to achieve this is with microscopy images. However, the
resulting data sets of images are large and their manual analysis is tedious, subjective,
and restrictive. Thus, an automated technology for analysis is needed urgently, of all
the image based research of cells, cell segmentation is the very first step.
Many image segmentation method have been used in cell segment, like
thresholding [1], gradient based methods [2], the watershed algorithm [3], level sets
[4], dynamic programming [5] and various other pattern analysis and machine
learning algorithms [6]. Ant colony clustering algorithm is a new swarm intelligence
method that has been used in many fields, like data mining, document retrieval, image
segmentation and so on. In this paper, an ant colony clustering algorithm is applied to
achieve cell pre-segmentation process, then, a noise filter process is implemented, and
after that a novel clustering algorithm is proposed to divide the adhering cells into
individuals.
254 F. Wang, B. Xu, and M. Lu
2 Framework of Cell Segmentation
In the process of cell segmentation, an ant cluster based method is applied to execute
rough segmentation, then another cluster algorithm and features filter step are used to
improve the performance of cell segmentation, its flow chart is shown as follows.
Fig. 1. Flow chart of cell segmentation
2.1 Ant Colony Clustering

Ant colony clustering algorithm is an improved ant colony method which was
proposed by Indian researcher P.S. Shelokar. As a globally optimized heuristic
method, it is used to address clustering problems and applied to integrate data
according to the information around the clustering centers [7]. During application of
the algorithm, every data sample is treated as an ant who has different attributes, and
the process for searching food by ants is regarded as clustering procedure.
Give the initial image X, and look every pixel Xi (i = 1, 2, …, n) as an ant, every
ant stands for the feature vector of pixel. Image segmentation is the process that these
ants with different feature vector searching food source. The distance of any pixel Xi
to Xj is dij, using Euclidean distance to calculate:
m
dij =  p (x
k =1
k ik − x jk ) 2 . (1)
where m is the number of feature vector, p is the weighted factor, which is set by its
attribute and subjected to equation: p k = 1, pk ≥ 0 . τ ij (t ) is the pheromone
information value of the arc linking the data Xi to data Xj at time t, and is given as:
1 dij ≤ r
τ ij (t ) =  . (2)
0 dij > r
r is the radius of clustering. The probability of path Xi to Xj is given by the following
equation:
τ ijα (t )ηij β (t )
Pij (t ) = .
τ ijα (t )ηij β (t )
s∈S
(3)
A Cluster Based Method for Cell Segmentation 255
ηij = r dij . (4)
where ηij is the Heuristic Guide function, which reflects the similarity between the
pixels and the clustering center, α and β are regulating factors, which respectively
reflect the accumulated information during the ants moving process, and the relative
importance of the heuristic information in the ants path selecting process. If
pij (t ) ≥ p0 , Xi would be clustered into the cluster where its Xj belongs to. Let
C j = { X k | d kj ≤ r , k = 1, 2,..., J } , C j denotes the collection of data which
converge into the neighborhood of Xj . Then, the clustering center is given as:
1 J
Cj =  Xk .
J k =1
(5)
With the ants moving, the pheromone information value on every path is changing.
Through one circulation, the pheromone information value on each path is adjusted
according the following formula:
τ ij (t ) = ρτ ij (t ) + Δτ ij . (6)
where ρ represents the evaporating degree of pheromone information value with the
elapse of time, Δτ ij is the augmentation of path pheromone information value in
this circulation.
n
Δτ ij =  Δτ ijk . (7)
k =1
Δτ ijk is the pheromone information value remained by the kth ant.

The following sections discuss a new way to solve the pre-segmentation image
problems by combining a novel clustering algorithm and a noise spots filtering
process.
2.2 Refinement Process

After ant colony clustering process, a rough cell segmentation image is obtained, from
which we can get adhering cells and some noise spots. So a refinement process is
needed in which the noise spots filtering process and adhering cells segment process
are included.
Aim at filtering noise spots, we use area feature of cells which expressed as:
1 x ∈ cell
S =  sx sx =  . (8)
0 x ∉ cell
sx is unit pixel and subjects to constraints: min( S ) < S < max( S ) , where
max( S ) and min( S ) are areas that corresponding to the maximum area and
minimum area of cells. Any objects that do not satisfied cell features are considered
as noise. Comparing with noise spots, cells have relatively stable feature, which
facilitates removing mismatch objects and reduces or eliminates noise spots.
After noise filtering process, a clustering method is used to find individuals of
cells, the clustering algorithm is shown below:
1. Initialize the set of clusters to the empty set, S = φ
2. Find a cluster C in S, such that for allCi in S dist (C , xi ) ≤ dist (Ci , xi )
3. If dist (C , xi ) ≤ w , then associate xi with the cluster C. otherwise a new cluster
is created S ← S  {Cn } , where Cn is a cluster with xi
4. Repeat step 2 and 3, until now instances are left.
where xi is data samples, Ci is the cluster in S. After this step all cells should be
identified.

The experiment was made under the environment of Inter I5 2410M, win7 64bit,
Matlab 2011b. As for parameters α = 2, β = 3, r = 10, p0 = 0.9 were selected.
Fig 2 is a simple cell image which has few cells. We use level set method, k-means
method and our method to process this image, and the results are shown as follows:
(a) Original image (b) K-means cluster
(c) Level set (d) Our method

Fig. 2. Simple cell image segmentation
A Cluster Based Method for Cell Segmentation 257
From the results shown above we can tell that at this degree, three methods share
the same detection results, but level set had the longest time consuming for its
iteration. We then choose an image from a frame of the sequence images which is
shown as Fig 3(a). From the image, we can see that there are many cells, some of
which are adhering cells or the noises of background. Fig 3(b) is the gray scale image
of Fig 3(a). The gray scale image is operated by ant colony clustering method and the
pre-segmentation image is obtained as Fig 3(c).
noise
(a) Original image (b)Gray-scale image (c)Pre-segmentation

Fig. 3. Pre-segmentation of multi-cell image
In Fig 3(c), most of cells are classified as white spots. Some spots are cells, some
are noise spots, which must be filtered. Then filtering process is carried out as
mentioned in section 2, the result is shown as Fig 4(a). From Fig 4(a), we can see that
many noise spots have been erased but still one big noise spot remained. After this
step, the cluster method that proposed in this paper is used to make refinement. We
first use this method find all possible cells in the image, as we can see from Fig 4(b),
each red dot denote one cell. Then, find the connected objects, we get the conclusion
that it is the adhering cell if more than one clusters are included in one connected
object. The clusters were found with the method of cut off the line between two
(a)After feature matching (b) Clustering label (c) Our method
(d) Level set (e) K-means cluster

Fig. 4. Multi-cell segmentation
clusters to divide adhering cells, which is shown as Fig 4(c). Comparing with Fig 4(b)
and Fig 4(d), we can see that some adhering cells are separated successfully.
Fig 4(d) is cell segmentation with level set, from this image we can see that some
of cells are identified but some connected cells are clustered as one cell and some
individuals are not detected. The last figure is k-means based segmentation method,
most of cells are classified with this method, but some of them are adhering cells that
can hardly be used at subsequent applications. In addition k- means based method still
remains more noise spots compared with our method.
4 Conclusions
The cell segmentation is an important problem in biological study. A cluster based

method is proposed in this paper, which combines ant colony cluster algorithm with
image refinement process. This method remains both the advantage of image segment
of ant colony cluster and the ability of further process of pre-segmentation, which
improves the performance of cell segmentation. The experiment results confirmed the
above point that the method proposed in this paper is effective.
Acknowledgments. This work is supported by National Natural Science Foundation

of china (No.61104186) and the Natural Science Foundation of the Jiangsu Higher
Education Institutions of China (No.14KJB510001).
References
1. Xiaobo, Z., Fuhai, L., Jun, Y.: A Novel Cell Segmentation Method and Cell Phase
Identification Using Markov Model. J. IEEE Transaction on Information Technology in
Biomedicine 13(2), 152–157 (2009)
2. Li, G., Liu, T., Nie, J., Guo, L., Wong, S.T.C.: Segmentation of touching cells using
gradient flow tracking. In: 4th IEEE International Symposium on Biomedical Imaging, pp.
77–80. IEEE Press, New York (2007)
3. Muhimmah, I., Kurniawan, R., Indrayanti, I.: Automated cervical cell nuclei segmentation
using morphological operation and watershed transformation. In: 2012 IEEE International
Conference on Computational Intelligence and Cybernetics, pp. 163–167. IEEE Press,
New York (2012)
4. Bergeest, J.-P., Rohr, K.: Efficient globally optimal segmentation of cells in fluorescence
microscopy images using level sets and convex energy functionals. J. Medical Image
Analysis 16(7), 1436–1444 (2012)
5. McCullough, D.P., Gudla, P.R., Meaburn, K., Kumar, A., Kuehn, M., Lockett, S.J.: 3D
Segmentation of whole cells and cell nuclei in tissue using dynamic programming. In: 4th
IEEE International Symposium on Biomedical Imaging, pp. 276–279 (2007)
6. Chankong, T., Theera-Umpon, N., Auephanwiriyakul, S.: Automatic cervical cell
segmentation and classification in Pap smears. J. Computer Methods and Programs in
Biomedicine 13(2), 539–556 (2014)
7. Shelokar, P.S., Jayaraman, V.K., Kulkarni: An ant colony approach for clustering. J.
Analytica Chimica Acta 509(2), 187–195 (2004)
Research on Lane Departure Decision Warning Methods
Based on Machine Vision
Chuncheng Ma, Puheng Xue, and Wanping Wang
CCCC First Highway Consultants Co., Ltd., Xi’an, Shaanxi, China

{496056424,1464925547,329833165}@qq.com
Abstract. To improve the driving safety of drivers, an effective lane detection

algorithm was proposed upon the research on lane departure decision warning
system based on machine vision. Firstly, the lane images were preprocessed to
adapt to various lighting conditions and improve the efficiency of the lane
detection. Then, by means of hough transform, actual lane line features were
extracted according to the different image lane line features. Finally, after the
study of different lane departure models based on lane line detection, this article
put forward a lane departure decision algorithm. Experimental results
demonstrate that the developed system exhibits good detection performances in
recognition reliability and warning decision. It has proved that this system has
high accuracy, large detection range and high practicability.
Keywords: machine vision, lane detection, hough transform, lane departure

decision, warning methods.
1 Introduction
Safety Driving Assist (SDA) is the current concern in the research of intelligent
transportation system, which mainly solves the traffic security problems. As an
essential component in the research field of security auxiliary driving, lane departure
decision warning system gets more and more attention in recent 20 years. If the
vehicle deviates from the lane or there is any trend of vehicle deviation, the system
will warn the tired or absent-minded drivers to alter driving directions, thus reduce
lane accidents. As a result, it is of great significance for driving safety.
2 Lane Departure Decision Warning System Based on Machine

Vision
Lane departure decision warning system based on machine vision mainly consists of
CCD camera, video capture card, PC processor, alarm unit, display equipment and
other components [1]. During the high speed running of vehicles, the system uses
camera and video capture card to get lane images, and computer processes the digital
images to detect real-time image of left and right lane, on which the lane departure
decision is made. If the vehicle deviates from the normal lane or has the trend to
260 C. Ma, P. Xue, and W. Wang
deviate, the system will send warning information through display and alarm circuit,
to remind or warn the driver to alter driving direction.
3 Research on Lane Detection Algorithm
Lane detection in road image is the basis of lane departure decision warning system.
The model of lane detection is shown in Figure 1. The road detection image includes
some interferential information, while lane line is mainly in the below of image. For
comprehensive feature analysis of road image, a threshold was set, and the region
below is our research part, and only that part is processed, so that we can enhance the
real-time of lane detection.
Researches show that the vehicle lane departure decision can be made by lane field
image of myopia [2]. In order to simplify the processing, we can select an appropriate
cut point to process the lane line locating in the below of as straight line Therefore,
the lane detection becomes the detection of straight line in the region of interest,
whose algorithm includes image equalization grayscale processing, edge detection,
Hough transform, lane feature extraction and so on.
0
x
yk
ym
Fig. 1. The model diagram of lane detection
3.1 Image Equalization Grayscale Processing

Lamps and lighting directly affect the quality of machine vision system. As for the
road images under different lights, the gray values vary greatly. In strong or low light
images, the lane line presents low contrast to the background, and the range of the
gray changes is narrow. To improve the adaptability and immunity of the system, the
research applies the method of histogram equalization grayscale processing for lane
image process.
Histogram equalization is expressed as the cumulative distribution function:
k k
ni
Sk =  =  pr ( ri ) (1)
i =0 N i =0
Research on Lane Departure Decision Warning Methods Based on Machine Vision 261
In this function, i = 0,1,…k, k = 0,1,…L, L is the gray level range, pr(ri) is the
appearance probability of i level gray in the image, N is the total pixels of the road
image.
3.2 Edge Detection and Binary Image Processing

Lane detection mainly focuses on the detection of the edge of lane line, and edge
detection is realized through the use of an algorithm to extract the lines between
objects and background in the image [3]. Changes in the gray image can be reflected
by the gradient of image gray distribution, so that differential technique based on
local image can be used to obtain edge detection operator [4].
To achieve the purpose of rapid detection, Sobel operator is applied to provide
relatively accurate information about edge direction [5].
After edge detection, the road image is processed by binary to reduce the image
storage capacity, thus raise lane detection speed. Figure 2 shows the image after edge
detection and binary image processing.
Fig. 2. Edge detection and binary image processing
3.3 Image Hough Transform

The key of lane detection is to detect the line representing real lane. The common
method for straight line extraction is Hough transform, the polar coordinates equation
of straight line based on Hough transform is as follows [6]:
ρ = x cos θ + y sin θ (2)
ρ is the distance from straight line to the origin in image space; θ is the angle
between the line and the x-axis. There are two planes: xy is the image plane, ρθ is
the parameter plane. Linear equation is used to transform the points of the same line
on the image planes to a coordinate point ( ρ ,θ ) on the parameter plane. The image
after Hough transform is shown in Figure 3.
Fig. 3. The image after Hough transform
3.4 Lane Feature Extraction

From Figure 3, it can be seen that there are still some interferential lines after Hough
transform, so that it is necessary to filter out such lines. After Hough transform,
parameters of the linear can be extracted, gaining all coordinate equations as well as
the starting and ending coordinates of line segments [7].
Generally speaking, left lane line locates in the left half of the image and right lane
line in the right half; accordingly, the interest image region is divided into two parts,
left and right, to identify the coordinating lane line. In the interest image region, the
middle part is the road which is relatively flat, with minimal disruption to the Hough
transform. Moreover, the road lines locate in the middle of the road, so line searching
algorithm from bottom to up and middle to left or right is employed to extract lane
features.
Steps for feature extraction of lane lines are as follows:
1) The angles formed by the left and right lane lines with x-axes are set to α and
β ; the range of α is (100°,170°) , and range of β is (10°,80°) . At the same time
the straight lines which do not meet the requirements are removed. Detection results
are shown in Figure 4.
Fig. 4. Lane detection
2) Extract three factors: the length of straight line itself L _ Line , the length from
the middle of the image to the line L _ Center , and the length from the bottom to the
line L _ Bottom .Calculate the value of the three factors with each range as Rl ，R ， c
Rb in the left and right half of the interest region.

3) Figure out the straight lines with the biggest value of comprehensive determine
coefficient in left and right half interest region respectively. Comprehensive
determine coefficient of lane lines comes to:
3
A =  wV
i i (3)
i =1
wi is the weight of factor values, which is set to W = (0.4,0.3,0.3) , Vi is the value

of factor values.
 L _ Line / Rl i =1

V =  1 − L _ Center / Rc i=2 (4)
 1 − L _ Bottom / R i=3
 b
The straight lines with the biggest value of comprehensive determine coefficient in
the left and right half interest region are the lane lines in left and right. Extracted lane
lines are shown in Figure 5.
Fig. 5. Extraction Lane Line
4 Lane Departure Decision Warning
When vehicles are on their ways, because of environment changes, lane turning and
changing, some errors may appear, resulting in lane line changes between the current
frame and former frame in the video, or detection failure of lane line. For such
situation, the current image frames are determined to be ineffective and current lane
departure warning decision is also eliminated.
Vehicles' driving states mainly contain normal driving, slight left skew, slight right
skew, vehicle departure from left lane, and vehicle departure from the right lane. The
model of lane departure is given in Figure 6. Two auxiliary horizontal straight lines are
set as y = yb and y = yt for the lane departure decision. In the Hough transform, the
upper and lower endpoints of right and left lane segments are set as ( xLt , y Lt ) ，
( xLb , yLb ) and ( xRt , yRt ) ，( xRb , yRb ) .Then the horizontal coordinates for
intersections of left and right lane line with yb are produced, X Lb and X Rb :
 xLb
 xLt = xLb
X Lb = yb − yLb (5)
 xLb − y − y ( xLt − xLb ) xLt ≠ xLb
 Lb Lt
 xRb
 xRt = xRb
X Rb =  yb − yRb (6)
 xRb + y − y ( xRb − xRt ) xRt ≠ xRb
 Rb Rt
 xLb
 xLt = xLb
X Lt =  yt − yLb (7)
 xLb − y − y ( xLt − xLb ) xLt ≠ xLb
 Lb Lt
 xRb
 xRt = xRb
X Rt =  yt − yRb (8)
 xRb + y − y ( xRb − xRt ) xRt ≠ xRb
 Rb Rt
The middle point A of the intersection between Left and right lane line with
X Lt + X Rt
y = yt is ( , yt ) , and the middle point B of the intersection between Left
2
X + X Rt
and right lane line with y = yb is ( Lt , yt ) . The slope of the straight line
2
AB is K AB .
 ∞ X Lt + X Rt X Lb + X Rb
 = (9)
K AB = yb − yt 2 2
 X Lt + X Rt − X Lb + X Rb X Lt + X Rt X Lb + X Rb
≠
 2 2 2 2
ym ym
yt yt
yb yb
ym ym
yt yt
yb yb
ym
yt
yb
Fig. 6. The model of lane departure
After analyzing a great quantity of vehicle driving states and lane images, setting a
left and a right threshold X Lb ' and X Rb ' as well as threshold K L and K R for the
slope of straight line AB are assigned to decide whether there is lane departure.
5 Experimental Results and Analysis
The lane departure decision system based on machine vision is developed by the
combination of Visual C++ and Open CV Open computer vision library. The
processed video image size is 640*480. After a large number of simulation
experiments on 1000 continuous frames image sequences from Video, through the
lanes detection and the analysis of lane departure decision, we can come to the
conclusion as follows: the rate of detection driveway is 98.23%; the success rate of
lane departure decision is 97.64%; the time that deciding a road image nearly need
30ms. The final results show that the algorithm of lane detection and lane departure
decision can effectively detect and track lane line and accurately make the right
decision of lane departure. It is more real-time and anti-jamming. Therefore the
system can improve the safety of drivers driving, playing the role of Vehicle
Auxiliary Navigation.
6 Conclusion
The paper studies and implements a kind of lane departure decision warning system
based on machine vision. It has shown that it can effectively and instantly detect the
left, right lane lines and make right decisions on the condition of lane deviations.
With favorable ability of lane detection and lane departure decision, this approach can
satisfy the requirements of the lane departure decision warning system. Still this
algorithm needs a further improvement so as to make sure that the system has the
ability to make more accurate decision and lane detection under different atrocious
weather circumstances and promotes the lane departure decision warning system to be
perfect.
References
1. Jiuqiang, H.: Machine vision technology and applications. Higher Education Press, Beijing
(2009)
2. Jung, C.R., Kelber, C.R.: A lane departure warning system using lateral offset with
uncalibrated camera. In: Proceedings of the 2006 IEEE Intelligent Transportation Systems,
pp. 102–107. IEEE (2005)
3. Huang, S.S., Chen, C.J., Hsiao, P.Y., Fu, L.C.: On-board vision system for lane recognition
and front-vehicle detection to enhance driver’s awareness. In: 2004 IEEE International
Conference on Robotics and Automation, Proceedings. ICRA 2004, vol. 3, pp. 2456–2461.
IEEE (2004)
4. Jung, C.R., Kelber, C.R.: A lane departure warning system based on a linear-parabolic lane
model. In: 2004 IEEE Intelligent Vehicles Symposium, pp. 891–895. IEEE (2004)
5. Li, X., Zhang, W.: Research on lane departure warning system based on machine vision.
Chinese Journal of Scientific Instrument 29, 1554–1558 (2008)
6. Yang, Y., Farrell, J.A.: Magnetometer and differential carrier phase GPS-aided INS for
advanced vehicle control. IEEE Transactions on Robotics and Automation 19(2), 269–282
(2003)
7. Zhang, R., Wang, Y., Yang, R.: Researches on Road Recognition in Landsat TM Images.
Journal of Remote Sensing 2, 016 (2005)
Searching Images in a Textile Image Database
Yin-Fu Huang and Sheng-Min Lin
Department of Computer Science and Information Engineering

National Yunlin University of Science and Technology
huangyf@yuntech.edu.tw, bhujmn2007@gmail.com
Abstract. In this paper, a textile image search system is proposed to query

similar textile images in an image database. Five feature descriptors about the
color, texture, and shape defined in the MPEG-7 specification, which are
relevant to textile image characteristics, are extracted from a dataset. First, we
tune the feature weights using a genetic algorithm, based on a predefined
training dataset. Then, for each extracted feature descriptor, we use K-means to
partition it into four clusters and combine them together to obtain an MPEG-7
signature. Finally, when users input a query image, the system finds out similar
images by combining the results based on MPEG-7 signatures and the ones in
three nearest classes. The experimental results show that the similar images
returned from an image database to a query textile image are acceptable for
humans and with good quality.
Keywords: CBIR, genetic algorithm, K-means, MPEG-7 specification, weight

tuning.
1 Introduction
At present, multimedia data have played an important role in our daily life. However,
querying a multimedia database by keywords is gradually insufficient to meet users'
needs. Thus, facing a huge amount of images in an image database, content-based
image retrieval has become a popular and required demand.
In the past years, many general-purpose image retrieval systems have been
developed [5, 6, 10], and these systems rely mainly on visual features. King and Lau
used MPEG-7 descriptors to retrieve fashion clothes [5]. In order to improve query
results, Lai and Chen proposed a user-oriented image retrieval system by iteratively
interacting with users about query results [6]. Smeulders et al. presented a review of
200 references in content-based image retrieval [10].
In this paper, we propose a textile image search system for querying similar textile
images in an image database. This system consists of an offline phase and an online
phase. In the offline phase, we tune the feature weights using a genetic algorithm [11],
based on a predefined training dataset. Then, for each extracted feature descriptor, we
use K-means to partition it into four clusters and combine them together to obtain an
MPEG-7 signature [3]. In the online phase, when users input a query image, the
system extracts its MPEG-7 visual features first, and then finds out similar images by
combining the results based on MPEG-7 signatures and the ones in three nearest
classes.
268 Y.-F. Huang and S.-M. Lin
The remainder of this paper is organized as follows. In Section 2, we present the

system architecture and briefly describe the procedure of searching similar textile
images. In Section 3, we introduce five feature descriptors relevant to textile image
characteristics. In Section 4, the methods of tuning the feature weights and generating
MPEG-7 signatures are proposed to facilitate searching similar images. In Section 5,
we present the experimental results to evaluate the effectiveness of three search
modes provided in our system. Finally, we make conclusions in Section 6.
2 System Architecture
In this paper, we propose a textile image search system consisting of an offline phase
and an online phase, as shown in Fig. 1. In the offline phase, five feature descriptors
about the color, texture, and shape defined in the MPEG7 specification are extracted
from training images; i.e., ColorLayout descriptor, ColorStructure descriptor,
EdgeHistogram descriptor, HomogeneousTexture descriptor, and RegionShape
descriptors including a total of 221 dimensions. Because each feature plays a different
role in distinguishing a textile image from others, feature weights should be
determined in order that the discrimination among textile images could be boosted.
Here, we use a genetic algorithm to determine feature weights. Then, we build
MPEG-7 signatures using k-means clustering on all textile images where the weighted
Euclidean distance calculated in k-means clustering takes the feature weights
determined in the genetic algorithm.
Offline Phase
MPEG-7 Descriptors
Color Layout
Color Structure
Training Images
Edge Histogram Feature Weight Tuning
Homogeneous Texture
Region Shape
Image Database
MPEG-7
K-means Clustering Images
Signatures
Online Phase MPEG-7 Descriptors
Color Layout
Color Structure
Query Images Result Images

Edge Histogram Similarity Calculation
Homogeneous Texture
Region Shape
Fig. 1. System architecture
In the online phase, we also extract the same features, as mentioned, from a query
image. Then, we find out 1) the images with the same MPEG-7 signature as the query
image as the first candidates and 2) the images in three nearest classes to the query
image as the second candidates. Finally, we can find out result images most similar to
the query image, which appear in both groups of candidates.
Searching Images in a Textile Image Database 269
3 Feature Extraction
In this paper, we adopt the bag-of-feature MPEG-7 [1, 2, 4, 7-9] defined by the
MPEG organization, which consists of color description (i.e., two color descriptors),
texture description (i.e., two texture descriptors), and shape description (i.e., one
shape descriptor), as shown in Table 1. Among them, the descriptors relevant to
textile image characteristics are as follows.
1) ColorLayout Descriptor describes the layout and variation of colors and this
reflects the color combinations in a textile image.
2) ColorStructure Descriptor counts the contents and structure of colors in a textile
image by using a sliding window.
3) EdgeHistogram Descriptor counts the number of edge occurrences in five different
directions in a textile image.
4) HomogeneousTexture Descriptor calculates the energies of a textile image in the
frequency space, which are the level of gray-scale uniform distribution and texture
thickness, and this reflects the texture characteristics in a textile image.
5) RegionShape Descriptor relates to the spatial arrangement of points (pixels)
belonging to an object or a region in a textile image.
Table 1. MPEG-7 visual descriptor features
Type Feature description Dim. Overall statistics Total number

1 ColorLayout 12 1 12
2 ColorStructure 32 1 32
3 EdgeHistogram 5 16 80
4 HomogeneousTexture 2 31 62
5 RegionShape 35 1 35
4 Method
K-means clustering is an effective approach to find similar images for a query image,
but it is usually dependent on how well stored images are clustered. In reality, the
Euclidean distance between two images, used in K-means clustering, plays a major
role in determining good clustering. In calculating the Euclidean distance between
two images, each kind of involved features (or descriptors) mentioned in Section 3
has its own semantic in measuring their similarity; thus, these features should be
assigned with weights in measuring the similarity between two images.
In our system, a finest weight set is tuned to represent the weight of each feature
involved in an image by using a genetic algorithm. Next, we generate MPEG-7
signatures based on K-means clustering with weighted features. Finally, in the online
phase, we find out result images most similar to a query image after the similarity
calculation.
4.1 Feature Weight Tuning

In this step, we use a genetic algorithm to tune the weights of features. A genetic
algorithm is a search heuristic used to generate useful solutions to optimization and
search problems. In general, a typical genetic algorithm requires 1) a genetic
representation of a solution domain and 2) a fitness function to evaluate the solution
domain. Here, we also use the weighted Euclidean distance as the fitness function to
measure the similarity between two images as follows.
, ∑ · (1)
where A, B are two images with feature sets A , , ,…, and B
, , ,…, .
Before starting the genetic algorithm, the values of all the 221 features extracted
from an image have been normalized in the range [0, 1]. For an initial population of
100 individuals in the genetic algorithm, each one is a set of weight values randomly
generated. Only the best 20 individuals with higher fitness values can be alive in the
next generation. Then, the 20 individuals are randomly selected to generate 40
children by crossover; in crossover, each feature weight of a child is selected from the
corresponding feature weight of parent A or parent B, respectively with 50%
probability. Furthermore, in order to avoid being trapped into a local optimal, we also
generate another 40 children by crossover plus mutation. In mutation, 10% probability
of the feature weights of a child are replaced with new random values.
To measure the fitness of each individual, 24 centroids are used to represent 24
pre-defined classes of 679 training textile images, which are calculated from the
images in each class. Then, the weighted Euclidean distance can be treated as a
classifier; if a training image has the shortest distance with the centroid of a class
using the weights of individual x and the matching class is indeed the class of the
training image, 1 point is added to the score of individual x, and this score is the
fitness of individual x.
By iteratively doing so, the best individual (or best feature weights) with the
highest score would be found. The genetic algorithm will terminate when the finest
weight set becomes stable; i.e., the finest weight set is always the best for 1000
iterations. During the iterations, if a new individual has higher fitness than the old
best one, the iteration counter is reset to 0 and the new individual will be examined
for the next 1000 iterations.
4.2 MPEG-7 Signatures Based on K-means Clustering

For each extracted visual descriptor, we use K-means to partition it into four clusters
respectively, and number them from 0 to 3. Then, we combine the cluster numbers
from the five visual descriptors together and obtain a 5-digit MPEG-7 signature.
Thus, an MPEG-7 signature can represent the characteristics of an image. An MPEG-
7 signature has 5 digits and each digit can be 4 different values, so that 4 1024
bins could be used to distinguish the characteristics of images. Since K-Means
compresses images into clusters, we would be able to build an index structure more
easily using these signatures. Besides, the centroids of K-means on the five visual
descriptors are also stored for the similarity measures in the online phase.
4.3 Similarity Calculation

First, we extract the MPEG-7 visual features of a query image. Then, the similarity
measures between the visual features extracted from the query image and the recorded
centroids of K-means on the five visual descriptors are performed to determine cluster
numbers, respectively. The cluster numbers of the most similar centroids are
combined together to become the query signature. Then, we can find out the images
with the same MPEG-7 signature as the query image as the first candidates. This
approach can be treated as the similarity measures based on local views. Next, we
find out the images in three nearest classes to the query image as the second
candidates where 24 centroids of pre-defined classes have been mentioned in Section
4.1. This approach can be treated as the similarity measures based on global views.
Finally, we can find out the most similar images to the query image, which appear in
both groups of candidates.
5 Implementation
We have implemented an “Image Search Engine” system to search similar images
from an image database to a query textile image. Totally the 4069 images in the
textile image database are from Globle-Tex Co., Ltd. [13], where 679 images are
training images with pre-defined classes and the others are input to their nearest
classes subsequently according to the weighted Euclidean distance as mentioned in
Section 4.1.
5.1 User Interface

The user interface of the ISE system is shown in Fig. 2. The “Initialization” button is
used to initialize the system (or to do data preprocessing in the offline phase). The
“Input” button can be clicked to input a query image, and after the query image is
input, Area 1 will record the path name of the query image. Furthermore, the
“Execution” button is used to show result images. The radius buttons shown in Area
2 are page switches used to display result images of different pages in Area 3. The
number of radius buttons is dynamic and dependent on the number of all result
images. The check boxes shown in Area 4 are used to select a search mode; i.e., full
(or default) mode, texture-concern mode, and color-concern mode. For the texture-
concern mode, we use only the EdgeHistogram Descriptor, the HomogeneousTexture
Descriptor, and the RegionShape Descriptor to search similar images. On the
contrary, for the color-concern mode, we use only the ColorLayout Descriptor and the
ColorStructure Descriptor to search similar images.
The result images displayed in Area 3 are ranked according to the similarity degree
to a query image. The most similar image is put at the upper-left and the others are
sequentially shown in the left-to-right and top-to-bottom way. The similarity degree
to a query image is also according to the weighted Euclidean distance.
272 Y.-F. Huang and S.-M
M. Lin
Fig. 2. User interface
5.2 Similarity Evaluatiion

Here, we invite ten evaluattors to rate the similar images returned by the ISE system,
to their query images. In orrder to observe the effectiveness of the ISE system, we use
the acceptable percentage measure
m defined by Zhang et al. [12] for rating each reesult
image as a 1-to-5 scale (1: not similar, 2: poorly similar, 3: fairly similar, 4: w well
similar, and 5: strongly simmilar). Besides, we also use the quality value measuree to
evaluate the quality of resullt images as follows.
The acceptable percentage measure:
m
n3 + n4 + n5
m1 =

5 (2)
i =1 i
n
The quality value measure:
 n ×i
5
m2 = i =1 i
(3)
 n
5
i =1 i
where n1, n2, n3, n4, and n5 are the number of result images with a score of 1, 2, 33, 4,
and 5, respectively.
5.3 Experiments and Discussions

D
In Experiment 1, each evaluator tests the effectiveness of three search modes ussing
images in the database. Sinnce the interpretations of the evaluators on textures in the
same image are diverse (i.e., someone focuses on patterns, but someone focusess on
tiny formations), their meassures on the full and texture-concern modes are with maajor
differences. However, on average,
a the acceptable percentages on the full and textuure-
concern modes are 81% and a 83%, and the quality values on the full and textuure-
concern modes are 3.6 and 3.7, respectively. Furthermore, since the interpretations of
the evaluators on colors are more consistent, the average acceptable percentage and
quality value on the color-concern mode are 92% and 4.1, respectively. Thus, the
system works well in all three modes, when using images in the database.
In Experiment 2, each evaluator tests the effectiveness of three search modes using
images out of the database. We found that their measures on these three modes are a
little decreased, when compared with using images in the database. The average
acceptable percentages on the three modes are 73%, 80%, and 73%, and the average
quality values are 3.1, 3.6, and 3.6, respectively. The reason could be that the query
images out of the database do not pertain to the pre-defined classes in the system, and
the system cannot but return the similar images in three nearest classes to the query
images.
6 Conclusions
In this paper, we propose and implement a textile image search system to search
similar images from an image database to a query textile image. In the system, a finest
weight set is tuned for the extracted features involved in an image by using a genetic
algorithm. Then, we generate MPEG-7 signatures based on K-means clustering with
weighted features. In the online processing, users can find out result images most
similar to a query image after the similarity calculation. The experimental results
show that the similar images returned from an image database to a query textile image
are acceptable for humans and with good quality in all three modes.
Although our content-based ISE system can work well for searching textile images,
there are still two issues to be overcome in the future. First, the descriptors we used
here are still not good enough to describe all the classes of images in our system so
that some of them cannot be well classified in the system. Second, for a query image
without a pre-defined class, this will lead the system to return unpredictable results.
For example, when users input a car image in the worst case, the system has no way
to exclude this situation.
Acknowledgments. This work was supported by National Science Council of R.O.C.

under grant MOST 103-2221-E-224-049.
References
1. Bober, M.: MPEG-7 visual shape descriptors. IEEE Transactions on Circuits and Systems
for Video Technology 11(6), 716–719 (2001)
2. Chang, S.F., Sikora, T., Puri, A.: Overview of the MPEG-7 standard. IEEE Transactions
on Circuits and Systems for Video Technology 11(6), 688–695 (2001)
3. Huang, Y.F., Chen, H.W.: A multi-type indexing CBVR system constructed with MPEG-7
visual features. In: Zhong, N., Callaghan, V., Ghorbani, A.A., Hu, B. (eds.) AMT 2011.
LNCS, vol. 6890, pp. 71–82. Springer, Heidelberg (2011)
4. ISO/IEC 15938-3, Information Technology – Multimedia Content Description Interface-
Part3: Visual (2002)
5. King, I., Lau, T.K.: A feature-based image retrieval database for the fashion, textile, and
clothing industry in Hong Kong. In: Proc. International Symposium on Multi-technology
Information Processing, pp. 233–240 (1996)
6. Lai, C.C., Chen, Y.C.: A user-oriented image retrieval system based on interactive genetic
algorithm. IEEE Transactions on Instrumentation and Measurement 60(10), 3318–3325
(2011)
7. Manjunath, B.S., Ohm, J.R., Vasudevan, V.V., Yamada, A.: Color and texture descriptors.
IEEE Transactions on Circuits and Systems for Video Technology 11(6), 703–715 (2001)
8. Martinez, J.M., Koenen, R., Pereira, F.: MPEG-7: the generic multimedia content
description standard, part 1. IEEE Multimedia 9(2), 78–87 (2002)
9. Martinez, J.M.: MPEG-7 overview (version 10). ISO/IEC JTC1/SC29/WG11 N6828
(2004)
10. Smeulders, A.W.M., Worring, M., Santini, S., Gupta, A., Jain, R.: Content-based image
retrieval at the end of the early years. IEEE Transactions on Pattern Analysis and Machine
Intelligence 22(12), 1349–1380 (2000)
11. Whitley, D.: A genetic algorithm tutorial. Statistics and Computing 4(2), 65–85 (1994)
12. Zhang, Y., Milios, E., Zincir-Heywood, N.: Narrative text classification for automatic key
phrase extraction in web document corpora. In: Proc. the 7th Annual ACM International
Workshop on Web Information and Data Management, pp. 51–58 (2005)
13. Globle-Tex Co., Ltd., http://www.globle-tex.com/
IIS: Implicit Image Steganography
K. Jithesh1,* and P. Babu Anto2

1.
Department of Computer Science, M.G College, Iritty, Kannur Uniersity, Kerala,India
2.
Department of Information Technology, Kannur University, Kerala, India
jithukotheri@gmail.com
Abstract. In steganography secrets are imposed inside the cover medium either
by replacing bits in the spatial domain or changing the frequency domain. In-
stead the proposed Implicit Image Steganography (IIS) scheme does not alter or
replace bits in the original cover image for hiding information. As the name im-
plies there is no explicit embedding of data inside the image. Before beginning
the communication, entities should agree upon a cover image with maximum
ranges of intensity values. At least, it should contain intensity values that can
represent ASCII of all characters. Coordinate positions of each pixel with inten-
sity which can be the ASCII of a letter in the secret will be the stego-key. In this
scheme, transferring of cover image is not done as in the case of usual proce-
dures. The communicating entities have to transfer only the key. The big advan-
tage of this technique is that the cover is not required to transmit each other.
Hence nobody can even know about the cover. So it is not only difficult but
impossible to attack this communication. Also it is not required to worry
whether distortion happens while embedding.
Keywords: Cover, Steganography, Stego-key, Stego-image, Steganalysis.
1 Introduction
The standard concept and practice of ‘‘What You See Is What You Get
(WYSIWYG)’’ which we encounter sometimes while printing images or other mate-
rials, is no longer precise and would not deceive a steganographer as it does not al-
ways hold true. Images can be more than what we see with our Human Visual System
(HVS); hence, an image can convey more than merely 1000 words. For decades
people have been trying to develop innovative methods for secret communication.
Networking and digitization have become part of the technological features in the
rapid economic development of the society. The convenient and timely acquisition of
on-line services through accessing the internet has assumed the proportion of a tidal
current for individuals and organizations in their pursuit of work. However, the relay
of sensitive information via an open Internet channel increases the risk of attacks.
Thus many techniques have been proposed to deal with this problem. Data hiding,
known as information hiding, plays an important role in the information security. For
*
276 K. Jithesh and P. Babu Anto
content authentication and perceptual transparency, the main idea of data hiding is to
conceal the secret data into the cover medium, and thereby to avoid attracting the
attention of attackers in the Internet channel. The growing numbers of internet-based
applications nowadays have made digital communication an essential part of infra-
structure. Confidentiality in some digital communication is absolutely necessary when
sensitive information is being shared between two communicating parties over a pub-
lic channel.
Steganography [1, 2, 3] and Cryptography [4] are two sides of the same coin for
providing confidentiality and protecting sensitive information. The former is the art
and science of hiding sensitive information within innocuous documents in an unde-
tectable way. A thorough history of steganography can be found in the literatures. The
latter is the art and science of writing sensitive information in such a way that no one
but the intended recipient can decode it. The innocuous documents (also known as
hosts/covers/carriers) can also be of any kind as long as they do not seem suspicious.
However, with the advent of the digital technology, digital hosts such as image, audio
and video files etc. have become nowadays the most commonly used host files. On
account of their insensitivity for the human visual system, digital images [5, 6] can be
regarded as an excellent choice for hiding sensitive information. One of the most
commonly used data hiding approaches is the substitution technique [7, 8], [12, 13].
The embedding algorithm may require a secret key, referred to as stego-key.
Each and every method introduced in the field of information security has the sole
purpose of achieving the triple pillars of information security. They are confidentiali-
ty, availability and authentication. The available algorithms of steganography have
been embedding secret data inside the cover medium. A few exceptions are there like
Quantum steganography protocol [9] and Multi-party covert communication with
steganography and quantum secret sharing [10]. Abbas Cheddad et al. reported the
current techniques of image stegnography with its pros and cones [11]. The problem
associated with these methods is that they cause image distortion. Also the current
techniques lose its security when the algorithm behind the communication is once
revealed or an intruder has hacked the very content. Once a new method is introduced
it is possible for many types of attacks to happen.
A different methodology has been introduced here for the purpose of accomplish-
ing the goals of information security. The proposed study is not intended to go along
with the usual way. Usually information is hidden inside the cover by replacing the
bits or changing the frequency level. But in this scheme we do not change or make
any distortion in the spatial or in the frequency domain of the cover medium. Here the
cover will not be transmitted. Instead, before beginning, they should agree upon the
stego-image and should keep a copy of the same. They need to send only the bit posi-
tions inside the cover which comprises the secret. They will just inform each other
about the bit positions inside the image. Since they hold the same copy of the cover,
they can easily extract the secret
From this perspective the authors have tried to develop a system which can give
maximum security by avoiding steganalysis and distortion to the cover medium. Our
primary objective here is to present a new and simple steganographic scheme that
gives high security. On account of the complexities in steganography and progressive
IIS: Implicit Image Steganography 277
power of steganalysis methods, it has turned out to be a challenge to systematically

develop techniques with much better performance. Key exchanging is also a big prob-
lem in the area of communication. This study aims to deal with these problems by
proposing a scheme to increase the non-detecting power of the secret. Since the origi-
nal cover image is not transmitted it can thwart all conventional attacks. The results of
our experiments illustrate that since there is no stego-image, the steg-analysis is not
possible. Hence the security of the system would be compromised only if communi-
cating entities cheated each other.
The remainder of this paper is organized as follows: Section 2 contains a brief dis-
cussion of related works. In Section 3, the proposed IIS scheme is discussed. Section
4 presents the experimental results and performance of the proposed method, in terms
of comparison with conventional schemes. Section 5 provides a comparative visual
analysis of cover images and finally section 6 gives the conclusion of this study.
2 Related Work
Some literature proposed varieties of data hiding techniques. Most of them are irre-
versible. It means that, after the secret data are extracted from the stego image, the
original image suffers some distortion and cannot be completely reconstructed. Nev-
ertheless, in some fields (i.e., medical, military applications), the restoration of origi-
nal image is essential after extracting the embedded secret data. Therefore, reversible
data hiding schemes, also called as distortion-free data hiding or lossless data hiding,
have drawn much attention of researchers. In principle, reversible data hiding
schemes can be classified into three types, i.e., spatial domain, frequency domain, and
compressed domain. In the spatial domain schemes, all pixel values are modified
directly to embed secret data. In the frequency domain schemes, the coefficient values
of image are computed by using some transformation methods (i.e., integer discrete
cosine transform, integer wavelet transform). In the compressed domain schemes, the
original image are first compressed based on some popular compression algorithms,
such as vector quantization, block truncation coding etc. Then, according to the pecu-
liarity of compressed codes, the compressed image is encoded to conceal secret data.
There are numerous techniques available in the literatures of steganography. The
Least Significant Bit [LSB] replacement steganographic methods [7, 8], [12, 13] are
the simplest one and are widely used in the fields of information security due to its
high hiding capacity and quality. It can embed a secret bit stream into the LSB plane
of an image. LSB replacement, LSB matching (LSBM), LSB matching revised
(LSBMR) [12], and LSBMR-based edge-adaptive (LSBMR-EA) [13] image stegano-
graphy are well-known LSB steganographic methods. The LSB-replacement embed-
ding method replaces the LSB plane with embedded message bits, but the others do
not. In LSB matching, if the embedded bit does not match the LSB of the cover im-
age, then the pixel value of the corresponding pixel is randomly changed by ±1. Un-
like LSB replacement and LSBM, which embed message bits pixel by pixel, LSBMR
deals with two pixels at a time and allows fewer changes to the cover image. The
steganalysis resistance and image distortion of LSBMR are better than those of the
previous two methods. In general, the choice of embedding positions within a cover
image depends on a pseudorandom sequence without consideration of the relation-
ship.
The proposed method is a new one of its kind. From the literature survey and to the
best of our knowledge there is no such steganography scheme similar to the proposed
one. In the available technologies of steganography the embedding is done inside the
cover image. A work which is an exception and which can be related to the proposed
work in terms of not embedding secret inside an image is introduced by Guo et al. [9]
in 2003 with the title Quantum Secret Sharing without Entanglement. Also, Xin Liao
et al. [10] in 2010 presented a novel multi-party covert communication scheme by
elegantly integrating steganography into Guo et al.’s QSS. This scheme is good in
terms of security but the payload of this method is comparatively very less. That is, it
communicates only one bit per transaction. Nevertheless, their idea of not embed-
ding secret inside a digital image motivated us to introduce a method which is better
in payload capacity and security. The proposed IIS scheme does nothing over the
cover image. It only keeps a location map. Another important feature of the proposed
stego system is that it does not require a secret key. Thus, the constructions presented
demonstrate that in order to achieve perfect steganographic security no secret has to
be shared between the communicating parties. The main idea behind the stegosystems
we propose is to conceal the cover from outside.
3 Proposed IIS Scheme
As stated earlier this is a simple but innovative technique. Here nothing is done with
the cover, but copy of the cover is shared by the communicating entities. The algo-
rithm is as follows:
1. Select appropriate cover image carefully.
2. It can be either grey-scale or color.
3. If possible it should contain intensity values that can easily represent ASCII of
each character.
4. Share the copy of the digital cover medium with the other end.
5. Exchanging of the cover should be made very confidential.
6. Hand to hand exchange is more reliable and secure, otherwise use a trusted third
party.
7. After the successful exchanging of the cover the communication can be started.
8. Take the secret.
9. Digitize it.
10. Find out pixels which can fully represent characters of the secret.
11. Mark up the coordinates of the respective pixels.
12. This is treated as the stego-key.
13. If pixel [intensity] values are not enough to represent all characters then try to find
consecutive bits of a pixel inside the image to represent such letters.
14. If consecutive bits of a pixel are used, the starting and ending position is required
to be kept.
15. In such case two or more pixels can be used to hold a character.
16. In such cases key should be the combination of coordinates and starting as well as
ending positions of the bits.
17. If color image is used coordinate position and RGB position should be the key.
18. If RGB is used a single pixel can represent 3 characters a time.
19. Stego-key can be sent to the other side with or without encryption.
20. With the key the receiving end can easily extract the secret from the copy of the
image.
Always there will be new threat in the field of communication. So no technique is

good for a long time. A method that combines both logic and craft can only survive in
the contest of constant changes in the field. In order to improve the security of the
secret communication, here we propose a technique which gives better performance.
In this section, we present the results from the experiments we conducted in order to
evaluate the reliability of this method. We have used both color and grey-scale im-
ages. Care has been taken while selecting images. That is, when grey-scale image is
the cover; it is assured that it contains maximum shades. Images with 0-255 ranges of
intensity are the optimal case. Any character of English language can be represented
with 0-255 ASCII values. So a grey scale image is enough. In a grey-scale cover,
coordinate positions will be the stego-key. When the cover is a color image it is poss-
ible to represent characters of all human languages. If range of color variations of the
image is more, it is very easy to represent secret.
As mentioned earlier, this method does not require anything to be done over the
cover image. It requires selecting a cover image of grey-scale or color with maximum
ranges. Then, hand over the copy of the same image to the other entity that belongs to
the same channel of communication. The exchanging should be very confidential. It is
better to exchange in a handshake mode. If it is not possible, trust a third party. The
secrecy and security of the communication rely on the cover transfer. If it fails, the
entire security will be compromised. As the title of this paper indicates, there is no
direct or explicit hiding of information inside the cover. In fact, it is a steganogrpahy
without steganography. Here follows a case in point.
Let the secret to be transmitted is “Pay”.
The binary form of the information is
P = 80  1010000, a = 97  1100001, y = 121 1111001
Fig.1 is the stego-image. Make a copy and hand it over to the other end. As men-
tioned the security of this approach lies in this step. Information about the stego-
image should not be revealed at any cost. If it is revealed, the entire essence of the
security will be easily broken.
In this experiment the first letter is P. Its binary is 101000. Select the appropriate
pixel position from the cover. We have a number of techniques to select the pixel
coordinate from an image with a single mouse click. For example matlab provides
impixelinfo function for the same. Keep the coordinate position as the stego-key. The
(x, y) coordinates of letters inside the image are (101, 63), (119, 60) and (149, 29)
respectively. Here the stego-key is 101631196014929. Send this stego-key to the
other end. We can use either public key or private key cryptography to transfer the
key. Public key cryptosystem is the best method to exchange the key. Even if the key
is lost or broken, it does not affect the security of the system. As far as the cover im-
age is not known to anybody else other than communicating entities, the confidentiali-
ty of the information is preserved at the maximum. It is possible to send the stego-key
without doing any encryption or hiding.
The proposed scheme is compared with the currently popular steganographic
schemes namely, Secure Bit-Plane Based Steganography for Secret Communication
[14], Adaptive Image Steganography Based on Depth-Varying Embedding [15] and A
Novel Technique for Image Steganography Based on a High Payload Method and
Edge Detection [16]. Table 1 shows the results of comparison. The default optimal
parameter settings suggested in the respective works are adapted for experimentation.
Fig. 1. Cover or Stego Image
Table 1. Comparison between conventional methods and IIS

Factors Proposed IIS Scheme Conventional schemes
Distortion NIL Caused due to hid-

ing
Payload Large Comparatively less
Security high Comparatively less
Key exchange It Can be revealed Big problem
Authentication Very easy Difficult.
5 Visual Analysis
As indicated earlier, the available steganography leads to visual distortions to the

original view of the cover image. Fig.2 shows the distortion caused to the image Pep-
per while using popular Chang et al.’s [17] steganography scheme. It is obvious that
using any scheme except the one proposed by Xiu-Bo Chen et al. will lead to image
distortion. But Xiu-Bo Chen et al.’s scheme has other limitation in that it can only
send one bit of data at a single transaction.
To evaluate the visual quality of stego images by using the human eye, we have en-
larged the partial area of original cover image and stego-image of Chang et al.’s
scheme, as shown in Fig.2. Fig.2 (a) shows the cropped area in the original Pepper
image and Fig.2 (b) shows the cropped area in the stego-image. The distortion be-
tween original cover image and stego-image is visually almost imperceptible. The
table.2 shows the MSE and PSNR of popular Yang et al.’s Scheme [18], Lin-Tsai
Scheme [19] and Chang et al.’s Scheme. The results reveal that even though they are
good techniques, they are not free from image distortion.
Steganalysis is an interesting topic that focuses on the detection of the presence of
embedded secret messages. RS attack [20], proposed by Fridrich et al., and x2 detec-
tion [21], proposed by Westfeld and Pfitzmann, are the two most effective LSB stega-
nalytic techniques. The RS-attack technique can detect both sequentially embedded
messages and randomly embedded messages. No current steganogrpahy method can
claim that they are free of image distortion while embedding secret inside because
PSNR never becomes infinity. Hence these techniques cannot escape from steganaly-
sis. They are not absolutely secure against attacks. Now that the proposed method
never embeds a secret directly or explicitly to an image, it can claim that it is free of
any such kinds of steganalysis. Enhancing the visual quality and authentication capa-
bility of stego-image are new challenges for researchers working on the development
of novel secret image sharing scheme. The results and discussion clearly indicate that
the proposed scheme provide high confidentiality and authentication.
Table 2. The PSNR and MSE of stego-images with the same payload capacity (The unit of
PSNR is db)
Secret image Stego-image Lin-Tsai Yang et al’s Chang et al.’s
(256×256) (512× 512) Scheme Scheme Scheme
MSE PSNR MSE PSNR MSE PSNR
Cameraman 7.74 39.25 4.43 41.66 5.49 40.73

Baboon 7.85 39.18 4.55 41.55 6.59 39.94
General test Lena 7.80 39.20 4.50 41.60 5.98 40.37
pattern Pepper 7.86 39.17 4.54 41.56 7.63 39.30
Sailboat 7.89 39.16 4.59 41.51 8.45 38.86
Average 7.83 39.19 4.52 41.58 6.83 39.84
(a) (b)
Fig. 2. Cover image and stego image of Change et al.’s scheme
6 Conclusion
Here a novel spatial steganography scheme realized with IIS paradigm is introduced.
It can be also referred as steganography without steganography. Nothing is done with
the cover image. Hence, as usual distortion to the medium is not caused. In fact, it is a
different strategy to achieve high security to communication. Its biggest advantage is
that it is impossible to break the confidentiality even if we lost the key. Since stego-
key alone can do nothing, it is not necessary to encrypt or hide the key. The greatest
risk of this method lies in transmitting the cover. Exchanging by hand is absolutely
secure. The additional concern of this approach is to get images with ranges varying
from 0 to 255. For any stego-system the next property to be considered after its secu-
rity is its capacity. The capacity of a stego-system can be defined as the number of
hidden bits transmitted per letter of the cover image. We show that our stego-system
has the maximum possible capacity: Though it can adopt large payload, it is optimal
for small secrets. In public key cryptography like RSA and others the length of the
key is very big. It is common to have a key length of 1024 bits and more. In decimal
notation it can easily exceed 300 digits. Here a key length with 300 digits can easily
represent a secret of about 150 letters. The length of the key is directly proportional to
the secret to be encoded. The common problem of recovering the original cover im-
age from the stego-image is also resolved here. Further studies can be undertaken to
increase the payload. This study has also taken advantages of human psychology in its
formulation so that it become more effective and practical. Since the cover image is
not transmitted all kinds of conventional threats become insignificant and ineffective
thereby making it more viable for its purpose.
References
1. Johnson, N.F., Jajodia, S.: Exploring steganography: seeing the unseen. IEEE Comput-
er 31(2), 26–34 (1998)
2. Shih, Y.F.: Digital Watermarking and Steganography Fundamentals and Techniques. CRC
Press, Taylor & Francis Group, Boca Raton (2008)
3. Petitcolas, F.A.P., Stefan, K.: Information Hiding Techniques for Steganography and Digi-
tal Watermarking, Canton Street, p. 685. Artech House, Inc., Norwood (2000)
4. Stalling, W. (ed.) Cryptography and Network Security-Principles and Practices, 4th edn.,
Pearson Prentice Hall, India P. Ltd., India (2000)
5. Salivahanan, S.: Digital Signal Processing. Tata McGraw-Hill, India (2000)
6. Gonzalez, Woods: Digital Image Processing, 3rd edn. Prentice Hall, USA (2008)
7. Hu, C.H., Lou, D.C.: LSB Steganographic method based on reversible histogram transfor-
mation function for resisting statistical steganalysis. Information Sciences 188, 346–358
(2012)
8. Chan, C.K., Chen, L.M.: Hiding data in images by simple LSB substitution. Pattern Rec-
ognition 37, 469–474 (2004)
9. Bo, X.C., Guo, Z.Q., Jie, X.Z., Xin, X.N., Xian, Y.Y.: Novel quantum steganography with
large payload. Optics Communications 283, 4782–4786 (2010)
10. Liao, X., Wena, Q.V., Suna, Y., Zhangb, J.: Multi-party covert communication with stega-
nography and quantum secret sharing. The Journal of Systems and Software 83, 1801–
1804 (2010)
11. Cheddad, A., Condell, J., Curran, K., McKevitt, P.: Digital image steganography: Survey
and analysis of current methods. Signal Processing 90, 727–752 (2010)
12. Mielikainen, J.: LSB Matching Revisited. IEEE Signal Process. Lett. 13(5), 285–287
(2006)
13. Huang, J., Luo, W.: Edge adaptive image steganography based on LSB matching revisited.
IEEE Trans. Inf. Forens. Security 5, 201–214 (2010)
14. Cong-Nguyen, B., Lee, H.Y., Joo, J.J.C., Lee, H.K.: Secure bit-plane based steganography
for secret communication. IEICE Transactions on Information and Systems E93, 79 (2010)
15. He, J., Tang, S., Wu, T.: An adaptive image steganography based on depth-varying em-
bedding. In: Image and Signal Processsing, CISP 2008, vol. 5, pp. 660–663 (2008)
16. Halkidis, S.T., Ioannidou, A., Stephanides, G.: A novel technique for image steganography
based on high payload method and edge detection. Expert Systems with Application 39,
11517–11524 (2012)
17. Chang, C.C., Hsieh, Y.P., Lin, C.H.: Sharing secrets in stego images with authentication.
Pattern Recognition 41, 3130–3137 (2008)
18. Yang, C.N., Chen, T.S., Yu, K.H., Wang, C.C.: Improvements of image sharing with ste-
ganography and authentication. Journal of Systems and Software 80, 1070–1076 (2007)
19. Lin, C.C., Tsai, W.H.: Secret image sharing with steganography and authentication. Jour-
nal of Systems and Software 73, 405–414 (2004)
20. Fridrich, J., Goljan, M., Du, R.: Reliable detection of LSB steganography in color and
grayscale images. In: Proceedings ACM Workshop Multimedia and Security, pp. 27–30
(2001)
21. Westfeld, A., Pfitzmann, A.: Attacks on steganographic systems. In: Proceedings of the
3rd International Workshop on Information Hiding, Dresden, Germany, pp. 61–76 (1999)
Humanized Game Design Based on Augmented Reality
Yanhui Su1,2, Shuai Li1, and Yongsong Zhan1

1
Guangxi Key Laboratory of Trusted Software, Guilin University of Electronic Technology,
Guilin, 541004, China
2
Institute of Computer, Zhejiang University, Hang Zhou, 310058, China
suyanhui@gmail.com
Abstract. Currently, the primary research of AR (Augmented Reality) is focus

on how to improve the accuracy of identification and reduce the dependency to
markers. Successful cases are scarce on how to play the special advantage of
augmented reality, combine it with mobile devices and generate practical value.
We start form the actuality of AR technology characteristics, achieved the first
blind game system which based on mobile platform Android. This system is not
only regard game`s emotion design as starting point and pay closely attention to
the vulnerable groups, especially the blind community, but also provide them
opportunities that compete with average person. Besides, it offers a
diversification choice of game against for blind users by set different difficulty
level and game player modes. The user experience indicates that the system
combines AR and the innovation of game design together, and help vulnerable
group achieve flexible user experience and operation.
Keywords: Augmented reality, Game design, Humanized design, Emotional

design.
1 Introduction
Augmented Reality (AR) is an emerging field and a hot spot of research in recent
years as an important branch of virtual reality technology. Based on information in
real scenarios, AR technology superposed virtual objects or other information that
computer generated to the real scenario and fused them. By this way, a bridge would
be set up between virtual and reality world, thus, we could realize the “Augment” to
reality world and presented a new environment with real sensory effect to users.
AR technology has a widely application and a more obvious superiority than VR
technology. Zhongwang Jiang`s article[1] introduced the development history and
application field of AR technology in detail. Sui Yi`s paper[2] introduced the AR
technology based on a handheld device. From 1990, Tom Caudell and David Mizell,
engineer of Boeing Co, proposed the concept of “Augmented reality” firstly when
they designed the auxiliary wiring system[3]. HIT laboratory at university of
Washington released a develop tool of AR system which is named ARToolKit in
1999. Now, just several decades, AR technology has formed a relatively complete
workflow and implementation system, its basic principle is shown as Fig.1.
Humanized Game Design Based on Augmented Reality 285
Games developed rapidly as an emerging industry and various types of games

emerged, but games’ control mode and scene effect still changed little. In recent
years, with the improvement of smartphone`s processing capacity and PDA, AR
technology applied to game gradually, so that the game scene is more realistic and
users` interactive experience is more immersive. In 2000, Bruce H. Thomas
developed the first outdoor mobile AR game which is named AR Quake. In 2004,
Mohring[4] et al developed the first application which could completely do all the AR
tasks by smart phone. In recent years, research of AR mobile game attracted more and
more scholars` attention, and kinds of handheld games based on AR technology
constantly emerge. Mobile Maze game[5] require user through the maze by push a car
ball. In 2005, Video Processing at VIT developed a multi-user table tennis game
which is named Symball[6]. HIT lab NZ developed an AR Tennis game[7]. In 2006,
Siemens developed a free throw game, AR Soccer[8]. Besides, Kurt D.Squire team
developed an AR scientific education game, Mad City Mystery[9]. From last year, a
series of mobile AR game appeared constantly, such as Drop Defender, Zombie
Room AR and so on. The application of AR technology made our visual information
more intuitive and richer, the real-time interactive experience more real, it will be a
development tendency for mobile application in the future.
Fig. 1. Augmented reality principle
2 Humanization of Game Design

Except the play ability, game design also needs to pay attention to players` emotional
needs. The book “Creating Emotion in Games”[10] told us that emotional design
could help game players to realize the additional value of game as an important part
of game design, especially for vulnerable groups. Such as: blind person, how to let
them to fully experience game`s joy as normal people, the key point is to provide
them the opportunities that competitive with normal players. So Shaking World, a
puzzle action game which we will introduce in this paper, used the third-person
design and AR technology, built a fair competitive platform for blind players.
Through the humanized touch board design and with the help of the blind touch
habits, blind community could fully experience fun of this game.
286 Y. Su, S. Li, and Y. Zhan
3 Game System and Mechanism
3.1 Game System

Shaking World is a fantastic third-person puzzle action game developed by Unity 3d.
It has two modes: online mode or console mode, where online mode set up two
characters in the same scene, player A and player B. Player A is responsible for shaking
the game world and player B should try best to control the character which named little
fool against the shaking. They can earn coins and buy props they like by play games.
Besides, Shaking World also designed custom mode what players can increase the fun
by building themselves level with game props. In a word, this game easy to play,
numerous levels plus thrilling scenes make it the best choice for recreation.
Fig. 2. Game system
3.2 Game Props Design and Mechanism Design

Game set in a lot of props in order to increase its playability, such as: bomb, alarm
clock, lollipop and so on. The relationship of every element in the game is as shown
in Fig. 2. Where, mostly props used to decorate the scene or hinder player B win the
game. For example, if character falling into the swimming pool, player B needs a
quick click on the character to climb back on the ground, but if player B’s clicking too
slowly, character will drown and the player A will win the game; The fan has a fixed
direction to blow, when the character close to the source of wind, he will be blown
away; If player A bought bombs and let player B touched it, player B will be killed,
and player A will win the game, but if player B bought the alarm clock and touched it,
player A will lose control of the game world 3 seconds, and he will win if he lets the
character get the lollipop. In a word, what Player A need to do is that he should try his
best to shake player B out of the game world and do the winner, as shown in the left
of Fig. 3. On the contrary, player B should make a great effort to overcome player A`s
shaking and get the lollipop successfully, as shown in the right of Fig. 3.
Fig. 3. Game mechanism design
3.3 Technology Implementation

Vuforia SDK is based on AR application and display mobile device as a “magic
camera” or show the scene as a world that coexistence of virtual and reality. We used
this SDK in Shaking World, adopted the AR technology which based on mobile
terminal and transplanted the style of AR to cellphone platform. The virtual scene
could be stacked to the paper when used the camera of cellphone to shoot it. Besides,
in order to assure the consistency between operation and braille, we used the style of
virtual button, set a touch area on the paper for blind person. Game runs on Android
platform, and its technical principle hierarchy Chart is showed as Fig. 4. Players take
photos by camera, find out the target image and determine its coordinate; besides,
identify the target images, overlay virtual image on real image and use virtual button
interactive with real world. Its bottom technology is as follows: bottom is on Android
system, around the operating system to build two big modules, including: camera and
rendering module; besides, on the basis of the camera module, we add target tracking
module, including: virtual buttons, multi-target tracking, target image recognition,
space target recognition and so on, and in the form of SDK provides the chance that
developers call low-level interface, can achieve all kinds of special effects effectively.
Fig. 4. Bottom framework of AR engine
Virtual Button System. Virtual buttons are developer-defined rectangular regions on

image targets that when touched or occluded in the camera view, can trigger an event.
Virtual buttons can be used to implement events such as a button press or to detect if
specific areas of the image target are covered by an object. Virtual buttons are
evaluated only if the button area is in the camera view and the camera is steady.
Evaluation of virtual buttons is disabled during fast camera movements. Define
virtual buttons in the Database Configuration XML as children of an image target. To
add virtual buttons, insert a section similar to the following:
<ImageTarget size="247 173" name="wood">
<VirtualButton name="red" rectangle="-108.68 -53.52 -
75.75 -65.87" enabled="true" />
<VirtualButton name="blue" rectangle="-45.28 -53.52 -
12.35 -65.87" enabled="true" />
</ImageTarget>
The Virtual Button state can be requested from active targets in the scene by
iterating through the button child objects:
// Iterate through this targets virtual buttons:
for (int i = 0; i< target->getNumVirtualButtons(); ++i)
{ constVirtual Button* button = target-
>getVirtualButton(i);
if (button->isPressed())
{ textureIndex = i+1;
break; }
}
User-defined the identified image. In this section we show how to use the user-
defined target feature to instantiate objects of classes from TrackableSource which
can be used to create new Trackables at runtime.
Two new classes, ImageTargetBuilder and ImageTargetBuilderState are
introduced: where, class ImageTargetBuilder exposes an API for controlling the
building progress, retrieving a TrackableSource for instantiating a new trackable upon
successful completion. The flow chart is as shown in Fig. 5.
Fig. 5. The process of instantiate object

Expression form of identified images. The expression forms of identified images are
as follows:
?xml version="1.0" encoding="UTF-8"?>
<QCARConfigxmlns:xsi="http://www.w3.org/2001/XMLSchema-
instance"
xsi:noNamespaceSchemaLocation="qcar_config.xsd">
<Tracking>
<ImageTarget size="247 173" name="stones" />
<ImageTarget size="247 173" name="chips" />
</Tracking>
</QCARConfig>
4 Game Results
Relative to other games, Shaking World has few bright spots: (1)Pay attention to
user`s emotional experiences, provide humanized attention and care design for
vulnerable groups. (2) Using augmented reality mode to increase fun of the game and
expand the original dimension of game experience which based on screen to a three-
dimensional game space. (3) Using the virtual button mode of AR technology,
provided game operation based on braille contact for blind user.(4) Combined the
play ways of virtual and reality, and provide an interaction with virtual game by
manipulating physical paper. (5) With the aid of AR technology and use the levels
mode of blind person book, blind man could touch the dots on graph to operate this
game. Besides, different maps represent different level, so with different maps, blind
person can take part in every level of game as normal people.
The game Shaking World which based on AR technology could achieve different
effects. The left of Fig. 3 stands that player B was shaking out the game world, the
right of Fig. 3 stands that the character eats the lollipop. Fig.6 is the whole scene of
this game. Fig.7 is the game mode that designed for blind person. They could play the
game by touch dots on the graph paper when run the game. Every elements of this
game drawn exquisite, character designed personality, and visual effect is preferably.
Fig. 6. Humanized design of game scene

Fig. 7. Blind person mode of the AR game
5 Conclusion
Shaking world game used the AR technology, fully considered humanized design, the
research result indicated that it has a certain appeal and also is a better choice for
entertainment. Beyond that, it also has much commercial value: (1) Convenient
development peripheral products: the “Shaking World” building blocks, gift boxes,
store content boxes, alarm clocks and so on. (2) Can implant the SDK advertising on
the top of houses in the game scenes and don't destroy the beautiful game. (3) Can
implant some products in the art style of the game for joint operations of products. (4)
Can promote to many operators, such as: App Store, CUCC, CM, CHA. (5)To
cooperate with shopping website: Play the game to earn the points or gold of the
shopping website. The more important is that the game also has much public value:
(1) Pay attention to vulnerable groups, to provide fair and competitive opportunities
for the blind, let them feel the game happiness. (2)With the help of a third party social
media channels, for example the game feeds, cause people concern for the vulnerable
groups and give them more help.
，
In the future AR technology is an emerging and active field of research. It could
bring people new visual experience, and has a broad market prospect and application
scope. For the game industry, apply the AR technology to game could greatly rich the
content of game frames and increase game`s entertainment.
Acknowledgments. This research work is supported by the grant of Guangxi science

and technology development project (No: 1355011-5), the grant of Guangxi Key
Laboratory of Trusted Software of Guilin University of Electronic Technology (No:
kx201309), the grant of Guangxi Education Department (No: SK13LX139) the，
grant of Guangxi Undergraduate Training Programs for Innovation and
Entrepreneurship (No: 20121059519).
References
1. Jiang, Z.: The development of Educational Augmented Reality Game. East China Normal
University, Shang Hai (April 2012)
2. Sui, Y.: Research and application of augmented reality technology based on handheld
device. Qing Dao University, Qing Dao (March 6, 2009)
3. Caudell, T.P., Mizell, D.W.: Augmented reality: An application of heads-up display
technology to manual manufacturing processes. In: 1992. Proceedings of the Twenty-Fifth
Hawaii International Conference on System Sciences, vol. 2, pp. 659–669. IEEE (1992)
4. Mohring, M., Lessig, C., Bimber, O.: Video see-through ar on consumer cell-phones. In:
Proceedings of the 3rd IEEE/ACM International Symposium on Mixed and Augmented
Reality, pp. 252–253. IEEE Computer Society (2004)
5. Bucolo, S., Billinghurst, M., Sickinger, D.: Mobile maze: a comparison of camera based
mobile game human interfaces. In: Proceedings of the 7th International Conference on
Human Computer Interaction with Mobile Devices & Services, pp. 329–330 (2005)
6. Hakkarainen, M., Woodward, C.: SymBall: camera driven table tennis for mobile phones.
In: ACM International Conference Proceeding Series, vol. 265, pp. 391–392 (2005)
7. Henrysson, A., Billinghurst, M., Ollila, M.: Face to face collaborative AR on mobile
phones. In: 2005. Proceedings. Fourth IEEE and ACM International Symposium on Mixed
and Augmented Reality, pp. 80–89. IEEE (2005)
8. Geiger, C., Paelke, V., Reimann, C.: Mobile entertainment computing. In: Göbel, S.,
Spierling, U., Hoffmann, A., Iurgel, I., Schneider, O., Dechau, J., Feix, A. (eds.) TIDSE
2004. LNCS, vol. 3105, pp. 142–147. Springer, Heidelberg (2004)
9. Squire, K.D., Jan, M.: Mad City Mystery: Developing scientific argumentation skills with
a place-based augmented reality game on handheld computers. Journal of Science
Education and Technology 16(1), 5–29 (2007)
10. Freeman, D.: Creating Emotion in Games. Beijing Hope Electronic Press, Beijing (2005)
A Hybrid PSO-DE Algorithm
for Smart Home Energy Management
Yantai Huang1,*, Lei Wang1,2, and Qidi Wu1

1
Department of Electronics and Information Engineering, Tongji University, Shanghai, China
huangyantai@sina.com
2
Shanghai Key Laboratory of Financial Information Technology, Shanghai, China
wanglei@tongji.edu.cn
Abstract. Home energy management system is an important part in smart

home, smart home is assumed to be equipped with smart meter which smart
control of generators, storages, and demand response programs. In this paper,
we study a versatile convex optimization framework for the automatic energy
management of various household loads in a smart home. The scheduling
algorithm determines how energy resources available to the end-users
considering a number of constraints. Hence, a hybrid PSO-DE algorithm
approach is proposed. We devise a model accounting for a typical household
user, and present computational results showing that it can be efficiently solved
in real-life instances.
Keywords: smart home, home energy management, hybrid PSO-DE algorithm.
1 Introduction
With the increase of the electric demand, energy crisis worldwide is one of the most
serious challenges in the 21st century. The residential sector is experiencing the
strongest increase on its electric demand. Therefore, the study of the electric demand
in the residential sector is an important task for the electric grid controllers as already
shown in [1, 2].
Home energy management system provides an opportunity for improving the
efficiency of energy consumption in residential sector [3, 4]. A typical Smart home
for residential sector integrates the operation of electrical and thermal energy supply
and demand.
Developing efficient demand response models of electrical appliances is a key
problem in an energy management system in a smart home, which have received
considerable attention recently [5-12]. In general, the main objective of Home energy
management system is to minimize the electricity bill or maximize their users’
satisfaction by allocating available resources and managing the load of appliances.
The home energy management in a smart home can be formulated as a complex
mathematical optimization problem. Dynamic programming may be used if the
*
A Hybrid PSO-DE Algorithm for Smart Home Energy Management 293
optimal schedule is required, however, the immense computational effort involved

might require a powerful computer and long computation times. On the other hand,
the complex optimization problem can be solved by the commercial solver such as
CPLEX and MOSEK. However, these commercial solvers are specialized
optimization package whose computational requirements are not suitable for a smart
meter. Thus, heuristic techniques have been applied to the original problem for
obtaining a global optimal solution to the problem. Thus, this paper proposes a
particle swarm optimization algorithm to solve the original problem for obtaining a
global optimal solution to the problem. In order to improve the performance of the
algorithm, we integrate particle swarm optimization with differential evolution (PSO-
DE) to solve constrained numerical and engineering optimization problems.
The rest of this paper is organized as follows. In Section 3, the mathematical
problem formulation is presented. In Section 4, the scheduling problem is resolved by
a hybrid PSO-DE algorithm. Section 5 includes necessary figures, results, and
discussions of the typical study case. The conclusion is drawn in Section 6.
References are appended last.
2 Mathematical Formulation
It is assumed that most electric appliances are networked together and are controlled
by a home energy management system. The smart home consists of micro-CHP unit,
storage devices, local loads, thermal loads and must run electrical loads. The output
power energy of micro-CHP unit is supplied to the kinds of appliance loads, the
surplus power energy is used to charge storage batteries or sail to the main grid. In
contrast, the deficit power energy is provided by batteries and main grid. In this paper,
the whole problem of the smart home energy management system is defined as an
optimization problem. In so doing, we have used simple models of the micro grid’s
components which will be described in the following subsections.
2.1 Objective Function

The objective function is to minimize the operation cost: the cost of the power
purchased from the main grid during a day, and the cost of gas consumed by the
micro-CHP unit (GCHP). In this study, the micro-CHP’s startup time and consequent
cost is neglected.
24
co s t =  [T O U ( h ) p
h =1
g rid ( h ) + G p G C H P ( h )] (1)
where TOU(h) is the main grid’s time-of-use price tariff; Pgrid(h) is the power
transferred between main grid and smart home; Gp is the natural gas price; And
Gchp(h) is the micro-CHP’s consumed gas.
2.2 Constraints
Electrical Demand Supply Balance. The loads contain thermal appliances (hot water
and air conditioner) and electrical appliances. We assume electrical appliances are
must run loads that can not be scheduled. While thermal appliances consumption can
be scheduled to avoid peak hours. For each household, let Pms(h) denote the total
294 Y. Huang, L. Wang, and Q. Wu
energy consumption of electrical(must run) appliances. while pheat(h) and pwater(h)

represent the air conditioner and hot water energy consumption one-hour respectively.
Pbatt(h) is the battery’s output power and Pchp(h) denote the micro-CHP’s electrical
output power. So, electrical demand-supply balance constraint would be as follow:
Pgrid (h) − pbatt (h) + pCHP (h) = pms (h) + pheat (h) + pwater (h) (2)
Constraint of Battery. The battery can charge or discharge at different condition, so
their can avoid peak time energy and can more efficient use of the overall energy.
ch
pbatt ( h) (3)
0≤ ≤ pchmax
ηch
0 ≤ pbatt
dch
(h)ηdch ≤ pdch
max
(4)
ch
p ( h) (5)
pbatt (h) = batt
− pbatt
dch
(h)ηdch
ηch
ch
pbatt (h) − pbatt
dch
( h) (6)
SOC (h + 1) = SOC (h) +
Ebatt
SOC min ≤ SOC (h) ≤ SOC max (7)
min max
where SOC(h) is the state of charge in the battery; SOC and SOC are the
minimum and the maximum SOC, respectively; Pbatt ch
(h) is the battery’s charging
power; Pbatt
dch
(h) is the batter’s discharging power; η ch
and ηdch are the battery’s
charging and discharging efficiency, respectively.
Desired Hot Water and Building Temperature. The building is modeled within the
context of desired hot water temperature and building temperature.
In modeling the hot water storage, the energy equivalent of the hot water storage at
each time step is taken in consideration and the dynamic of the water flow is not
considered. So, the water storage temperature at each hour is calculated as the
following equation [13], [8]:
Vcold (h ) *(Tstcold − Tst ( h)) + Vtotal * Tst ( h) H air (h ) (8)
Tst ( h + 1) = +
Vtotal Vtotal * Cwater
T min
st ≤ Tst (h) ≤ T max
st
(9)
where Tstmin and Tstmax are the minimum and the maximum desired hot water
temperature respectively.
According to the thermal modeling for a building presented in [13], the building
temperature at each hour is obtained by:
− −
Tin ( h + 1) = Tin (h) * e τ + ( R * H air ( h) + Tout ( h))*(1 − e τ ) (10)
T min
in ≤ Tin ( h) ≤ T max
in
(11)
where Δ=1h and τ = RC . The values used are R=18◦C/kW, C=0.525kWh/◦C, and the
initial room temperature=20◦C.
Constraint of Micro-CHP Operation. Electrical and thermal output power limits for
micro-CHP:
min
VCHP ( h ) * PCHP ≤ PCHP ( h ) ≤ VCHP ( h ) * PCHP
max
(12)
min
VCHP ( h ) * H CHP ≤ H CHP ( h ) ≤ VCHP ( h ) * H CHP
max (13)
Micro-CHP electrical and thermal efficiency:

ηe (14)
PCHP ( h) = H CHP ( h) *
ηth
Micro-CHP output power ramp rate:
− Prr ≤ PCHP (h) − PCHP (h − 1) ≤ Prr (15)
ηth ηth
− ηe * Prr ≤ H CHP (h) − H CHP (h − 1) ≤ ηe * Prr (16)
where Hchp(h) is the micro-CHP’s thermal output power. Vchp(h) is the micro-
CHP’s status. ηe and ηth are the micro-CHP’s electrical and thermal efficiency,
respectively.
3 Scheduling Using Hybrid PSO-DE Algorithm

The objective of the scheduler is to maximize the net benefits. Based on (1), the
corresponding optimization problem is to find the real-value variable for battery,
micro-CHP and thermal appliances. Therefore, it can be considered as a constraint
real-value optimization problem. This paper mixes particle swarm optimization and
differential evolution (PSO-DE) to solve the constraint optimization problem.
3.1 Overview of PSO

Particle swarm optimization is a stochastic global optimization method inspired by the
choreography of a bird flock. PSO relies on the exchange of information between
individuals, called particles. In PSO, each particle adjusts its trajectory stochastically
towards the positions of its own previous best performance and the best previous
performance of its neighbors or the whole swarm. At the each iteration, the velocity
and position updating rules are given by:
vi,t+1j = wvi,t j + c1r1 ( pbesti,t j − xi,t j ) + c2 r2 (gbest tj − xi,t j )
(17)
xi,t+1j = xi,t j + vi,t j
(18)
where w is an inertia weight factor, c1 is a cognition weight factor, c2 is a social
weight factor, r1 and r2 are two random numbers uniformly distributed in the range of
[0,1]. In this version, the variable Vi,j is limited to the range ±Vmax [17-19] analyzed and
introduced the velocity adjustment as
vi,t+1j = R1 ( pbesti,t j − xi,t j ) + R2 (gbest tj − xi,t j )
(19)
where R1 and R2 are generated using abs(N(0,1)) According to the statistical
knowledge, the mean of abs(N(0,1))is 0.798 and the variance is 0.36.
3.2 Overview of DE
Different evolution (DE) [15] has become a popular algorithm in global optimization.
DE starts the search with an initial population containing NP individuals, which are
randomly sampled from the search space. Then, one individual called the target vector
in the population is used to generate a mutant vector by the mutation operation. So
far, several mutation strategies have been proposed [14].
Subsequently, DE applies a crossover operator to generate the offspring individual,

the crossover is employed and executed as follows:
 yit, j if rand ≤ CR OR j = jrand (20)
ui , j =  t
 xi , j otherwise
where i ∈{1, 2,..., NP}, j ∈{1, 2,..., n} ,rand is a uniformly distributed rand number
between [0,1], jrand is a randomly selected integer from [1,n],CR is the crossover
control parameter, u it, j is the jth component of the trial vector .
Finally, the target vector xi is compared with the trial vector u i in terms of the
objective function value and the better one survives into the nest generation:
u t , if f (u ti ≤ f(x ti )) (21)
xit +1 =  it
 xi , otherwise
3.3 PSO-DE for Real-Value Constraint Optimization

The PSO-DE’s main procedure can be summarized in Fig.1 [19].
j = R1 ( pbest i, j − xi, j ) + R2 (gbest j − xi, j )

vi,t+1 t t t t
xi,t +1
j = xi , j + vi, j
t t
0.5 * (l (j) + x ti , j ) , xit,+j1 < l (j) ,


xit,+j1 = 0.5 * (u(j) + x ti , j ) , xit,+j1 > u (j)
 xit,+j1 , otherwise

rand / 1 : yit, j = xrt [1], j + F (x tr [2], j − x tr [3], j )

current − to − best/ 1: y ti , j = xit, j + F (x best
t
, j − x i , j ) + F (x r [1], j − x r [ 2], j )
t t t
rand/ 2 : y ti , j = xr[1],
t
j + F (x r[2], j − x r[3], j ) + F (x r [4], j − x r [5], j )
t t t t
 2 * l (j) − z ti , j , if (z ti , j < l (j))


zit, j =  2 * u (j) − z ti , j , if (z ti , j > u (j))
 zit, j , otherwise

 zt , if (f(z ti < f(pbest ti )  G(z ti ) ≤ G (pbest ti ))
pbestit +1 =  i
 pbestit , otherwise
Fig. 1. PSO-DE main procedure

As described above, the paper is to solve a constrained optimization problem, the

strategy for handling constraints is usually the use of penalty function methods.
However, the main problem is that they require a careful fine tuning of the penalty
factors .In order to overcome the drawback of choice penalty factors, this paper
applies PSO-DE method to handle constraints, which does not require setting any
additional parameters in comparison to the original PSO. The feasibility-based rules
are applied to handle constraints of the problem.
In order to handle the constraints, we minimize the original objective function f (x)
as well as the degree of constraint violation G(x). At each generation, pop is sorted
according to the degree of constraint violation in a descending order. Only the first
half of pop are evolved by using Krohling and dos Santos Coelho’s PSO [18].
In order to compensate the convergence speed and supply more valuable
information to adjust the particles’ trajectories, the pBest is updated by Different
evolution. DE-based search process motivates the particles to search for new regions
including some lesser explored regions and enhance the particles capability to explore
the vast search space [19].
4 Simulation Result
The case study is a typical residential building. A micro-CHP with 3kW capacity is
considered for the building. The water storage capacity is 80L. The building’s hot
water demand is shown in Fig.2. The building loads include must run electrical
appliances and thermal loads. Fig.3 shows the total electrical demand by the must run
appliances and the price of electricity supplied to terminal loads [8, 20]. Which the
must run include lights, cook, fridge, computers, washing machine, dryer, dish washer
and pool pump.
Fig. 2. Hot water demand in building
Fig. 3. Total electrical demand and the price of electricity

In this case the price of natural gas is 2.05RMB/m3, and the price of electricity fed
into grid is 0.457 RMB/kWh [20]. Table I shows assumed parameters in solving(1).
Table 1. Assume values for parameters
Parameter Value Unit

cold
Tinmin , Tinmax Tstmin , Tstmax Tst 22,24,72,74,10
°C
kWh
Cwater 11.61*10-4
L°C
min max min max
PCHP , PCHP H CHP , H CHP Pchmax , Pdch
max
0.3,3,0.5,5,3,3 kW
kW
Prr 2.5 h
ηe ,ηth η c h , η d c h 30,50,0.9,0.9 %
SOCmin, SOCmax 0.3,1 p.u.
Gref -3 m3
92.59*10 h
The PSO-DE algorithm is programmed using the C++ programming language

using a PC with an Intel dual core processor 2.6GHz on a Windows XP operation
system. A visual C++ compiler was used in this work.
The proposed algorithm takes 5 trials to get the final best cost. And in each trial,
the population size and maximum iteration take 300 and 5000 respectively. The best
result with minimum cost is shown in Table II.
Table 2. Five trials of algorithm

Generation Fitness Vilolation
93 19.21 0.0
80 15.89 0.0
0 17.90 0.0
132 17.48 0.0
72 16.81 0.0
The simulation results show the algorithm can achieve optimum result under kinds
of constraints. As been shown in Fig. 4, the building’s temperature have been set
within desired temperature (ie.,22-24), the hot water temperature also been set to
comfort temperature (ie.,74-76) in the demand time (ie.,4:00-9:00,16:00-22).
Fig. 4. Hot water temperature and building temperature

The micro-CHP output power, battery output power and its SOC is depicted in Fig.
5. The positive and negative values represent battery charging and discharging
respectively.
Fig. 5. Micro-CHP& Buy_grid& Battery
As expected, the battery coordinates with the micro-CHP output power to achieve
economic operation of micro grid. The battery charges and discharges during low
price and peak price hours, respectively. While the micro-CHP operation at its
maximum capacity of peak load price hours.
Also ,as shown in the Fig.4 it achieve optimum cost operation by buying minimum
electrical power during peak hours and selling its extra electrical power to the main
grid at off-peak price hours.
5 Conclusion
This paper pioneers a problem of a residential smart home user equipped with kinds
of appliances. We developed an optimal control algorithm for the smart home. The
objective of the optimal control algorithm is to reduce the total electricity cost over a
billing period (i.e., a day).This proposed PSO-DE algorithm can be used in home
energy management systems and help in realization of optimum cost operation. As
future study, it is suggested to look into issues such as optimization under price
uncertain environment.
Acknowledgments. This work was supported in part by the Natural Science

Foundation of China (61075064, 61034004, 61005090), Program for New Century
Excellent Talents in University of Ministry of Education of China. Ministry of
Education (NCET-10-0633), the Fundamental Research Funds for the Central
Universities, and the Research Fund of State Key Lab. of Management and Control
for Complex systems.
References
1. Mansouri, I., Newborough, M., Probert, D.: Energy consumption in UK households:
impact of domestic electrical appliances. Appl. Energy 54(3), 211–285 (1996)
2. Hamidi, V., Li, F., Robinson, F.: Demand response in the UK’s domestic sector. Electr.
Power Syst. Res. 79(12), 1722–1726 (2009)
3. Katz, J.S.: Educating the smart grid. Presented at the IEEE En- ergy 2030 Conf., Atlanta,
GA (November 17-18, 2008)
4. Colson, C.M., Nehrir, M.H.: A review of challenges to real-time power management of
microgrids. Presented at the 2009 Power Energy Soc. Gen. Meet., Calgary, AB, Canada,
PESGM2009-001250 (2009)
5. Mohsenian-Radand, A., Leon-Garcia, A.: Optimal residential load control with price
prediction in real-time electricity pricing environments. IEEE Trans. Smart Grid 1(2), 120–
133 (2010)
6. Conejo, A.J., Morales, J.M., Baringo, L.: Real-time demand response model. IEEE Trans.
Smart Grid 1(2), 236–242 (2010)
7. Li, N., Chen, L., Low, S.H.: Optimal demand response based on utility maximization in
power networks. In: Proc. IEEE Power Energy Soc. Gen. Meet. (July 2011)
8. Pedrasa, M.A.A., Spooner, T.D., MacGill, I.F.: Coordinated scheduling of residential
distributed energy resources to optimize smart home energy services. IEEE Trans. Smart
Grid 1(2), 134–143 (2010)
9. Mohsenian-Rad, A., Wong, V.W.S., Jatskevich, J., Schober, R., Leon-Garcia, A.:
Autonomous demand-side management based on game- theoretic energy consumption
scheduling for the future smart grid. IEEE Trans. Smart Grid 1(3), 320–331 (2010)
10. Sane, H., Guay, M.: Minimax dynamic optimization over a finite- time horizon for
building demand control. In: Proc. Amer. Control Conf., pp. 1469–1474 (June 2008)
11. Samadi, P., Mohsenian-Rad, A., Schober, R., Wong, V.W.S., Jatskevich, J.: Optimal real-
time pricing algorithm based on utility maximization for smart grid. In: Proc. IEEE Int.
Conf. Smart Grid Commun., pp. 415–420 (2010)
12. Wi, Y.-M., Lee, J.-U., Joo, S.-K.: Electric Vehicle Charging Method for Smart
Homes/Buildings with a Photovoltaic System. IEEE Transactions on Consumer
Electronics 59(2) (May 2013)
13. Tasdighi, M., Ghasemi, H., Rahimi-Kian, A.: Residential Microgrid Scheduling Based on
Smart Meters Data and Temperature Dependent Thermal Load Modeling. IEEE Trans.
Smart Grid (2013)
14. Price, K., Storn, R., Lampinen, J.: Differential Evolution: A Practical Approach to Global
Optimization. Springer-Verlag (2005) ISBN:3-540-20950-6
15. Neri, F., Tirronen, V.: Recent advances in differential evolution: a survey and
experimental analysis. Artificial Intelligence Review 33(1-2), 61–106 (2010)
16. Mohsenian-Rad, A.-H., Wong, V.W.S., Jatskevich, J., Schober, R.: Optimal and
autonomous incentive-based energy consumption scheduling algorithm for smart grid. In:
Proc. IEEE Conf. Innov. Smart Grid Technol., Gaithersburg, MD, USA (2010)
17. Clerc, M., Kennedy, J.: The particle swarm-explosion, stability, and convergence in a
multidimensional complex space. IEEE Transactions on Evolutionary Computation 6(1),
58–73 (2002)
18. Krohling, R.A., dos Santos Coelho, L.: Coevolutionary particle swarm optimization using
Gaussian distribution for solving constrained optimization problems. IEEE Transactions
on Systems, Man, Cybernetics Part B: Cybernetics 36(6), 1407–1416 (2006)
19. Liu, H., Cai, Z., Wang, Y.: Hybridizing particle swarm optimization with differential
evolution for constrained numerical and engineering optimization. Applied Soft
Computing 10, 629–640 (2010)
20. Guan, X., Xu, Z., Jia, Q.-S.: Energy-Efficient Buildings Facilitated by Microgrid. IEEE
Trans. Smart Grid 1(3) (December 2010)
A Multiobjective Large Neighborhood Search
for a Vehicle Routing Problem
Liangjun Ke and Laipeng Zhai
The State Key Laboratory for Manufacturing Systems Engineering,

Xian Jiaotong University, Xi’an, 710049, China
keljxjtu@xjtu.edu.cn
Abstract. In this paper, a multiobjective adaptive large neighborhood

search is proposed for a vehicle routing problem (VRP) of which the
objectives are the total travel time and the cumulative time, i.e., the
total arrival time at all customers. It hybrids destroy-repair operators
with local search for generating new solutions. An adaptive probabilistic
rule based on Pareto dominance is proposed to select a combination of
destroy-repair operator. The effectiveness of the proposed algorithm is
supported by the experimental study.
Keywords: multiobjective optimization, vehicle routing problem, adap-

tive large neighborhood search.
1 Introduction
Vehicle routing problem (VRP) is one of the most important combinatorial opti-
mization problems. In this problem, a set of customers are dispersed in a graph.
Each customer is associated with a demand. vehicles are scheduled to serve these
customers so as to achieve one or more optimal objectives whilst the route of
each vehicle must satisfy specific requirements. The most common studied ob-
jectives are the total travel time, the number of vehicles, makespan, balance, and
others [1].
Recently, cumulative VRP becomes a hot topic [2, 3]. It aims at minimizing
the cumulative time, that is, the total arrival time of all customers. This problem
was extended from the delivery man problem [4, 5], and can be used to model
many problems such as the routing schedule during the disaster aids [2].
Although many researchers considered VRP with only one single objective,
VRP is multi-objective in nature [1]. Multiobjective VRP (MVRP) has attracted
great research interests [1]. A lot of approaches have been used to deal with
various MVRPs. A popular approach is the scalar approach, which transforms a
multiobjective problem into a single objective problem by weighted sum method
or other methods, then solves it by a single objective heuristic or exact algorithm
[6,7]. Based on the concept of Pareto dominance, Pareto methods are also widely
used [8–10]. Other approaches, e.g., genetic algorithm [11], lexicographic method
[12] and ant colony optimization [13], etc, were also adopted. Since VRP is

302 L. Ke and L. Zhai
a NP-hard problem, metaheuristics are widely adopted for finding satisfactory

solutions within acceptable computational time.
In this paper, we consider an MVRP. In the MVRP, there are n customers.
Customer i, staying at node i has a demand di . A fleet of vehicles start from
node 0 and each node n + 1 to serve these customers. The travel time between
every two nodes i, j(i, j ∈ {0, 1, · · · , n, n + 1}) is wij . A feasible solution consists
of a set of paths such that each customer is served by one path, and the total
demand of a path is not more than the vehicle capacity. The objectives are the
total travel time and the cumulative time at all customers. Intuitively, a solution
with smaller total travel time may have smaller cumulative time. Nevertheless,
these two objectives may be not always compatible [2]. Therefore this problem
can not be reduced to a single objective problem. By optimizing this problem,
one can obtain a solution for minimizing the total travel time or the cumulative
time in a single run. Moreover, one can obtain a set of tradeoff solutions for
decision making.
To deal with this problem, we propose a new algorithm extended from adap-
tive large neighborhood search (ALNS). ALNS has been proven to be a powerful
metaheuristic for many variants of single objective VRP. It was firstly proposed
by Ropke and Pisinger [14]. It is closely related to the large neighborhood search
developed by Shaw [15]. ALNS generates new solutions by using destroy and
repair operators. Local search is optionally adopted for exploiting the neighbor-
hood of those solutions. An adaptive probabilistic rule is used to schedule destroy
and repair operators according to their weights, which are renewed based on the
previous search experience.
The remainder of this paper is structured as follows. Section 2 describes the
details of the proposed multiobjective ALNS algorithm. The experimental results
are presented in section 3. Finally, the main results are concluded in section 4.
2 Multiobjective Adaptive Large Neighborhood Search
Our algorithm, deonted as MALNS, evolves a population of N solutions over time

where N is population size, and employs an external archive EA to store the
nondominated solutions. It works as follows. At first, a population of solutions
are initialized. After that, starting from each individual in the population, a new
solution is generated by a combination of destroy-repair operator. A combina-
tion is probabilistically chosen from a set of candidate combinations, depending
on their weights. Subsequently, local search is adopted to improve each new so-
lution. By using the obtained solutions, our algorithm updates the population,
the external archive EA, and the weights of the combinations of destroy-repair
operators. It terminates when a stopping condition is satisfied.
2.1 The Single Objective Function
The single objective function is used to evaluate the quality of a move in destroy-
repair operators or local search. Two things are considered to define the single
A Multiobjective Large Neighborhood Search 303
objective function: 1) it is a weighted aggregation of the objectives in MVRP;

2) to obtain a wide and evenly nondominated front, it is desirable to guide the
destroy-repair operator or local search to explore different directions. Therefore,
the weight vector (or search direction) should be carefully assigned.
Formally, the single objective function is given as follows.
f1 (x) f2 (x)
F (x, λ) = λ1 + λ2 ∗ (1)
f1∗ f2
where λ = (λ1 , λ2 ) is a weight vector, λ1 +λ2 = 1, λ1 , λ2 ≥ 0. Since the difference

between f1 and f2 is very big, scale transformation is carried out in (1). f1∗ and
f2∗ are the minimal values of these two objectives respectively. In practice, they
are approximated by the minimal values obtained so far. Each weight vector is
randomly selected.
2.2 Destroy and Repair Operators

Many different destroy and repair operators have been proposed in the literature.
An interesting survey is available in [14]. MALNS heavily depends on simpler
destroy-repair operators for exploring new competitive search areas, and uses
local search for exploitation. The destroy operators we used are random removal,
worst removal, relatedness removal, and cluster removal [14,16]. These operators
remove some customers from the routes. The set of unvisited customers, called
request bank [14], is denoted as U . During the running, removed customers are
saved in U . The maximal size of U is denoted as u. Parameter u significantly
affects the behavior of MALNS. With a larger value, the operators are able
to explore larger search area. The repair operators consist of the basic greedy
insertion heuristic and regret insertion heuristic.
2.3 Selecting a Combination of Destroy-Repair Operator

There are eight combinations of destroy-repair operators. During the search,
a roulette-wheel selection method is employed to select one combination every
time. Each combination ci is associated with a weight wi which is used to measure
how well combination ci has performed in past iterations. At the beginning of
the algorithm, the initial value
8 of wi is set to 1. The combination ci is selected
with probability pi = wi / j=1 wj .
2.4 Local Search

Local search plays an very important role in the design of metaheuristics for
VRP [17]. A local search operator iteratively improves a solution by exploring
its neighborhood in terms of the single objective function given in (1). Given a
feasible solution to VRP represented by a set of routes x = {R1 , . . . , Rl , . . . , Rv },
where Rl is the set of customers serviced by route (or vehicle) l. Its neighborhood
is denoted as N (x) .
The following components are critical in the implementation of a local search

operator [17]. The first is the starting solution. The second is the mechanism
to generate neighboring solutions of a given solution. The third is the accep-
tance criterion. Two popular acceptance strategies are first-accept (FA) and
best-accept (BA). The FA strategy chooses the first neighboring solution that
satisfies the pre-specified acceptance criterion (e.g., the objective variation after
a move). The BA strategy checks all neighboring solutions which satisfy a cri-
terion and chooses the best among them. The fourth is the condition when to
stop the local search operator.
The 2-opt, exchange, cross and relocation operators [17] are adopted to gener-
ate neighborhood. FA strategy is used. These operators are invoked one by one.
Local search will be ended when no more improvement can be achieved.
Unlike 2-opt, exchange, cross and relocation operators are inter-route oper-
ators. To speed up these inter-route operators, we first propose the concept of
neighborhood of routes based on polar angle. Polar angle is the basic tool in
the famous sweep algorithm [18]. For a route Rl , its polar angle is defined as
the polar angle of its center of gravity. Two routes are said to be close if their
polar angles are close. For a route, only the N closest neighboring routes are
permitted to inter-change. When to select a move, a local search operator only
checks its neighborhood for each route. By interchanging with the routes out of
smaller neighbor, it is more likely to find better solution with smaller travel time
or cumulative time.
2.5 Update of Weights

In MALNS, a combination ci is associated with a score, denoted by ϕi . At each
iteration, ϕi is initialized to 0. After an iteration, ϕi of the chosen combination ci
will be renewed based on the quality of the solutions constructed at the iteration.
In detail, starting from each solution (in current population) xs , a new solution
xn is generated by a combination ci and improved by local search. If xn is
nondominated by solutions found so far, the score of ci is increased by 15; If xn
dominates xs , the score of ci is increased by 10; If xn is nondominated by xs ,
the score of ci is increased by 5; Otherwise, no score is obtained. Formally, it is
updated as follows:
⎧
⎪
⎪ ϕi + 15 if the new solution xn is nondominated
⎪
⎪
⎪
⎪ by solutions found so far
⎪
⎪
⎨ ϕi + 10 if the new solution xn dominates
ϕi = its starting solution xs (2)
⎪
⎪
⎪
⎪ ϕ i + 5 if the new solution xn is nondominated
⎪
⎪
⎪
⎪ by its starting solution xs
⎩
ϕi otherwise
Every iteration, each weight wi is updated based on the scores obtained.

ϕi
wi = (1 − ρ)wi + ρ (3)
max(F reqi , 1)
where F reqi denotes the times the combination ci has been applied in the past
iteration. ρ is a parameter which controls the forgotten rate of the past ex-
perience. ρ is set to 0.05. MALNS re-initializes the weights to 1 once no new
nondominated solutions can be found during consecutive 50 iterations.
2.6 Population Initialization
N solutions are constructed by the initial procedure. At first, each customer

is inserted in request bank, and then each customer is inserted by the regret
insertion heuristic. Note that the single objective function is given by (1). To
construct the the lth solution (l ∈ {1, · · · , N }), the weight vector is ((l − 1)/(N −
1), (N − l)/(N − 1)).
2.7 Update of Population
At each iteration, we only accept a new generated solution of which request

bank is empty and the number of routes is |R| . The population is updated by
using nondominance ranking and crowding distance in [19]. From the last popu-
lation and new generated solutions, a set of fronts are obtained by nondominance
ranking. The first front F1 consists of the nondominated solutions. The second
front is the set of solutions which are only dominated by the first front, and so
on. Crowding distance is the average side length of the cuboid formed by the
objective values of these nearest neighboring solutions [19].
2.8 Update of External Archive
Once a solution x is accepted, the external archive EA is renewed as follows: If

no vector in EA dominates F (x), F (x) will be added to EA. At the mean time,
all the vectors dominated by F (x) will be removed from EA.
MALNS was coded in C++ and tested on a PC with Pentium 4, 2.4G CPU, and
4GB RAM. It was tested on 20 large-scale instances with 240 to 483 customers
in [20]. The travel time between every two nodes is their Euclidean distance. All
travel time is rounded to double precision [2, 3].
Based on the preliminary test, the population size was set to 30. As done
in [3], the maximal size u of request bank was randomly chosen from [10,60]. For
each instance, MALNS was test the same times in [3] (i.e., 5) independently and
stopped when a given time limit was achieved. In our experiment, the time limit
was chosen as follows. At first the computational time in [3] was transformed,
then the transformed time T was set. For example, the computational time of
GWKC1 in [3] is 1038, then the transformed time T is 865, since our CPU is 1.2
times as fast as the one of [3].
3.1 Performance Metrics

It is widely accepted that, given two nondominated fronts, the better front has
the following properties: it is closer to Pareto-optimal front; it distributes more
evenly and widely in objective space [21]. Many performance metrics have been
proposed to evaluate the solution quality. Among them, hypervolume is a very
nice and popular metric [22], therefore we adopted it in this paper. The Hyper-
volume value of a nondominated front is the volume of the area in the objective
space which is bounded by a reference point and dominated by the front itself.
Let ζ be all the nondominated solution sets found in a test, the reference point
was chosen as (max{f1 (x)|x ∈ ζ}, max{f2 (x)|x ∈ ζ}). Larger hypervolume value
indicates better solution quality.
In order to pictorially illustrate the nondominated fronts, a statistical tool,
called summary-attainment surface [23], is employed. Summary-attainment sur-
face refers to the union of all tightest points in the objective space obtained
by an algorithm. If an algorithm is tested l times, there will be l summary-
attainment surface. The first, l/2th, lth summary-attainment surface is called
the best, median, worst summary-attainment surface respectively.
3.2 Comparison with a Weighted Sum ALNS

To study MALNS, we implemented a weighted sum ALNS which works as fol-
lows: the same procedure of MALNS with only one single weight vector is per-
formed N times (N = 30). In the lth time (l = 1, · · · , N ), only weight vector
((l − 1)/(N − 1), (N − l)/(N − 1)) is used and the computational time is T /N .
As seen from Fig. 1, MALNS provides better hypervolume than the weighted
sum ALNS. According to the summary-attainment surfaces shown in Fig. 2,
MALNS can provide wider nonodimianted front. We also note that the weighted
sum ALNS performs better in some central parts. The reason may be that the
computational resource in the weighted sum ALNS is biased to search some
specific areas.
6
x 10
1.25
1.24
1.23
1.22
1.21
1.2
1.19
1.18
1.17
1 2
Fig. 1. The hypervolume values obtained by MALNS and weighted sum ALNS (shown
from left to right) are tested for GWKC20.
best surface median surface worst surface

10000 10000 10000
MALNS MALNS MALNS
weighted sum ALNS weighted sum ALNS weighted sum ALNS
9500 9500 9500
9000 9000 9000
8500 8500 8500
8000 8000 8000
7500 7500 7500
7000 7000 7000

1800 1900 2000 2100 2200 2300 2400 2500 1800 1900 2000 2100 2200 2300 2400 2500 1800 1900 2000 2100 2200 2300 2400 2500
Fig. 2. The best, median, and worst summary-attainment surfaces obtained by

MALNS and weighted sum ALNS for GWKC20
4 Conclusion
This paper presented a multi-objective adaptive large neighborhood search for

an MVRP of which the objectives are the total travel time and the cumulative
time. Although the total travel time and the cumulative time has been sepa-
rately studied before, this paper investigated these objectives together. It selects
a combination of destroy-repair operator for generating new solutions and im-
proves them by local search. According to the yielded solutions, the preference
of each combination is renewed based on the Pareto dominance. Compared with
a weighted sum ALNS, the proposed algorithm performs better.
References
1. Jozefowiez, N., Semet, F., Talbi, E.: Multiobjective vehicle routing problems. Eu-
ropean Journal of Operational Research 189, 293–309 (2008)
2. Ngueveu, S.U., Prins, C., Wolfler-Calvo, R.: An effective memetic algorithm for
the cumulative capacitated vehicle routing problem. Computers and Operations
Research 37, 1877–1885 (2010)
3. Ribeiro, G., Laporte, G.: an adaptive large variable neighborhood search heuristic
for cumulative capacitated vehicle routing problem. Computers and Operations
Research 39(3), 728–735 (2012)
4. Lucena, A.: Time-dependent traveling salesman problem:The deliveryman case.
Networks 20, 753–763 (1990)
5. Fischetti, M., Laporte, G., Martello, S.: The delivery man problem and cumulative
matroids. Operations Research 41, 1055–1064 (1993)
6. Bowerman, R., Hall, B., Calamai, P.: A multi-objective optimization approach
to urban school bus routing: Formulation and solution method. Transportation
Research Part A 29, 123–197 (1995)
7. Riera-Ledesma, J., Salazar-Gonzalez, J.: The biobjective travelling purchaser prob-
lem. European Journal of Operational Research 160, 599–613 (2005)
8. Geiger, M.: Genetic algorithms for multiple objective vehicle routing. In: Meta-
Heuristic International Conference 2001 (MIC 2001), pp. 348–353 (2001)
9. Lacomme, P., Prins, C., Sevaux, M.: A genetic algorithm for a bi-objective capac-
itated arc routing problem. Computers and Operations Research 33, 3473–3493
(2006)
10. Jozefowiez, N., Semet, F., Talbi, E.G.: The bi-objective covering tour problem.
Computers and Operations Research 34, 1929–1942 (2007)
11. Ombuki, B., Ross, B., Hanshar, F.: Multi-objective genetic algorithm for vehicle
routing problem with time windows. Applied Intelligence 24, 17–30 (2006)
12. Current, J., Schilling, D.: The median tour and maximal covering tour problems:
Formulations and heuristics. European Journal of Operational Research 73, 114–
126 (1994)
13. Doerner, K., Focke, A., Gutjahr, W.: Multicriteria tourplanning for mobile health-
care facilities in a developing country. European Journal of Operational Re-
search 179, 1078–1096 (2007)
14. Ropke, S., Pisinger, D.: A unified heuristic for a large class of vehicle routing
problems with backhauls. European Journal of Operational Research 171, 750–775
(2006)
15. Shaw, P.: A new local search algorithm providing high quality solutions to vehicle
routing problems. In: Technical report, University of Strathclyde, Glasgow (1997)
16. Ropke, S., Pisinger, D.: An adaptive large neighborhoodsearch heuristic for the
pickup and delivery problem with time windows. Transportation Science 40, 455–
472 (2006)
17. Braysy, O., Gendreau, M.: Vehicle routing problem with time windows, part I: route
construction and local search algorithms. Transportation Science 39(1), 104–118
(2005)
18. Gillett, B., Miller, L.: A heuristic algorithm for the vehicle dispatch problem. Op-
erations Research 22, 340–349 (1974)
19. Deb, K., Agrawal, S., Pratap, A., Meyarivan, T.: A fast and elitist multiobjec-
tive genetic algorithm: NSGA?II. IEEE Trans. Evolutionary Computation 6(2),
182–197 (2002)
20. Golden, B.L., Wasil, E.A., Kelly, J.P., Chao, I.M.: Metaheuristics in vehicle routing.
In: Fleet Management and Logistics, pp. 33–56. Springer, Heidelberg (1998)
21. Coello, C., Veldhuizen, D.V., Lamont, G.: Evolutionary Algorithms for Solving
Multi-objective Problems. Kluwer (2002)
22. Fonseca, C.M., Paquete, L., Lopez-Ibanez, M.: An improved dimension-sweep al-
gorithm for the hypervolume indicator. In: Proceedings of the 2006 Congress on
Evolutionary Computation (CEC 2006), pp. 1157–1163. IEEE Press, Piscataway
(2006)
23. Knowles, J.: A summary-attainment-surface plotting method for visualizing the
performance of stochastic multiobjective optimizers. In: Proceedings of the Fifth
International Conference on Intelligent Systems Design and Applications (ISDAV),
pp. 552–557. IEEE Computer Society (2005)
A Self-adaptive Interior Penalty Based Differential
Evolution Algorithm for Constrained Optimization
Cui Chenggang*, Yang Xiaofei, and Gao Tingyu
Shanghai Advanced Research Institute, Chinese Academy of Sciences, Shanghai, China

cuicg@sari.ac.cn
Abstract. A self-adaptive interior penalty method is proposed for the

constrained optimization problems by using interior penalty method to handle
constraints. A set of interior penalty rules are designed to evaluate feasible
solutions and infeasible solutions separately. A self-adaptive penalty factor
method is proposed to prevent the interior penalty method from being sensitive
to the values of penalty factor and to minimize the interior penalty function
value of the optimal solution. As an instance of implementation, a different
evolution algorithm is improved by means of the method proposed in this paper,
based on which 10 benchmark problems are tested. The numerical solution
results indicate that the performance of the method is better than four existing
state-of-the-art techniques.
Keywords: Constrained optimization, Evolutionary algorithm, Interior penalty

method, Differential evolution.
1 Introduction
Evolutionary algorithms (EAs) have been widely used to solve constrained optimization
problems (COPs). However, EAs are normally used as “blind heuristics” in the sense of
lacking an explicit mechanism to bias the search in constrained search spaces [1].
Several researchers have proposed different mechanisms to incorporate constraints into
the fitness function of an EA [2]. Penalty functions are the most common approaches
used to handle constraints with EAs [3]. There are two basic types of penalty functions:
exterior penalty functions, which penalize infeasible solutions, and interior penalty
functions, which penalize feasible solutions. Compared to interior penalty functions,
exterior penalty functions are more common in EAs. The main reason is that there is no
need to start with a feasible solution in exterior penalty functions. Another category of
constraint handling techniques involves the preference of feasible solutions over
infeasible solutions [4]. In these methods, a heuristic rule that feasible solutions are
preferred over infeasible ones is used to process infeasible solutions. Multiobjective
optimization techniques have also been used in the solution of constrained single
objective optimization problems [5]. These techniques can be classified based on the
way they transform the COP into a multiobjective optimization problem.
*
Correspondig author.
310 C. Chenggang, Y. Xiaofei, and G. Tingyu
In this paper, a set of interior penalty (IP) based selection rules is proposed to
balance the avoidance of the constraint boundaries and the minimization of the
objective function in the search process of Differential Evolution (DE) algorithms.
Feasible solutions closed to the constraint boundaries are penalized to balance the
conflict aims of reducing the objective function and approaching the constraint
boundaries; infeasible solutions are evaluated by constraint violations to reach
feasible region quickly. Three elements are employed to make these rules more
effective in DEs: (1) Logarithmic penalty functions are used to make DEs yield a
rapid convergence; (2) Penalty factors are updated according to the types of
constraints which are determined by the Spearman's rank-order correlation
coefficients; (3) Equality constraints are handled by an adaptive relaxing rule. In this
paper, a self-adapt interior penalty based differential evolution algorithm is
implemented as an example of this constraint handling approach. Finally, the
efficiency and effectiveness of the proposed method are evaluated on 10 benchmark
problems.
2 Interior Penalty Based Selection Rules
2.1 Statement of the Problem

Generally, a COP can be expressed as follows:
min f ( x ) x = ( x1 ,..., xn ) ∈ R n
, (1)
s.t. g j ( x ) ≤ 0 , j ∈ {1,..., m} ,
h j ( x ) = 0 , j = q + 1,..., m .
where x is the decision vector, f(x) is the objective function, q is the number of
inequality constraints, and m-q is the number of equality constraints. Let S⊂R n define
the search space bounded by the parametric constraints xi ≤ xi ≤ xi , i∈{1,2,…,n},
where xi and xi are the lower bound and the upper bound of xi, respectively.
2.2 Interior Penalty Method

The interior penalty method is very popular in the traditional mathematical
programming techniques, motivated by minimizing a composite function that reflects
the original objective function as well as the influence of the constraints [6].
Given the constrained optimization problem (1), the interior penalty function can
be formulated as follows:
φ ( x, r (t )) = f ( x) + r (t ) B ( x) . (2)
B(x) is a penalty term that is nonnegative and approaches ∞ as the constraint
boundaries are approached from the interior. r(t) is a penalty factor.
The interior penalty method replaces a COP with a sequence of unconstrained
optimization problems, defined as:
A Self-adaptive Interior Penalty Based Differential Evolution Algorithm 311
lim min φ ( x, r (t )) . (3)

t →∞
Solving problem (3) sequentially for a monotonously decreasing sequence {r(t)}

such that limt →∞ r (t ) = 0 gives a sequence {x(r(t))} yielding limt →∞ f ( x(r (t ))) = f ( x* ) ,
where x* is the optimal solution of problem (1).
2.3 Implementation of the IP Based Selection Rules

In order to balance the avoidance of the constraint boundaries and the minimization of
the objective function in the search process, a set of interior penalty based selection
rules is proposed to improve the efficiency of search and avoid the violations.
The IP rules can be formulated as follows:
1) Between two feasible solutions, the one with a better interior penalty function
value is preferred.2) If both solutions candidates are infeasible, the one with a
smaller constraint violation is preferred.3) A feasible solution is always preferred
to an infeasible one.
For the IP based selection rules, there are three important properties:
1) The interior penalty term is defined only if a solution is feasible. Thus, the
interior penalty function cannot handle infeasible solutions. Therefore, a
preference of feasible solutions to infeasible ones is used in rule (3). In this way,
feasible solutions and infeasible solutions can be evaluated by different methods.
2) In the interior penalty method, feasible solutions are penalized in order to
avoid the boundaries of feasible region. Rule 1) penalizes feasible solutions
approaching to the boundaries of feasible region. In this way, the search process
of EAs will avoid the boundaries of feasible region.
3) The objective function is completely disregarded in rule 2). Therefore, the
entire search effort is directed toward finding a feasible solution. This rule is
especially suitable for highly constrained problems wherein finding a feasible
solution may be extremely difficult [3]. In this way, the penalty rules can be
applied to highly constrained problems without initial feasible solutions.
3 Self-adapt Interior Penalty
3.1 Form of Interior Penalty Function

The logarithmic penalty function is used as the interior penalty function in this paper
since it has a superlinear convergence in the traditional mathematical programming
[7]. The penalty term B(x) in Eq.(2) can be formulated as follows:
m
B( x) = − ln ( −vi ( x) ) , (4)
i =1
where vi(x) is the normalized constraint value, defined as:

gi ( x)
vi ( x) = . (5)
| min gi ( x) |
The scaling factor |min gi(x) | for each constraint is taken as the minimal value of
constraint value gi(x) in the search process.
3.2 Self-adapt Penalty Factor

The candidate solutions are selected based on interior penalty function value in the IP
based selection rules. Therefore, the optimal solution must be the one with minimum
internal penalty value of all the candidates. According this rule, we proposed a self-
adapt interior penalty method as follows.
Given the constrained optimization problem (1), the penalty factor r must make
the penalty value of the optimal solution is less than the one of any other solution.
The Penalty factor selection rules can be expressed as the following formula:
f ( x*) + r (t ) B( x*) ≤ f ( xi ) + r (t ) B( xi ) , (6)
where xi is any candidate, x* is the optimal solution.

The Eq.(6) can be converted to:
( B( x*) − B( xi ))r (t ) ≤ f ( xi ) − f ( x*)
. (7)
Considering three conditions as follows:
(1) B(x*)-B(xi)=0
The left of Eq.(7) is 0 in this condition. Therefore, for any r(t), the formula
f(xi)≥f(x*)is established, i.e. Eq. (7) is established. This condition is not considered.
This condition is not considered.
(2) B(x*)-B(xi)<0
The Eq.(7) can be convert to:
f ( x*) − f ( xi )
r (t ) ≥ . (8)
B( xi ) − B( x*)
The right of Eq.(8) is less than or equal to 0 since f(xi)≥f(x*). Therefore, for any
r(t)≥0, Eq.(8) is established. This condition is also not considered.
(3) B(x*)-B(xi)>0
We can get the upper bounder of r(t) by Eq.(7) in this condition:
f ( x*) − f ( xi )
r (t ) ≤ . (9)
B( xi ) − B( x*)
We can get the follows since the optimal solution x* satisfies all the constraints, i.e.
，
∀ j gj(x*)≤0:
q q q q
B(x* ) = −ln | g j (x* ) −ε j |= −ln(ε j − g j (x* )) ≤ −ln(ε j ) Let Bd = − ln(ε j ) , then:
j =1 j =1 j =1 j =1
Bd − B ( xi ) ≥ B( x* ) − B( xi ) > 0 . (10)
Alternative B(x*) with Bd in Eq.(10). We can get an upper bounder of r(t) :

f ( xi ) − f ( x*)
r (t ) ≤ min . (11)
i∈ N Bd − B( xi )
Further, the optimal solution can’t be obtained before the original constrained
optimization problem solved. However, we can use the best feasible solution in the
current population instead of the optimal solution in the search process, i.e.
f ( xi ) − f ( x *)
r (t ) ≤ min , (12)
i∈N Bd − B ( xi )
where x* is the feasible solution with minimum objective in the current population.
f ( xi ) − f ( x *) f ( xi ) − f ( x*)
r (t ) ≤ min ≤ min is established since f ( x *) ≤ f ( x*) .
i∈ N Bd − B( xi ) i∈ N Bd − B( xi )
Therefore, the penalty factor obtained by Eq. (12) satisfies Eq.(6). We can get an
appropriate penalty factor without the optimal solution.
According to the above analysis, we can obtain the penalty factor by Eq.(12).
There may not be a feasible solution when the algorithm starts. We use a larger
penalty factor in the early search process to ensure evolutionary algorithm can quickly
find a feasible solution.
4 Self-adaptive Interior Penalty Based Differential Evolution

To illustrate validity of the interior penalty based selection rules, we introduce them
to Differential Evolution algorithm called a self-adaptive interior penalty based
differential evolution algorithm. The DE algorithm proposed by Storn and Price [8] is
a heuristic method for real parameter optimization problems.
Let xti denote an individual in the population of the DE algorithm and NP the size
of the population, where i indicates the index of the individual, j the index of the
variable, and t the current generation. A new mutated individual v ij,t+1 is generated
according to the following equation:
vij ,t +1 = x dj ,3t + η ( x dj 1,t − x dj ,2t ), (13)
where the random indexes d1, d2, d3∈[0, NP] are mutually different integers and also
different from the running index i, and η∈(0, 1] is called the scaling factor or the
amplification factor.
According to Eq. (13), a crossover operator is used to generate the trial individual
u j,t+1 based on the original individual x dj,3t and the new individual v ij,t+1 .
i
v ij ,t +1 , if Rand[0, 1) ≤ CR or j = randint(1, D) ,

u i
j , t +1 = i (14)
 x j ,t , otherwise,
where Rand[0, 1) is a function that returns a real number between 0 and 1,

randint(min, max) is a function that returns an integer between min and max, CR∈[0,
1] is a crossover factor. The probability of the mutated individuals being preserved in
the next generation is determined by the crossover factor CR.
A selection operator is used to choose an individual for the next generation (t+1)
according to the following rule:
uti+1 , if uti+1 is better than xti ,

xti+1 =  i (15)
 xt , otherwise,
where u it+1 and xti are compared by the IP based selection.
In this way, an individual will replace the one with a lower IP with respect to it; an
individual will replace the one with the same IP depending on different conditions,
where an infeasible individual will replace the one with a larger violation of the
maximal non-common satisfied constraint and a feasible individual will replace the
one with a worse objective with respect to it, respectively.
The capability of finding the global minimum and a fast convergence speed of DE
are both highly sensitive to the control parameters CR and η [9]. Therefore, a self-
adaptive approach is developed to adjust these parameters based on the success rate
φt, where φt is defined by the percentage of original individuals replaced by trial
individuals in the population at every generation, through the following updating law:
η + rand1ηu , φt ≤ 0.5,
ηt =  l (16)
ηt −1 , otherwise,
 rand 2 , φt ≤ 0.5,
CR t =  (17)
CR t −1 , otherwise,
where ηt and CRt are the scaling factor η and the crossover factor CR at generation t,
respectively; rand1 and rand2 are uniformly distributed random numbers in [0, 1];
ηl=0.1, ηu=0.9. The updating of ηt and CRt is conducted before the mutation is
performed. Eqs. (14) and (15) ensure that ηt∈[0.1, 1]⊂(0, 1], CRt∈[0, 1], ∀t.
The pseudo code of the DE with the IP based selection rules is shown as follows,
the rules keep the operators of DE algorithms unchanged.
Begin
t=0;
Create NP random solutions for the initial population;
Evaluate all individuals;
For t=1 to MAX_GENERATION Do
For i=1 to NP Do
Select randomly d1•d2•d3;
d2•d3;
If (Rand[0, 1]•CR
CR or j=randint(1, D)) Then
u ij ,t +1 = vij ,t +1
;
Else
u ij ,t +1 = xij ,t
;
End If
End For
u ij ,t +1 x ij ,t
Compare and by the IP based selected rules;
uti+1 xti
If is better than Then
xti+1 = uti+1
;
Else
xti+1 = xti
;
End If
t=t+1;
update interior penalty factior;
End For
End

We performed the self-adapt interior penalty based differential evolution algorithm
(SIPDE) algorithm 30 independent runs for ten benchmark problem described in
Runarsson and Yao [9]. Equality constraints were transformed into inequalities using
a tolerance value of 0.0001. The parameters were set the same as those of Mezura-
Montes et al. [10]: NP=60, MAX_ GENERATIONS=5800. The control parameters
CR and η were adjusted using a self-adaptive method. We compared our approach
against four state-of-the-art approaches: the stochastic ranking (SR) algorithm [9], the
simple multimembered evolution strategy (SMES) algorithm [11], the adaptive
tradeoff model evolution strategy (ATMES) algorithm [12], and the constraint
handling differential evolution (CHDE) algorithm [10]. The best, mean, worst results,
and the standard deviations obtained by each approach are shown in Table 1.
5.1 Statement of the Problem

As described in Table 1, our approach was able to find the global optimum in ten
benchmark problems. For problems g01, g03, g04, g05, g06, g07, g08, g09 and g10,
the optimal solutions were consistently found in all 30 runs. For problems g02, the
optimal solutions were not consistently found since the ratio of the size of the feasible
region to the size of the search space was very large. Furthermore, feasible solutions
were continuously found for all the benchmark problems in 30 runs. These results
reveal that SIPDE has the substantial capability to deal with various kinds of COPs.
Table 1. Comparison of the best, the mean, the worst solutions, and the standard deviations
found by our SIPDE against SR, SMES, ATMES, and CHDE
methods
Prob optimal stat
SR SMES ATMES CHDE SIPDE
g01 best -15.000 -15.000 -15.000 -15.000 -15.000
-15.000 mean -15.000 -15.000 -15.000 -14.792 -15.000
worst -15.000 -15.000 -15.000 -12.743 -15.000
g02 best -0.803515 -0.803601 -0.803388 -0.803619 -0.803619
-0.803619 mean -0.781975 -0.785238 -0.790148 -0.746236 -0.801758
worst -0.726288 -0.751322 -0.756986 -0.302179 -0.780843
g03 best 1.000 1.000 1.000 1.000 1.000
1.000 mean 1.000 1.000 1.000 0.640326 1.000
worst 1.000 1.000 1.000 0.029601 1.000
g04 best -30665.539 -30665.539 -30665.539 -30665.539 -30665.539
-30665.539 mean -30665.539 -30665.539 -30665.539 -30592.154 -30665.539
worst -30665.539 -30665.539 -30665.539 -29986.214 -30665.539
g05 best 5126.497 5126.599 5126.498 5126.497 5126.498
5126.497 mean 5128.881 5174.492 5127.648 5218.729 5126.498
worst 5142.472 5304.167 5135.256 5502.410 5126.498
g06 best -6961.814 -6961.814 -6961.814 -6961.814 -6961.814
-6961.814 mean -6875.940 -6961.284 -6961.814 -6367.575 -6961.814
worst -6350.262 -6952.482 -6961.814 -2236.950 -6961.814
g07 best 24.307 24.327 24.306 24.306 24.306
24.306 mean 24.374 24.475 24.316 104.599 24.306
worst 24.642 24.843 24.359 1120.541 24.306
g08 best -0.095825 -0.095825 -0.095825 -0.095825 -0.095825
-0.095825 mean -0.095825 -0.095825 -0.095825 -0.091292 -0.095825
worst -0.095825 -0.095825 -0.095825 -0.027188 -0.095825
g09 best 680.630 680.632 680.630 680.630 680.630
680.630 mean 680.656 680.643 680.639 692.472 680.630
worst 680.763 680.719 680.673 839.783 680.630
g10 best 7054.316 7051.903 7052.253 7049.248 7049.248
7049.248 mean 7559.192 7253.047 7250.437 8442.657 7049.255
worst 8835.655 7638.366 7560.224 15580.370 7049.399
5.2 Comparison with Four State-of-the-Art Approaches

The performance of SIPDE was compared in detail with four state-of-the-art
techniques using the selected performance metrics (Table 1). For benchmark
problems g01, g03, g04, and g08, RFDDE, SR, SMES, and ATMES consistently
found the optimal solutions in all 30 runs. For problem g06, the optimal solutions
were consistently found by SIPDE and ATMES in all 30 runs. For problem g05, SR,
SMES, and ATMES found better ‘mean’ and ‘worst’ results than SIPDE. However,
SIPDE was also able to find the optimal solution in 30 runs and the ‘mean’ results
were very close to the optimal solution. For all the other 4 problems, SIPDE found
better ‘best’, ‘mean’, and ‘worst’ results than SR, SMES, and ATMES. As against
CHDE, our approach found “similar” best results in all the problems, and furthermore
located better ‘mean’ and ‘worst’ results in all the problems.
In summary, we can conclude that SIPDE outperforms or has similar performances
to SR, SMES, ATMES, and CHDE in all the problems.
6 Conclusion
In order to combine constraints into the evaluation of feasible solutions, a set of
interior penalty rules for handling COPs was proposed in this paper. In these rules,
interior penalty functions are used to evaluate feasible solutions and constraint
violations are used to evaluate infeasible solutions. Three elements are proposed to
make these rules effective in an EA: (1) a logarithmic penalty function is used to
make the algorithm convergence quickly; (2) the penalty factors are updated
according to the type of constraints which determined by a Spearman's rank-order
correlation coefficient; (3) the equalities are handled by an adaptive relax method.
Furthermore, the interior penalty rules are implemented based on a DE, namely,
SIPDE. Finally, the experiment results show that the proposed approach is
competitive with four other state-of-the-art techniques.
References
1. Mezura-Montes, E., Coello, C.: Constraint-handling in nature-inspired numerical
optimization: past, present and future. Swarm and Evolutionary Computation 1(4), 173–
194 (2011)
2. Kramer, O.: A review of constraint-handling techniques for evolution strategies. Applied
Computational Intelligence and Soft Computing 1, 1–11 (2010)
3. Tessema, B., Yen, G.: An adaptive penalty formulation for constrained evolutionary
optimization. IEEE Transactions on Systems, Man, and Cybernetics, Part A: Systems and
Humans 39(3), 565–578 (2009)
4. Deb, K.: An efficient constraint handling method for genetic algorithms. Computer
Methods in Applied Mechanics and Engineering 186(2-4), 311–338 (2000)
5. Mezura-Montes, E.: Constraint-handling in evolutionary optimization. Springer,
Heidelberg (2009)
6. Wright, M.: The interior-point revolution in optimization: history, recent developments,
and lasting consequences. Bulletin of the American Mathematical Society 42(1), 39–56
(2005)
7. Wright, M.: The interior-point revolution in constrained optimization. High-Performance
Algorithms and Software in Nonlinear Optimization, 359–381 (1998)
8. Storn, R., Price, K.: Differential Evolution–A Simple and Efficient Heuristic for global
Optimization over Continuous Spaces. Journal of Global Optimization 11(4), 341–359
(1997)
9. Runarsson, T., Yao, X.: Stochastic ranking for constrained evolutionary optimization.
IEEE Transactions on Evolutionary Computation 4(3), 284–294 (2000)
10. Mezura-Montes, E., Coello Coello, C.A., Tun-Morales, E.I.: Simple feasibility rules and
differential evolution for constrained optimization. In: Monroy, R., Arroyo-Figueroa, G.,
Sucar, L.E., Sossa, H. (eds.) MICAI 2004. LNCS (LNAI), vol. 2972, pp. 707–716.
11. Mezura-Montes, E., Coello, C.: A simple multimembered evolution strategy to solve
constrained optimization problems. IEEE Transactions on Evolutionary Computation 9(1),
1–17 (2005)
12. Wang, Y., Zixing, C.: An Adaptive Tradeoff Model for Constrained Evolutionary
Optimization. IEEE Transactions on Evolutionary Computation 12(1), 80–92 (2008)
A Novel Hybrid Algorithm for Mean-CVaR
Portfolio Selection with Real-World Constraints
Quande Qin1,2 , Li Li1 , and Shi Cheng3,4

1
Department of Management Science, Shenzhen University, Shenzhen, China
2
Research Institute of Business Analytics & Supply Chain Management,
Shenzhen University, Shenzhen, China
3
Division of Computer Science, University of Nottingham Ningbo, China
4
International Doctoral Innovation Centre, University of Nottingham Ningbo, China
qinquande@gmail.com, llii318@163.com, shi.cheng@nottingham.edu.cn
Abstract. In this paper, we employ the Conditional Value at Risk

(CVaR) to measure the portfolio risk, and propose a mean-CVaR port-
folio selection model. In addition, some real-world constraints are con-
sidered. The constructed model is a non-linear discrete optimization
problem and difficult to solve by the classic optimization techniques.
A novel hybrid algorithm based particle swarm optimization (PSO) and
artificial bee colony (ABC) is designed for this problem. The hybrid al-
gorithm introduces the ABC operator into PSO. A numerical example is
given to illustrate the modeling idea of the paper and the effectiveness
of the proposed hybrid algorithm.
Keywords: Conditional Value at Risk, CVaR, Hybrid algorithm,

Portfolio selection.
1 Introduction
Portfolio selection is concerned with the allocation of a limited capital to a
combination of securities in order to trade off the conflicting objectives of high
profit and low risk [13,17]. Since the introduction of mean-variance (MV) model
developed by Markowitz, variance has become the most popular risk measure
in portfolio selection. Variance considers high returns as equally undesirable as
low returns because high returns will also contribute to the extreme of variance.
Both theory and practice indicate the variance is not a good risk measure. Some
alternative risk measures have been proposed [11, 18]. Value at Risk (VaR) is
widely used by financial institution. However, it has its limitations, such as it
is not a coherent risk measure [1]. Rockafellar and Uryasev [15] proposed the
Conditional Value at Risk (CVaR), which is the conditional expectation of losses
above the VaR.
In practice, problem of portfolio selection has some real-world constraints,
which exacerbates the complexity. For example, it assumes that there exists a
perfect market with no tax or transaction cost. In the present study, we will con-
sider transaction cost, and floor and ceiling constraints. In addition, the least

320 Q. Qin, L. Li, and S. Cheng
unit of trading is 100 shares in stock market of China, and shares must be sub-
scribed a round lot. The modeling of such constraints involves the introduction
of integer variables. We employ CVaR to measure the risk of portfolio, and a
Mean-CVaR (MC) portfolio selection model with real-world constraints is pro-
posed. In view of the difficulty to solve this model using classical optimization
techniques, a hybrid meta-heuristics algorithm based Particle Swarm Optimiza-
tion (PSO) and Artificial Bee Colony (ABC) is designed to handle this problem.
The hybrid algorithm introduces the ABC operator into PSO. The added ABC
operator is used to evolve personal experience of the particles. The hybrid ap-
proach elegantly combines the exploitation ability of PSO with the exploration
ability of ABC.
The rest of the paper is organized as follows. Section 2 presents the back-
grounds including PSO, ABC and CVaR. Section 3 the proposed MC portfolio
selection model with real-world constraints. A hybrid algorithm based on PSO
and ABC is provided in Section 4. In Section 5, a numerical example is given.
The conclusions are drawn in Section 6.
2 Backgrounds
2.1 Particle Swarm Optimization
PSO was originally developed to emulate the flocking behavior of birds and fish
schooling [5, 9]. Each individual, called a particle, in the PSO population repre-
sents a potential solution of the optimization problem [2, 19]. The population of
PSO is referred to as a swarm, which consists of a number of particles. Particle
i at iteration t is associated with a velocity vector v ti = [vi1 t t
, vi2 , · · · , viD
t
] and
a position vector xi = [xi1 , xi2 , · · · , xiD ] where i ∈ {1, 2, · · · , N P }, N P is the
t t t t
population size. xid ∈ [ld , ud ], d ∈ {1, 2, · · · , D}, where D is the number of di-
mensions, and ld and ud are the lower and upper bounds of the dth dimension
of search space, respectively. Each particle flies through space with a velocity.
The new velocities and the positions of the particles for the next iterations are
updated using the following two equations [3–5, 9]:
t+1
vid t
= wvid + c1 r1 (pbesttid − xtid ) + c2 r2 (gbesttd − xtid ) (1)
xt+1
id = xtid + t+1
vid (2)
where w is the inertia weight; pbesti = [pbesti1 , pbesti2 , · · · , pbestiD ] is the best
position has been found by particle i, gbesti = [gbesti1 , gbesti2 , · · · , gbestiD ]
is the historically best position has been found by the whole swarm so far; c1
and c1 are acceleration coefficients. The inertia weight w is used to trade off
the exploration and exploitation; r1 and r2 represent two independently random
numbers uniformly distributed on [0, 1].
2.2 Artificial Bee Colony

ABC algorithm was proposed by simulating waggle dance and intelligent for-
aging behaviors of honeybee colonies [7]. In the ABC algorithm, there are two
A Novel Hybrid Algorithm for Mean-CVaR Portfolio Selection 321
components: the foraging artificial bees and the food source [8]. The position
of the a food source, xi = [xi1 , xi2 , · · · , xiD ], represents a possible solution and
the nectar amount of a food source corresponds to the fitness of the associated
solution. The colony of artificial bees contains three groups of bees: employed
bees, onlookers and scouts [14].
The ABC algorithm consists of four phases: initialization, employed bee, on-
looker bee and scout bee. In the initialization phase of the ABC, SN food source
positions are randomly produced with the search space. After producing food
sources and assigning them to the employed bees. In the employed bee phase of
ABC, each employed bee tries to find a better quality food source based on xi .
The new food source, denoted as ui = [ui1 , ui2 , · · · , uiD ], is calculated from the
equation below.
uij = xij + φ(xij − xsj ) (3)
where i ∈ {1, 2, · · · , SN }, where SN denotes the number of food source; j is a
randomly generated integer number in the range [1, D], φ is a randomly number
uniformly distributed in the range [−1, 1], and s is the index of a randomly chosen
solution. ABC changes each position in only one dimension at each iteration. The
source position xi in the employed bee’s memory will be replaced by the new
candidate food source position ui if the new position has a better fitness value.
Each onlooker bee chooses one of the proposed food sources depending on the
probability value pi associated with the fitness value, where

SN
pi = f iti / f itj (4)
j=1
where f iti is the fitness of the food source i. After the food source is selected,
a new candidate food source can be expressed by Eq. (3). If a food source, xi ,
cannot be improved for a predetermined number of cycles, referred to as limit,
this food source is abandoned. Then, the scout produces a new food source
randomly to replace xi .
2.3 Conditional Value at Risk
Let L(x, y) be the loss function with weight vector x and the return rate vector
y. Let p(r) be the density function of the return rate vector y. Then L(x, y)
is random variable dependent on x. The probability of L(x, y) not exceeding a
threshold α is given by
ψ(x, α) = p(y)dy (5)

L(x,y)≤α
The VaR of the loss associated with x and a specified probability level β in
(0, 1) is the value
V aRβ (x) = min{α ∈ Rm : ψ(x, α) ≥ β} (6)

As an improved risk measure, CVaR, is the expected portfolio return, conditioned

on the portfolio returns being lower than VaR. It is defined as the Eq. (7).
Compared with VaR, CVaR has some superior mathematical properties.
CV aRβ (x) = E[L(x, y)|L(x, y)) ≥ V aRβ (x)]
= (1 − β)−1 L(x, y)p(y)dy (7)

L(x,y)≥V aRβ (x)
CVaR can be obtained by the following equation based on reference [15]
Fβ (x, α) = α + (1 − β)−1 [L(x, y) − α]+ p(y)dy (8)

y∈RM
where (a)+ is defined as max(a, 0).
3 The Proposed Portfolio Selection Model

In this section, we discuss the MC portfolio selection model. Assume there n
risky asset and one risk-free asset in a financial market for trading. An investor
hopes to allocate his/her initial wealth m0 . For notational convenience, we first
introduce the following notations:
– ri : the return of risky asset i.
– rf : the return of risk-free asset.
– ti : the transaction cost of risky asset i;
– s(x): the total return of the portfolio.
– pi : the price of risky asset i each round lot;
– ki : the round lot of risky asset i invested;
– σi : the highest limits on risky asset i;
– εi : the lowest limits on risky asset i;
– λ the acceptable return of the portfolio.
n
The
n capital invested in risk assets is i=1 ki pi and the remaining n capital m0 −
i=1 ki p i invested in the risk-free asset. Obviously, it holds
n that i=1 ki pi ≤ m0 .
The transaction cost are consider, and it denotes as i=1 ti ki . Thus, the total
return s(x) of the portfolio can be described as follows:

n
n
n
s(x) = ki pi ri + rf (m0 − ki pi ) − ti ki
i=1 i=0 i=1

n
= rf m0 + [ki pi (ri − rf ) − ti ki ] (9)
i=1
The intention of the proposed model is to minimize the CVaR in the case of the
return of the portfolio is equal or greater than λ.
min z = CV aR (10)
⎧
⎪
⎪ εi ≤ xi ≤ σi i = 1, 2, · · · , n
⎨
s(x)/m ≥ λ
s.t. n 0
⎪
⎪ ki pi ≤ m0
⎩ i=1
ki ≥ 0, integer, i = 1, 2, · · · , n
where x = (k1 p1 /m0 , k2 p2 /m0 , · · · , kn pn /m0 ) is the weight vector. In practice,

asset i is chosen to be invested and the weight lies in [εi , σi ], where 0 ≤ εi ≤
σi ≤ 1. The first constraint is called floor and ceiling constraints. The second
constraint is used to ensure the return of the portfolio.
4 A Hybrid Algorithm Based on PSO and ABC

Due to the simple concept and efficiency of converging to reasonable solution
fast, PSO has been successfully applied to a wide range of real-world problems.
Despite the competitive performance of PSO, researchers have noted a major
problem associated with the PSO is its premature convergence when solving
complex problems [12]. ABC algorithm is good at exploration but poor at ex-
ploitation [20]. From the analysis of the merits and demerits of PSO and ABC,
it is intuitive that hybridizing the PSO and ABC is a potential way to design
an effective algorithm.
Generally, the locality of personal best position in PSO algorithm is distant
from the global optimum. Once the swarm aggregates to such position, little
opportunity is afforded for the swarm to explore for other solution and find the
global optimum. This leads to the swarm suffer from premature convergence
easily, especially when solving complicated multimodal problems. Thus, the evo-
lution of the personal experience will promote the exploration of the personal
experience space, which could potentially enhance PSO’s performance. ABC has
better ability to explore, which is beneficial to global search, but poor ability of
exploitation. In this paper, we utilize the ABC operator to evolve the personal
best position when the personal best position stagnated. It is expected that the
proposed hybrid algorithm, PSOABC, combines the merits of PSO and ABC,
and have capabilities of escaping from local optima and converge fast.
In PSOABC algorithm, we use PSO in the main loop. When the fitness of
pbesti , denoted as f it(pbesti ), has not improved within a predefined num-
ber of successive iterations, denoted as k, it is considered to be stagnated and
trapped into local optima. The setting of k is set to 3 in this paper. We only
use the employed bee operator in ABC algorithm to evolve pbesti in this work.
The pseudo-code of the PSOABC algorithm is described in Algorithm 1. When
pbesti stagnated, we can use the employed bees operator to evolve pbesti . The
mathematical expressions of this ABC operator described as follows:
zij = pbestij + φ(pbestij − pbestsj ) (11)
where s are randomly selected integers from the index of all solution with s = i.
j is a randomly selected dimension number. φ is a randomly number uniformly
distributed within the interval [−1, 1].
5 Numerical Example
The portfolio selection model constructed is a non-linear discrete optimization
problem. The proposed hybrid algorithm based on PSO and ABC is suitable
Algorithm 1. The pseudo-code of PSOABC algorithm

1 Initialization: set up all parameters;
2 Set the maximum iteration number F Es; t = 1, Stop = 0;
3 Evaluate the fitness of the swarm and determine pbesti and gbest ;
4 while the stopping criteria is not satisfied do
5 for i = 1 : N P do
6 for d = 1 : D do
7
t+1
vid = wvidt
+ c1 r1 (pbesttid − xtid ) + c2 r2 (gbesttd − xtid );
t+1 t+1
8 xid = xtid + vid ;
9 i = i + 1;
10 Evaluate the fitness of the particle i; Update pbesti and gbest ;
11 if f it(pbestti ) − f it(pbestt−1
i ) = 0 then
12 Stop(i) = Stop(i) + 1 ;
13 else
14 Stop(i) = 0;
15 for i = 1 : N P do
16 if Stop(i) ≥ k then
17 zij = pbestij + φ(pbestij − pbestsj );
18 if f it(z i < f it(pbesti ) then
19 pbesti = z i ;
20 t=t+1
for real-valued problems. Kitayama et al. utilized penalty function approach

handle the discrete decision variables [10]. In this approach, the discrete decision
variables are handled as the continuous ones by penalizing at the intervals. The
penalty function is given as the following the Eq. (12).
n " #
1 2π{xcm+i − 0.25(di,j+1 + 3di,j )}
φ(x) = sin +1 (12)
i=1
2 di,j+1 − di,j
where di,j and di,j+1 represents the discrete decision variables. xcm+i is the con-
tinuous decision variables between di,j and di,j+1 .
We select 20 stocks from Chinese security market, as shown in Table 1. The
symbol of m(%) in Table 1 denotes the expected return. The requirement of
selecting the average yield is greater than 0. This paper selected raw data for
the weekend’s closing price.
Assuming the investor has 500 million investment funds. According to the tax
and commission in Chinese securities market, the transaction cost rate is set to
0.4%. The minimum invest weigh of each stock is 0, and the maximum weight is
10%. The risk-free return rate is equal to 4.14% based on one-year deposit rate
in China, and λ is 4.5%.
Experimental results among genetic algorithm (GA), PSO-w [16], basic ABC
[6] and PSOABC are compared. For a fair comparison, the population size is
Table 1. Stocks selected and expected return rate
Ticker m(%) Ticker m(%)

000002 0.45 600631 0.25
000039 0.46 600642 0.5
600058 0.77 600649 0.18
600098 0.63 600663 0.1
600100 0.12 600688 0.26
600115 0.35 600690 0.09
600183 0.4 600776 0.22
000541 0.26 600811 0.3
000581 0.53 600812 0.29
600600 0.37 600887 0.18
Table 2. Experimental results comparison
β = 90% β = 95% β = 99%

Algorithm
Mean SD Mean SD Mean SD
GA 0.0454 0.0019 0.0527 0.0064 0.0737 0.0042
PSO-w 0.0428 0.0014 0.0519 0.0044 0.0743 0.0057
ABC 0.0412 0.0015 0.0479 0.0027 0.0632 0.0024
PSOABC 0.0336 0.0009 0.0343 0.0016 0.0443 0.0013
set to 40 for all algorithms, the maximum iteration is 3500. The selection rate,
crossover rate and mutation rate is set to 0.9, 0.7 and 0.03, respectively. Other
parameter settings in each algorithm are used according to their original refer-
ences. All algorithms run 30 times independently. The experimental results are
shown in the Table 2. In Table 2, “Mean” indicate the mean values of CVaR,
and “SD” stands for the standard deviation. From Table 2, it can be seen that
PSOABC has a good performance and is a good alternative for the proposed
portfolio selection model.
6 Conclusions
In this work, we proposed a MC portfolio selection model. In this model, the

portfolio risk is measured by CVaR and some real-world constraints are added.
Note that the round lot, which involves the introduction of integer variables,
is considered. We have proposed a novel hybrid algorithm to solve the portfolio
selection problem. The proposed algorithm introduces the ABC operator to PSO
in order to balance exploration and exploitation. A penalty function is adopted
to transform the discrete portfolio selection model into a continuous one. A
numerical example is given to illustrate the modeling idea of the paper, and the
experimental results show that the proposed hybrid algorithm outperforms is
highly competitive for this portfolio problem.
Acknowledgment. This work is partially supported by Natural Science Foun-

dation of China under grant NO.71240015, 60975080, 61273367, 51305216, Natu-
ral Science Foundation of Guangdong Province under grant No.S2011010001337,
Foundation for Distinguished Young Talents in Higher Education of Guang-
dong, China, under grant 2012WYM 0116 and the MOE Youth Foundation
Project of Humanities and Social Sciences at Universities in China under grant
13YJC630123, and Ningbo Science & Technology Bureau (Science and Technol-
ogy Project No.2012B10055). This work was carried out at the International
Doctoral Innovation Centre (IDIC). The authors acknowledge the financial sup-
port from Ningbo Education Bureau, Ningbo Science and Technology Bureau,
China’s MOST and The University of Nottingham.
References
1. Artzner, P., Delbaen, F., Eber, J.M., Heath, D.: Coherent measures of risk. Math-
ematical Finance 9(3), 203–228 (1999)
2. Chen, X., Li, Y.: A modified pso structure resulting in high exploration ability with
convergence guaranteed. IEEE Transactions on Systems, Man, and Cybernetics,
Part B: Cybernetics 37(5), 1271–1289 (2007)
3. Cheng, S.: Population Diversity in Particle Swarm Optimization: Definition, Ob-
servation, Control, and Application. Ph.D. thesis, Department of Electrical Engi-
neering and Electronics, University of Liverpool (2013)
4. Cheng, S., Shi, Y., Qin, Q.: Population diversity of particle swarm optimizer solving
single and multi-objective problems. International Journal of Swarm Intelligence
Research (IJSIR) 3(4), 23–60 (2012)
5. Eberhart, R., Kennedy, J.: A new optimizer using particle swarm theory. In: Pro-
ceedings of the Sixth International Symposium on Micro Machine and Human
Science, pp. 39–43 (1995)
6. Karaboga, D.: An idea based on honey bee swarm for numerical optimization. Tech.
rep., Erciyes University, Engineering Faculty, Computer Engineering Department
(October 2005)
7. Karaboga, D., Basturk, B.: A powerful and efficient algorithm for numerical func-
tion optimization: artificial bee colony (ABC) algorithm. Journal of Global Opti-
mization 39(3), 459–471 (2007)
8. Karaboga, D., Basturk, B.: On the performance of artificial bee colony (ABC)
algorithm. Applied Soft Computing 8(1), 687–697 (2008)
9. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of IEEE
International Conference on Neural Networks (ICNN), pp. 1942–1948 (1995)
10. Kitayama, S., Arakawa, M., Yamazaki, K.: Penalty function approach for the mixed
discrete nonlinear problems by particle swarm optimization. Structural and Mul-
tidisciplinary Optimization 32(3), 191–202 (2006)
11. Konno, H., Yamazaki, H.: Mean-absolute deviation portfolio optimization model
and its applications to tokyo stock market. Management Science 37(5), 519–531
(1991)
12. Liang, J.J., Qin, A.K., Suganthan, P.N., Baskar, S.: Comprehensive learning parti-
cle swarm optimizer for global optimization of multimodal functions. IEEE Trans-
actions on Evolutionary Computation 10(3), 281–295 (2006)
13. Markowitz, H.: Portfolio Selection. The Journal of Finance 7(1), 77–91 (1952)
14. Qin, Q., Cheng, S., Li, L., Shi, Y.: Artificial bee colony algorithm: A survey. CAAI
Transactions on Intelligent Systems 9(2), 127–135 (2014)
15. Rockafellar, R.T., Uryasev, S.: Optimization of conditional value-at-risk. Journal
of Risk 2(3), 21–41 (2000)
16. Shi, Y., Eberhart, R.: A modified particle swarm optimizer. In: Proceedings of the
1998 Congress on Evolutionary Computation (CEC1998), pp. 69–73 (1998)
17. Yoshimoto, A.: The mean-variance approach to portfolio optimization subject to
transaction costs. Journal of the Operations Research Society of Japan 39(1), 99–
117 (1996)
18. Young, M.R.: A minimax portfolio selection rule with linear programming solution.
Management Science 44(5), 673–683 (1998)
19. Zhang, G., Li, Y.: Orthogonal experimental design method used in particle swarm
optimization for multimodal problems. In: The Sixth International Conference on
Advanced Computational Intelligence (ICACI 2013), pp. 183–188 (October 2013)
20. Zhu, G., Kwong, S.: Gbest-guided artificial bee colony algorithm for numerical
function optimization. Applied Mathematics and Computation 217(7), 3166–3173
(2010)
A Modified Multi-Objective Optimization
Based on Brain Storm Optimization Algorithm
Lixia Xie, Yali Wu
Xi’an University of Technology, Xi’an Shaanxi

710048, China
Abstract. In recent years, many evolutionary algorithms and population-based

algorithms have been developed for solving multi-objective optimization
problems. In this paper, A new Multi-objective optimization algorithm-Modified
Multiobjective Brain Storm Optimization (MMBSO) algorithm is proposed. The
clustering strategy acts directly in the objective space instead of in the solution
space and suggests potential Pareto-dominance areas in the next iteration. A
Density-Based Algorithm for Discovering Clusters in Large Spatial Databases
with Noise (DBSCAN) clustering and Differential Evolution (DE) mutations are
used to improve the performance of MBSO. A group of multi-objective problems
with different characteristics were tested to validate the usefulness and
effectiveness of the proposed algorithm. Experimental results show that
MMBSO is a very promising algorithm for solving these tested multi-objective
problems.
Keywords: Brain Storm Algorithm, Clustering Technique, Multi-objective

Optimization, Pareto-dominance.
1 Introduction
Many real world problems are commonly looked at from a variety of perspectives, and
therefore are represented as multiple objectives which usually conflict with each other.
These problems are called Multi-objective problems, which have gained much
attention in the study of sciences, economic, engineering, etc. The optimum solution for
a multi-objective optimization problem is not unique but a set of candidate solutions. In
the candidate solution set, no solution is better than any other one with regards to all
objectives. This set is named as Pareto-optimal set, and the associated objective vectors
form the trade-off surface, also called Pareto-front, in the objective space.
During the last decades, a number of evolutionary algorithms and population-based
methods have been successfully used to solve multi-objective optimization problems.
For example, there are Multiple Objective Genetic Algorithm (MOGA) [1],
Nondominated Sorting Genetic Algorithm (NSGA, NSGA II)[2][3] , Strength Pareto
Evolutionary Algorithm (SPEA, SPEA II) [4][5], Multi-objective Particle Swarm
Optimization (MOPSO) [6], to name just a few. Most of the above algorithms can
improve the convergence and distribution of the Pareto-front more or less.
A Modified Multi-Objective Optimization Based on Brain Storm Optimization Algorithm 329
Human beings, as one kind of social animals, are the most intelligent in the world.
When we face a difficult problem which every single person cannot solve, group
person, especially with different background, get together to brain storm, the problem
can usually be solved with high probability. Being inspired by this human idea
generation process, Shi [6] proposed a novel optimization algorithm - Brain Storm
Optimization (BSO) algorithm. The simulation results on two single-objective
benchmark functions validated the effectiveness and usefulness of the BSO to solve
optimization problems. In [8], two novel component designs were proposed to modify
the BSO algorithm and it has significantly enhanced the performance of BSO. In [9]
and [10], a multi-objective optimization algorithm based on the brainstorming process
was developed. Simulation results illustrated that it can be a good optimizer for solving
multi-objective optimization problems.
In this paper, a modified Multi-objective BSO (MMBSO) algorithm with clustering
strategy in the objective space is proposed to solve multi-objective optimization
problems. Instead of action on the population and on the obtained Pareto front, in the
MMBSO, the clustering strategy acts directly on the objective vectors in the objective
space. Then this operation gives a feedback to the decision space to decide which
candidate solution should survive. The novel using of the clustering technique,
especially for the multi-objective optimization problems with high dimensional
decision vectors could reduce computational burden. Clustering and mutation, the main
operators of BSO were analyzed by using a Density-Based Algorithm for Discovering
Clusters in Large Spatial Databases with Noise (DBSCAN) clustering and Differential
Evolution (DE) mutation which is different from the previous operator. Then the
different dimensions of bench functions that named ZDT [3] were tested. The
simulation results showed that MMBSO would be a promising algorithm in solving
multi-objective optimization problems.
The remaining paper is organized as follows. Section 2 briefly reviews the related
works about the BSO and the MOP. In Section 3, the Modified Multi-objective BSO
(MMBSO) is introduced and described in detail. Section 4 contains the simulation
results and discussion. Finally, Section 5 provides the conclusions and some possible
paths for future research.
2 Related Work
2.1 Multi-Objective Optimization Problem (MOP)

Without loss of generality, all of the multi-objective optimization problems can be
formulated as minimization optimization problems. Let us consider a multi-objective
optimization problem:
Minimize F( X) = ( f1 ( X) , f 2 ( X) , , f M ( X)) (1)
Where X = ( x1 ,  , xD ) ∈ ℜ D is called the decision vector in the D dimensional
search space and F ∈ Ω M is the objective vector with M objectives in the M
dimensional objective space. The basic concepts of a minimization MOP can be
described in [11].Two goals of a multi-objective optimization are the convergence to
330 L. Xie and Y. Wu
the true Pareto-optimal set, and the maintenance of diversity of solutions in the Pareto
front set. Many performance metrics have been suggested to measure the performance
of multi-objective optimization algorithms. In this paper, we use the metric ϒ and
metric Δ , which were defined by Deb et al. in [3] , to measure the performance of the
MBSO algorithm.
2.2 Brainstorm Optimization Algorithm

The BSO algorithm is designed based on the brainstorming process [12] . In the
brainstorming process, the generation of the idea obeys the Osborn’s original four rules
[12] . The people in the brainstorming group will need to be open-minded as much as
possible and therefore generate more diverse ideas. Any judgment or criticism must be
held back until at least the end of one round of the brainstorming process, which means
no idea will be ignored. The algorithm is described as follows. In the initialization, N
potential individuals were randomly generated. During the evolutionary process, BSO
generally uses the clustering technique, mutation operator and selection operator to
create new ideas based on the current ideas, so as to improve the ideas generation by
generation to approach the problem solution. In the clustering technique, BSO uses a
k-means clustering [11] . In the mutation operator, BSO creates N new individuals one
by one based on the current ideas. To create a new individual, BSO first determines
whether to create the new individual based on one selected cluster or based on two
selected clusters. After the cluster(s) have been selected, BSO then determines whether
create the new idea based on the cluster center(s) or random idea(s) of the cluster(s). No
matter to use the cluster center or to use random idea of the cluster, we can regard the
selected based idea as X selected which can be expressed
as X selected = ( x1selected , xselected
2
,  , xselected
d
) , then applying a mutation of the X selected to get
the new idea X new which can be expressed as X new = ( x1new , xnew
2
,  , xnew
d
) . After the new
idea X new has been created, BSO evaluates X new and replaces X selected if X new has a better
fitness than X selected . The procedure of the BSO algorithm is shown in [6].
In fact, there have been several recent proposals to extend BSO to handle
optimization problems. For example, A Modified Brain Storm
Optimization[8],Predator–Prey Brain Storm Optimization for DC Brushless Motor
[15],Brain Storm Optimization [13],Solution Clustering Analysis in Brain Storm
Optimization Algorithm [14], Brain Storm Optimization Algorithm for Multi-objective
Optimization Problems[9], and Multi-objective Optimization Based on Brain Storm
Optimization Algorithm(MBSO)[10]. In [10], The BSO is used to solve
multi-objective problems in which the k-means cluster and Gaussian mutation was
used. Besides, Cauchy mutation was also utilized.
3 Modified Multi-Objective Brain Storm Optimization

Algorithm (MMBSO)
The paper makes improvements about clustering and mutation operations for the paper
[10]. DBSCAN clustering and differential mutation was used to improve the original
algorithm. Also a probability of generating a random individual is added to increase the
diversity of algorithm.
3.1 Clustering Technique

In the Multi-objective Brain Storm Optimization Algorithm, the k-means cluster
algorithm was used in the clustering technique, but the k-means cluster algorithm has
two disadvantages: First, it must specify the cluster center in advance, thus it could
have been the idea of the same class was assigned to a different class and then makes
the clustering lost its original role in the algorithm. And the second is that it must
determine the number of cluster, and the number of cluster is fixed which is not
changed with the idea of change in each iteration. In MMBSO, the clustering technique
is implemented by a clustering method based on density named DBSCAN(A
Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with
Noise)[19] .We can image that when the similar ideas beyond a certain number, they
can be put together as a class, then the ideas generated in each iteration can be fully
used and thus makes different number of clusters according to the different ideas in
each iteration. There are a lot of density-based clustering algorithms currently, in this
article, in order to make the algorithm simple we use the simplest and most classical
density-based clustering algorithm which is named DBSCAN. The procedure of it is
shown in [19].
3.2 Generation Progress

The generation progress of MMBSO contains the mutation and selection operator
which can be referred as follows:
 Mutation Operator:
The mutation operator plays an important role in the generation progress. Gaussian
and Cauchy mutation are typical mutation operators which were used in [10]. In solving
some multimodal or multi-objectives optimization problems, Gaussian mutation as one
of the main operators in the algorithm may lead to a result with a slow convergence to a
good near-optimum. In MMBSO, Differential Mutation is used to improve the
performance of the algorithm.
As to BSO, a new idea X new is created by adding Gaussian random noise to a based
idea X selected . It also can be known that the noise is large in early evolutionary phase and
gradually become smaller during the running by the control of the logarithmic sigmoid
transfer function according to [8]. Such a time varying noise strategy is consistent with
the commonly intuition that large noises are needed in the early phase for global search
while small noises are needed in the late phase for local fine-tuning.
In this paper, we propose to use the Differential Mutation to produce the noise value.
The Differential Mutation is based on such a consideration. In the human being’s
brainstorming process, we can image that at the beginning of the process, everyone’s
idea would be much different. When they create new ideas based on the current ideas,
they should take the differences of the current ideas into consideration. For example,
when creating a new idea X new based on a current idea X selected , two distinct random
ideas X a which can be expressed as X a = ( x1a , xa2 ,  , xad ) and X b which can be
expressed as X b = ( xb1 , xb2 , , xbd ) from all the current ideas are taken to represent the
idea difference, and the X new is created as:
d
xnew = xselected
d
+ rand (0,1) d × ( xad − xbd )
(2)
Where rand (0,1)d is a random number between (0, 1).
Using Eq. (2) to create new ideas, there are two advantages. Firstly, the
computational burden of (2) is much lighter than that of the mutation of BSO that
involves logarithmic sigmoid transfer function, Gaussian distribution function, random
function, addition, subtraction, multiplication, and division, while (2) involves random
function, multiplication, and subtraction for making up the noise value. Secondly, Eq.
(2) can match the search environment of the evolutionary process. Be consistent with
the brainstorming process for human being in solving problem, the ideas are much
different from each other in the beginning, therefore the term ( xad − xbd ) in (2) is larger
and the new created ideas can keep the diversity in the early phase. In the late phase of
the brainstorming process, the people may reach a consensus and the idea difference
may be smaller. In this condition, the term ( xad − xbd ) in (2) is also smaller to help refine
the ideas. Therefore, Eq. (2) may be good at balance the global search and local search
abilities according to the search information during the evolutionary process.
 Selection Operator:
It is also quite important to decide whether any newly generated solution should
survive to the next generation. The selection based on Pareto dominance is utilized in
this paper.
3.3 Update the Pareto Set

The Pareto Set is updated by the new non-dominated solutions. In this step, each new
non-dominated solution obtained in the current iteration will be compared with all
members in the Pareto Set. If the size of the Pareto Set exceeds the maximum size limit,
it is truncated using the diversity consideration. In this paper, the circular crowded
sorting operator [18] is adopted to guide the points toward a uniformly spread-out
Pareto-optimal front.
3.4 The Procedure of the MMBSO

The MMBSO contains clustering, mutation, selection, and updating the Pareto set,
which have been described above. The whole procedure of the MMBSO is shown in
Fig. 1.
In the process of MMBSO algorithm, population size N imitate the number of ideas
generated during the course of each round; randomly selects an elite cluster imitate the
process of generating new ideas which excited by a single thought. Randomly selects
two clusters was imitated by the idea of two people from different clusters inspired by
the process of generating new ideas. Select an individual from the archive set was used
to keep a few good ideas which imitate to pick up several good ideas in the brain storm
process. The clustering technique is used to classify the idea of people with different
backgrounds and ideas. And in each cluster, the cluster center acts as a facilitator of the
problem to solve problems with better idea. The mutation operator is used to generate
new ideas on the basis for the existing ideas. Also different from MBSO, a probability
of generating a random individual is added to increase the diversity of algorithm.
4 Experiments and Discussions
In this section the MMBSO will be tested. Without loss of generality, all the
multi-objective optimization problems tested in this paper are minimization problems.
4.1 Test Problems

In order to evaluate the performance of MMBSO, the ZDT test functions [3] are used in
this paper. The ZDT suite is comprised of six problems, each one presenting a specific
characteristic that generally cause difficulties to major evolutionary optimization
strategies [20]. The bi-objective test functions used to examine the effect of the
introduced MMBSO are the ZDT1, ZDT2, ZDT3, ZDT4 and ZDT6. Test problem ZDT1
and ZDT3 have convex Pareto fronts while other test problems have non-convex Pareto
fronts; ZDT3 also possesses a disconnected Pareto front; the Pareto front of ZDT6 is
non-uniformly spaced; and ZDT4 as a complex multimode problem is difficult to find the
global Pareto front. The information of these test functions in detail can be seen in [3].
4.2 Parameter Settings

During the test, a lot of parameters are used to test the algorithm. Finally a set of
parameter that is relatively good for these test functions is used. In all the simulation
runs, the population size is set to be 200 and the maximum size of the Pareto set is fixed
at 100. After conducting a series of experiments, the pre-determined probability values
P1 is set to be 0.99, P2 and P3 are set to 0.8, P4 and P5 are set to 0.2. For the DBSCAN
clustering, MinPts is set to 7 while ε is set according to [21]. All of the algorithms are
implemented in MATLAB using a real-number representation for decision variables.
For each experiment, 30 independent runs were conducted to collect statistical results.
Each test problem will be run with different dimension, 5, 10, 20, and 30, respectively.
Begin
Initialization N
individuas
Clustering and seek for

cluster centers
N
P < P1
Y
N
P < P2
Y
Y N
P < P3
Randomly select Randomly select two clusters,

an elite cluster one for the elite cluster
Randomly
Randomly generate
generate p,0<p<1
p,0<p<1
N P < P5
P < P4 N
Y
randomly select an indi-
select two cluster centers
Y vidual from two clusters
x1selected , x2selected x1selected , x2 selected
Randomly select an
Selece a cluster center denoted as x selected
individual denoted as
xselected
xselected = C1 × x1selected + (1 − C1 ) × x2 selected
Where C1 is random generated from 0<C1<1
Randomly select an individual from archive set and denoted as x selected
Randomly Generate an indiv-idual and denoted as x selected
d
xnew = xselected
d
+ rand*(xad -xbd )
f ( xnew ) dominates f ( xselected )

N Y
xi = xselected xi = xnew
Update the archive set
N
Termination condition
Y
End
Fig. 1. The procedure of the MMBSO
4.3 Results
In all simulation runs, the metric ϒ [3]and metric Δ [3]will be calculated and recorded
for all the test problems. Table 1 compares the best and mean values of the convergence
metric ϒ obtained using MMBSO (denoted as DE), MBSO-G (MBSO with Gaussian
mutation)[10] and MBSO-C (MBSO with Cauchy mutation)[10]. The diversity

metric Δ about the test problems are listed in Table 2. The best results are marked with
italics and bold.
Table 1. The comparisons of best and mean value of γ between MMBSO,MBSO-G and
MBSO-C
Dime- Algo- ZDT1 ZDT2 ZDT3 ZDT4 ZDT6

nsion rithm best mean best mean best mean best mean best mean
5 DE 0.0010 0.0011 0.6806 0.7904 0.0010 0.0012 0.0062 0.1898 0.0037 0.0040
e-003 e-003
G 0.0011 0.0016 0.0007 0.0008 0.0012 0.0015 0.0009 0.0019 0.0040 0.0048
C 0.0011 0.0017 0.0009 0.0011 0.0012 0.0014 0.0010 0.0048 0.0039 0.0050
10 DE 0.0009 0.0011 0.6803 0.7888 0.0010 0.0012 0.0015 1.9661 0.0036 0.0041
e-003 e-003
G 0.0050 0.0079 0.0014 0.0050 0.0024 0.0033 0.0029 2.4179 0.0046 0.0079
C 0.0031 0.0062 0.0016 0.0029 0.0018 0.0027 0.0011 0.1335 0.0046 0.0072
20 DE 0.0010 0.0011 0.7047 0.8018 0.0011 0.0012 0.0016 6.7146 0.0037 0.0040
e-003 e-003
G 0.0328 0.0472 0.0294 0.0498 0.0168 0.0248 2.8367 18.768 0.0323 0.0412
C 0.0193 0.0312 0.0186 0.0311 0.0111 0.0150 1.4905 4.4188 0.0151 0.0241
30 DE 0.0010 0.0011 0.0007 0.0008 0.0011 0.0012 2.9322 13.8379 0.0037 0.0040
G 0.1060 0.1347 0.1006 0.1308 0.0863 0.1136 10.3624 38.2581 0.0928 0.1385
C 0.0695 0.0912 0.0725 0.0905 0.0443 0.0589 6.4966 15.2905 0.0580 0.0813
Table 2. The comparisons of best and mean value of Δ between MMBSO,MBSO-G and
MBSO-C

5 DE 0.1074 0.1266 0.1031 0.1228 0.4128 0.4180 0.6160 1.0436 0.5294 0.5422
G 0.3478 0.4187 0.3363 0.4016 0.5667 0.6013 0.3550 0.5058 0.6635 0.7021
C 0.3325 0.4066 0.3422 0.4126 0.5016 0.5763 0.3444 0.4786 0.6627 0.7002
10 DE 0.1087 0.1259 0.1031 0.1213 0.4113 0.4195 0.1413 1.2555 0.5325 0.5473
G 0.3765 0.4370 0.3475 0.4258 0.5420 0.6054 0.5159 0.8372 0.6751 0.7050
C 0.3882 0.4542 0.3682 0.4115 0.5357 0.5931 0.4330 0.7865 0.6633 0.6989
20 DE 0.1098 0.1275 0.0955 0.1208 0.4117 0.4173 0.1138 1.2331 0.5339 0.5443
G 0.4238 0.4789 0.3975 0.4869 0.5381 0.5868 0.9342 0.9748 0.6823 0.7103
C 0.4215 0.4717 0.4371 0.4855 0.5073 0.5795 0.8660 0.9569 0.6724 0.7038
30 DE 0.1008 0.1257 0.0997 0.1253 0.4126 0.4188 1.2958 1.3968 0.5322 0.5436
G 0.4823 0.5340 0.4834 0.5494 0.6026 0.6364 0.9619 0.9890 0.7080 0.7452
C 0.5105 0.5529 0.4898 0.5588 0.5708 0.6293 0.8362 0.9699 0.6969 0.7425
In the Table 1 and 2, DE is represents MMBSO,G is MBSO-G and C is MBSO-C.

The result from table 1 ~ 2 showed that MMBSO has better convergence and diversity
than MBSO-G and MBSO-C except ZDT4. For the ZDT4, it is a complex multimode
problem which is difficult to find the global Pareto front. DBSCAN which used in
MMBSO uses the current ideas to get the clustering result, so it is limited the ZDT4
find the global optimum to some extent while k-means cluster can made a slightly good
performance for ZDT4.
To further verify the performance of the algorithm, the hypervolume(HV)[22] which
can evaluate convergence and diversity at the same time is added to illustrate the
performance of the algorithm. The Table 3 compares the best and mean values of the
ratio of HV that get by the solution set of algorithm and the corresponding true solution
on the Pareto front. The ratio is between 0-1. If the ratio equals 1, then the solution
obtained by the algorithm is on the true Pareto frontier. The more the ratio is close to 1,
the solution obtained by algorithm is closer to the true Pareto frontier. The best results
are marked with italics and bold.
Table 3. The comparisons of best and mean value of reaching the target HV value between
MMBSO,MBSO-G and MBSO-C

5 DE 0.9945 0.9945 0.9892 0.9891 0.9975 0.9974 0.9988 0.9944 0.9945 0.9943
G 0.9908 0.9891 0.9811 0.9795 0.9958 0.9951 0.9952 0.9915 0.9938 0.9909
C 0.9976 0.9910 0.9986 0.9874 1.5494 1.3569 0.9961 0.8582 0.9989 0.9939
10 DE 0.9945 0.9944 0.9931 0.9890 0.9975 0.9974 0.9998 0.9962 0.9944 0.9938
G 0.9861 0.9831 0.9773 0.9700 0.9922 0.9904 0.9890 0.7215 0.9869 0.9825
C 0.9939 0.9779 0.9924 0.9476 0.9939 0.9774 0.9797 0.7009 0.9942 0.9802
20 DE 0.9950 0.9944 0.9956 0.9888 0.9975 0.9973 0.9999 0.9983 0.9942 0.9933
G 0.9729 0.9579 0.9493 0.9146 0.9706 0.9553 0.7565 0.4163 0.9786 0.9518
C 0.9720 0.8955 0.9450 0.8078 1.1744 1.1016 0.9775 0.7821 0.9577 0.9130
30 DE 0.9945 0.9943 0.9891 0.9889 0.9974 0.9973 0.9991 0.9989 0.9943 0.9930
G 0.9393 0.9069 0.8898 0.8203 0.9333 0.8824 0.7124 0.3065 0.9455 0.8990
C 0.9298 0.7910 0.8709 0.6943 0.9847 0.9403 0.9732 0.7143 0.8725 0.7973
In the Table 3, DE is represents MMBSO,G is MBSO-G and C is MBSO-C.As can

be seen from Table 3, the results obtained by MMBSO are better than MBSO-G and
MBSO-C for the dimensions 5, 10, 20 and 30. Most of the obtained solution reached by
MMBSO is more than 0.99, which illustrates the algorithm has good performance.
The Fig. 2-5 shows the simulation result. In the figure, the fine line represents the
true Pareto edge and the point represents the solution get by the algorithm.
1 1.4 2 1.2
ZDT1--D--DE--5D ZDT2--D--DE--5D 1 ZDT3--D--DE--5D ZDT4--D--DE--5D ZDT6--D--DE--5D
1.2
0.8 TruePareto TruePareto TruePareto TruePareto 1 TruePareto
1.5
1 0.5
0.8
0.6 0.8
f2
f2
f2
f2
1
f2
0.6
0
0.4 0.6
0.4
0.4
-0.5 0.5
0.2
0.2 0.2
0 0 -1 0 0
0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1 0.2 0.4 0.6 0.8 1
f1 f1 f1 f1 f1
Fig. 2. Pareto-optimal front of ZDT1, 2, 3, 4 and 6 obtained by the MMBSO (5 dimension)
1 1.4 1 1.2
1.2 TruePareto 1
0.8 TruePareto TruePareto 0.8 TruePareto TruePareto
1 0.5
0.8
0.6 0.8 0.6
f2
f2
f2
f2
f2
0.6
0
0.4 0.6 0.4
0.4
0.4
-0.5
0.2 0.2
0.2 0.2
0 0 -1 0 0
0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1 0.2 0.4 0.6 0.8 1
f1 f1 f1 f1 f1
1 1.4 1 1.2
1.2 TruePareto 1
0.8 TruePareto TruePareto 0.8 TruePareto TruePareto
1 0.5
0.8
0.6 0.8 0.6
f2
f2
f2
f2
f2
0.6
0
0.4 0.6 0.4
0.4
0.4
-0.5
0.2 0.2
0.2 0.2
0 0 -1 0 0
0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1 0.2 0.4 0.6 0.8 1
f1 f1 f1 f1 f1
1 1.4 120 1.2

1.2 TruePareto 100 1
0.8 TruePareto TruePareto TruePareto TruePareto
1 0.5
80 0.8
0.6 0.8
f2
f2
f2
f2
f2
60 0.6
0
0.4 0.6
40 0.4
0.4
-0.5
0.2
0.2 20 0.2
0 0 -1 0 0
0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1 0.2 0.4 0.6 0.8 1
f1 f1 f1 f1 f1
The result from table 1~2, also with the Fig. 2-6 show that MMBSO does better than
MBSO on all functions except some result of ZDT4. This may be due to that DBSCAN
clustering can make good use of the ideas generated in each iteration and get a
reasonable clustering result according to the current idea, thus get a good current
optimal solution. In addition, Differential Evolution mutation can match the search
environment to provide suitable noise to create better ideas around the global optimal
region to refine the solution for high accuracy. In MBSO, the new created idea was
disturbed based on the current idea by Gaussian noise. However, this noise may be
coarse. In the contrast, MMBSO uses the difference between two ideas as the disturbed
noise. This way, the disturbed noise can be within a comparable order of magnitude
with the current ideas. With the combination of DBSCAN clustering and Differential
Evolution mutation, MMBSO get a good convergence and the diversity.
5 Conclusions and Discussion
In this paper, a modified multi-objective brain storm optimization algorithms, called

MMBSO, was introduced. In the MMBSO, A clustering operator named DBSCAN has
been utilized to cluster the different individuals and a different mutation operator is
utilized to generate new individuals in the generation progress of the algorithm.
The results of the MMBSO have been evaluated according to three performance
measures. From the simulation results, it was observed that MMBSO can be a good
optimizer for solving multi-objective optimization problems. Although MMBSO
performs better than the MBSO on most test problems in this study, more problems
need to be tested to fully confirm this observation.
The results in Table 1 and 2 showed that MMBSO has a bad preference in solving
ZDT4. In order to improve the shortcoming of the MMBSO, adaptive and mixing
mutations based on niching techniques should be investigated. Another interesting area
is to exploit the MMBSO for solving multi-objective optimization problems with
constraints , many-objective problems and the other new test functions.
References
1. Fonseca, C.M., Fleming, P.J.: Genetic Algorithms for Multiobjective Optimization:
Formulation, Discussion and Generalization. In: Proceedings of the Fifth International
Conference on Genetic Algorithms. University of Illinois at Urbana-Champaign, pp.
416–423. Morgan Kaufmann Publishers (1993)
2. Srinivas, N.: Deb. K.: Multiobjective Optimization Using Nondominated Sorting in Genetic
Algorithms. Evolutionary Computation 2(3), 221–248 (1994)
3. Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A Fast and Elitist Multiobjective Genetic
Algorithm: NSGA–II. IEEE Transactions on Evolutionary Computation 6(2), 182–197
(2002)
4. Zitzler, E., Thiele, L.: Multiobjective Evolutionary Algorithms: A Comparative Case Study
and the Strength Pareto Approach. IEEE Transactions on Evolutionary Computation 3(4),
257–271 (1999)
5. Zitzler, E., Laumanns, M., Thiele, L.: SPEA2: Improving the Strength Pareto Evolutionary
Algorithm. In: Evolutionary Methods for Design, Optimization and Control with
Applications to Industrial Problems, pp. 95–100 (2001)
6. Coello, C.A.C., Pulido, G., Lechuga, M.: Handling multi-objective with particle swarm
optimization. IEEE Transactions on Evolutionary Computation 8(3), 256–279 (2004)
7. Shi, Y.: Brain Storm Optimization Algorithm. In: Tan, Y., Shi, Y., Chai, Y., Wang, G. (eds.)
ICSI 2011, Part I. LNCS, vol. 6728, pp. 303–309. Springer, Heidelberg (2011)
8. Zhan, Z., Zhang, J., Shi, Y., Liu, H.: A Modified Brain Storm Optiization. In: IEEE World
Congress on Computational Intelligence, pp. 10–15 (2012)
9. Xue, J., Wu, Y., Shi, Y., Cheng, S.: Brain Storm Optimization Algorithm for
Multi-objective Optimization Problems. In: Tan, Y., Shi, Y., Ji, Z. (eds.) ICSI 2012, Part I.
LNCS, vol. 7331, pp. 513–519. Springer, Heidelberg (2012)
10. Shi, Y., Xue, J., Wu, Y.: Multi-objective Optimization Based on Brain Storm Optimization
Algorithm. Journal of Swarm Intelligence Research (IJSIR) 4(3) (2013)
11. Coello, C.A.C., Becerra, R.L.: Evolutionary Multiobjective Optimization using a Cultural
Algorithm. In: Proceedings of IEEE Swarm Intelligence Symposium (SIS 2003), pp. 6–13
(2003)
12. Smith, R.: The 7 Levels of Change, 2nd edn. Tapeslry Press (2002)
13. Zhan, Z., Chen, W., Lin, Y., Gong, Y., Li, Y., Zhang, J.: Parameter Investigation in Brain
Storm Optimization. In: IEEE Symposium on Swarm Intelligence (SIS), pp. 103–110
(2013)
14. Cheng, S., Shi, Y., Qin, Q., Gao, S.: Solution Clustering Analysis in Brain Storm
Optimization Algorithm. In: IEEE Symposium on Swarm Intelligence (SIS), pp. 111–118
(2013)
15. Duan, H., Li, S., Shi, Y.: Predator–Prey Brain Storm Optimization for DC Brushless Motor.
IEEE Transactions on Magnetics 49(10), 5336–5340 (2013)
16. Xu, D., Wunsch II, D.: Survey of Clustering Algorithms. IEEE Transactions on Neural
Networks 16(3), 645–677 (2005)
17. Jain, A.K.: Data clustering: 50 years beyond K-means. Journal of Pattern Recognition
Letters 31, 651–666 (2010)
18. Luo, C., Chen, M., Zhang, C.: Improved NSGA-II algorithm with circular crowded sorting.
Control and Decision 25(2), 227–232 (2010)
19. Ester, M., Kriegel, H., Sander, J., Xu, X.: A Density-Based Algorithm for Discovering
Clusters in Large Spatial Databases with Noise. In: Proceedings of 2nd International
Conference on Knowledge Discovery and Data Mining, KDD 1996 (1996)
20. Adra, S.F., Dodd, T.J., Griffin, I.A., Fleming, P.J.: Convergence Acceleration Operator for
Multiobjective Optimization. IEEE Transactions on Evolutionary Computation 13(4),
825–847 (2009)
21. Daszykowski, M., Walczak, B., Massart, D.L.: Looking for Natural Patterns in Data. Part 1:
Density Based Approach, Chemmon Intell. Lab. Syst. 56, 83–92 (2001)
22. Zitzler, E., Thiele, L.: Multiobjective evolutionary algorithms: a comparative case study and
the strength Pareto approach. IEEE Transactions on Evolutionary Computation 3(4),
257–271 (1999)
Modified Brain Storm Optimization Algorithm
for Multimodal Optimization
Xiaoping Guo, Yali Wu, and Lixia Xie
Xi’an University of Technology, Xi’an Shaanxi, China

710048
Abstract. Multimodal optimization is one of the most challenging tasks for

optimization. The difference between multimodal optimization and single
objective optimization problem is that the former needs to find both multiple
global and local optima at the same time. A novel swarm intelligent method,
Self-adaptive Brain Storm Optimization (SBSO) algorithm, is proposed to solve
multimodal optimization problems in this paper. In order to obtain potential
multiple global and local optima, a max-fitness grouping cluster method is used
to divide the ideas into different sub-groups. And different sub-groups can help
to find the different optima during the search process. Moreover, the
self-adaptive parameter control is applied to adjust the exploration and
exploitation of the proposed algorithm. Several multimodal benchmark functions
are used to evaluate the effectiveness and efficiency. Compared with the other
competing algorithms reported in the literature, the new algorithm can provide
better solutions and show good performance.
Keywords: Brain Storm Algorithm ，

Max-fitness Grouping Cluster,
Self-adaptive Parameter Control, Multimodal Optimization.
1 Introduction
Multimodal optimization is a most challenging task in the area of optimization.

Unlike the single objective optimization problems, multimodal optimization needs to
provide multiple optimal solutions simultaneously. Since in the practical optimization
problems, it is very common that the best solution cannot be realized at times due to
the physical constraints. If multiple satisfaction solutions are known, the
implementation can be quickly switched to an alternative solution while the system
performance is still maintaining. As the name suggests, multimodal optimization
requires optimization algorithms to find multiple optimal solutions (both local and
global) and not just one single optimum as is done in a typical optimization study.
Evolutionary algorithms (EAs), due to their population-based approach, provide a
population of possible solutions processed at every iteration. If multiple solutions can
be preserved over all these iterations, we can have multiple good solutions at
termination of the algorithm. Although evolutionary algorithm shows the potential
power on multimodal optimization, two difficulties may be faced in the multimodal
optimization algorithm. The first is how to locate multiple global and local optima
Modified Brain Storm Optimization Algorithm for Multimodal Optimization 341
during the evolutionary process. And the second is that how to maintain the identified
optima until the end of the search.
With the development of computer science and technology, numerous techniques
have been developed for locating multiple optima (global or local). Multipopulation
based method can be incorporated into a standard EA to promote and maintain
formation of multiple stable subpopulations within a single population to locate
multiple optimal or suboptimal solutions. One of the multipopulation techniques is
commonly referred as “niching” methods. R. Thomsen [1] uses crowding metric to
force new individuals entering a population to replace similar individuals. But the
algorithm suffers from higher computational complexity. And the performance relies
on prior knowledge of some niching parameters. Fitness Euclidean distance ratio PSO
(FERPSO) [2] and speciation-based PSO (SPSO) [3] are two commonly effective
niching PSO algorithms. It is designed only for locating all global optima, while
ignoring local optima. In the literature [4], Thomsen proposed a Crowding DE (CDE)
to solve multimodal problems. It is generally difficult to select suitable trial vector
generation strategies and control parameters for CDE that can generate satisfactory
performance over all test functions. Most of existing niching methods have
difficulties to be overcome before they can be applied successfully to real-world
multimodal problems [5]. In recent years, the clustering technique, as another
multipopulation method, is used successfully to solve multi-modal optimization by
more and more scholars. Yin and Germay [8] proposed an Adaptive Clustering
algorithm (ACA) to avoid the a priori estimation of σ share . ACA adopts the identified
cluster instead of sharing the fitness function, but this method introduces two
additional variables at the same time, which need to set a reasonable maximum and
minimum value as the radius of the cluster. Hanagandi and Nikolaou [10] also use the
clustering method to find the global optima in a genetic search framework. However,
the main difficulty of all the methods lies in how to define the area for each
sub-region in the search space and how to determine the number of sub-populations.
Brain Storm Optimization (BSO) algorithm, inspired by human idea generation
process, is proposed by Shi in [11] to solve single objective optimization problem.
But two features make it perform well in multimodal problems. One is the clustering
operator that divides all the ideas generated in the current generation into some
different groups, which is possible to maintain multi optimal solutions. And the other
is the creating operator that creates new idea by learning from the self-group or other
group, which can maintain the diversity of each group. While the classical clustering
method such as k-means cannot solve the multimodal problem well. So in this paper,
a new clustering method named Max-fitness Clustering Method (MCM) is cooperated
with BSO for multimodal problem. And the self-adaptive parameter control is used
for the creating operator. The clustering method enables the algorithm to assign
individuals to different promising sub-regions. And the self-adaptive parameter
control methods are effective in maintaining the diversity of the population.
The rest of the paper is organized as follows. Section II briefly reviews the related
works about BSO. The improved algorithm is described in detail in section III. The
parameter setting and results are given in Section IV. And the conclusion and further
research are detailed finally in Section V.
342 X. Guo, Y. Wu, and L. Xie
2 Brain Strom Optimization
In any swarm intelligence algorithm, each individual cooperatively and collectively

move toward the better and better areas in the solution space. When Human being face
a difficult problem which every single person cannot solve, group person, especially
with different background, get together to brain storm, the problem can usually be
solved with high probability. Being inspired by this human idea generation process, in
2011, Shi proposed a novel algorithm named Brain Storm Optimization(BSO). In the
brainstorming process, the generation of the idea obeys the Osborn’s original four rules
(Smith, 2002). The people in the brainstorming group will need to be open-minded as
much as possible and therefore generate more diverse ideas. Any judgment or criticism
must be held back until at least the end of one round of the brainstorming process,
which means no idea will be ignored. The algorithm is given in Fig.1 and is described
as follows:
In the initialization, N potential ideas were randomly generated. BSO uses a k-means
clustering [12] as its clustering technique. In the selection operator, BSO creates N new
ideas one by one based on the current ideas. To create a new idea, BSO first determines
whether to create the new idea based on one selected cluster or based on two selected
clusters. After the cluster(s) have been selected, BSO then determines the selected ideas
whether create the new idea based on the cluster center(s) or random idea(s) of the
cluster(s).
Algorithm BSO
01 Begin
02 Randomly generate N ideas ( X i ,1 d i d N ) and evaluate their fitness;
03 While (Not stop)Do
04 Cluster the N ideas into M clusters;//According to the positions
05 Record the best idea in each cluster as the cluster center,
//Probability of replacing a randomly selected cluster center,0.2
06 If(random(0,1)< p_replace )
07 Randomly selected a cluster and replace the cluster center
with a randomly generated idea;
08 End of If
09 For( i =1 to N )
//Probability of generating new idea based on one cluster ,0.8
10 If(random(0,1)< p_one )
11 Randomly select a cluster j with a probability p j ;
//Probability of using the cluster center, p_one_center =0.4
12 If (random(0,1)< p_one_center )
13 Add random values to the selected cluster center

to generate a new idea Yi ;
14 Else
15 Add random values to a random idea of the selected cluster
to generate a new idea Yi ;
16 End of If
17 Else//Generating new idea based on two clusters
18 Randomly select two clusters j1 and j2 ;
//Probability of using the cluster center, p_ two _center =0.5
19 If (random(0,1)< p_ two _center )
20 Combine the two selected cluster centers and add with
random values to generate a new idea Yi ;
21 Else
22 Combine two random ideas from the two selected clusters
and add with random values to generate a new idea Yi ;
23 End of If
24 End of If
25 Evaluate the idea Yi and replace X i if Yi has better fitness than
Xi ;
26 End of For
27 End of While
28 End
Fig. 1. Pseudo-code of the BSO algorithm
If the selected idea X selected = ( xselected

1 2
,xselected d
,...,xselected ) is gotten, a mutation of
the X selected is applied to get the new idea expressed as X new = ( xnew
1 2
,xnew d
,...,xnew ) . After the
new idea X new has been created, BSO evaluates X new and replaces X selected if
X new has a better fitness than X selected .
In the mutation process, the Gaussian mutation will be used as random values which
are added to generate new ideas; it can be represented as follows:
x d = xd + ξ ∗ N( μ ,σ ) (1)
new selected
ξ = logsig((0.5 ∗ max_iternation - current_iteration) / K ) ∗ rand() (2)
In the equation (1) and (2), x d is the d-dimensional of the idea selected to
selected
d
generate new idea; xnew is the d-dimensional of the idea newly generated; N( μ ,σ ) is
the Gaussian random function with mean μ and σ ; ξ is a coefficient that weights the
contribution of the Gaussian mutation; log sig() is a logarithmic sigmoid transfer
function; max_iternation and current_iteration are the maximum iteration number
and the current iteration number, K is for changing log sig() function’s slope, and
rand() is a random value within (0,1).
3 Self-adapted Brain Strom Optimization
In the multimodal optimization problem, the challenging issue is how to find all the
global optimum, local optimum while maintaining the diversity of the population. The
experimental results reported in [6]-[10] show that clustering operation is an ideal
technique. Although many common evolutionary algorithms (EAs) are used to solve
multimodal problems, most of them introduce a variety of clustering strategies into the
evolutionary process. Different from other EAs, BSO adopts the clustering operation as
their converging process, which makes it very suitable to solve multimodal problem.
As we all know, different clustering strategies show different advantages. The K-means
method in traditional BSO is a typical clustering based on distance. The biggest
drawback is that the choice of initial cluster center has a great influence on its clustering
result. Once the initial cluster center is not chosen well, it may have a bad influence on
its clustering result. On the other hand, the preservation mechanism of traditional BSO
is just for comparison of the fitness values of simple problems. For multimodal
problem, this mechanism is not able to maintain all individuals which have found
extreme point. In this paper we improved the classical BSO to solve multimodal
optimization in two aspects. Firstly, a new clustering operation named Max-fitness
Clustering method (MCM) replaces the classical k-means clustering method to assign
individuals in different promising subregions. Secondly, a self-adaptive
parameter control technique is used to ensure retention of individual diversity and
convergence. Detailed explanation will be showed as the followings subsections.
3.1 Max-fitness Clustering Method
The K-means clustering method in traditional BSO overly depends on the selection of
the initial cluster centers, which cannot solve multimodal optimization. The maximum
clustering method is proposed to solve the different clustering center selection in this
paper. The operation produces are listed as follows. Firstly, the largest individual
fitness value is selected as the first category center from an individual original
population. Secondly, the nearest to the center of each individual is emptied. Finally,
the process is repeated for the remaining individuals until all of them are classified. The
difference between the clustering and other clustering methods is that each category
center is the best individual of all the remaining individuals, so each clustering center
has a larger probability that is distributed in the extreme point, which makes the
individual learning more direction. This clustering algorithm makes full use of the
information about each individual, and the solution space is effectively combined with
the target space. A schematic illustration of the clustering partition is shown in Fig 2.
The replacing operator of BSO is given as line 4&5 in Fig.1.
Algorithm Max-fitness clustering method

Step 1 Initialize a population P of Np ideas {X i | i = 1,2,...,Np}
in the search region randomly
Step2 Find the best(fitness value) idea as seed X
Step3 Combine M − 1 ideas of the population P , which are
nearest to X , with X to form a subpopulation. Any
tie will be broken randomly.
Step4 Eliminate these M ideas from P
Step5 Execute Step 1- Step 3 repeatedly until the population
P is divided
into P / M subpopulation
Fig. 2. Pseudo-code of the max-fitness clustering
3.2 Self-adaptive Parameter Control

According to Eq.(1), a new idea is created by adding Gaussian random noise to the
selected idea. By this way, all ideas in the group will quickly converge to the direction
along the previous best idea. In this paper, a self-adaptive parameter control is used to
make the algorithm quickly converge to different optima by different clusters. The
crossover probability Cr is updated dynamically according to the evolution process.
This is similar to the adaptation strategy used in JADE [13].
In the mutation operation, the new idea ui , j is generated by making use of a
binomial crossover operation on the last iteration idea xi , j and the after mutation
idea vi , j .
 vi , j if (rand j (0,1) ≤ C r ) or ( j = jrand )
ui , j =  (3)
 xi , j otherwise
Where i = 1,2,..., Np , j = 1,2,...,n _ d , jrand is a randomly chosen integer from
{1,2,,...,n _ d } , rand j ( 0,1) is a random value within (0,1). Due to the use of jrand , ui , j
is guaranteed to differ from xi , j . At each generation, the crossover rate Cri of each
idea is independently generated according to a normal distribution of mean Crm and
standard deviation 0.1
Cri = randn ( Crm ,0.1) (4)
C rm is updated as follow:
Crm = mean ( Scr ) (5)

Where Scr is the set of all successful crossover rates at previous generation. The
initial value of Crm is 0.5 and Scr = φ .
The selection operation is conducted by comparing ui and the closest individual

xs in population. The better one will enter the next generation. For the maximization
problem:
u if f (ui ) > f( xs )
xs =  i (6)
 xs otherwise
A schematic illustration of the self-adaptive parameter control is shown in
algorithm 2.
Algorithm2 self-adaptive parameter control

For i = 1 : Np
Step 1 Use the Eq.3 to produce new idea ;
Step 2 Evaluate offspring using fitness function;
Step 3 According to Eq.(6) to update idea and save the
respective
crossover probability in Scr
End for
Step 4 Update Crm = mean ( Scr ) .
Fig. 3. Pseudo-code of the self-adaptive parameter control
4 Experimental Studies
4.1 Parameter Setting

A set of multimodal optimization benchmark functions [14] is used to evaluate the
ability of the proposed algorithm in this paper. A level of accuracy
(typically 0 < ε < 1 ) indicates the error between the fitness value and the given
extreme point. The extreme point will be considered been found if the error is less
than ε . In order to compare with other algorithms fairly, the parameters will follow the
guideline given in [15] and [16] and apply these parameters to all compared algorithms.
The level of accuracy ( ε ), niching radius ( γ ), population size and maximal number of
function evaluations (FES) allowed are listed in Table II. From the table 2, we can see
that a larger number of optima require a large population size and more function
evaluations.
Table 1. Parameter setting of different benchmark functions
Test function Population size M H J No. of function evaluations

F1 50 5 0.05 0.5 10,000
F2 50 5 0.05 0.5 10,000
F3 50 5 0.05 0.5 10,000
F4 50 5 0.000001 0.01 10,000
F5 50 5 0.000001 0.01 10,000
F6 50 5 0.000001 0.01 10,000
F7 50 5 0.000001 0.01 10,000
F8 50 5 0.0005 0.5 10,000
F9 50 5 0.000001 0.5 10,000
F10 250 25 0.00001 0.5 10,000
F11 250 10 0.05 0.5 100,000
F12 100 10 0.0001 0.2 20,000
F13 500 50 0.001 0.2 200,000
F14 1000 100 0.001 0.2 400,000
4.2 Experimental Result

To give a clearer view, SBSO is compared with the original BSO. The distributions of
population of the two algorithms at different iterations are plotted in Fig. 4. the （
function of F6 and Fig. 5. the function of F7 . ）（）
BSO
Generation=50 Generation=100 Generation=150 Generation=200
1 1 1 1
0.9 0.9 0.9 0.9
0.8 0.8 0.8 0.8
0.7 0.7 0.7 0.7
0.6 0.6 0.6 0.6
0.5 0.5 0.5 0.5
0.4 0.4 0.4 0.4
0.3 0.3 0.3 0.3
0.2 population 0.2 population 0.2 population 0.2

population
optima optima optima optima
0.1 0.1 0.1 0.1
0 0 0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
SBSO
1 1 1 1
0.9 0.9 0.9 0.9
0.8 0.8 0.8 0.8
0.7 0.7 0.7 0.7
0.6 0.6 0.6 0.6
0.5 0.5 0.5 0.5
0.4 0.4 0.4 0.4
0.3 0.3 0.3 0.3
0.2
population 0.2 population 0.2 population 0.2 population
0.1 0.1 0.1 0.1
0 0 0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 4. Distributions of population of BSO and SBSO on different stages （F6）

BSO
1 1 1 1
0.9 0.9 0.9 0.9
0.8 0.8 0.8 0.8
0.7 0.7 0.7 0.7
0.6 0.6 0.6 0.6
0.5 0.5 0.5 0.5
0.4 0.4 0.4 0.4
0.3 0.3 0.3 0.3

population population population population
0.2 0.2 0.2 0.2
0.1 0.1 0.1 0.1
0 0 0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
SBSO
1 1 1 1
0.9 0.9 0.9 0.9
0.8 0.8 0.8 0.8
0.7 0.7 0.7 0.7
0.6 0.6 0.6 0.6
0.5 0.5 0.5 0.5
0.4 0.4 0.4 0.4
0.3 0.3 0.3 0.3
0.2 population 0.2 0.2 population 0.2

population
population optima
optima optima
0.1 0.1 optima 0.1 0.1
0 0 0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 5. Distributions of population of BSO and SBSO on different stages （F7）

From the above figure, we can clearly see the original BSO have accurately found all
the extreme points of function F6 which has the same extreme points, but for the
function like F7 which has different extreme points, only the SBSO can find and save
all the extreme points. With the number of iterations increasing, the SBSO can
converge to the extreme points. Therefore, it presents that the effectiveness of SBSO in
dealing with multimodal optimization problem. In order to further demonstrate the
superior performance of this algorithm, we compare the performance indicators of the
proposed algorithm with the other algorithms in one of the latest multimodal works
[17]. As the same as the literature [17], we adapts the following indicator as our
assessment criteria.
Peak Accuracy: For each desired peaki ,( i = 1, 2,...,# peaks ) to be located, the closet
idea X in the population is taken and absolute difference in objective values is
calculated. If the objective value of idea X is denoted by f ( X ) ,the peak accuracy is
calculated using
# peaks f ( peaki ) − f ( X )
peak accuracy = 
i =1 # peaks
On all the benchmark functions, all algorithms are run until all known peaks are
found or maximum number of FEs is exhausted. The proposed algorithm is compared
with following standard multimodal algorithms which are generally cited in literature
[17]:
1) CDE: Crowding DE;

2) SDE: Speciation-based DE
3) FER-PSO: Fitness –Euclidean distance Ratio PSO
4) SPSO: Speciation-based PSO
5) r3pso: a local best PSO with a ring topology•each

each member interacts with its
immediate member on its left and right.
6) r2psolhc: no overlapping neighborhoods, hence acting as multiple local hill
climbers, more suitable for finding global as well as local optima.
7) r3psolhc: Basically multiple PSOs search in parallel, like local hill climbers.
This variant is more appropriate if the goal of optimization is to find global
ad well as local optima.
The simulation results are presented in Table 3.
Table 2. Peak Accuracy comparisons of different algorithms

Fun SBSO CED[18] SDE[19] FER-PSO[1 SPSO[14] r3pso[14] r2psolhc[14] r3psolhc[14]
4]
F1 4.33e-05(3) 9.46e-08 1.23e-08(1) 5.24e-02(5) 8.74e-02(7) 3.54e-02(4) 9.78e-02(8) 8.68e-02(6)
(2)
F2 1.78e-05(3) 8.76e-06 3.43e-07(1) 9.65e-04(4) 9.45e-02(8) 1.43e-02(5) 5.23e-02(7) 4.54e-02(6)
(2)
F3 5.49e-04(3) 9.76e-05 5.73e-05(1) 41999,.34e- 9.54e-01(8) 7.93e-02(6) 5.52e-01(7) 7.64e-02(5)
(2) 02(4)
F4 2.63e-10(1) 7.43e-05 9.53e-07(7) 5.65e-07(6) 3.12e-07(5) 2.24e-07(4) 5.43e-09˄2˅ 9.43e-08(3)
(8)
F5 8.24e-09(3) 9.43e-06 4.03e-09(2) 8.34e-09(4) 2.16e-09(1) 9.61e-07(7) 8.58e-07(6) 5.36e-07(5)
(8)
F6 7.79e-10(1) 5.39e-05 8.27e-07(7) 5.45e-09(2) 9.58e-08(4) 8.79e-08(3) 8.05e-07(6) 7.54e-08(5)
(8)
F7 1.23e-06(7) 8.97e-05 4.51e-07(3) 7.41e-07(4) 2.98e-07(1) 9.29e-07(5) 3.43e-07(2) 9.88e-07(6)
(8)
F8 4.53e-06(1) 4.27e-02 8.57e-04(2) 8.69e-04(3) 5.21e-02(8) 5.63e-03(4) 6.12e-03(5) 9.19e-03(6)
(7)
F9 2.26e-08(1) 3.42e-04 5.33e-08(2) 7.38e-08(3) 3.58e-04(8) 6.92e-05(5) 4.33e-05(4) 8.28e-05(6)
(7)
F10 5.91e-08(1) 3.96e-03 9.92e-02(8) 5.50e-06(2) 9.77e-04(6) 8.31e-05(3) 9.43e-05(4) 6.87e-04(5)
(7)
F12 1.10e-005( 5.24e-04 8.33e-04(8) 4.53e-04(5) 1.23e-04(2) 3.87e-04(4) 2.54e-04(3) 4.56e-04(6)
1) (7)
F13 7.93e-004( 9.87e-04 9.85e-03(8) 9.69e-03 8.45e-03 8.25e-03(4) 8.52e-03 6.32e-03
1) (2) (7) (5) (6) (3)
F14 1.4e-003(1) 9.23e-02 7.89e-01(5) 5.95e-01 7.86e-01 8.45e-01 9.68e-01 8.12e-01(6)
(2) (3) (4) (7) (8)
Total 27 70 52 56 67 61 63 68
ranks
Please note that function F11 is a 2-D inverted Shubert function, which is not
reported in[19]. Thus, the simulation results of function F11 is not included in Table III.
Above the table, the numbers in brackets present the rank of Peak Accuracy, which is
obtained by the different algorithms dealing with the same function. The smaller the
rank, the higher the accuracy. The last line of the table is the total rank (i.e. summation
of all the individual ranks). The lower the total rank, the better the performance of the
algorithm. From the result, in terms of peak accuracy, we can see that the proposed
algorithms show better performance. Besides, it also indicates that the proposed
algorithm present a good exploitative behavior in convergence to different global and
local optima.
5 Conclusion
In this paper, there are two parts being proposed to modify the BSO algorithm in
multimodal optimization. The clustering strategy drives populations to search the
different sub-regions to obtain the potential multiple global and local optima. At the
same time, it also reduces the calculation of complexity in the proposed algorithm.
Self-adaptive parameter control maintains population diversity by allowing
competition to limit resources among similar idea in subpopulation. Our future works
will focus on testing the performance of the algorithm on much more massive
multimodal problems with high dimensionality and constraints. In some degree, the
subpopulation size M affects the performance of the algorithm, so how to design an
adaptive strategy to control M through the process is the future work.
References
1. Thomsen, R.: Multimodal optimization using crowding-based differential evolution. In:
Proc. IEEE Congr. Evol. Comput., pp. 1382–1389 (June 2004)
2. Wang, H.F., Moon, I., Yang, S.X., Wang, D.W.: A memetic particle swarm optimization
algorithm for multimodal optimization problems. IEEE Trans. Cyber. 43(2), 634–647
(2013)
3. Parrott, D., Li, X.: Locating and tracking multiple dynamic optima by a particle swarm
model using speciation. IEEE Trans. Evol. Comput. 10(4), 440–458 (2006)
4. Wang, Y.J., Zhang, J.S., Zhang, G.Y.: A dynamic clustering based differential evolution
algorithm for global optimization. Journal of Operational Research 183(1), 56–73 (2007)
5. Li, X.: Niching without niching parameters: Particle swarm optimization using a ring
topology. IEEE Transaction on Evolutionary Computation 14(1), 150–169 (2010)
6. Rigling, B., Moore, F.: Exploitation of subpopulations in evolutionary strategies for
improved numerical optimization. In: Proc. 11th Midwest Artif. Intell. Cogn. Sci. Conf., pp.
80–88 (1999)
7. Rumpler, J., Moore, F.: Automatic selection of subpopulations and minimal spanning
distances for improved numerical optimization. In: Proc. Congr. Evol. Comput., pp. 38–43
(2001)
8. Zaharie, D.: A multi-population differential evolution algorithm for multimodal
optimization. In: Proc. 10th Mendel Int. Conf. Soft Comput., pp. 17–22 (June 2004)
9. Hendershot, Z.: A differential evolution algorithm for automatically discovering multiple
global optima in multidimensional discontinuous spaces. In: Proc. 15th Midwest Artif.
Intell. Cogn. Sci. Conf., pp. 92–97 (April 2004)
10. Zaharie, D.: Extensions of differential evolution algorithms for multimodal optimization. In:
Proc. SYNASC, pp. 523–534 (2004)
11. Shi, Y.: Brain storm optimization algorithm. In: Tan, Y., Shi, Y., Chai, Y., Wang, G. (eds.)
ICSI 2011, Part I. LNCS, vol. 6728, pp. 303–309. Springer, Heidelberg (2011a)
12. Zhan, Z.H., Shi, Y., Zhang, J.: A modified brain storm optimization. In: IEEE World Congr.
Comput. Intell. (Jule10-15, 2012)
13. Zhang, J., Sanderson, A.C.: JADE: Adaptive differential evolution with optional external
archive. IEEE Trans. Evol. Comput. 13(5), 945–958 (2009)
14. Li, X.: Niching without niching parameters: Particle swarm optimization using a ring
topology. IEEE Trans. Evol. Comput. 14(1), 150–169 (2010)
15. Qu, B.Y., Suganthan, P.N., Liang, J.J.: Differential evolution with neighborhood mutation
for multimodal optimization. IEEE Trans. Evol. Comput. 16(5), 601–614 (2012)
16. Qu, B.Y., Suganthan, P.N., Das, S.: A distance-based locally informed particle swarm
model for multi-modal optimization. IEEE Trans. Evol. Comput. 17(3), 387–402 (2013)
17. Roya, S., Islama, S.M., Dasb, S., Ghosha, S.: Multimodal optimization by artificial weed
colonies enhanced with localized group search optimizers. Appl. Soft Comput. 13(1), 27–46
(2013)
18. Thomsen, R.: Multimodal optimization using crowing-based differential evolution. In: Proc.
IEEE Congr. Evol. Comput., pp. 1382–1389 (June 2004)
19. Li, X.: Efficient differential evolution using speciation for multimodal function
optimization. In: Proc. Conf. Genetic Evol. Comput., pp. 873–880 (2005)
Classification of Electroencephalogram Signals Using
Wavelet Transform and Particle Swarm Optimization
Nasser Omer Ba-Karait1,2, Siti Mariyam Shamsuddin1,2, and Rubita Sudirman3

1
UTM Big Data Centre, Universiti Teknologi Malaysia
2
Faculty of Computing
Universiti Teknologi Malaysia, 81310 Skudai, Johor Bahru, Malaysia
3
Faculty of Electrical Engineering
Universiti Teknologi Malaysia, 81310 Skudai, Johor Bahru, Malaysia
bakarait@yahoo.com, mariyam@utm.my, rubita@fke.utm.my
Abstract. The electroencephalogram (EEG) is a signal measuring activities of

the brain. Therefore, it contains useful information for diagnosis of epilepsy.
However, it is a very time consuming and costly task to handle these subtle
details by a human observer. In this paper, particle swarm optimization (PSO)
was proposed to automate the process of seizure detection in EEG signals.
Initially, the EEG signals have been analysed using discrete wavelet transform
(DWT) for features extraction. Then, the PSO algorithm has been trained to
recognize the epileptic signals in EEG data. The results demonstrate the
effectiveness of the proposed method in terms of classification accuracy and
stability. A comparison with other methods in the literature confirms the
superiority of the PSO.
Keywords: Particle swarm optimization, machine learning, discrete wavelet

transform, EEG, epileptic seizure.
1 Introduction
The electroencephalogram (EEG) is a highly complex signal recording the brain's

neural activities through changes in electrical potentials at multiple locations over the
scalp. It conveys valuable clinical information about the state of the brain. Therefore
in recent decades, the EEG signals (EEGs) have been used intensively to study brain
function and neurological disorders. For this reason, the EEG has long been an
important clinical tool in diagnosing, monitoring and managing of neurological
disorders, especially those related to epilepsy [1, 2]. Epileptic seizures are caused by
temporary electrical disturbance of the brain. The occurrence of a seizure seems
unpredictable and its course of action is still very poorly understood. Research is
therefore needed to gain a better understanding of the mechanisms causing epileptic
disorders. Careful analysis of EEGs could provide valuable insight into this
widespread brain disorder [3, 4].
A continuous EEG recording is required to study the epileptic seizures for pre-
surgical evaluation. It provides essential information for locating the brain regions
Classification of Electroencephalogram Signals Using Wavelet Transform 353
that generate epileptic activity [5, 6]. Clearly, the analysis of the recorded EEG based
on visual inspection is a very time consuming and costly task. In some cases, the
seizures are uncontrollable. Recently, methods have started being developed to treat
medically resistant epilepsy through delivering a local therapy to the affected regions
of brain. Automatic seizures detection forms an integral part of such methods [6, 7].
It is therefore worthwhile to propose an effective algorithm for EEG changes
recognition. In literature, most methods deal with this problem based on well-known
classification techniques [8-12]. However, classification can be seen in
multidimensional space as optimization problem where a class is identified by a
centroid. In this case, particle swarm optimization (PSO) can be employed effectively
to optimize coordinates of the centroids [13, 14]. To the best of our knowledge, the
PSO has not been used for EEGs classification. However in many works, the
classifier of EEGs is trained and/or its parameters are optimized by PSO [15, 16].
Also, it was employed to estimate the locations of sources of electrical activity, e.g.
epileptic, in the brain based on the scalp EEGs [17-19]. Other EEGs issues have been
addressed by PSO such as feature selection [20, 21] and optimal selection of
Electrode Channels [22, 23]. In this research, the PSO algorithm is studied to evaluate
its performance in detecting the epileptic seizures in EEGs using discrete wavelet
transform (DWT) for feature extraction.
2 Particle Swarm Optimization
Particle swarm optimization (PSO) algorithm was originally designed by Kennedy

and Eberhart [24] in 1995. The idea was inspired by the social behavior of flocking
organisms, as a bird flock. PSO uses a population (called swarm) of candidate
solutions (called particles) to probe promising regions of the search space. Each
particle moves in the search space with a velocity that is dynamically adjusted
according to its own flying experience and its companions’ flying experience and it
retains the best position it ever encountered in memory. The best position ever
encountered by all particles of the swarm is also communicated to all particles.
The popular form of PSO algorithm is defined as:
vi , d (t + 1) ← w * vi , d (t ) + c1 r1 ( pi , d (t ) − xi , d (t )) + c2 r2 ( p g , d (t ) − xi , d (t )) (1)
xi , d (t + 1) ← xi , d (t ) + vi , d (t + 1) (2)
where vi,d is the velocity of particle i along dimension d, xi,d is the position of particle i
in d, c1 is a weight applied to the cognitive learning portion, and c2 is a similar weight
applied to the influence of the social learning portion. r1 and r2 are separately
generated random numbers in the range between zero and one. pi,d is the previous best
location of particle i; it is called the personal best (pbest). Pg,d is the best location
found by the entire population; it is called the global best (gbest). w is the inertia
weight [25, 26].
354 N.O. Ba-Karait, S.M. Shamsuddin, and R. Sudirman
Velocity values must be within a range defined by two parameters -vmax and vmax. A
PSO with an inertia weight in the range (0.9, 0.4) has the better performance on
average. To get a better searching pattern between global exploration and local
exploitation, researchers recommended decreasing w linearly over time from a
maximal value wmax to a minimal value wmin [26-28].
wmax − wmin
w = wmax − ∗t (3)
tmax
where, tmax is the maximum number of iterations allowed and t is the current iteration
number.
3 Materials and Methods
3.1 EEG Data
The current study used the publicly available EEG data described by Andrzejak et al.
[29]. The complete dataset contains five different sets (denoted A-E), each containing
100 single channel EEG segments of 23.6 second duration. Segments in sets A and B
were recorded from five healthy volunteers. They were relaxed in an awake state with
eyes open (set A) and closed (set B). The EEG archive of presurgical diagnoses was
used to originate sets C, D and E by selecting EEGs from five patients. Signals in sets
C and D were measured in seizure free intervals from within the epileptogenic zone
and opposite the epileptogenic zone of the brain, respectively. Set E were obtained
from within the epileptogenic zone during seizure activity. Fig. 1 shows typical EEG
segments, one from each category.
Fig. 1. Samples of five different sets of EEG data

3.2 Discrete Wavelet Transform: Feature Extraction
Discrete wavelet transform (DWT) has been particularly successful in the area of
epileptic seizure detection due to its ability to capture transient features and localize
them in both time and frequency domains accurately [30]. The DWT analyses a signal
s(n) at different frequency bands by decomposing the signal into an approximation
and detail information using two sets of functions known as scaling functions and
wavelet functions, which are associated with low-pass g(n) and high-pass h(n) filters,
respectively. The DWT decomposition process is described in Fig. 2.
Fig. 2. Sub-band decomposition of DWT
When DWT is used to analyse the signals, two important aspects should be
considered: the number of decomposition levels and the type of wavelet. The
decomposition level number is selected based on the dominant frequency components
of the signal. According to Subasi [31], the levels are selected such that those parts of
the signal that correlate well with the frequencies required for the signal classification
are retained in the wavelet coefficients. Therefore, level 4 wavelet decomposition was
selected for the present study. Accordingly, the EEGs have been decomposed into the
details D1-D4 and one final approximation, A4. The smoothing feature of the
Daubechies wavelet of order 2 (db2) made it more suitable to detect changes in EEGs
[32]. In this research, db2 has been used to compute the wavelet coefficients of the
EEGs.
The computed coefficients of discrete wavelet provide a compact representation
that shows the energy distribution of the signal in time and frequency. In order to
decrease dimensionality of the extracted feature vectors further, statistics over the set
of the wavelet coefficients are used [32]. The following statistical features were used
to represent the time-frequency distribution of the EEGs: Maximum, Minimum,
Mean, and Standard deviation of the wavelet coefficients in each sub-band.
3.3 PSO for EEG Classification

The PSO algorithm has been introduced in this study to classify EEGs for diagnosis
purposes. In PSO for classification, each class is identified by a centroid. Therefore,
for a dataset Z = (z1, z2,…, zp, …, zNp) with Nc classes, where zp is a pattern with Nd
features, and NP is the number of patterns in Z. The PSO is applied to optimize
positions of Nc centroids in an Nd-dimensional space. Fig. 3 highlights model of
encoding PSO for EEGs classification. It includes three main steps: particle
encoding, defining the fitness function and optimization process.
Start EEG signals

(training set)
Encode the particle Fitness Evaluation

into centroids
Classify data using
centroids of particle
Initialize swarm of
particles
Calculate the
objective function
No
Maximum
number of Optimization process
iterations
reached?
Yes
Output centroids of
gbest as classifier End
Fig. 3. PSO-based classification model for EEG signals
Each particle is represented by Nc centroids. Consequently, the position of a

particle i is encoded as (xi(1), …, xi(j)…, xi(Nc)) where xi(j) refers to the jth centroid of
the ith particle. The position of the jth centroid is constituted by Nd real numbers
representing its Nd coordinates in the problem space:
x ( j ) = { x ( j ), x ( j ), ..., x ( j )} (4)
i i ,1 i ,2 i,Nd
As above, the velocity of each particle i is encoded as (vi(1), …, vi(j)…, vi(Nc))

where the velocity of the jth centroid, vi(j) is made up of Nd real numbers representing
its Nd velocity components in the problem space:
v ( j ) = {v ( j ), v ( j ), ..., v ( j )}
i i ,1 i ,2 i, N d (5)
In classification problem, the objective is to assign any pattern to its correct class.
Therefore, the performance of a classification algorithm is evaluated by its accuracy,
defined as the percentage of patterns correctly assigned to their classes. This study
uses accuracy measure as a fitness function to evaluate the quality of solutions. The
fitness of the ith particle is computed based on the dataset portion ZD (training set) as
in Eq. 6.
ZD
 A ssess ( z
p =1
p ,i )
A ccuracy ( i , z D ) = (6)
ZD
1, iiiiiif iClassify (z p , i ) = z p .c

Assess (z p , i ) =  (7)
0, iiiiotherwiseiiiiiiiiiiiiiiiiiiiii
where zp is a pattern in ZD, zp.c is the class of zp and Classify(zp,i) returns the class
assigned to zp by the particle i according to the nearest centroid based on Euclidean
distance.
With the above premises, optimization mechanism of PSO algorithm is used to
update coordinates of the centroids toward the best solution as summarized in Alg.1.
Alg.1. PSO for classification

1. Initialize each particle i to contain Nc centroids
2. For t=1 to tmax
a. For each particle i
i. Calculate fitness value using Eq. 6.
ii. Update the personal best solution, pbest
b. Update the global best solution, gbest
c. For each particle i
i. Update the centroids using Eq. 1 and Eq. 2.
d. Update the inertia weight, w using Eq. 3.
4.1 Performance Measures

In medical diagnosis tasks, the common performance measures are sensitivity,
specificity and classification accuracy. Sensitivity is defined as the percentage of
correctly detected epileptic EEG patterns to the total number of patterns in epileptic
EEG. On the other hand, specificity is defined as the percentage of correctly detected
normal EEG patterns to the total number of patterns in normal EEG. Finally, the
percentage of all correctly classified patterns to the total number of patterns in both
normal and seizure EEG dataset represents the accuracy. Formally, the performance
of a diagnostic system is measured as
TP
Sensitivity = (8)
TP + FN
TN
Specificity = (9)
TN + FP
where TP, TN, FP and FN denote true positives, true negatives, false positives and
false negatives respectively.
Accuracy: Eq.6 is calculated for testing set ZT using the final gbest;
Accuracy (gbest, ZT).
4.2 Results and Discussion
The EEG dataset used consists of three categories of signals: healthy (sets A and B),
seizure-free (sets C and D) and seizure (set E). Therefore, the three sets: A, D, and E
of the above-described dataset are used to analyse the performance of PSO. Sets A
and D are gathered to form the normal class against set E which represents the
epileptic class. This is similar to real medical applications in which the EEG segments
are classified into non-seizures and seizures.
In each set of EEG data, there are 100 EEGs of 4096 samples. In this research,
each signal is further divided by a rectangular window composed of 256 samples.
Therefore, the dataset of the considered EEG problem was formed of 4800 patterns;
i.e., each set has 1600 vectors. Consequentially, the epileptic class contains 1600
patterns, while the number of patterns in the normal class is 3200. The DWT
coefficients at the fourth level (D1-D4 and A4) were computed for each pattern. The
statistical features that were calculated over the set of wavelet coefficients reduce the
dimensionality of feature vector to 20.
It is common to partition the dataset into two separate sets: a training set and a
testing set. Additionally, k-fold cross validation is often used by the researchers to
evaluate the behavior of the algorithm in the bias associated with the random
sampling of the training data. In this study, the EEG dataset (sets A, D and E) was
randomly divided into training-testing as 50-50%, 70-30%, and with a 10-fold cross
validation. The values of the PSO parameters are as follows: vmax=0.05, c1=2.0,
c2=2.0, wmax=0.9, wmin=0.4. 50 particles were trained for 1000 iterations to evolve two
centroids for the normal and epileptic classes. The centroids produced are then used to
classify the patterns in the testing set in order to assess the effectiveness of the
proposed method.
Table 1 presents the results achieved by the PSO algorithm with respect to the
sensitivity, specificity and accuracy. The results are reported in terms of average, and
standard deviation (SD) of ten runs for each partition of the dataset. As can be seen
from Table 1, the PSO on average classified the EEGs of training-test datasets
partitions: 50-50%, 70-30%, and 10-fold cross validation with accuracies of 96.91%,
97.08%, and 96.53% respectively. The results using all training-test datasets partitions
are depicted in Fig. 4. These overall results illustrate that the PSO has good
performance and stable behaviour for EEGs classification with accuracy of 96.84%,
and standard deviation of 0.90.
Table 1. Sensitivity, specificity, and accuracy of the PSO algorithm on EEG signals
Training-testing dataset Performance measures (%)

partitions (%) Sensitivity Specificity Accuracy
50-50 Average 96.50 97.12 96.91
SD 1.04 0.57 0.30
70-30 Average 96.49 97.38 97.08
SD 0.83 0.63 0.37
10-fold cross validation Average 95.75 96.92 96.53
SD 3.77 2.35 1.45
100.00 2.50
97.14 96.84 2.30
97.00 96.25 2.25
94.00 2.00
Perf ormance Measure (%)
91.00 1.75
Standard Deviation
88.00 1.50 1.43
85.00 1.25
82.00 1.00 0.90
79.00 0.75
76.00 0.50
73.00 0.25
70.00 0.00
Sensitivity Specif icity Accuracy Sensitivity Specif icity Accuracy
(a) (b)
Fig. 4. Overall results of PSO on EEG signals: average (a), and standard deviation (b)
Table 2 illustrates a comparative study of the proposed algorithm with other studies
in the literature. For a feasible comparison, only the studies that used the same EEG
dataset with the three mentioned sets (A, D and E), and DWT for features extraction
are considered. It can be concluded from this comparison that the PSO algorithm
showed a promising performance compared to other methods, with a difference in
accuracy varies from 0.17% to 6.46%. This proves its ability to compete with well-
known classification techniques. In fact, the results reveal that combination of PSO
and DWT can produce an efficient automated system for diagnosing epileptic seizures
in EEGs.
Table 2. Comparison of the PSO accuracy on EEG signals with methods in the literature
Study Method Accuracy

Übeyli [8] Mixture of expert 93.17
Übeyli [9] Combined neural network 94.83
Hsu and Yu [10] Genetic algorithm, support vector machine 90.38
Guo et al. [11] Genetic programming, k-Nearest Neighbor 93.50
Orhan et al. [12] K-means clustering, artificial neural network 96.67
This study Particle swarm optimization 96.84
5 Conclusion
In the present study, discrete wavelet transform and particle swarm optimization have
been hybridized to process EEGs for automatic diagnosis of epileptic seizures. The
DWT is used to extract the features of signals for the PSO which separates epileptic
signals from others in the EEG data. The ability of the proposed method was tested on
EEG recordings with healthy, seizure-free, and seizure data. The results indicate that
the PSO has very good performance in discriminating the EEGs compared to
algorithms reported in the literature. Therefore, the proposed system could be a
powerful tool to assist experts in facilitating the analysis of a patient's information and
reducing the time and effort required to make accurate decisions on their patients.
Acknowledgment. This work is supported by Research University Grant (Vot

03H72) and Long Term Research Grant (LRGS/TD/2011/UTM/ICT/03/03). The
authors would like to thanks Research Management Centre (RMC), Universiti
Teknologi Malaysia (UTM) for the support in R & D, Soft Computing Research
Group (SCRG) for the inspiration in making this study a success. The authors would
also like to thank the anonymous reviewers who have contributed enormously to this
work.
References
1. Adeli, H., Zhou, Z., Dadmehr, N.: Analysis of EEG records in an epileptic patient using
wavelet transform. Journal of Neuroscience Methods 123, 69–87 (2003)
2. Übeyli, E.D.: Least squares support vector machine employing model-based methods
coefficients for analysis of EEG signals. Expert Systems with Applications 37, 233–239
(2010)
3. Subasi, A.: EEG signal classification using wavelet feature extraction and a mixture of
expert model. Expert Systems with Applications 32, 1084–1093 (2007)
4. Nigam, V.P., Graupe, D.: A neural-network-based detection of epilepsy. Neurological
Research 26, 55–60 (2004)
5. Ocak, H.: Optimal classification of epileptic seizures in EEG using wavelet analysis and
genetic algorithm. Signal Processing 88, 1858–1867 (2008)
6. Patnaik, L.M., Manyam, O.K.: Epileptic EEG detection using neural networks and post-
classification. Computer Methods and Programs in Biomedicine 91, 100–109 (2008)
7. Gardner, A.B.: A novelty detection approach to seizure analysis from intracranial EEG.
PhD Thesis, Georgia Institute of Technology. Georgia, United States (2004)
8. Übeyli, E.D.: Wavelet/mixture of experts network structure for EEG signals classification.
Expert Systems with Applications 34, 1954–1962 (2008)
9. Übeyli, E.D.: Combined neural network model employing wavelet coefficients for EEG
signals classification. Digital Signal Processing 19, 297–308 (2009)
10. Hsu, K.-C., Yu, S.-N.: Detection of seizures in EEG using subband nonlinear parameters
and genetic algorithm. Computers in Biology and Medicine 40, 823–830 (2010)
11. Guo, L., Rivero, D., Dorado, J., Munteanu, C.R., Pazos, A.: Automatic feature extraction
using genetic programming: An application to epileptic EEG classification. Expert Systems
with Applications 38, 10425–10436 (2011)
12. Orhan, U., Hekim, M., Ozer, M.: EEG signals classification using the K-means clustering
and a multilayer perceptron neural network model. Expert Systems with Applications 38,
13475–13481 (2011)
13. Sousa, T., Silva, A., Neves, A.: Particle Swarm based Data Mining Algorithms for
classification tasks. Parallel Computing 30, 767–783 (2004)
14. De Falco, I., Cioppa, A.D., Tarantino, E.: Facing classification problems with Particle
Swarm Optimization. Applied Soft Computing 7, 652–658 (2007)
15. Hema, C.R., Paulraj, M.P., Nagarajan, R., Yaacob, S., Adom, A.H.: Application of particle
swarm optimization for EEG signal classification. Biomedical Soft Computing and Human
Sciences 13, 79–84 (2008)
16. Chai, R., Ling, S., Hunter, G., Tran, Y., Nguyen, H.: Brain Computer Interface Classifier
for Wheelchair Commands using Neural Network with Fuzzy Particle Swarm
Optimization. IEEE Journal of Biomedical and Health Informatics (in Press)
17. Qiu, L., Li, Y., Yao, D.: A feasibility study of EEG dipole source localization using
particle swarm optimization. In: 2005 IEEE Congress on Evolutionary Computation,
Edinburgh, Scotland, UK, vol. 1, pp. 720–726 (2005)
18. Xu, P., Tian, Y., Lei, X., Yao, D.: Neuroelectric source imaging using 3SCO: A space
coding algorithm based on particle swarm optimization and l0 norm constraint.
NeuroImage 51, 183–205 (2010)
19. Shirvany, Y., Mahmood, Q., Edelvik, F., Jakobsson, S., Hedstrom, A., Persson, M.:
Particle Swarm Optimization Applied to EEG Source Localization of Somatosensory
Evoked Potentials. IEEE Transactions on Neural Systems and Rehabilitation
Engineering 22, 11–20 (2014)
20. Nakamura, T., Ito, S., Mitsukura, Y., Setokawa, H.: A Method for Evaluating the Degree
of Human’s Preference Based on EEG Analysis. In: Fifth International Conference on
Intelligent Information Hiding and Multimedia Signal Processing, Kyoto, Japan, pp. 732–
735 (2009)
21. Zhiping, H., Guangming, C., Cheng, C., He, X., Jiacai, Z.: A new EEG feature selection
method for self-paced brain-computer interface. In: 10th International Conference on
Intelligent Systems Design and Applications, pp. 845–849. IEEE, Cairo (2010)
22. Jin, J., Wang, X., Zhang, J.: Optimal selection of EEG electrodes via DPSO algorithm. In:
7th World Congress on Intelligent Control and Automation, pp. 5095–5099. IEEE,
Chongqing (2008)
23. Kim, J.-Y., Park, S.-M., Ko, K.-E., Sim, K.-B.: A Binary PSO-Based Optimal EEG
Channel Selection Method for a Motor Imagery Based BCI System. In: Lee, G., Howard,
D., Ślęzak, D., Hong, Y.S. (eds.) ICHIT 2012. CCIS, vol. 310, pp. 245–252. Springer,
Heidelberg (2012)
24. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: IEEE International Conference
on Neural Networks, Perth, Australia, pp. 1942–1948 (1995)
25. Ghosh, S., Das, S., Kundu, D., Suresh, K., Abraham, A.: Inter-particle communication and
search-dynamics of lbest particle swarm optimizers: An analysis. Information
Sciences 182, 156–168 (2012)
26. Samal, N.R., Konar, A., Das, S., Abraham, A.: A closed loop stability analysis and
parameter selection of the Particle Swarm Optimization dynamics for faster convergence.
In: IEEE Congress on Evolutionary Computation, Singapore, pp. 1769–1776 (2007)
27. Eberhart, R.C., Shi, Y.: Particle swarm optimization: developments, applications and
resources. In: 2001 Congress on Evolutionary Computation, Seoul, Korea, pp. 81–86
(2001)
28. Lin, C.-L., Mimori, A., Chen, Y.-W.: Hybrid Particle Swarm Optimization and Its
Application to Multimodal 3D Medical Image Registration. Computational Intelligence
and Neuroscience 2012, 7 (2012)
29. Andrzejak, R.G., Lehnertz, K., Mormann, F., Rieke, C., David, P., Elger, C.E.: Indications
of nonlinear deterministic and finite-dimensional structures in time series of brain
electrical activity: Dependence on recording region and brain state. Physical Review E 64,
061907 (2001)
30. Subasi, A.: Automatic detection of epileptic seizure using dynamic fuzzy neural networks.
Expert Systems with Applications 31, 320–328 (2006)
31. Subasi, A.: Epileptic seizure detection using dynamic wavelet network. Expert Systems
with Applications 29, 343–355 (2005)
32. Güler, İ., Übeyli, E.D.: Adaptive neuro-fuzzy inference system for classification of EEG
signals using wavelet coefficients. Journal of Neuroscience Methods 148, 113–121 (2005)
FOREX Rate Prediction Using Chaos, Neural
Network and Particle Swarm Optimization
Dadabada Pradeepkumar1,2 and Vadlamani Ravi1,∗

1
Center of Excellence in CRM and Analytics,
Institute for Development and Research in Banking Technology,
Hyderabad-500057, India
2
SCIS, University of Hyderabad, Hyderabad-500046, India
dpradeepphd@gmail.com, rav_padma@yahoo.com
Abstract. This paper presents two two-stage intelligent hybrid FOREX

Rate prediction models comprising chaos, Neural Network (NN) and
PSO. In these models, Stage-1 obtains initial predictions and Stage-2
fine tunes them. The exchange rates data of US Dollar versus Japanese
Yen (JPY), British Pound (GBP), and Euro (EUR) are used to test the
effectiveness of hybrid models. We conclude that the proposed intelligent
hybrid models yield better predictions compared to the baseline neural
networks and PSO in terms of MSE and MAPE.
Keywords: FOREX Rate Prediction, Hybrid model, Chaos, MLP,

GRNN, GMDH, PSO.
1 Introduction
The accurate prediction of Foreign Exchange (FOREX) rates helps a country to

make its economy stronger [1]. The prediction process is carried out using time
series analysis. A time series is said to be chaotic time series if and only if it is non-
linear, deterministic and sensitive to initial conditions [2]. The exchange rates
are inherently noisy, non-stationary and deterministically chaotic [34]. Chaotic
time series prediction involves the prediction of chaotic system behavior in future
based on information of current and past states of that system and is always a
complex problem.
In order to predict today’s value accurately one has to ask the following ques-
tions: (i) How many steps of relevant past values are to be considered in the
process of prediction? (ii) How many such past values are needed? The answer
to the first question is lag and to the second one is embedding dimension.
The following two studies form background to the current work. Ravi et al. [20]
presented a forecasting model using a number of computational intelligent tech-
niques. According to them, a variable Yt is predicted using a vector of lagged vari-
ables Yt−1 , Yt−2 ...Yt−m where m is count of considered lagged variables. Then,
Hadvandi et al. [21] presented the PSO-based autoregression time series model
∗

364 D. Pradeepkumar and V. Ravi
to forecast gold price, where only last two days’ gold price is considered to pre-
dict today’s gold price. They employed PSO to estimate optimal coefficients of
the model.
The drawbacks of these studies are: (i) Both studies considered sequential
lagged variables without using any scientific method to determine the optimal
lag. (ii) Both studies took arbitrary count of lagged variables which is not a
scientific approach. (iii) Both studies did not check the presence of chaos in
dataset. (iv) Hadvandi et al. [21] modeled only characteristic information of
gold price time series but they did not model residual informaton. Hence the
model needs to be extended.
The proposed two-stage prediction models have the following features:
1. The models check for presence of chaos at both stages.
2. The methodology determines both optimal lag and optimal embedding di-
mension scientifically as opposed to guessing them arbitraily.
3. They model both characteristic information and residual information in or-
der to yield better predictions.
The remainder of this paper is organized as follows: A review of literature is
presented in Section 2. In section 3, we presented overview of proposed hybrid
models. In section 4, experimental methodology is presented. Section 5 discusses
the obtained results. The paper is then concluded in Section 6.
2 Literature Survey
It is known that combining many forecasting models yields better estimates than
single model [35,36,37,38] in general and in time series [6] in particular. Many
previous researches had presented various hybrid FOREX rate prediction mod-
els. In this direction, Ni and Yin [7] proposed a hybrid of various regressive
neural networks and trading indicators moving average convergence/divergence,
relative strength index and genetic algorithm, Zhang [8] hybridized ARIMA and
MLP models, Zhang and Wan [9] proposed a statistical fuzzy interval neural
network, Donate et. al. [10] proposed a weighted cross-validation evolutionary
artificial neural network (EANN) ensemble, Gheyas et. al. [11] proposed novel
neural network ensemble, Yu et. al. [12] proposed a multistage nonlinear radial
basis function (RBF) neural network ensemble, Sermpins et.al. [13] proposed hy-
brid neural network architecture of Particle Swarm Optimization and Adaptive
Radial Basis Function (ARBF-PSO), Chang [14] proposed hybrid (PSOBPN)
that is composed of particle swarm optimization and back propagation network
(BPN), Huang et. al. [15] implemented a two-stage chaos and Support Vector
Machines (SVMs), Aladag et. al. [16] proposed a time invariant fuzzy time series
forecasting method based on PSO, Rout et. al. [17] proposed a hybrid predic-
tion model by combining an adaptive ARMA and Differential Evolution (DE)
based training of its feed-forward and feed-back parameters, Chen and Leung
[18] proposed a hybrid comprising a time series model and GRNN in tandem and
Ince and Trafalis [19] proposed a hybrid two-stage model consisting of ARIMA,
FOREX Rate Prediction Using Chaos, Neural Network 365
VAR and SVR and NN to predict foreign exchange rates. All of these researches
presented that hybrid forecasting models yielded better predictions than stand-
alone forecasting models. However comparing all of them is out of the scope of
the paper.
3 Proposed Models
3.1 Notations
Let l1 and m1 be lag and embedding dimensions that are used in Stage-1; l2 and
m2 be lag and embedding dimensions that are used in Stage-2 respectively; e(t)
be error at time t and ė(t) be predicted error at time t; α0 , α1 , α2 , ... be coeffi-
cients to be optimized. Finally, let ẏ(t) be the predicted value of Stage-1 at time
t and ÿ(t) be the final predicted value at time t; f (.) be a non-linear function for
obtaining predictions using Multi-Layer Perceptron(MLP)/General Regression
Neural Network (GRNN)/ Group Method for Data Handling(GMDH).
3.2 Hybrid Model-1

The proposed Hybrid Model-1 consists of two stages. Each stage, in turn, works
in both training and test phases. The detailed description is as follows. Let
Y = {y(1), y(2), ..., y(k), y(k+1), ..., y(n)} be a dataset of n observations at times
1, 2, ..., n respectively. First, check for the presence of chaos in Y and rebuild Y
using l1 and m1 , if chaos is present. Later, divide this dataset as Training set
Y1 = {y(1), y(2), ..., y(k)} and Test set Y2 = {y(k + 1), y(k + 2), ..., y(n)}. The
two-stage prediction proceeds as follows:
I. Stage-1: Using NN (MLP/GRNN/GMDH)
A. Training Phase
1. Input Y1 to NN.
2. Train NN and obtain initial training set predictions using eq.(1) and
errors using eq. (2):
y˙1 (t) = f (y(t − l1 ), y(t − 2l1 ), y(t − 3l1 ), ..., y(t − m1 l1 ))

t = l1 m1 + 1, l1 m1 + 2, ..., k (1)
e(t) = y(t) − y˙1 (t); t = l1 m1 + 1, l1 m1 + 2, ..., k (2)
B. Test Phase
1. Input Y2 to trained NN and obtain initial test set predictions using
eq.(3) and errors using eq.(4) :
y˙2 (t) = f (y(t − l1 ), y(t − 2l1 ), y(t − 3l1 ), ..., y(t − m1 l1 ))

t = k + 1, k + 2, ..., n (3)
e(t) = y(t) − y˙2 (t); t = k + 1, k + 2, ..., n (4)

After Stage-1 ends, Initial training set predictions

Ẏ1 ={ ẏ1 (l1 m1 + 1),ẏ1 (l1 m1 + 2),...,ẏ1 (k)} and Initial test set predictions
Ẏ2 = {ẏ2 (k + 1),ẏ2 (k + 2),...,ẏ2 (n)} are obtained.
II. Stage-2: Using PSO-Based Autoregression model

A. Training Phase
1. Check e(t) for the presence of chaos. Rebuild e(t) using l2 and m2 .
Obtain ė(t) using PSO-based Autoregression model as in eq. (5),
if chaos present; otherwise, apply Polynomial Regression to obtain
ė(t).
ė(t) = α0 + α1 e(t − l2 ) + α2 e(t − 2l2 ) + ... + αm2 e(t − m2 l2 )

t = l1 m1 + l2 m2 + 1, l1 m1 + l2 m2 + 2, ..., kl2 (5)
2. Compute final training set predictions using eq. (6):
y¨1 (t) = y˙1 (t) + ė(t); t = l1 m1 + l2 m2 + 1, l1 m1 + l2 m2 + 2, ..., k

(6)
B. Test Phase
1. Obtain ė(t) using PSO-based Autoregression model as in eq. (7), if
chaos present; otherwise, use Polynomial Regression to obtain ė(t).
ė(t) = α0 + α1 e(t − l2 ) + α2 e(t − 2l2 ) + ... + αm2 e(t − m2 l2 )

t = k + 1, k + 2, ..., n (7)
2. Compute final test set predictions using eq. (8):
y¨2 (t) = y˙2 (t) + ė(t); t = k + 1, k + 2, ..., n (8)
After Stage-2 ends,Final training set predictions

Ÿ1 ={ ÿ1 (l1 m1 + l2 m2 + 1),ÿ1 (l1 m1 + l2 m2 + 2),...,ÿ1 (k)} and Final test set
predictions Ÿ2 ={ ÿ2 (k + 1),ÿ2 (k + 2),...,ÿ2 (n)} are obtained.
3.3 Hybrid Model-2

The proposed Hybrid Model-2 also consists of two stages. In this hybrid, PSO is
invoked in Stage-1 and NN/PR is invoked in Stage-2 (if chaos is present/absent).
Accordingly, in this hybrid, in place of eq. (1) and eq. (3), eq. (9) and eq. (10)
are used to obtain initial predictions and in place of eq. (5) and eq. (7), eq. (11)
and eq. (12) are used to obtain final predictions.
y˙1 (t) = α0 + α1 y(t − l1 ) + α2 y(t − 2l1 ) + ... + αm1 y(t − m1 l1 )

t = l1 m1 + 1, l1 m1 + 2, ..., k (9)
y˙2 (t) = α0 + α1 y(t − l1 ) + α2 y(t − 2l1 ) + ... + αm1 y(t − m1 l1 )

t = k + 1, k + 2, ..., n (10)
ė(t) = f (e(t − l2 ), e(t − 2l2 ), ..., e(t − m2 l2 ))

t = l1 m1 + l2 m2 + 1, l1 m1 + l2 m2 + 2, ..., k (11)
ė(t) = f (e(t − l2 ), e(t − 2l2 ), ..., e(t − m2 l2 ))

t = k + 1, k + 2, ..., n (12)
4 Experimental Design
The foreign exchange data used in our study are obtained from US Federal Re-
serve System(http://www.federalreserve.gov/releases/h10/hist/). The
sets of data collected are of daily US dollar exchange rates with respect to
three currencies- JPY, GBP and EUR. The daily data of USD-JPY and USD-
GBP from 1st January 1993 to 31st December 2013 (6036 observations each)
and USD-EUR from 3rd January 2000 to 31st December 2013 (3772 observa-
tions), are used as datasets. From both USD-JPY and USD-GBP datasets,80%
of dataset is used as training set (4829 observations) and 20% of dataset is used
as test set (1207 observations) and from USD-EUR dataset,80% of dataset is
used as training set (3018 observations) and 20% of dataset is used as test set
(754 observations).
In the proposed hybrid models, Saida’s Method [22,23] implemented in MAT-
LAB is used for checking the presence of chaos. Akaike Information Criterion
(AIC) [24,31] available in Gretl tool is used to obtain optimal lag. Cao’s Method
[25,32] implemented in MATLAB is used to obtain minimum embedding dimen-
sion. Various Neural Networks (MLP/GRNN/GMDH) available in NeuroShell
tool [26,27,28,29,33] are used to obtain predictions. PSO [30] implemented in
Java is used to obtain coefficients of the autoregression model. Finally, Polyno-
mial Regression available in Microsoft Excel is used to obtain predictions in the
abscence of chaos.
While conducting experiments over the datasets, different user-defined param-
eters are tweaked in order to obtain the best performance from the techniques.
While training on USD-JPY data set using MLP, the learning rate is 0.6,the
momentum rate 0.9 and number of hidden nodes is 10, and using GRNN, the
smoothing factor is 0.1144531 and using PSO, number of particles is 50, dimen-
sions are 11, inertia is 0.8, iterations are 40000 and c1 = c2 = 2 are tweaked.
Similarly, while training on USD-GBP data set using MLP, the learning rate
is 0.5,the momentum rate 0.7 and number of hidden nodes is 30, and using
GRNN, the smoothing factor is 0.0915294 and using PSO, number of particles
is 60, dimensions are 16, inertia is 0.6, iterations are 40000 and c1 = c2 = 2 are
tweaked.Similarly, while training on USD-EUR data set using MLP, the learning
rate is 0.6,the momentum rate 0.8 and number of hidden nodes is 20, and using
GRNN, the smoothing factor is 0.0179688 and using PSO, number of particles
is 60, dimensions are 16, inertia is 0.8, iterations are 40000 and c1 = c2 = 2 are
tweaked. Since the performance of machine learning techniques in general and
Neural Networks techniques, in particular, is, by and large dataset dependent,
we need to tweak parameter values in order to get best results for that Neural
Network architecture in a given dataset.Mean Squared Error (MSE) and Mean
Absolute Percentage Error (MAPE) are used as performance measures as in (13)
and (14) :
k
(y(t) − ẏ(t))2
M SE = t=1 (13)
k
k $ $
100 $$ y(t) − ẏ(t) $$
M AP E = (14)
k t=1 $ y(t) $
where k is number of forecasting observations, y(t) is actual observation at time

t, ẏ(t) is predicted value at time t.

In tables 1, 2 and 3, the MSE and MAPE values for both training and test sets
after modeling chaos using stand-alone models are presented. In Tables 4, 5 and
6, the MSE and MAPE values for both training and test sets after modeling
chaos using hybrid models are presented along with the optimal lag and the
optimal embedding dimension. The experiments #1,#2 and #3 are variants of
Hybrid Model-1 and #4 is Hybrid Model-2. From Tables 4, 5 and 6, it is observed
that the GRNN, GMDH and PSO were extremely effective in modeling the chaos
present in the datasets in Stage-1 whereas MLP was ineffective in doing so. The
observation is corroborated by the figures 1-6 where the winning hybrid variant
and corresponding neural network in stand-alone mode (Stage-1) behaved pretty
closely. Further, it is observed that Stage-2 fine tunes the predictions only when
chaos is completely modeled in Stage-1.
Moreover, GMDH turned out to be the best technique in Stage-1, for obtaining
initial predictions of USD-JPY and USD-GBP datasets, while PSO-based auto-
regression model turned out to be the best in the case of USD-EUR dataset. The
best hybrid models in tables 4,5 and 6 are highlighted in boldface. The variant
of Hybrid Model-1, involving GMDH, turned out to be the best hybrid model in
the case of both USD-JPY and USD-GBP. whereas the Hybrid Model-2 yielded
the best result in the case of USD-EUR. Figures 1-6 depict the predictions of
proposed hybrid models for both training and test sets of all datasets along
with actual values, MLP, GRNN, GMDH and PSO predictions. It can be easily
observed that the proposed hybrid models outperformed the standalone MLP,
GRNN, GMDH and PSO.
Table 1. Results of Stand-alone models for USD-JPY data
Model Training set MSE (MAPE) Test set MSE (MAPE)

MLP 486.5760 (15.6530) 10.7855 (3.0728)
GRNN 5.7430 (1.6291) 33.4999 (5.7749)
GMDH 0.6581 (0.5064) 0.3560 (0.4793)
PSO 2.4686 (1.0316) 2.0915 (1.3491)
Table 2. Results of Stand-alone models for USD-GBP data

MLP 0.0417 (8.7604) 0.0027 (2.6912)
GRNN 0.0001765 (0.6146) 0.000540 (1.1062)
GMDH 0.0001024(0.4324) 0.0000856 (0.4530)
PSO 2.5353 (0.7056) 2.1005 (0.7044)
Table 3. Results of Stand-alone models for USD-EUR data

MLP 0.0527 (17.4361) 0.0024 (2.9137)
GRNN 0.00004707 (4.4992) 0.00010758 (0.5918)
GMDH 0.00006252 (0.4880) 0.000064089 (0.4473)
PSO 0.00006741 (0.5115) 0.000073068 (0.4842)
Fig. 1. Predictions of Training set of USD-JPY
Fig. 2. Predictions of Test set of USD-JPY

Table 4. Results of proposed hybrid models for USD-JPY data

Experi Training set Test set
-ment Stage # Technique / Chaos paramters MSE MSE
# (MAPE) (MAPE)
Stage -1
MLP (l1 = 4, m1 = 20) 0.7093 0.5211
#1 (Chaos present)
(0.5374) (0.6147)
Stage-2 PSO-based Model
(Chaos present) (l2 = 10, m2 = 11)
Stage-1
GRNN (l1 = 4, m1 = 20) 5.8905 5.0712
(Chaos present)
(1.6450) (7.9670)
#2 2nd degree polynomial
Stage-2 regression
(Chaos absent) 3rd degree polynomial 5.9475 2.9160
regression (1.6660) (4.4905)
4th degree polynomial 7.0218 1.8247
regression (1.8343) (3.4829)
Stage-1
GMDH (l1 = 4, m1 = 20) 0.7314 0.3552
(Chaos present)
(0.5487) (0.4757)
Stage-2 regression
(Chaos absent) 3rd degree polynomial 0.7314 0.3548*
regression (0.5487) (0.4757)
regression (0.5501) (0.4759)
2.6083 1.4666
(1.0825) (1.0676)
Stage-2 regression
regression (0.5147) (0.9866)
regression (1.3870) (0.4975)
Fig. 3. Predictions of Training set of USD-GBP

Table 5. Results of proposed hybrid models for USD-GBP data

# (MAPE) (MAPE)
Stage -1
MLP (l1 = 5, m1 = 16) 0.0002266 0.0001967
#1 (Chaos present)
(0.6717) (0.6933)
Stage-1
GRNN (l1 = 5, m1 = 16) 0.0008580 0.0008841
(Chaos present)
(0.6335) (1.5039)
Stage-2 regression
regression (1.6660) (1.6830)
regression (0.6409) (2.3914)
Stage-1
GMDH (l1 = 5, m1 = 16) 0.0001097 0.000094*
(Chaos present)
(0.4595) (0.4773)
Stage-2 regression
regression (0.4623) (0.4773)
regression (0.4595) (0.4959)
0.0002641 0.0002058
(0.7346) (0.7013)
Stage-2 regression
regression (0.7114) (0.9866)
regression (0.7114) (0.9993)
Fig. 4. Predictions of Test set of USD-GBP

Table 6. Results of proposed hybrid models for USD-EUR data

# (MAPE) (MAPE)
Stage -1
MLP (l1 = 1, m1 = 10) 0.0002328 0.0001922
#1 (Chaos present)
(0.9916) (0.7970)
Stage-1
GRNN (l1 = 1, m1 = 10) 0.0000522 0.0001245
(Chaos present)
(0.4807) (0.6445)
Stage-2 regression
regression (0.4888) (0.6843)
regression (0.5069) (0.7084)
Stage-1
GMDH (l1 = 1, m1 = 10) 0.0000697 0.0000744
(Chaos present)
(0.5226) (0.4899)
Stage-2 regression
regression (0.5226) (0.4941)
regression (0.5226) (0.4941)
0.0000684 0.0000736*
(0.5147) (0.4870)
Stage-2 regression
regression (0.5147) (0.4870)
regression (0.7114) (0.4975)
Fig. 5. Predictions of Training set of USD-EUR

Fig. 6. Predictions of Test set of USD-EUR
6 Conclusion
For predicting FOREX rates, the paper proposes two 2-stage hybrid models
comprising chaos theory, various neural network architectures viz. MLP, GRNN
and GMDH and PSO or Polynomial Regression. The results of the hybrids in
terms of MSE and MAPE on test datasets indicate that the proposed hybrid
models outperformed the stand-alone forecasting models: MLP, GRNN, GMDH
and PSO. This is the significant outcome of this study. And also, systematic
modeling of chaos present in the datasets along with the application of powerful
neural networks and PSO for prediction is the single most advantage of the
current research. Future directions include applying Multi-objective-PSO and
other competing techniques.
References
1. Hoag, A.J., Hoag, J.H.: Introductory Economics. 4th edn. World Scientific Pub-
lishing Co. Ptc. Ltd., Singapore (2006)
2. Dhanya, C.T., Nagesh Kumar, D.: Nonlinear ensemble prediction of chaotic daily
rainfall. Advances in Water resources 33, 327–347 (2010)
3. Yu, L., Wang, S., Lai, K.K.: Foreign-Exchange-Rate with Artificial Neural Net-
works. Springer, USA (2007)
4. Zhang, G., Patuwo, B.E., Hu, M.Y.: Forecasting with artificial neural networks:
The state of the art. Intl. J. of Forecasting 14, 35–62 (1998)
5. Ozkan, F.: Comparing the forecasting performance of neural network and purchas-
ing power parity: The case of Turkey. Elsevier, Economic Modelling 31, 752–758
(2013)
6. Temizel, T.T., Casey, M.C.: A comparitive study of autoregressive neural network
hybrids. Neural Networks 18, 781–789 (2005)
7. Ni, H., Yin, H.: Exchange rate prediction using hybrid neural networks trading
indicators. Neurocomputing 72, 2815–2823 (2009)
8. Zhang, G.P.: Time sereis forecasting using a hybrid ARIMA and neural network
model. Neurocomputing 50, 159–175 (2003)
9. Zhang, Y.Q., Wan, X.: Statistical fuzzy interval neural networks for currency ex-
change rate time series prediction. Applied Soft Computing 7, 1149–1156 (2007)
10. Donate, J.P., Cortez, P., Sanchez, G.G., Migue, A.S.: Time sereis forecasting using
a weighted cross-validation evolutionary artificial neural network ensemble. Neu-
roComputing 109, 27–32 (2013)
11. Gheyas, I.A., Smith, L.S.: A novel neural network ensemble architecture for time
sereis forecasting. NeuroComputing 74, 3855–3864 (2011)
12. Yu, L., Lai, K.K., Wang, S.: Multistage RBF neural network ensemble learning for
exchange rates forecasting. NeuroComputing 71, 3295–3302 (2008)
13. Sermpins, K., Theofilatos, K., Karanthanpoulos, A., Georgopoulos, E.F., Dunis,
C.: Forecasting foreign exchange rates with adaptive neural networks using radia-
basis functions and particle swarm optimization. European Journal of Operational
Research 225, 528–540 (2013)
14. Chang, J.F., Hsieh, P.Y.: Particle Swarm Optimization based on BackPropagation
Network Forecasting Exchange Rates. ICIC International 7(12), 6837–6847 (2011)
15. Huang, S.C., Chuang, P.J., Wu, C.F., Lai, H.J.: Chaos-based support vector regres-
sions for exchange rate forecasting. Expert Systems with Applications 37, 8590–
8598 (2010)
16. Aladag, C.H., Yolcu, U., Egrioglu, E., Dalar, A.Z.: A new time invariant fuzzy
time series forecasting method based on particle swarm optimization. Applied Soft
Computing 12, 3291–3299 (2012)
17. Rout, M., Majhi, B., Majhi, R., Panda, G.: Forecasting of currency exchange rates
using an adaptive ARMA model with differential based evolution. Journal of King
Saud University-Computer and Information Sciences 26(1), 7–18 (2014)
18. Chen, A.S., Leung, M.T.: Regression neural network for error correction in for-
eign exchange forecasting and trading. Computers and Operations Research 31(7),
1049–1068 (2004)
19. Ince, H., Trafils, T.B.: A Hybrid model for exchange rate prediction. CDecision
Support Systems 42, 1054–1062 (2004)
20. Ravi, V., Lal, R., Raj Kiran, N.: Foreign Exchange Rate Prediction using com-
putational Intelligence methods. Int. J. of Computer Information Systems and
Industrial Management Applications 4, 659–670 (2012)
21. Hadavandi, E., Ghanbari, A., Naghneh, S.A.: Developing a Time Series Model
Based On Particle Swarm Optimization for Gold Price Forecasting. In: IEEE Third
Int. Conf. on Business Intelligence and Financial Engineering (2010)
22. Saida, A.B.: Using the Lyapunov exponent as a practical test for noisy chaos (Work-
ing paper), http://ssrn.com/abstract=970074
23. http://www.mathworks.in/matlabcentral/fileexchange/22667-chaos-test
24. Akaike, H.: A new Look at the Statistical Model Identification. IEEE Transactions
on Automatic Control AC-19(6) (1974)
25. Cao, L.: Practical Method for determining the minimum embedding dimension of
a scalar time series. Physica D 110, 43–50 (1997)
26. Rumelhart, G.E., Hinton, G.E., Williams, R.J.: Learning internal representations
by error propagation (1). MIT Press, Cambridge (1986)
27. Specht, D.F.: A General Regression Neural Network. IEEE Transactions on Neural
Networks 2(6), 568–576 (1991)
28. Ivakhnenko, A.G.: The GMDH: A rival of stochastic approximation. Sov. Autom.
Control 3(43) (1968)
29. Farlow, S.J.: Self-Organizing Methods in Modeling: GMDH type Algorithm, Bazel,
Marcel Dekker Inc. Newyork (1984)
30. Kennedy, J., Eberhart, R.: Particle Swarm Optimization. In: Proc of IEEE Inter-
national Conference on Neural Network, Perth, Australia, pp. 1942–1948 (1995)
31. http://gretl.sourceforge.net/
32. http://www.mathworks.in/matlabcentral/fileexchange/
36935-minimum-embedding-dimension/content/cao_deneme.m
33. http://www.neuroshell.com/
34. Yao, J., Tan, C.L.: A case study on using neural networks to perform technical
forecasting of forex. Neurocomputing 12(4), 79–98 (2000)
35. Bates, J.M., Granger, C.W.J.: The combination of forecasts. Oper. Res. Quart. 20,
451–468 (1969)
36. Clemen, R.: combining forecasts: a review and annotated bibliography with dis-
cussion. Int. J. Forecast 5, 559–608 (1989)
37. Makridakis, S., Anderson, A., Carbone, R., Fildes, R., Hibdon, M., Lewandowski,
R., Newton, J., Parzen, E., Winkler, R.: The accuracy of extraploation (time series)
methods: results of a forecasting competition. J. Forecast. 1, 111–153 (1982)
38. Pelikan, E., De Groot, C., Wurtz, D.: Power consumption in West-Bohemia: im-
proved forecasts decorrelating connectionist networks. Neural Network World 2,
710–712 (1992)
Path Planning Using Neighborhood
Based Crowding Differential Evolution
Boyang Qu1,3, Yanping Xu1, Dongyun Wang1, Hui Song2, and Zhigang Shang2
1
School of Electric and Information Engineering, Zhongyuan University of Technology,
Zhengzhou, China
2
School of Electrical Engineering, Zhengzhou University, Zhengzhou, China
3
School of Information Engineering, Zhengzhou University, Zhengzhou, China
qby1984@hotmail.com, 120828633@qq.com, wdy1964@aliyun.com,
hsong320@163.com, zhigang_shang@zzu.edu.cn
Abstract. Path planning problems are known as one of the most important
techniques used in robot navigation. The task of path planning is to find several
short and collision-free paths. Various optimization algorithms have used to
handle path planning problems. Neighborhood based crowding differential
evolution (NCDE) is an effective multi-modal optimization algorithm. It is able
to locate multiple optima in a single run. In this paper, Bezier curve concept and
NCDE are used to solve path planning problems. It is compared with several
other methods and the results show that NCDE is able to generate satisfactory
solutions. It can provide several alternative optimal paths in one single run for
all the tested problems.
Keywords: Evolutionary Computation, Path Planning, Different Evolution,

Constrained Optimization.
1 Introduction
In recent years, due to the development of artificial intelligence and electrical

integration technologies, robots have been used in various fields such as lunar
exploration, navigation, underground exploration, rescue, etc. Mobile robot is an
integrated system which combines different functions. Path planning is one of the
key technologies used in mobile robot. The performance of the robot highly depends
on the path planning method used. Therefore, developing efficient path planning
method has attracted many researchers’ attention. The aim of path planning is to
generate a collision-free path between an initial location and a desired destination in
an environment which is full of obstacles. The planned path refers to the optimal or
suboptimal trajectory under certain specific conditions, such as the shortest path or the
safest path. A good strategy of robot path planning can make a robot fulfill a desired
task safely and effectively.
In an environment full of obstacles, the mobile robot must arrange its trajectory to
avoid obstacles and find the shortest path when it travels form starting point to target
point [1]. In literature, many algorithms have been used to solve the path planning
Path Planning Using Neighborhood Based Crowding Differential Evolution 377
problems, such as self-adjusting fuzzy control algorithm [2], genetic algorithms [3]
and [4], ant colony optimization [5] and particle swarm optimization [6]. However,
only a few works use differential evolution (DE) to solve this problem [7] and [8]. DE
is one of the most powerful stochastic real-parameter optimization algorithms in
current use. It is effective in solving single global optimization problems. However,
the canonical DE is not suitable to solve multimodal problems or locate multiple
peaks in one run. Therefore, a newly developed niching DE algorithm called
Neighborhood based Crowding Differential Evolution (NCDE) is used to handle path
planning problem in this paper. This algorithm is able to generate multiple optimal
paths simultaneously.
The remainder of this paper is organized as follows: Section 2 introduces the
definition of Bezier curves and how to use it to solve path planning problem. Section
3 presents the novel DE algorithm which is used to optimize the path. The
experimental preparation and simulation results are presented and discussed in section
4 and 5 respectively. Finally, section 6 concludes the paper.
2 Bezier Curves
For designing automobile bodies, French engineer Pierre Bezier invented the Bezier
Curves in 1962, which is a new parameter curve [9]. Bezier Curves have become an
essential tool in many areas, especially that it has been widely used in computer
graphics and animation. Bezier curve is suitable to describe the path because of its
space properties. Through controlling the anchor points, different Bezier curves can
be obtained. The path planning problem can be transformed into an optimization
problem with limited control points to be optimized [10].
2.1 The Definition and Properties of Bezier Curves

In Bezier Curves, n+1 vertices can define polynomials of degree n, and a Bezier
Curve of basis polynomial of degree n can be expressed using the following equation
[11]:
n
P(t ) =  Pi Bi , n (t )
i
，
t ∈ [0 1]. （1）
where Pi represents the coordinates of ith vertex, while Bi,n(t) stands for a Bernstein
polynomial, which is given as:
Bi , n (t ) = Cn i t i (1 − t ) n −i (i = 0,1,..., n). 2 （）
i
Cn is the binomial coefficient. The parameter equation of three times Bezier curve
could be obtained by combining formula (1) and (2) as follows:
P ( t ) = P0 (1 − t ) + 3Pt
3
1 (1 − t ) + 3P2 t (1 − t ) + P3 t .
2 2
3
（3）
From formula (3), we can observe that the three times Bezier curve starts at t=0
and ends at t=1.
The properties of Bezier curves can be described as follows [12] and [13]:
378 B. Qu et al.
1. Bezier curve is decided by four points. The curve goes through the first point
and the last point. However, the shape of the curve is determined by the two
other points.
2. The curve is a straight line if and only if all the control points are in a straight
line.
3. The start/end of the curve is tangent to the first/last section of the Bezier polygon
(The lines are used to connect the Bezier curve points, and they start from the
first point and end at the last point. The lines formed the Bezier polygon.). this
property can be described using the following formula:
P ′ ( 0 ) = 3 × ( P1 − P0 ) , P′ ( n ) = 3 × ( Pn − Pn −1 ) . (4)
Where P’(0) is the first derivative of the start point and P’(n) is the first derivative
of the end point.
3 DE and Neighborhood Based Differential Evolution
3.1 DE
Differential Evolution (DE) is a stochastic search technique developed by Storn and

Price in 1995 [14]. It is a simple but efficient search algorithm. It starts with a
randomly initialized population. The optimal solution is obtained through a cycle of
three stages known as mutation, crossover and selection. These three stages are
applied to every solution member xi in each generation to create a new solution. In
each generation, DE employs the mutation operation to produce a mutant vector vp
associated with each parent xp at each generation.
After the mutation phase, the binary (or uniform) crossover operation is applied
to each pair of the generated mutant vector and its corresponding parent vector. The
process can be represented as:
 v p ,i if randi ≤ CR
u p ,i =  (5)
 p ,i
x otherwise.
Where u p ,i is the offspring vector. The crossover rate CR is a user-specified
constant within the range [0,1], which determines what parameter dimensions of
u p ,i are copied from.
F and CR are the two most important control parameters in the DE. These two
parameters can significantly influence the optimization performance of the DE.
Therefore, to successfully solve an optimization problem, it is generally required to
perform a trial-and-error search for the most appropriate values for these parameters
[14].
Different from traditional evolutionary algorithms, DE employs the difference of
the parameter vectors to explore the objective function landscape. DE has been
proved to be one of most powerful evolutionary algorithm for solving the real-valued
optimization problems.
3.2 NCDE
The standard DE searches for a global optimum in a D-dimensional hyperspace. It is

efficient when searching for single global solution. However, it is often desirable to
locate multiple optimal solutions in real world problems. Taking path planning
problem as an example, it is always better to find several alternative optimal paths for
the decision makers to choose, as the prime objective for different decision makers
can be different. Therefore, the neighborhood based crowding differential evolution
(NCDE) is applied to solve the path planning problem. Different from canonical DE,
vector generations are limited to a number of similar individuals as measured by
Euclidean distance. In this way, individuals are evolved towards its nearest optimal
point and the possibility of between niche difference vector generation is reduced
[15]. The neighborhood concept facilitates multiple convergences, because it allows a
higher exploitation of the areas which pilot the moves.
4 Experiment Preparation
4.1 Experiment Setup
For the experiment, Matlab R2008a is used as the programming language and the
computer configurations are Intel Pentium® 4 CPU 3.00 GHZ, 4 GB of memory. The
DE parameters used are list as below:
Population size=30, F=0.5, CR=0.5
4.2 Cost Function

In order to obtain a satisfying path, an objective function for the path planning
problem need to be defined. Security and length of the path are the two most
important criteria for path planning problems. Therefore, these two criteria are used to
form the objective function in this work.
(1) Security
Considering that the path cannot intersect with the obstacles, a punishment function
fsafe can be designed as below:
 0 if d min > Dsafe
f = (6)
safe
d min if 0≤ d min ≤ Dsafe.
where dmin is the minimum distance between the path and all obstacles. Dsafe is a
predefined security distance.
(2) Length of the path
The length of the curve should be as short as possible. The length function is defined
as:
1
flen = L =  ( x '(t ))2 +( y '(t ))2 dt . (7)
0
Where x(t), y(t) are the coordinates of points.
380 B. Qu et al.
(3) The overall objective function

Considering the two objectives above weight coefficient a, the final cost function can
be defined as:
f = f len + af safe . (8)
where a is the weight factor to balance the two objectives and it is chosen as 1000 in
this experiment.
To assess the performance of the proposed algorithm, four predefined path planning
problems are used. The results are plotted in Fig.1. The green circles in the figure
mean the dangerous areas around the obstacles. The red path describes the best path
while the black paths illustrate the possible paths for every problem. Four algorithms
are tested on these problems with two Bezier curves (n=2, D=4 n=8):
1. PSO+: Classical Particle Swarm Optimizer with crossover operator
2. DMS-PSO+: Dynamic Multi-Swarm Optimizer with crossover operator
3. DE: Classical Differential Evolution
4. NCDE: Neighborhood Based Crowding Differential Evolution
The max fitness evaluation is 20000 for all algorithms. The population size of PSO
with crossover is set to 30 and the number of sub-swarms of DMS-PSO with
16 16
14 14
12 12
10 10
y
y
8 8
6 6
4 4
5 6 7 8 9 10 11 12 13 14 15 4 6 8 10 12 14 16
x x
F1 F3
16 18
16
14
14
12
12
10
y
10
8
8
6
6
4 4
4 6 8 10 12 14 16 4 6 8 10 12 14 16 18
x x
F2 F4
Fig. 1. Landscapes of the test problems
crossover is set to 10 (with 3 particles in each sub-swarm). All performances are

calculated and averaged over 25 runs with the random initialization. The results are
presented in Table 1. And to evaluate the difference of the two algorithms,
Nonparametric statistical method ttest is used. h=1 indicates a rejection of the null
hypothesis at the 5% significance level. h=0 indicates a failure to reject the null
hypothesis at the 5% significance level.
In Fig.1, F1, F2, F3 and F4 show that the distributions of the solutions achieved by
NCDE for each test problem, where the red line describes the best path. We can
observer that it is easy for NCDE to find the best suitable path for every problem as
well as other alternative satisfying paths which are not around the best solution at the
same time.
Table 1. Results for F1-F4
Problems PSO+ DMS-PSO+ DE NCDE
F1 Mean 14.7698 14.7465 14.8774 14.7237

Std. 0.0016 0.0006 0.1717 0.0000
Min 14.7007 14.7290 14.6842 14.7000
Max 14.8418 14.7797 15.9563 14.7404
h 1 1 1 0
F2 Mean 15.9863 15.8101 16.5603 15.6768
Std. 0.5050 0.1726 0.9181 0.1010
Min 15.6221 15.0865 15.2764 15.3144
Max 17.3518 17.2887 17.5268 16.7618
h 1 0 1 0
F3 Mean 17.0764 16.8388 17.4209 16.6135
Std. 0.2402 0.2563 0.1594 0.1398
Min 16.4046 16.3139 16.2838 16.3900
Max 18.1084 17.8895 17.8122 17.0463
h 1 1 1 0
F4 Mean 15.6119 15.1971 16.5413 14.9887
Std. 0.5745 0.4369 0.1743 0.2086
Min 14.8479 14.8024 14.8028 14.7138
Max 16.6730 16.6179 17.6099 16.6336
h 1 1 1 0
In [16], compared with PSO and DMS-PSO, PSO+ and DMS-PSO+ perform better
respectively. For every problem, every algorithm is compared with the best algorithm
with ttest. From the table, some conclusions could be made as follows:
1. DMS-PSO+ performs better than PSO+.
2. NCDE outperforms DE on all the four problems.
3. NCDE performs best among all algorithms on mean value.
382 B. Qu et al.
4. Except for problem 2, NCDE is better than the other three algorithms
obviously which can be seen from the result of ttest. The result of ttest on
problem 2 is no difference between NCDE and DMS-PSO+, which means
that the result of NCDE accepts DMS-PSO+ at the 5% significance level.
However, from the mean value we could know that NCDE has a smaller
value which means NCDE has a stable ability to find good solutions during
the process of search compared with DMS-PSO+.
In the process of searching, NCDE has a better local search ability, which makes
multi-paths possible, so various satisfied paths could be acquired for every problem in
this task. From the above four points we can know that, NCDE performs betters.
6 Conclusion
In this work, Bezier curves and neighborhood based crowding differential evolution
algorithm are used to tackle path planning problem. To assess the performance of the
neighborhood based crowding differential evolution algorithm in solving path
planning problem, four different path problems are tested. The experiments show that
neighborhood based crowding differential evolution is effective in solving all four
problems. In future work, dynamic environment and constraints will be added to
increase the complexity of the path planning problems. High order Bezier curves will
also be used to improve the quality of the solutions.
Acknowledgments. This work was supported in part by National Natural Science

Foundation of China (61305080, U1304602), Postdoctoral Science Foundation of
China (Grants 20100480859, 2014M552013), Specialized Research Fund for the
Doctoral Program of Higher Education (20114101110005), Scientific and
Technological Project of Henan Province (132102210521, 122300410264), and Key
Foundation of Henan Educational Committee (14A410001).
References
1. Chakraborty, J., Konar, A., Chakraborty, A.K., Jain, L.C.: Distributed Cooperative Multi–
Robot Path Planning Using Differential Evolution. In: 2008 IEEE Congress on
Evolutionary Computation, CEC 2008, Hong Kong, China, pp. 718–725 (2008)
2. Wu, C.D., Zhang, Y., Li, M.X.: A Rough Set GA-based Hybrid Method for Robot Path
Planning. Journal of International Automation and Computing, 29–34 (2006)
3. Sugihara, K., Smith, J.: Genetic algorithms for adaptive motion planning of an autonomous
mobile robot. In: Proc. of IEEE Intl. Symposium on Computational Intelligence in
Robotics and Automation, Monterey, CA, USA, pp. 138–143 (1997)
4. ALtaharwa, S.A., Alweshah, M.: A mobile robot path planning using genetic algorithm in
static environment. Journal of Computer Science 4, 341–344 (2008)
5. Bell, J.E., McMullen, P.R.: Ant colony optimization techniques for the vehicle routing
problem. Advanced Engineering Informatics, 41–48 (2004)
6. Saska, M., Macas, M., Preucil, L., Lhotska, L.: Robot path planning using particle swarm
optimization of ferguson splines. In: IEEE Conference on Emerging Technologies and
Factory Automation, Prague, pp. 833–839 (2006)
7. Jayasree, C., Amit, K., Jain, L.C., Chakraborty, U.K.: Cooperative multi-robot path
planning using differential evolution. Journal of Intelligent and Fuzzy Systems, 13–27
(2009)
8. Mo, H.W., Li, Z.Z.: Bio-geography based differential evolution for robot path planning.
In: 2012 IEEE International Conference on Information and Automation, ICIA 2012, pp.
1–6. Inner Mongolia, China (2012)
9. Bashir, Z.A., Hawary, M.E.: Short-term Load Forecasting using Artificial Neural Network
based on Particle Swarm Optimization Algorithm. In: 2007 Canadian Conference on
Electrical and Computer Engineering, CCECD, Canadian, pp. 272–275 (2007)
10. Ho, Y.J., Liu, J.S.: Collision-free Curvature-bounded Smooth Path Planning using
Composite Bezier Curve based on Voronoi Diagram. In: Proceedings of IEEE
International Symposium on Computational Intelligence in Robotics and Automation,
CIRA, pp. 463–468 (2009)
11. Yang, L.Q.: Path Planning Algorithm for Mobile Robot Obstacle Avoidance Adopting
Bezier Curve Based on Genetic Algorithm. In: Control and Decision Conference,
pp. 3286–3289 (2008)
12. Liu, H.G., Qin, G.L.: A Bezier Curve Based on The Military Arrow Mark Realized.
Ordnance of Sichuan 30, 67–68 (2009)
13. Gao, S., Zhang, Z.Y., Cao, C.G.: Particle Swarm Algorithm for The Shortest Bezier Curve.
In: International Workshop on Intelligent Systems and Applications, pp. 1–4 (2009)
14. Storn, R., Price, K.: Differential Evolution—a Simple and Efficient Adaptive Scheme for
Global Optimization over Continuous Spaces. Journal of Global Optimization 11, 22–25
(1995)
15. Qu, B.Y., Suganthan, P.N., Liang, J.J.: Differential Evolution with Neighborhood
Mutation for Multimodal Optimization. IEEE Transactions on Evolutionary Computation,
601–614 (2012)
16. Liang, J.-J., Song, H., Qu, B.-Y., Mao, X.-B.: Path Planning Based on Dynamic Multi-
Swarm Particle Swarm Optimizer with Crossover. In: Huang, D.-S., Ma, J., Jo, K.-H.,
Gromiha, M.M. (eds.) ICIC 2012. LNCS, vol. 7390, pp. 159–166. Springer, Heidelberg
(2012)
Neural Network Based on Dynamic Multi-swarm Particle
Swarm Optimizer for Ultra-Short-Term Load
Forecasting
Jane Jing Liang1, Hui Song1, Boyang Qu1,2, Wei Liu3, and Alex Kai Qin4
1
School of Electrical Engineering, Zhengzhou Univerisity, China
2
School of Electric and Information Engineering, Zhongyuan University of Technology, China
3
State Grid Henan Economic Research Institute, Zhengzhou, China
4
School of Computer Science and Information Technology RMIT University,
Melbourne 3001, Victoria, Australia
LIANGJING@zzu.edu.cn, qby1984@hotmail.com, liuwei830610@163.com,
{hsong320,kai.qin}@rmit.edu.au
Abstract. Ultra-Short-Term Load Forecasting plays an important role in Power

Load Forecasting. Back Propagation Neural Network(BPNN) has become one
of the most commonly used methods in Power System Ultra-Short-Term Load
Forecasting for its ability of computing complex samples and training large-
scale samples. However, traditional BPNN algorithm needs to set up a large
amount of network training parameters, and it is easy to be trapped into local
optima. A new algorithm which is Neural Network based on Dynamic Multi-
Swarm Particle Swarm Optimizer (DMSPSO-NN) is proposed for Ultra-Short-
Term Load Forecasting in this paper. DMSPSO-NN overcomes the shortage of
traditional BPNN and has a good global search and higher accuracy which
shows that it is suitable to be used for Ultra-Short-Term Load Forecasting.
Keywords: Ultra-Short-Term Load Forecasting; Back Propagation Neural

Network; Dynamic Multi-Swarm Particle Swarm Optimizer
1 Introduction
Power load forecasting is an indispensable part for managing and researching power
system, and it can make the full use of electricity and ease the conflict between supply
and demand based on the analysis of the existing electric energy [1]. Power system
load forecasting method based on electric power, economic, social and meteorological
factors and so on. According to the time length of the prediction, power load
forecasting can be classified as ultra-short-term load forecasting, short-term load
forecasting, medium long-term load forecasting and long-term load forecasting. In
terms of power system dispatching and management, ultra-short-term load forecasting
which varies from an hour to a week is the most important. Accurate ultra-short-term
load forecasting is very important in maintaining ultra-short-term analysis for electric
power, power exchange, trading evaluation as well as the analysis of the network
function, security and trend, the safety strategy of reducing load and so on[2].
Neural Network Based on Dynamic Multi-swarm Particle Swarm Optimizer 385
In recent years, there are various methods to solve this problem, such as Expert
Systems(ES)[3], Support Vector Machines(SVM)[4][5], Back Propagation Neural
Network(BPNN)[6][7] and so on. The main drawback of the ES is that it learns
nothing from the environment and has ambiguous relationship between rules, low
efficiency and adaptability. The main disadvantage of SVM is difficult to achieve
large-scale training samples and solve multi-classification problems. BPNN is widely
used in power load forecasting in recent years, to calculate large-scale complex
training samples as well as to slow the speed of convergence during training process.
With the development of Evolutionary Algorithm (EA), some researchers have
found its advantage in handling large-scale, non-differentiable and complex multi-
mode problem without any information about optimized problems for its global
convergence ability and strong robustness. EA can optimize the weight, structure and
learning rules of NN by searching for optimal solutions in search space with the help
of evolutionary strategy, genetic algorithm or evolutionary programming. Genetic
Algorithm (GA) is the most widely used in EA because it can deal with many
complex problems. However, GA is easy to be trapped into local optimum and the
process is difficult to control when GA is used to train NN.
Recently, the emergence of Swarm Intelligence such as Particle Swarm Optimizer
(PSO) has overcome the drawbacks of EA. Compared with other algorithms, Particle
Swarm Optimizer is simple, easy to realize, and need less parameter to adjust, making
it an effective optimal tool. So PSO has been used to optimize NN widely such as
power load forecasting [8], Fault Diagnosis of Power Transformer [9], Reservoir
Parameter Dynamic Prediction [10], Modeling and Simulation of Screw Axis [11]and
so on[12]. However, the traditional PSO can’t maintain the diversity of particles and
is difficult to reach global optimum. So Dynamic Multi-Swarm Particle Swarm
Optimizer (DMSPSO) which not only overcomes the drawback of the Particle Swarm
Optimizer, but also has strong global search ability, is proposed to optimize NN in
this paper. The result shows DMSPSO is easy to find global optimum when it is used
to optimize NN.
The rest of this paper is organized as follows. Section II gives a brief introduction
about the basic Particle Swarm Optimizer and describes the search process of the
Dynamic Multi-Swarm Particle Swarm Optimizer. The Back Propagation Neural
Network and Neural Network Based on Dynamic Multi-Swarm Particle Swarm
Optimizer model employed in this work are described in detail in Section III. Section
IV introduces the experimental setup and presents the results. Conclusions and future
work are given in Section V.
2 Dynamic Multi-swarm Particle Swarm Optimizer
Particle Swarm Optimizer was proposed by Kennedy and Eberhart in 1995[13][14].

The basic idea of PSO simulates the behavior of flying birds. In PSO, each bird is
regarded as a potential solution in search space which is called a “particle”. There
exists a fitness value in each particle obtained by the fitness function. Each particle
386 J.J. Liang et al.
adjusts the distance and its flying direction according to its velocity. The model of
PSO is shown as:
Vi d = ω *Vi d + c1 * rand1id * ( pbestid − X id ) + c2 * rand 2id * ( gbest d − xid ) (1)
X id = X id + Vi d (2)
Where, ω is the inertia weight and the range is [0.4 0.9]; c1 and c 2 are balance
factors which are set 2.05 in general; rand is a random number in [0, 1].The weakness
of standard PSO is premature and easy to be trapped into local optimum.
Dynamic Multi-Swarm Particle Swarm Optimizer (DMS-PSO) is developed from
local Particle Swarm Optimizer, where neighborhood structure is used in a small
population [15][16]. In order to increase the distribution of population and accelerate
the speed of convergence, the entire population is divided into sub-swarms equally in
DMS-PSO and each sub- swarm searches in space with its own particles. The
population is regrouped randomly every L generation (L is called regroup period),
then the population starts the search with new topology structure. Due to this method,
the information obtained from sub-swarms exchanges among them, and the diversity
of the population is also increased. The updating formula is given as follows:
Vi d ← ω ∗ Vi d + c1 ∗ rand1i d ∗ ( pbesti d − X i d ) + c2 ∗ rand 2i d ∗ (lbestk d − X i d )
Vi d = min(Vmax
d
, max(−Vmax
d
,Vi d )) (3)
X i ← X i + Vi
d d d
where, Vid represents the velocity of ith particle in dimension d; Xid is the position of
th
i particle in dimension d. lbestkd is the position of local optimum in dimension d of
kth sub-swarm; pbestid is the best personal position in dimension d of ith particle.
3 Back Propagation Neural Network Based on Dynamic

Multi-swarm Particle Swarm Optimizer
3.1 Back Propagation Neural Network
Back Propagation Neural Network can solve learning problems for connecting
weights between hidden units in multi-layers network, so it has become one of the
most important modal of Artificial Neural Network. The BPNN is made up of three
layers: input layer, hidden layer and output layer. In order to get satisfied forecast,
backward transmission error and error correction methods are used to adjust the
network parameters (weights and threshold)[17].
Input layer neurons are responsible for receiving input information from the
outside world, and transmitting to the middle layer neurons (which is hidden layer).
The hidden layer is the internal information processing layer, which is responsible for
transforming information. According to the requirements of changed information,
hidden layer can be designed for single hidden layer or several hidden layers[18][19].
The structure of the BP Neural Network is shown as follows:
Where, x0 , x1 ,..., x j ..., xn is the input value of BPNN, o1 , o2 ,..., ok ..., ol is predictive
value, vij and w jk (i = 1, 2,..., n, j = 1, 2,..., m, k = 1, 2,..., l ) are the input and output
weights of BPNN respectively. If the input node is n and the output node is l, BPNN
expresses the mapping relationship from n independent input variables to l
independent output variables. BPNN acquires associative memory and ability to
predict through training the network.
x x x
i n−1 n
Fig. 1. Structure of BPNN
3.2 The Process of Neural Network Based on Dynamic Multi-swarm Particle

Swarm Optimizer
Dynamic Multi-Swarm Particle Swarm Optimizer is used to optimize the input

weights, input thresholds, output weights output thresholds of NN. The process of
DMSPSO which is used to train the parameter of NN includes the following steps:
a. Ensure the parameters of NN
According to the input and output(x, o) determine the number of input layer nodes n,
hidden layer nodes m and output layer nodes l. Then the total dimension can be
calculated through the following formula:
D=n*m+m+m*l+l (4)
Sigmoid is used as the excitation function.
1
f ( x) = (5)
1 + e− x
At the same time, we normalize the original data input and output samples in order
to increase the learning efficiency of NN.
0.1 + 0.8 × ( xkold − min xkold )
xknew =
max xkold − min xkold
(6)
0.1 + 0.8 × (okold − min okold )
ok =
new
max okold − min okold

where, K=1, 2,…m is the number of samples. xkold and ykold, xknew and yknew
represent the input and output of the network which are unprocessed and processed
respectively.
b. Initialize the parameters
Initialize the weight between input layer and hidden layer vij and between hidden layer
and output layer w jk , hidden threshold a and output threshold b, regroup period L,
number of Sub-warms P, population of every sub-swarm ps and every particle’s
velocity.
c. Calculate fitness value
Firstly, calculate the output of hidden layer H according to the input data x,
weights vij between input and hidden layer and threshold a.
n
H j = f ( vij xi − a j ) j = 1, 2,..., m (7)
i =1
m is the nodes of hidden layer and f is the excitation function of hidden layer.
Secondly, according to the output of hidden layer H, weights w jk between hidden
layer and output layer, threshold b calculate o which is the predicted output of NN.
m
ok =  H j w jk − ak k = 1, 2,..., l (8)
j =1
Then, according to predicted output o and expected output p calculate predicted

error e of neural network.
etk = Pt k − otk k = 1, 2,..., l t = 1, 2,... ps * P (9)
The fitness function is set as:
l
fit (t ) =  abs (etk ) (10)
k =1
Calculate every particle’s fitness value by (10), search for each particle’s best
position achieved so far.
d. Get sub-swarm
Divide the whole population into sub-swarms equally and get the local optimum
lbestP of every sub-swarm according to the idea of DMSPSO.
e. Update
Update every particle’s position and velocity, and then enter into b.
f. Judge the times of loop
If the iteration of regroup is satisfied, all the sub-swarms will be regrouped again.
g. If iteration ends, stop, else return e.
This data of experiment is got through the system of monitoring and analyzing key
power industry. Two models which are BPNN and DMSPSO-NN are used to predict
the power load about one day which is based on the data achieved 29 day previously.
According to the recorded load data of October, November, December, precious 29

days in December, related days (related days of everyday are 20 days) and related
sampling points(related points for every sampling point are 42 points) of these 29
th
days are regarded as train data to predict the load of 30 , December. The prediction of
each algorithm contains the whole day’s prediction and every point’s prediction
during one day. They are defined as follows:
BPNN1: The data of the whole day is regarded as a whole prediction output (the
dimension of the output is 96) and the predicted mode is BPNN.
DMSPSO-NN1: The data of the whole day is regarded as a whole prediction
output and the predicted mode is DMSPSO-NN.
BPNN2: Every point in one day is regarded as a whole prediction output (the
dimension of the output is 1) and the predicted mode is BPNN.
DMSPSO-BPNN2: Every point in one day is regarded as a whole prediction output
and the predicted mode is DMSPSO-NN.
The parameters used in this task are set as follows:
The structure of NN1: 62 input nodes, 20 hidden nodes, 96 output nodes
The structure of NN2: 62 input nodes, 20 hidden nodes, 1output node
Dimension: 1281
Population size: 30
Number of sub-swarms:5
MaxFES (Maximum Fitness Evaluation): 40000
x 10
4
DMSPSO-NN1 x 10
4 BPNN1
2.3 2.3
forecast output forecast output
2.2 real output 2.2 real output
2.1
2.1
2
output
output
2
1.9
1.9
1.8
1.8
1.7
1.7 1.6
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
sample sample
Fig. 2. The whole day as a sample to be optimized
x 10
4
DMSPSO-NN2 x 10
4 BPNN2
2.3 2.3
forecast output forecast output

2.2 real output 2.2 real output
2.1 2.1
output
output
2 2
1.9 1.9
1.8 1.8
1.7 1.7
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
sample sample
Fig. 3. Every point in one day as a sample to be optimized

x 10
-3 Comparison of error rate x 10
-3 Comparison of error rate
6 12
DMSPSO-NN1 DMSPSO-NN2
4 BPNN1 10
BPNN2
8
2
6
error rate
error rate
0
4
-2
2
-4
0
-6 -2
-8 -4
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
sample sample
Fig. 4. Comparison of error rate under different conditions
Table 1. The error rate which is below 0.3%

BPNN1 DMSPSO-NN1 BPNN2 DMSPSO-NN2
Error rate 0.9796 1 0.9892 1
Some conclusions could be made from the result as follows:

• DMSPSO-NN performs better than BPNN which shows that dynamic sub-
swarm has improve the diversity of the solutions;
• To predict every point has a better result than regard the whole day as a
sample;
• The accuracy of DMSPSO-NN is higher than PSO.
From these three points, we could know that DMSPSO has a better global search
ability which helps it avoid being trapped into local optimum. After combining the
NN with the DMS-PSO, the prediction is much more better which shows that
DMSPSO-NN is suitable for Ultra-Short-Term Load Forecasting.
5 Conclusions
In this paper, an improved PSO(DMSPSO) is employed to optimize NN. What’s
more, two methods are used to test the property of DMSPSO-NN. The result shows
that DMSPSO-NN has better global search ability when it is used in Ultra-Short-Term
Load Forecasting problem. The result of error rate also makes us know that when
every point is regared as training sample, the result is much more better. In the future,
more algorithms will be used to predict load, and the more better and fast algorithm
will be used for online forcast.
Acknowledgment. This research is partially supported by The Second Batch Project

of Science and Technology of Henan Electric Power Company in 2013
(5217L0135029)and National Natural Science Foundation of China (61305080) and
Postdoctoral Science Foundation of China (20100480859) and Specialized Research
Fund for the Doctoral Program of Higher Education (20114101110005) and
Postdoctoral Science Foundation Grand (2014M552013).
References
1. Gross, G., Galiana, F.D.: Short-term load forecasting. Proceedings of the IEEE 75(12),
1558–1573 (1987)
2. Abdel-Aal, R.E.: Improving electric load forecasts using network committees. Electric
Power Systems Research 74(1), 83–94 (2005)
3. Hsu, Y.Y.: Fuzzy expert systems: an application to short-term load forecasting. IEE
Proceedings C: Generation Transmission and Distribution 139(6), 471–477 (1992)
4. Francis, E., Tay, H.: Application of support vector machines in financial time series
forecasting. Omega 29(4), 309–317 (2001)
5. Rüping, S.: Incremental learning with support vector machines. In: Proceedings-IEEE
International Conference on Data Mining, ICDM, pp. 641–642 (2001)
6. Luo, X., Zhou, Y.H., Zhou, H.: Forecasting the daily load based on ANN. In: Control
theory and Application, pp. 1–4 (2007)
7. Kim, C.I., Yu, I.K.: Kohonen neural network and transform based approach to short-term
load forecasting. Elect. Elecr. Power Syst. Res. 63(3), 169–176 (2002)
8. Hu, J., Zeng, X.: A hybrid PSO-BP algorithm and its application. In: 2010 Sixth
International Conference on Natural Computation (ICNC), vol. 5, pp. 2520–2523. IEEE
(2010)
9. Li, H., Yang, D., Ren, Z.: Based on PSO-BP network algorithm for fault diagnosis of
power transformer. In: 2010 International Conference on Computer, Mechatronics, Control
and Electronic Engineering (CMCE), vol. 4, pp. 484–487. IEEE (2010)
10. Zhang, L., Ma, J., Wang, Y.: PSO-BP neural network in reservoir parameter dynamic
prediction. In: 2011 Seventh International Conference on Computational Intelligence and
Security (CIS), pp. 123–126. IEEE (2011)
11. Zhang, P.Y., Sheng, Y.L., Wan, L.L.: Modeling and simulation of screw axis based on
PSO-BP neural network and orthogonal experiment. In: Second International Symposium
on Computational Intelligence and Design, ISCID 2009, pp. 272–275. IEEE (2009)
12. Ren, J., Yang, S.: An Improved PSO-BP Network Model. In: 2010 International
Symposium on Information Science and Engineering (ISISE), pp. 426–429. IEEE (2010)
13. Kennedy, J., Eberhart, R.C.: Particle Swarm Optimization. In: IEEE International
Conference on Neural Networks, Piscataway, NJ, pp. 1942–1948 (1995)
14. Shi, Y., Eberhart, R.C.: A Modified Particle Swarm Optimizer. In: The IEEE Congress on
Evolutionary Computation, pp. 69–73 (1998)
15. Liang, J.J., Suganthan, P.N.: Dynamic Multi-Swarm Particle Swarm Optimizer with Local
Search. In: IEEE Congress on Evolutionary Computation, vol. 1, pp. 522–528 (2005)
16. Liang, J.J., Suganthan, P.N.: Dynamic Multi-Swarm Particle Swarm Optimizer. In: IEEE
International Swarm Intelligence Symposium, pp. 124-129 (2005)
17. Gavrilas, M., Ciutea, I., Tanasa, C.: Short-term Load Forecasting with Artificial Neural
Network Models. Proceedings of IEE CIRED 25(12), 28–31 (2001)
18. Wang, Q., Zhou, B., Li, Z.: Forecasting of short-term load based on fuzzy clustering and
improved BP algorithm. In: International Conference on Electrical and Control
Engineering (ICECE), pp. 4519–4522. IEEE (2011)
19. Yao, S.J., Song, Y.H., Zhang, L.Z., Cheng, X.Y.: Wavelet transform and neural networks
for short-term electrical load forecasting. Energy Conversion and Management 41(18),
1975–1988 (2000)
Dynamic Differential Evolution for Emergency
Evacuation Optimization
Shuzhen Wan
School of Computer and Information Technology,

Three gorges University of ChinaYichang, China
wanshuzhen@163.com
Abstract. Emergency evacuation in public places has become the hot area of
research in recent years. Emergency evacuation route assignment is one of the
complex dynamic optimization problems in emergency evaluation. This paper
proposed the modified dynamic differential evolution algorithm and studied the
emergency evacuation, then applied the multi-strategy dynamic differential
evolution for emergency evacuation route assignment in public places. We use
the Wuhan Sport Center in Wuhan China as the experiment scenario to test the
performance of the proposed algorithm. The results show that the proposed
algorithm can effectively solve the complex emergency evacuation route
assignment problem.
Keywords: dynamic differential evolution, emergency evacuation assignment,

dynamic optimization.
1 Introduction
Evolutionary Algorithms (EAs) have been applied to solve dynamic optimization

problems. Many real world optimization problems are dynamic optimization
problems (DOPs)[1-3]. Dynamic optimization algorithms are different from the
traditional one which focuses on the static conditions, while the former one admits
that both the problems and the solutions may be changed in time.
More and more scholars pay special attention to the study of applying the EAs to
solve dynamic optimization problems in recent years. Several approaches have been
developed in EAs to address DOPs, such as maintaining diversity during the run via
random immigrants [4, 5], using memory to store and reuse useful information[3, 4,
6], multi-population approach[1, 7, 8], increasing diversity after a change[2, 9, 10].
Although all the above approaches are effectively for solving some dynamic
optimization problems, there are some points of criticism, Such as: how to trace the
different changing optimums in searching space. We use the prediction based multi-
strategy differential evolution algorithm [11]to meet the challenges. We use a hybrid
method that combines population core based multi-population strategy and prediction
strategy and new local search scheme to enhance differential evolution (DE)
performance for solving DOPs. The population core based multi-population strategy
Dynamic Differential Evolution for Emergency Evacuation Optimization 393
is useful to maintain the diversity of the population by using the multi-population and
population core concept. The prediction strategy is helpful to rapidly adapt to the
dynamic environment by using the prediction area. The local search scheme is useful
to improve the searching accuracy by suing the new chaotic local search method.
Experimental results on the moving peaks benchmark show that the proposed
schemes enhance the performance of DE in the dynamic environments. In this paper,
we apply the proposed algorithm to solve the real world dynamic optimization
problems-emergency evacuation route assignment optimization problem.
Emergency evacuation is an important issue for the large public spaces. There have
been a lot of literatures focus on the problem of emergency evacuation[12, 13]
Emergency evacuation is the study of how to evacuate the people from dangerous
locations to safe places in hurry.[3] Evacuation planning is a very complex problem
which needs to satisfy the consideration of many aspects. From the perspective
operation research, evacuation planning is a dynamic optimization problem.
Studies on evacuation in buildings mainly focus on the simulation [13] some other
researches focus on the evacuation of the traffic network [14].
This paper applied the proposed dynamic differential evolution algorithm to solve
the emergency evacuation route assignment optimization problem and achieved good
results comparing with other algorithms.
The remaining sections of this paper are organized as follows: Section II describes
the proposed algorithm. Section III details the emergency evacuation route
assignment model. Section IV presents the applying of the proposed algorithm on the
emergency evacuation route assignment. Section V introduces the experimental study
and discussion based on the experimental results. Finally, Section VI draws
conclusions.
2 Prediction based Multi-Strategy Differential Evolution
In order to enhance the performance of differential evolution for dynamic

optimization problems, we use the population core based multi-population strategy to
maintain the population diversity and avoid the premature convergence. And the new
local search scheme is employed in our algorithm to improve the exploitation
precision. Moreover, the prediction strategy is utilized to trace the optima when the
environment changes. Some of the details of the three schemes will be given as
follows. This algorithm is also detailed in our previous work [11].
2.1 Multi-Population Strategy

The multi-population method can be used to maintain multiple populations on
different optima when dealing with DOPs. To make this method work, the search
space is divided into several local search spaces, and each local search space can
contain more than one peak. Each subpopulation covers one local search space and
searches the local optimum in it. However, there are some important issues to be
considered such as how to divide the local search space, how to decide the number of
394 S. Wan
the subpopulation, how to generate the subpopulations. In this paper, we use the
hierarchical clustering method to achieve the goal of dividing subpopulations. The
population core concept is used also to maintain the diversity of populations after the
dividing.
2.2 Tent Map Based Local Search

Because of the ergodicity and randomicity, a chaotic system changes randomly, but
eventually goes through every state if the time duration is long enough. This
characteristic of chaotic systems can be utilized to build up a search operator for
optimizing objective functions. To improve the local search ability, we introduce the
tent map based local search method in the proposed algorithm. The tent map based
local search scheme is described in [11].
2.3 Prediction Strategy

Tracing the new optima in the changed environment is the most important issue for
the algorithms which are applied to deal with DOPs. The changes of environments
might be very complex, and the width, the height and the location of the optima can
all be changed. So, how to find the optima accurately and quickly have become the
challenges for all the dynamic algorithms. We propose the prediction strategy to
predict the location of the new optima before the environments are changed. The
prediction strategy is based on the idea that we create the prediction areas before the
changes occur, if the prediction areas just cover the areas containing the new optima,
the algorithm will quickly find the optima, if not, the searching speed can also be
improved due to the prediction areas are nearer to the new optima, and the probability
of finding the optima will be increased. The details of the prediction scheme are
shown in [11].
3 Evacuation Route Assignment Model
In this paper, we utilize the model [15] as the base of our evacuation route assignment
model. In literature [15], three optimization objects are presented: minimizing the
evacuation time, minimizing the total travel distance of all evacuees and minimizing
the congestion during the evacuation process. The evacuation time is the most
important in the three objects because if all evacuees can be evacuated within the set
time, the evacuation route assignment will be effective. Thus, in this model, we select
the evacuation time as the single optimization object, as the same, we take the
congestion as the constraints. If the congestion degree during the evacuation process
exceeds a threshold value which is defined according to the extent of the evacuation
environment can afford, the evacuation route assignment proves infeasible. This
object is defined as follows:
min T = max{t D1 , t D 2 ,...t Dr ,...t Dn }, r ∈ M

(1)
s.t. f max < f pre
Where, f max is the max congestion of the evacuation passageway in the system during
the evacuation, f pre is threshold of the congestion. The congestion in the evacuation
passageway during the evacuation is changed with time, so the evacuation route
assignment is the dynamic optimization problem.
4 Modified Dynamic DE for Evacuation Route Assignment
4.1 Frame of the Proposed Algorithm for Evacuation Route Assignment

In this paper, we apply the prediction based multi-strategy differential evolution
algorithm [11]to solve the evacuation route assignment optimization problem. We
employ the k-shortest paths algorithm [16] to initialize the evacuation routes. The
frame of the proposed algorithm for evacuation route assignment is detailed as
follows:
Step 1. Initialize the evacuation routes with k-shortest paths algorithm.
Step 2. Initialize the population randomly in the searching space and the archive
for prediction strategy
Step 3. Use the clustering method to create the subpopulations
Step 4. Apply DE operators (mutation, crossover and selection) on each individual
in each subpopulation.
Step 5. Apply the diversity control operator to maintain the population diversity.
Step 6. If optimization environments are changed, Set the prediction zone then go
back to Step 3. Else, goto next step.
Step 7. If a termination condition is met then goto end, else go back to Step 4.
4.2 Frame of the Proposed Algorithm for Evacuation Route Assignment

The goal of the evacuation route assignment is to construct an effective and safe
evacuation scheme to guide the evacuees evacuating from the dangerous orderly.
Considering the complexity of the evacuation route assignment, we adopt the sorting
based precoding method to encode each individual.
Firstly, because the evacuees often select the shortest paths while evacuating, so
we apply k-shortest paths algorithm to initialize the paths from the starting point to
the destination of evacuation in the evacuation zone.
Next, encoding each route with an integer and then we assign the coded paths to
the individuals randomly. The individual of the population is coded as follows:
{
individual = λ1λ2...λN1 | λN1 +1...λN 2 | ...
}
(2)
| λNn−1 +1...λNn , λ1λ2...λNn ∈ N
396 S. Wan
where λNi is the evacuation passageway assigned to i individual. Each evacuee can
randomly select these paths to evacuate, all the evacuation paths selected by evacuees
can be constructed the evacuation routes assignment scheme.
4.3 Algorithm Design

For the evacuation route assignment, the dynamic DE[11]should be modified to adopt
this optimization problem. The mutation operation and the crossover operation of DE
are modified as follows.
For the mutation operation, firstly, we set a mutation factor F (value range 0~1),
then mutate the individual A with the mutation factor F , thus, the evacuation paths
change. The change of evacuation paths is to reselect the evacuation path whose start
point is corresponding to variables of the individual A and being calculated by the k-
shortest paths algorithm. After the muting, we can obtain the mutated individual A ' .
For the crossover operation, we perform the crossover operation on each variable
of the mutated individual A ' with the crossover rate ( CR ). The variable will be
retained if that variable is selected, otherwise, it will be replaced by the variable
selected from the individual A . The process will continue till the crossover operation
performs on all variables. Where, the variables are referred to the evacuation paths.
The mutation operation and crossover operation are shown in Fig. 1 and Fig. 2.
Individual A= La1 La2 La3 La4 La6
mutated individual A’= L’a1 L’a2 La3 La4 L’a5 La6
Selecting new path L’a1 L’a2 L’a5
Fig. 1. Mutation operation for evacuation route assignment with F = 0.5
A’= L’a1 L’a2 La3 La4 L’a5 La6
A’’= La1 L’a2 L’a3 La4 La5 La6
A= La1 La2 La3 La4 La5 La6
Fig. 2. Crossover operation for evacuation route assignment with CR = 0.5

5 Experimental Study and Discussion
5.1 Experimental Settings
The settings for the proposed dynamic evolution algorithm are shown in Table 1.
Table 1. Settings for the proposed algorithm
Parameter Setting
Population size 100
Maximum subpopulation size 15
Mutation factor( F ) 0.5
Crossover rate( CR ) 0.9
Radius of the subpopulation core rcore 5.0
Number of optional paths for each start point 50
Threshold of congestion degree f pre 0.75
The Wuhan Sport Center in Wuhan city China is taken as the experimental area to
test the performance of the proposed algorithm. Wuhan Sport Center can hold about
60000 people, and there are 42 grandstands to accommodate the spectators. Suppose
in a massive activity, the spectators should be evacuated to the safe area as quickly as
possible for some reasons such as fire disaster and horrible attack. The number of
evacuees is about 24727. The spectator is assigned to each grandstand randomly
according to the number limitation of each grandstand.
The evacuation network of this stadium which contains 158 nodes and 224 arcs is
converted by its structure (Fig. 3).The original locations of the spectators are the 42
grandstands, and the exits are the 5 ticket entrances, final destinations of evacuation in
this scenario, as it is shown below.
Fig. 3. Experiment area and evacuation network
In order to measure the performance of the proposed algorithm, we compare our

proposed algorithm with DynDE[17]which is the classic dynamic differential
evolution. In the experiments, each algorithm runs independently for 20 times with
the 200 initialized paths for each start point. Though, the object is the minimum
evacuation time, we also calculate the congestion degree and the total length of all
evacuation paths.
398 S. Wan
The results are shown in Table 2.
Table 2. Results of the two algorithms for the evacuation route assignment
Algorithm Evacuation Congestion Total length of

time( T ) evacuation paths
(L)
Proposed Avg _ best 721.101 0.2731 34240.3
algorithm Avg _ mean 735.476 0.4353 34266.2
Avg _ worst 767.150 0.5252 34301.5
STD 9.2346 0.0863 36.07
DynDE Avg _ best 732.119† 0.2907† 34304.8†
Avg _ mean 740.167† 0.4410 34329.6
Avg _ worst 770.234† 0.5723† 34378.3†
STD 15.3028† 0.2057† 33.482‡
†indicates Proposed algorithm is significantly better than DynDE by the Wilcoxon signed-rank test at α = 0. 05.
‡means that DynDE is significantly better than our algorithm.
The convergence curve of the evacuation time and congestion is shown in Figure 4.
760 0.7
Proposed Algorithm Proposed Algorithm
DynDE 0.6 DynDE
750
Congestion
0.5
Time
740
0.4
730
0.3
720 0.2
0 50 100 150 200 0 50 100 150 200
Generations Generations
Fig. 4. The convergence curve of the evacuation time (left) and the convergence curve of the
congestion (right)
From Table 2 can be seen that the proposed algorithm is superior to DynDE for the
evacuation route assignment optimization. The satisfactory results are achieved by the
proposed algorithm. It has the shorter evacuation time than DynDE with the better
congestion. The total length of evacuation paths achieved by the proposed in the
optimization process is shorter than the DynDE’s.
It is clearly seen from Figure 4 that the proposed algorithm outperforms the
DynDE not only on the evacuation time and but also on the evacuation congestion.
6 Conclusions
Making a feasible and effective evacuation plan in the emergency situation is an

important issue. In this paper, a prediction based multi-strategy differential evolution
is adopted to solve this evacuation route assignment optimization problem based on
the evacuation model. The experimental results of the evacuation scenario of a large
public activity in Wuhan Sport Center in Wuhan City of China show that the
proposed algorithm can solve the real world complex dynamic optimization problems
such as the evacuation route assignment optimization problem. The analyses of the
results imply that the congestion value restricts the performance of an evacuation
plan. Thus, the congestion should be given a high priority.
Acknowledgments. This work was partially supported by the Scientific Research

Foundation for Introduced Excellent Scholars, China Three Gorges University (Grant
No. KJ2012B055).
References
1. Blackwell, T., Branke, J.: Multiswarms, exclusion, and anti-convergence in dynamic
environments. IEEE Transactions on Evolutionary Computation 10(4), 459–472 (2006)
2. Branke, J., Saliho, U. G., Lu, E., Uyar, C.S.: Towards an analysis of dynamic
environments. In: Proceedings of the 2005 Conference on Genetic and Evolutionary
Computation, pp. 1433–1440 (2005)
3. Branke, J.: Memory enhanced evolutionary algorithms for changing optimization
problems. In: Proceedings of the 1999 Congress on Evolutionary Computation,
Washington, DC, USA, pp. 1875–1882 (1999)
4. Mavrovouniotis, M., Yang, S.: Memory-based immigrants for ant colony optimization in
changing environments. In: EvoCOMPLEX, EvoGAMES, EvoIASP,
EvoINTELLIGENCE, EvoNUM, and EvoSTOC, EvoApplications, pp. 324–333. Springer
Verlag (2011)
5. Grefenstett, J.J.: Genetic algorithms for changing environments. In: Proceedings of the 2nd
Conference on Parallel Problem Solving from Nature, Brussels, Belg, pp. 137–144 (1992)
6. Yang, S.: Explicit Memory Schemes for Evolutionary Algorithms in Dynamic
Environments. In: Evolutionary Computation in Dynamic and Uncertain Environments,
vol. 51, pp. 3–28 (2007)
7. Novoa-Hernández, P., Corona, C., Pelta, D.: Efficient multi-swarm PSO algorithms for
dynamic environments. Memetic Computing 3(3), 163–174 (2011)
8. Mendes, R., Mohais, A.S.: DynDE: a differential evolution for dynamic optimization
problems. In: Proceedings of the 2005 IEEE Congress on Evolutionary Computation, pp.
2808–2815. IEEE Press (2005)
9. Blackwell, T.M., Bentley, P.: Dynamic Search with Charged Swarms. In: Proceedings of
the 2002 Genetic and Evolutionary Computation Conference, pp. 19–26. Morgan Kaufman
(2002)
10. Cobb, H.G.: An Investigation into the Use of Hypermutation as anAdaptive Operator in
Genetic Algorithms Having Continuous, Time-Dependent Nonstationary Environments
Naval Res. Lab., Washington, DC: Tech.Rep. AIC-90-001 (1990)
400 S. Wan
11. Shuzhen, W., Shengwu, X., Yi, L.: Prediction based multi-strategy differential evolution
algorithm for dynamic environments. In: 2012 IEEE Congress on Evolutionary
Computation (CEC), pp. 1–8. IEEE Press, Brisbane (2012)
12. Lämmel, G., Grether, D., Nagel, K.: The representation and implementation of time-
dependent inundation in large-scale microscopic evacuation simulations. Transportation
Research Part C: Emerging Technologies 18(1), 84–98 (2010)
13. Zheng, X., Zhong, T., Liu, M.: Modeling crowd evacuation of a building based on seven
methodological approaches. Build Environ. 44(3), 437–445 (2009)
14. Yamada, T.: A network flow approach to a city emergency evacuation planning. Int. J.
Syst. Sci. 27(10), 931–936 (1996)
15. Qiuping, L., Zhixiang, F., Qingquan, L., Xinlu, Z.: Multiobjective evacuation route
assignment model based on genetic algorithm. In: 18th International Conference on
Geoinformatics, Beijing, pp. 1–5. (2010)
16. Eppstein, D.: Finding the k shortest paths. In: Proceedings of 35th Annual Symposium on
Foundations of Computer Science, Santa Fe, NM, pp. 154–165, (1994)
17. Mendes, R., Mohais, A.S.: DynDE: a differential evolution for dynamic optimization
problems. In: Proceedings of the 2005 IEEE Congress on Evolutionary Computation,
vol. 3, pp. 2808–2815. IEEE Press (2005)
Centralized Charging Strategies of Plug-in Electric
Vehicles on Spot Pricing Based on a Hybrid PSO
Jiabao Wang1, Qi Kang1, Hongjun Tian1, Lei Wang1,2, and Qidi Wu1
1
Department of Control Science and Engineering, Tongji University, Shanghai, China
wangjiabao0316@163.com, qkang@tongji.edu.cn
2
Shanghai Key Laboratory of Financial Information Technology, Shanghai, China
wanglei@tongji.edu.cn
Abstract. This work proposes an efficient charging regulation strategy based on

optimal charging priority and location of plug-in electric vehicles (PEVs). It
employs a hybrid particle swarm optimization for optimal charging priority and
location of PEVs in distribution networks, with the objectives of minimization
of charging cost, power loss reduction and voltage profile improvement. The
algorithm is executed on IEEE 30-bus test system. The results are compared
with those that are gained by executing sample genetic algorithm (SGA) with
diverse parameters on the same system. The results indicate the effectiveness
and promising application of the proposed methodology.
Keywords: PEVs, spot pricing, centralized charging strategy, HPSO, optimal

charging priority and location.
1 Introduction
Nowadays, more and more vehicles are on roads, thereby increasing the consumption
of fossil fuel. Consequently, the environment is being seriously polluted. Under this
circumstance, many countries have proposed their energy policies with objectives of
economic effectiveness improvement, achievement of energy security, and
environment pollution reduction, which promotes electrification of transportation and,
especially, the rapid development of plug-in electric vehicle (PEV) industry [1].
However, PEVs, a new kind of power load, would exert a tremendous influence on
the daily residential load curve of distribution network if they widely connected to the
power grid for battery charging [2]. Due to the uncertainty of their charging
behaviors, uncoordinated random charging of PEVs may lead to unforeseen effect on
normal operation of distribution system, such as aggravating the load peak and off-
peak difference in network, etc. Meanwhile, taking the spot pricing into consideration,
the owners of PEVs may afford much higher cost for battery charging. Therefore, the
appropriate dispatch of PEVs in a distribution system will be a challenging demand
side management (DSM) [3]. Fortunately, PEVs are more flexible than traditional
load, because majority of PEVs owners usually return home early in the evening and
have no request for the special time that their vehicle will be charged, as long as the
402 J. Wang et al.
batteries are full by the next morning [4]. According to the statistical data of National
Household Travel Survey (NHTS), more than 90% of vehicles are parked at home
between 9 P.M and 6 A.M [4]. Taking this opportunity into account, several
centralized charging strategies of PEVs have been researched for utilizing less
expensive electricity and shifting the PEVs load to off-peak hours. For example,
under a spot pricing based on electricity market environment, a demand side response
based charging strategy is proposed, and a dynamic estimation interpolation based
algorithm is designed to optimize the mathematical model which is established taking
into account the valley-filling effect of supply side and the users’ cost [1]. Wu et al.
[4] proposes a novel minimization of charging cost based heuristic approach by
analyses of PEVs travel pattern and spot pricing. The results show that the strategy
can lower the peak-valley difference and save users’ cost effectively, but they ignore
the influence of PEVs charging behavior on power qualities, such as power loss,
voltage fluctuation, etc. Deilami et al. [5] proposes a real-time smart load
management control strategy which is developed for the coordination of PEVs
charging based on real-time minimization of total cost of generating the energy plus
the associated grid energy losses. The results indicate that the approach can reduce the
power losses and improve the voltage profile by considering the maximum
sensitivities selection optimization based priority charging. Lan et al. [6] presents a
nonlinear electric vehicle (EV) battery model, and a dynamic programming-based
algorithm for optimizing an EV’s charging schedule with given electricity price and
driving pattern. But only one EV’s charging schedule is researched here.
In this paper, taking advantage of the flexibility of the PEVs, we arrange them to
charge at the relatively inexpensive electricity which occurs during off-peak hours at
night. An efficient charging regulation based on optimal charging priority and
location of PEVs is proposed under a spot pricing based electricity market
environment. And then, PSO-GA [7], a hybrid particle swarm optimization by
incorporating genetic algorithm, hereinafter to be referred as HPSO, is used for
optimal charging priority and location of PEVs in distribution networks system. We
take power quality and economy objectives into account to define the optimization
objectives in this paper, including power losses, voltage profile and charging cost.
The proposed approach is executed on the IEEE 30-bus test system.
2 Problem Formulation
Under a spot pricing based electricity market environment, uncoordinated charging of

many PEVs may exert negative impacts on the security and economy of power system
operation, such as power losses increment, overload, voltage fluctuation aggravation
and charging cost increment. To minimize such impacts, we may proceed as follows:
1) Different locations for PEVs charging may have different influences on power
quality. Appropriate locations for PEVs charging benefit power quality improvement.
2) In order to abate the vacancy of electric power, PEVs are scheduled for charging
during the off-peak period of the day.
Centralized Charging Strategies of Plug-in Electric Vehicles 403
3) The charging cost should be as low as possible. Because different periods of the
day correspond to different electricity price under a spot pricing based electricity
market environment, if all PEVs owners had a preference for the exact time of which
electricity price is lowest, a new peak demand would occur. Hence, maximum power
consumption should be set for every time slot in order to avoid overload. Under this
circumstance, PEVs charging priority that determined a sequence of PEV choosing
charging slots impacts on daily residential load curve and total cost of electricity
heavily.
2.1 Charging Rule
In this paper, the maximum demand level has been defined as the maximal value of
residential load during a scheduling period. The power for PEV charging is
constrained by:
P , 1,2, … . . , T . (1)
P P P . (2)
where i and T are the time slot number and total number of slots, P is the
maximum residential load demand level with PEVs being charged, P is the total
residential power consumption at the th time slot without PEVs plug in, P is
the maximum permissible power consumption for PEVs charging at ith time slot,
is the total power consumption for PEVs charging at th time slot.
If the PEVs charging priority and locations are known, load scheduling is
transformed into a PEVs charging rule. The basic idea is to charge each vehicle in the
time slots where the lowest electricity price occur and power consumption meet the
Eq. (1).
The flowchart of PEVs charging rule is shown as Fig.1. And the relevant
parameters can be define as follow: is the charging priority number of th
PEV , is the charging location number of th PEV , Duration is the number
of time slots which th PEV need for charging , n is the total number of PEVs, Price
is the electricity price at th time slot, P is the maximum permissible power
consumption for PEVs charging at th time slot, T is the total number of time slots,
is the serial number of PEV whose priority rank is k, is the serial number of
slot whose price rank is l, P is the rated power of PEV, is the total power
consumption of PEVs at th time slot and th charging node, N is the number of
locations, is the set of time slots where th PEV is being charged.
2.2 Objective Outline
The objective of problem model includes minimization of charging cost, power loss
reduction, voltage profile improvement. Therefore, a three-fold objective function is
given by
404 J. Wang et al.
OBJ ： f min σ γ . (3)
where is active power loss, σ and γ are the non-negative weighting factor
used to indicate the relative important of three items, (here σ=20, γ=0.01),
denotes load bus voltage deviations from 1.0 per unit, is the total electricity
cost.
can be obtained with power flow calculation and is represented as
∑T ∑L | | , where T is the number of slots, L is the number of lines
in the power system, is the current of th line at ith time slot, is the resistance
of th line.
Fig. 1. The flowchart of PEVs charging rule

can be denoted as ∑T ∑N 1.0 p. u. , where T plays the

same role in , N is the number of buses in power system, and is node
voltage p. u. of th line at th time slot [8].
is denoted as ∑T P price ∆T, where T plays the same role in
, P is the power consumption of PEVs at th time slot, price is the electricity
price at th time slot, ∆T is time span of a time slot.
3 Scheduling Algorithms
3.1 HPSO: An Improved Particle Swarm Optimization
Particle swarm optimization (PSO) was proposed by Kennedy and Eberhart in 1995.
In [9], dynamic weights are integrated into a standard PSO to improve its global and
local convergence.
HPSO [7] discards the method with which particles update their positions by
tracking the individual and group optimal position in PSO. Instead, it introduces the
crossover and mutation operation of a genetic algorithm (GA) into
PSO. Its new particles are refreshed by crossover and mutation operators
according to optimal solutions of entire population and individual.
The steps of HPSO are as follows:
Step 1. Set the optimal position of individual ( , ) as the initial position of th
particle ( =1, 2,…, M), where M is the population size. Select the best one among
, ,…, ,M as optimal position of population ( ).
Step 2. Calculate the objective function values of all members.
Step 3. Update the , and according to the fitness of all members.
Step 4. Execute a crossover operation with , and . Update the position
of the th particle by executing a crossover operation with , and ,
respectively.
Step 5. Perform a mutation operation on each particle in the swarm.
Step 6. If the termination condition is met, stop. Otherwise, return to Step 2.
4 Simulation and Result
4.1 Simulation Data
In order to explain the problem more specifically and clearly, some assumptions are
given and listed below:
1) All information of EVs and control signals generated by aggregators can be
delivered immediately between EVs and aggregators [6].
2) PEV battery capacities typically range from a few kWh to over 50 kWh [5]. The
capacity and rated power of PEV can be defined as 50kWh and 10kW, respectively.
3) The number of charging time slots of PEV is assumed to follow
approximately Gaussian distribution whose mean value and standard deviation equal
to 20 and 5 respectively.
406 J. Wang et al.
4) 1000 PEVs are scheduled as a whole, thereby they can be treated as a PEV set,
hereinafter still referred to as “PEV”. And all vehicle batteries in PEV have same state
of charge (SOC).
In this paper, PEVs are dispatched for charging during off-peak hours from 9 P.M.
to 7 A.M., because most vehicles are vacant and the electricity price is generally low
during this period. The daily residential load curve [5] shows in Fig.2. Furthermore,
other relevant data are given as follow:
1) The total time for charging dispatch is segmented into 40 time slots, where each
time slot has a duration of 15 minutes.
2) Data of spot pricing of electricity is released by Long Island, New York on Jan.
1, 2010 [2]. The value of electricity price ($/MW) during off-peak hours from 9 P.M.
to 7 A.M is given as {71.50, 71.16, 63.46, 58.86, 62.67, 39.84 44.78, 53.03, 65.34,
57.82}.
3) The total number of PEVs which participate in scheduling is 50.
4) The distribution system used for simulation and analysis of PEVs ordered
charging strategy in this paper is the IEEE 30-bus test system [10].
Fig. 2. Daily residential load curve Fig. 3. Optimal fitness dynamics of system
IEEE 30-bus test system [11] is adopted for simulation and analysis in this work. In
this paper, 10 buses, i.e., buses 3-12, are chosen as PEVs charging nodes. The HPSO-
based method scheme can be described as follows, which is used as a solver of this
optimization problem of charging priority and location of PEVs.
Step 1. Initialization
Set the time count t=0, dimension D (here D=50), maximum iteration number
Iter , and population size M. The current position of all members is 0
, … , , … , M , where
, , … ,
… (i=1,2,…, M).
, , ,
represents the coding scheme of th particle at th iteration, , (j=1,2,…, D) is
the charging priority index of th PEV of th particle at th iteration,
, (j=1,2,…, D) is the charging location (bus index) of th PEV of th particle at th
iteration. The first line of is a random permutation of integers 1~50 and , is
a random integer in {3,4,…,12}.
PEVs charging rule and power flow calculation are applied to all members, by the
results of which the fitness of value of is obtained. For each individual,
set , = and , i=1,2,…, M . Select the best one among
, ,…, ,M as , and set .
Step 2. Update the time counter.
t=t+1
Step 3. Execute crossover operation.
(1) Execute a crossover operation with ,
Update the position of (i=1,2,…, M) by executing crossover operation with
, . If < Pbest , then update individual best as and set
, = .
(2) Execute a crossover operation with
Update the position of (i=1,2,…, M) by executing crossover operation with
. If < , then update individual best as and set
, = .
Step 4. Execute mutation operation
Execute a mutation operation by exchanging any two elements of first line of
and change corresponding element of second line using a random integer in
{3,4,...,12}. If < , and then update individual best as
and set , = .
Step 5. Carry out fitness evaluation and update individual best and population best.
For all members, PEVs charging rule and power flow calculation are applied, and
then evaluates the fitness of every member (i=1,2,…, M) according to Eq.(3). If
< , and then update individual best as and set , =
. Select the best one among , ,…, ,M as g , and set
.
Step 6. Check the stopping criteria
If Iter , then go to Step 2; Else, stop the algorithm.
4.2 Parameters Setting
The parameters setting of SGA and HPSO are given in Table 1. In SGA•aa population
of M solutions is maintained and two probability-based operations, i.e., crossover
operator and mutation operator are employed. Whether the two operators work
depends on crossover rate Rate and mutation rate Rate , respectively. Meanwhile,
the SGAs with different crossover and mutation parameters are denoted as SGAa,
SGAb and SGAc.
Table 1. The parameters setting of SGA and HPSO
a M=20; Iter =300; Rate =0.8; Rate =0.05.

SGA b M=20; Iter =300; Rate =0.9; Rate =0.005.
c M=20; Iter =300; Rate =0.7; Rate =0.025.
HPSO M=20; Iter =300.
408 J. Wang et al.
By executing two types of algorithms 20 times separately, the Monte Carlo

simulation results are shown in Table 2, where iteration represents the number of
iterations. The evolutionary trajectory of four algorithms is shown in Fig.3.
The simulation results show that the HPSO outperforms SGA on the metrics in
P , Cost , and fitness value. Further, HPSO have better global search ability.
However, the superiority of HPSO in searching the metric V is not outstanding.
The proposed PEVs charging rule is applied to the optimal individual sought out by
HPSO, and yields the power consumption of every time slot. Fig.4 shows the impact
of HPSO-based coordinated PEVs charging on total system power demand.
Compared with the results of uncoordinated PEVs charging, HPSO-based coordinated
PEVs charging strategy effectively shift electricity use from on-peak to off-peak period.
Fig. 4. The impact of HPSO-based Fig. 5. The impact of uncoordinated PEVs

coordinated PEVs charging on total system charging on total system power demand
power demand
The impact of uncoordinated PEVs charging on total system power demand is

shown in Fig.5, in which many peak power demands appear. With
the increase of the gap between peak load and valley load in power system, it is
uneconomic to install units considering the capacity of peak load. Besides, frequent
start and stop are detrimental to the life of generators.
The impact of HPSO-based coordinated PEVs charging is further compared with
Min-Cost Load Scheduling (MCLS) approach [4]. The basic idea of MCLS is that
each vehicle is charged in the time slots where the lowest electricity price occurs.
Fig.6 shows that the impact of MCLS-based PEVs charging on total system power
demand. The minimum electricity cost can be achieved by carrying out the MCLS
approach, but the impact of PEVs charging on system power loss and voltage
deviation is ignored by MCLS approach. Besides, as we can see in Fig.6, there are
still three peak loads and large gap between peak load and valley load. HPSO-based
voltage (p.u.) of 30 buses at worst time slot is compared with MCSL-based voltage of
30 buses at random time slot and stochastic charging nodes, as shown in Fig.7. The
HPSO-based coordinated PEVs charging improve the security and reliability of
networked distribution network by minimizing voltage deviations, overloads, and
power losses that would otherwise be impaired by MCSL-based PEVs charging.
Table 2. Simulation Results

Metric SGAa SGAb SGAc HPSO
Optimum value 1374.8 1453.7 1402.4 1331.7
Worst value 1480.5 1518.5 1461.6 1363.5
Median value 1439.0 1474.1 1455.6 1336.8
Mean value 1432.2 1478.1 1437.9 1344.5
Standard deviation 41.7753 21.5690 26.3989 12.3578
Optimum value 132.9018 133.6457 133.6895 133.9468
Worst value 136.2426 136.6502 136.5378 135.3387
$ 10 Median value 135.4281 134.9455 135.1636 135.2335
Mean value 134.8389 135.2212 135.1725 134.8535
Optimum value 38.1942 40.3422 38.9862 37.3199
Worst value 40.8019 41.5776 40.5185 38.0455
Median value 39.2931 40.5403 40.1329 37.6670
Mean value 39.6493 40.6993 39.8711 37.6834
Optimum value 3501.9 3606.8 3536.9 3429.5
Worst value 3648.9 3716.5 3628.3 3477.3
Median value 3585.0 3630.4 3606.7 3440.1
Mean value 3573.5 3644.3 3587.0 3457.0
Minimum value 172 82 123 252
Maximum value 281 275 287 286
Median value 269 245 270 274
Mean value 249.8 218.6 244.8 269.4
Fig. 6. The impact of MCSL-based Fig. 7. Voltage deviation at 30 buses

coordinated PEVs charging on total system
power demand
410 J. Wang et al.
5 Conclusion
This paper presents a centralized charging model of PEVs under a spot pricing-based
electricity market environment. All PEVs taking part in centralized charging
scheduling are equally divided into multiply, and an efficient charging rule which is
based on charging priority and location of PEVs is proposed in the model. Further, in
this paper, HPSO is employed to optimally determine the PEVs charging priority and
location to be plugged in distribution system. Combining the HPSO with Newton
method of power flow calculation can solve the problem of optimal charging priority
and location of PEVs well with the objective of minimization of charging cost, power
loss reduction and voltage profile improvement. The results are compared with those
are obtained using SGA, and validate the superiority and effectiveness of the
approach.
Under the intelligent centralized charging strategy, PEV owners are able to pay the
electricity bills only in line with the number of charging time slots they use. The
electricity costs resulting from the use of the proposed approach are slightly higher
than those by MCLS algorithm whose objective is the maximization of energy trading
profits, but it effectively reduces the gap between peak and valley
load and simultaneously improves the power quality. In this case, a discount
electricity price scheme can be introduced by power supplier in order to encourage
PEV owners to take part in centralized charging mechanism.
The further research will be oriented to dispatching PEVs for charging under the
assumption of stochastic vehicles’ arrival, with the same objectives as discussed in
this work.
Acknowledgments. This work was supported in part by the Natural Science

Foundation of China (71371142, 61005090, 61034004), the Program for New
Century Excellent Talents in University of Ministry of Education of China. Ministry
of Education (NCET-10-0633), the Fundamental Research Funds for the Central
Universities, and the Research Fund of State Key Lab. of Management and Control
for Complex systems.
References
1. Zhao, J., Wen, F., Yang, A., Xin, J.: Impacts of electric vehicles on power systems as well
as the associated dispatching and control problem. Automation of Electric Power
Systems 14, 2–9 (2011)
2. Zou, W., Wu, F., Liu, Z.: Centralized charging strategies of plug-in hybrid electric vehicles
under electricity markets based on spot pricing. Dianli Xitong Zidonghua (Automation of
Electric Power Systems) 35(14), 62–67 (2011)
3. Masoum, A.S., Deilami, S., Moses, P.S., Masoum, M.A.S., Abu-Siada, A.: Smart load
management of plug-in electric vehicles in distribution and residential networks with
charging stations for peak shaving and loss minimisation considering voltage regulation.
Generation, Transmission & Distribution, IET 5(8), 877–888 (2011)
4. Wu, D., Aliprantis, D.C., Ying, L.: Load scheduling and dispatch for aggregators of plug-
in electric vehicles. IEEE Transactions on Smart Grid 3(1), 368–376 (2012)
5. Deilami, S., Masoum, A.S., Moses, P.S., Masoum, M.A.: Real-time coordination of plug-
in electric vehicle charging in smart grids to minimize power losses and improve voltage
profile. IEEE Transactions on Smart Grid 2(3), 456–467 (2011)
6. Lan, T., Hu, J., Kang, Q., Si, C., Wang, L., Wu, Q.: Optimal control of an electric
vehicle’s charging schedule under electricity markets. Neural Computing and
Applications 23(7-8), 1865–1872 (2013)
7. Min, X.I.E.: An Improved Hybrid Particle Swarm Optimization Algorithm for TSP.
Journal of Taiyuan University of Technology 4, 023 (2013)
8. Kang, Q., Lan, T., Yan, Y., Wang, L., Wu, Q.: Group search optimizer based optimal
location and capacity of distributed generations. Neurocomputing 78(1), 55–63 (2012)
9. Jian-Hua, L., Rong-Hua, Y., Shui-Hua, S.: The analysis of binary particle swarm
optimization. Journal of Nanjing University (Natural Sciences) 5, 003 (2011)
10. AlRashidi, M.R., El-Hawary, M.E.: Hybrid particle swarm optimization approach for
solving the discrete OPF problem considering the valve loading effects. IEEE Transactions
on Power Systems 22(4), 2030–2038 (2007)
11. AlRashidi, M.R., El-Hawary, M.E.: Hybrid particle swarm optimization approach for
solving the discrete OPF problem considering the valve loading effects. IEEE Transactions
on Power Systems 22(4), 2030–2038 (2007)
A New Multi-region Modified Wind Driven Optimization
Algorithm with Collision Avoidance
for Dynamic Environments
Abdennour Boulesnane1,2 and Souham Meshoul1

1
Computer Science Department, Constantine 2 University, Algeria
2
MISC Laboratory, Constantine, Algeria
abdennour.boulesnane@gmail.com,
souham.meshoul@univ-constantine2.dz
Abstract. This paper describes a new approach to deal with dynamic

optimization that uses a multi-population. Its main features include the use of a
modified wind driven optimization algorithm that aims to foster impact of
pressure on velocities of particles. Moreover, a concept of multi-region inspired
from meteorology has been introduced along with a new collision avoidance
technique to maintain good diversity while preventing collision between sub-
populations. The method has been assessed using Moving Peaks Benchmark
and compared to state of the art methods. Preliminary results are very
encouraging and show viability of the method.
Keywords: Dynamic optimization, Swarm intelligence, Wind driven

optimization, collision, multiple population methods, Moving Peaks Benchmark.
1 Introduction
In everyday life and in almost all domains, each type of optimization problem has
features that make it different from the others. However, these problems usually have
a common property that is their dynamic nature. Such problems are difficult to solve
because the challenge is not only to locate global optima, but also to track them in
environments that change over time. Therefore, a crucial requirement a dynamic
optimization algorithm should fulfill is to achieve a balance between exploitation and
exploration of the search space to handle optimization over time. This requires
fostering diversity while ensuring very fast convergence to global optima throughout
the search process because the time between two successive changes may be
insufficient to converge and to follow optima at the same time. Moreover, dynamic
optimization is faced to the challenge to solve both issues of outdated memory due to
changes in environment and diversity loss due to traps of local optima. Outdated
memory problem is usually solved by clearing the memory when a change is detected
however the matter is what to do with the knowledge acquired once a change in the
environment occurs: should it be reused for next changes or discarded? In [1], a study
showed that the reuse of information lead to faster adaptation to changes, and thus, to
A New Multi-region Modified Wind Driven Optimization Algorithm 413
better solutions. Many algorithms have been proposed to address dynamic

optimization problems (DOPs). A comprehensive survey can be found in [2]. More
particularly, using multiple populations has been shown to be a suitable way to keep
up with changes over time [7,10].
Recently, a new swarm based metaheuristic inspired from atmospheric motion has
been proposed in [3] and termed as Wind Driven Optimization (WDO). WDO’s
model is based on the definition of trajectories of small air parcels within the earth
atmosphere according to the Newton’s second law of motion. WDO has been
designed for static optimization and applied to electromagnetics optimization
problems. In this paper, we propose investigating the potential of WDO to solve
dynamic optimization. Intuitively, the motivation can be explained by the fact that the
movement of air from high pressure zones to low pressure zones at velocities
proportional to pressure gradient force would lead to simulation models that can be
used in dynamic environments. Taking inspiration from WDO, a modified version is
developed and used within a multi-population framework with the aim to adaptively
detect promising regions in the search space. Therefore, we refer to the developed
approach as Multi-Region Modified WDO (MR-MWDO). Furthermore, in order to
maintain several sub-populations on several peaks (promising regions) and to avoid
collisions between sub-populations a new strategy called Collision Avoidance
Technique is introduced.
The rest of the paper is organized into four sections. Section 2 presents the WDO
algorithm. The proposed Multi-Region Modified WDO is described in section 3.
Section 4 is devoted to the experimental study. Finally, conclusions and perspectives
are given in section 5.
2 Wind Driven Optimization
Recently, a new approach to deal with multi-dimensional and multi-modal

optimization problems has been proposed by Bayraktar [3] and termed as Wind
Driven Optimization. As the name suggests, WDO is inspired by the earth’s
atmosphere in the Troposphere layer and more specifically by the contribution of
wind in the equalization of horizontal imbalances in the air pressure. In his study,
Bayraktar [3] used the physical equations that govern atmospheric motion. This later
is generally described by the movement of air which is a consequence of pressure
gradient due to temperature differences. It is observed that wind blows from a high
pressure zone to low pressure zone with a velocity proportional to the pressure
gradient force. For in-depth insight into the physical model, the reader can refer to [4].
Starting from the Lagrangian model that describes atmospheric motion, Bayraktar
derived WDO as an iterative metaheuristic. WDO’s dynamics is similar to that of
Particle Swarm Optimization (PSO). Particles in WDO refer to small air parcels that
are assumed dimensionless and weightless for simplification. The trajectories of these
parcels are defined according to the Newton’s second law of motion. Like PSO, these
air parcels are described by a position and a velocity that refer to a candidate solution
and the amount of position displacement respectively. The pressure at each air parcel
414 A. Boulesnane and S. Meshoul
is used as information about the related solution quality. Updating positions and
velocities of air parcels is governed by the following equations where the variable i
refers to the particle and the variable t to iteration [3].
1 1 (1)
(2)
Where and are the current and the new velocity of the air parcel
respectively, is the global best position, and are the current and the new
positions of the air parcel, parameters , , , and T are related respectively to the
friction coefficient, gravity, universal gas constant and temperature in the physical
model. The variable represents the rank of the air parcel where all air parcels are
ranked in descending order based on their pressure. An in-depth description WDO is
available at [3].
3 The Proposed Algorithm for Dynamic Optimization
For sake of clarity, we first describe the modifications brought to WDO to properly
handle optimization in dynamic environments then we present the proposed MR-
MWDO for dynamic optimization.
3.1 Modified WDO Algorithm

In nature, atmospheric pressure is influenced by several factors such as temperature,
humidity and altitude among others. These factors contribute in raising and lowering
the air pressure according to the spatial distribution and topography on the surface of
the earth. This leads to the formation of high pressure regions and low pressure
regions and the gradient of the pressure force gives rise to air motion. As shown in
equation (1) of WDO model, the influence of pressure on velocity of air parcels is not
expressed in terms of actual values of pressure but is implicitly represented in the
third and fourth terms by the rank of the particle among other particles based on their
fitness values. By another side, it is known that altitude is inversely proportional to
pressure. Moreover, one natural way to establish analogy between optimization and
atmospheric models is to relate altitude to fitness. Therefore, the matter is how to
express in a convenient manner the relationship between pressure and fitness. In [5],
Chao et al. proposed a Tropical Cyclone-based Method (TCM), for solving global
optimization problems with box constraints. Chao and al expressed the relationship
between the pressure pi of a particle i and its fitness f(xi) by the following equation:
∑
(3)
Where represents the fitness of the worst position, n is the problem

dimension and m the number of particles. According to equation (3), a particle with a
higher fitness possesses a lower pressure, and the pressure is scaled to be 1 at .
Therefore, in the modified WDO we propose to use the value of pressure as given by
equation (3) instead by the rank of the particle to better reflect the influence of
pressure on velocities. In other words, parameter r in equation (1) is replaced by
parameter .
3.2 Features for Handling Dynamic Optimization
3.2.1 Concepts of Multi-region

Multi-population methods maintain multiple sub-populations concurrently. Each sub-
population may handle a separate area of the search space. Each of them may also
handle a specific task. For example, some sub-populations may focus on searching for
the global optimum while some others may concentrate on tracking any possible
changes. These two types of populations then may communicate with each other to
bias the search [6].
In meteorology, a region is set as a low-pressure area (respectively a high-pressure
area) if the atmospheric pressure is lower (respectively higher) than that of
surrounding locations. Wind blows in an attempt to equalize horizontal imbalances in
the air pressure. Meteorologists use the value of pressure at sea level as a threshold to
decide the type of region: if the pressure values in a region are lower than 1013.25,
then the region is low pressure else it is high pressure. The idea behind MR-MWDO
is then classifying regions of a search space into low and high pressure regions could
be used to guide the search process and better identify promising areas that may
include optima.
In the proposed Multi-Region MWDO algorithm, this principle is implemented
through the use of two different types of sub-populations with different numbers of
particles. The aim is to better control the level of diversity inside and outside sub-
populations knowing that after occurrence of a change in dynamic environment, the
new optimum can be located near or very far compared to the previous optimum. The
first type is a sub-population of observer particles (SOP), which role is exploring and
calculating pressure at various search areas. The second type is a sub-population of
air particles (SAP) that navigate within the search space according to MWDO
equations. For example, in the case of a maximization problem, a new low pressure
region is detected (a low pressure region in this case represents a promising area) if
the best value of pressure in SOP is less than or equal to a threshold value and all SAP
are converged. The threshold value is the median of all pressure values of all
particles.
When a new low pressure region is found, a new SAP is generated within this
region. Accordingly, a radius is assigned to each SAP in order to avoid the re-
exploration of already visited regions by SOP.
3.2.2 Collision Avoidance Technique

In order to ensure that each promising region (each peak) is explored by only one sub-
population, researchers proposed to delimit regions using a radius. For example,
Blackwell et al [7] proposed the technique of exclusion which consists in using a
simple competition between the swarms that are close to each other. The swarm with
the best function value is kept (best solution found so far) whereas the other will be
expelled and reinitialized. Two swarms collide if and only if their attractors (best
solutions found so far by each swarm) are located within an exclusion radius of
each others [7]. Within the same spirit, we introduce a new technique called Collision
Avoidance Technique to prevent collision between sub-populations and therefore
maintain several sub-populations on several peaks. For each sub-population, the
distance to the closest other sub-population is recorded. To determine whether this
proximity between these sub-populations is desirable or not, the midpoint of the
recorded distance is considered. Figure 1 shows two cases where proximity between
subpopulations is desirable and one case where it is undesirable.
Fig. 1. Proximity between sub-populations: plots (a) and (b) show cases of desirable
proximities between two SAPs related to different peaks. In plot (c), the proximity between the
two SAPs is undesirable (both SAPs are related to the same peak).
From Figure 1, we can identify two major cases depending on the relative fitness
of the midpoint compared to the best solutions in the two SAPs; the first case is when
this value is less than the best solutions found by SAPs Figure 1.a. In this case, both
SAPs will be kept. While in the second case, if this value is better than one of them as
in Figure 1.b and 1.c, a new position of the midpoint must be calculated iteratively
between the previous midpoint and the weakest SAP’s solution. At each iteration, if
the fitness of the new midpoint is worse than the weakest SAP’s solution the iterative
process is stopped and both SAPs will be retained Figure 1.b. Otherwise, the process
will pursue until the distance between the midpoint and the lowest SAP is less than
. At this stage, if the fitness of the midpoint remains better than the weakest
SAP’s solution, then this SAP is removed to avoid the collision as in Figure 1.c.
3.3 Outline of the proposed Multi-region MWDO Algorithm
The algorithm starts with single sub-population of n observer particles (SOP) with the
aim to explore only the search space and to find promising regions. Then the
algorithm goes through an iterative process as described on Figure 2. For each

generated SAPi the convergence test is performed in the following manner:
For each SAPi
If best value of pressure is better than threshold
& not Convergedi then
Convergedi=true;
Elseif best value of pressure is lower than threshold
& Convergedi then
Convergedi=false;
Endif
Endfor
To ensure a reasonable number of sub-populations in the search space without

degrading the performance of the algorithm and performing unnecessary functions
evaluations knowing that the number of evaluation should be limited because the time
of the next change is unknown, two successive phases are applied. The first phase is
to generate a new sub-population (SAP) when a new promising region is detected as
seen previously, while in the second phase, a sub-population with low pressures or
being in collision with another one is removed.
Fig. 2. Flow chart of proposed algorithm
The phase of change test is a significant step as it allows the algorithm to adapt to a
new change in the environment by comparing, at each iteration, the best solutions
found by the sub-populations (SAPs) to their best old achieved values. Then in the
next phase, particles positions are updated using modified equations (1) and (2) for
the sub-population of air particles (SAP) and random displacement for sub-population
of observer particles (SOP). Once new positions are recorded, Collision Avoidance
procedure is performed to keep several sub-populations on different peaks as
described above. The algorithm evolves in this manner till a termination criterion is
satisfied.
In order to assess the performance of MR-MWDO, Moving Peaks Benchmark (MPB)

has been used [8]. All experiments have been performed using the Scenario 2 of the
MPB, proposed by Branke in [8]. This scenario has also been used by several authors
and allows comparison of results obtained by different methods. The settings for this
scenario are as follows: the search space has five dimensions 0,100 there are
10 peaks, the peak heights vary randomly in the interval 30,70 and the peak
width parameters vary randomly within 1,12 . The peaks change position every
5000 evaluations by a distance of 1 in a random direction, and their
movements are uncorrelated (the MPB coefficient 0). The algorithm was run for
100 consecutive changes, and each run was independently repeated 30 times, with
different random seeds. The performance measure used to evaluate the algorithm is
the offline error [9] given by the following equation:
∑ ∑ (4)
Where is the total number of changes in the objective function, is the

number of evaluations performed during the change, is the value of the
global optimum in the change and is the value of the best solution
obtained by the algorithm in the evaluation of change.
First experiment has been conducted to find suitable settings of the algorithm’s
parameters that consist in the size of a SOP (n), the size of a SAP (m), the radius
assigned to each SAP to avoid collisions and parameters used for the update
of air particles’ positions that is , , , seen in the equation (1). These latter
have been set to 1, 0, 0.4, 3 . The results of this experiment are
shown on tables 1 and 2.
Table 1. Offline error and Standard deviation for varying particles number. The data is for 30
runs of MPB (Scenario 2) and =5.0.
Offline error (Std error) Offline error (Std error)

5 5 6.19 0.07 10 5 4.92 0.09
5 10 2.26 0.08 10 10 3.76 0.10
5 20 1.78 0.04 10 20 2.66 0.08
5 30 1.92 0.03 10 30 2.23 0.06
5 50 2.13 0.03 10 50 1.85 0.03
Table 2. Offline error and Standard deviation depending on parameter
Offline error (Std) Offline error (Std)

0.5 1.91 0.03 5.0 1.78 0.04
1.0 1.88 0.04 8.0 1.83 0.04
3.0 1.79 0.03 10.0 1.84 0.04
As can be seen, the algorithm achieves the best performance with

and . . With these settings, the algorithm was found to converge
quickly after each change while maintaining good diversity level within and outside
sub-populations. Furthermore, it has been observed that sub-populations work within
their regions delimited by radius without collision problem.
4.1 Impact of the number of Sub-populations

The number used of sub-populations plays an important role in maintaining good
diversity and occupying the promising areas that can include global optima after each
change. However, using large numbers of sub-populations impacts negatively the
performance of the algorithm by making unnecessary evaluations of the objective
function. Figure 3 shows the total number of sub-populations during a single run of
the 10 peaks MPB environment with 5 20 and 5.0. As can
be seen, the algorithm was able to maintain the appropriate number of sub-
populations (SAP) on different peaks (10 peaks in this case), which is achieved at
35000 evaluations while improving the offline error.
Fig. 3. Total number of sub-populations for a single instance of the 10 peaks MPB
environment. Upper plot shows offline error, lower plot shows number of converged sub-
populations (SAPs) and explorer sub-population (SOP).
4.2 Comparison with Other Algorithms

In order to compare MR-MWDO algorithm to other algorithms from the literature, an
experiment was performed during which the offline error and its standard deviation
were recorded for different shift severities (s) and different number of peaks (p) for all
algorithms namely: HmSO [10], CPSO [11], mQSO [7], SPSO [12] and CPSOR [13].
Results are reported on tables 3 and 4. The shift length s is the severity of the problem
dynamics. Whenever it increases, localization and tracking of peaks becomes more
difficult.
Table 3. Offline error Standard error for different algorithms on the MPB problem with
different shift severities
MR-MWDO CPSOR HmSO CPSO mQSO SPSO

0 1.01 1.57 2.68 0.899 0.601 0.807
0.04 1.23 1.26 1.05 0.439 0.972
1 1.78 . 3.49 1.35 1.69 2.28
0.04 0.626 1.65 1.12 0.784 1.49
2 2.53 . 3.88 1.31 1.84 2.81
0.03 0.654 1.66 1.15 0.723 1.54
3 3.57 . 4.22 1.41 2.06 3.24
0.03 0.70 1.74 1.08 0.775 1.34
Table 4. Offline error Standard error for different algorithms on the MPB problem with
different numbers of peaks
MR-MWDO CPSOR HmSO CPSO mQSO SPSO

1 1.48 0.245 1.80 . 3.15 1.86
0.06 0.253 2.09 5.9e 4 1.21 1.45
5 1.46 . 3.99 0.97 1.56 1.52
0.06 0.788 1.92 0.755 0.691 0.769
10 1.78 . 3.49 1.35 1.69 2.28
0.04 0.626 1.65 1.12 0.784 1.49
100 2.60 2.19 3.50 . 3.39 3.50
0.03 1.11 0.953 0.752 1.53 1.47
From the results shown above, we can see that MR-MWDO achieves better results
than SPSO, HmSO and mQSO for all numbers of peaks except for number of peaks
p=10 where mQSO achieves slightly better values with highest standard deviation. In
general, MR-MWDO achieves intermediate results compared to the other algorithms.
The best results are almost shared between CPSOR and CPSO using a hierarchical
clustering method to locate and track multiple peaks. These results are very promising
and show viability of the proposed approach.
5 Conclusion
In this paper, we described a Multi-Region Modified Wind Driven Optimization

algorithm (MR-MWDO) to solve dynamic optimization problems. MR-MWDO
employs a new technique called Multi-Region technique inspired from meteorology
to detect adaptively promising regions in search space and track multiple peaks. A
new Collision Avoidance Technique has been introduced as well. Sub-populations
evolve according to a modified WDO algorithm. The method has been assessed using
Moving Peak Benchmark and compared to state of the art methods. Obtained results
are very promising and show viability of the method. As future work, we intend to
foster algorithm search abilities by introducing prediction models to better track
moving peaks.
References
1. Calderín, J.F., Masegosa, A.D., Suárez, A.R., Pelta, D.A.: Adaptation Schemes and
Dynamic Optimization problems: A Basic Study on the Adaptive Hill Climbing Memetic
Algorithm. In: Terrazas, G., Otero, F.E.B., Masegosa, A.D. (eds.) NICSO 2013. SCI,
2. Yang, S., Yao, X. (eds.): Evolutionary Computation for Dynamic Optimization Problems.
SCI, vol. 490. Springer, Heidelberg (2013)
3. Bayraktar, Z., Komurcu, M., Bossard, J.A., Werner, D.H.: The Wind Driven Optimization
Technique and its Application in Electromagnetics. IEEE Transactions on Antennas and
Propagation 61(5), 2745–2757 (2013)
4. James, R.H.: An Introduction to Dynamic Meteorology, 4th edn., USA, vol. 88 (2004)
5. Chao, C.W., Fang, S.C., Liao, C.J.: A Tropical Cyclone-Based Method For Global
Optimization. Journal of Industrial And Management Optimization 8, 103–115 (2012)
6. Nguyen, T.T.: Continuous Dynamic Optimisation Using Evolutionary Algorithms. PhD
thesis, School of Computer Science, University of Birmingham (2011)
7. Blackwell, T., Branke, J.: Multiswarms, exclusion, and anti-convergence in dynamic
environments. IEEE Trans. Evol. Comput. 10(4), 459–472 (2006)
8. Branke, J.: The moving peaks benchmark,
http://www.aifb.uni-karlsruhe.de/~jbr/MovPeaks/
(viewed November 8, 2008)
9. Branke, J., Schmeck, H.: Designing evolutionary algorithms for dynamic optimization
problems. In: Advances in Evolutionary Computing: Theory and Applications, pp. 239–
262 (2003)
10. Kamosi, M., Hashemi, A.B., Meybodi, M.R.: A hibernating multi-swarm optimization
algorithm for dynamic environments. In: Proc. World Congr. on Nature and Biologically
Inspired Computing, NaBIC 2010, pp. 363–369 (2010)
11. Yang, S., Li, C.: A clustering particle swarm optimizer for locating and tracking multi-ple
optima in dynamic environments. IEEE Trans. Evol. Comput., 959–974 (2010)
12. Parrott, D., Li, X.: Locating and tracking multiple dynamic optima by a particle swarm
model using speciation. IEEE Trans. Evol. Comput. 10(4), 440–458 (2006)
13. Li, C., Yang, S.: A general framework of multipopulation methods with clustering in
undetectable dynamic environments. IEEE Trans. Evol. Comput. 16(4), 556–577 (2012)
Evaluating a Hybrid DE and BBO with Self

Adaptation on ICSI 2014 Benchmark Problems
Yu-Jun Zheng1 and Xiao-Bei Wu2

1
College of Computer Science & Technology, Zhejiang University of Technology,
Hangzhou 310023, China
2
College of Electronics and Information Engineering,
Tongji University, Shanghai 201804, China
yujun.zheng@computer.org, xwu4@StateStreet.com
Abstract. The paper presents a new hybrid differential evolution (DE)

and biogeography-based optimization (BBO) algorithm and tests its per-
formance on the benchmark set for the ICSI 2014 Competition. The al-
gorithm tends to perform more DE mutations in early search stage and
more BBO migrations in later stage, in order to provide a good balance
of exploration and exploitation. It also uses a trial-and-error method
inspired by the self-adaptive DE (SaDE) to choose appropriate muta-
tion/migration schemes during the search. Computational experiment
shows that the algorithm outperforms DE, SaDE, and blended BBO on
the benchmark set.
Keywords: single objective optimization, differential evolution (DE),

biogeography-based optimization (BBO), self-adaptation.
1 Introduction
Evolutionary algorithms (EAs) are stochastic search methods drawing inspira-

tion from biological evolution for optimization problems. Most EAs are initially
proposed for solving single objective optimization problems, which are the basis
of a wide range of real-world optimization problems.
The tradeoff between exploration and exploitation is the sticking point in
search processes, having a great effect on convergence speed and accuracy of
EAs [4]. Among various EAs, differential evolution (DE) [11] is a popular algo-
rithm known for its very competitive exploration ability. Another relatively-new
EA, biogeography-based optimization (BBO) [10], has shown strong exploitation
ability on a variety of optimization problems.
The purpose of the paper is to present a hybrid algorithm that combines
the DE’s exploration ability with BBO’s exploitation ability, and to evaluate
the performance of the hybrid algorithm on the benchmark set for the ICSI
2014 Competition on Single Objective Optimization [12]. The hybrid method,
named HSDB (Hybrid Self-adaptive DE and BBO), prefers to make more use of

This work was supported by Natural Science Foundation (No. 61105073) of China.

Evaluating a Hybrid DE and BBO with Self Adaptation on ICSI 2014 423
DE mutations in exploration in early search stage, and is more likely to adopt

BBO migrations in exploitation in later search stage. Since the performance of
DE/BBO heavily depends on the mutation/migration operators, here we embed
two DE mutation schemes and two BBO migration schemes in HSDB, and bor-
row the ideas from the SaDE algorithm [9] to dynamically choose the schemes
during the search. We also compare the performance of HSDB with the DE,
SaDE, and an improved BBO on the benchmark set.
The rest of this paper is as follows: Section 2 and Section 3 respectively intro-
duce DE and BBO, Section 4 describes our HSDB algorithm, Section 5 presents
the numerical experiments on the benchmark set, and Section 6 concludes.
2 Differential Evolution
DE is a population-based EA that simultaneously evolves a population P of
floating-point solution vectors towards the global optimum to the given opti-
mization problem. The most important operator in DE is mutation, which pro-
duces a mutant vector vi for each individual xi in the population (1 ≤ i ≤ |P |).
The most-widely used mutation scheme is the DE/rand/1/bin scheme that adds
the weighted difference between two randomly selected vectors to a third one:
vi = xr1 + F · (xr2 − xr3 ) . (1)
where F is the scaling factor typically in the range [0,1], r1 , r2 and r3 are three
mutually exclusive random indexes in [1, |P |].
Afterwards, a trial vector ui is generated by mixing the components of the mu-
tant vector and the original one, where each dth component of ui is determined
as follows: % d
d vi if rand(0, 1) < cr or d = ri
ui = . (2)
xdi else
where cr is the crossover rate ranged in (0, 1) and ri is a random integer within
[1, |P |] for each i, ensuring that the trial vector gets at least one component
from the mutant vector.
In the last step of each iteration, the selection operator chooses the better one
for the next generation by comparing the fitness of ui with xi :
%
ui if f (ui ) > f (xi )
xi = . (3)
xi else
Nevertheless, there are other mutation schemes that can be implemented in

DE. The following presents three other mutation schemes frequently used:
– DE/best/1:
vi = xbest + F · (xr1 − xr2 ) . (4)
– DE/rand-to-best/2:
vi = xi + F · (xbest − xi ) + F · (xr2 − xr3 ) . (5)

424 Y.-J. Zheng and X.-B. Wu
– DE/best/2:
vi = xbest + F · (xr1 − xr2 ) + F · (xr3 − xr4 ) . (6)
The performance of DE heavily depends on its mutation scheme and con-

trol parameter settings. Unfortunately, due to the variety of problems, various
conflicting conclusions have been drawn with regard to parameter settings in
DE literatures [9]. To avoid manual tuning, Abbass [1] proposed a self-adaptive
mechanism that encodes the crossover rate cr into each solution and simultane-
ously evolves the parameter and the solution. Omran et al. [8] developed a similar
mechanism for adapting the scaling factor F . Qin et al. [9] proposed the SaDE
algorithm, which performs a trial-and-error search for the most appropriate mu-
tation scheme during the search. Let K be the total number of mutation schemes
in the pool, at the Gth generation, the probability of choosing the kth scheme
is proportional to its success rate sk (G) within the previous LP generations:
G−1
g=G−LP nsk (g)
sk (G) = G−1 G−1 . (7)
g=G−LP nsk (g) + g=G−LP nfk (g)
where nsk (g) is the number of trial vectors generated by the kth strategy at
the gth generation that successfully enter the next generation, and nfk (g) is the
number of trial vectors that fail to do so. SaDE also uses a similar strategy for
adapting the cr values, and sets F as a Gaussian random number with mean 0.5
and standard deviation 0.3, denoted by N (0.5, 0.3).
3 Biogeography-Based Optimization
BBO is also a population-based EA, which was proposed by Simon [10] based
on the mathematics of island biogeography. A solution is analogous to a habitat
the solution components are analogous to the habitat’s suitability index variables
(SIVs), and the solution fitness is analogous to the species richness or habitat
suitability index (HSI) of the island. High HSI habitats tend to share their
features with low HSI habitats, and low HSI habitats are likely to accept many
new features from high HSI habitats. For example, in a simple linear migration
model, the immigration rate λi and the emigration rate μi of each habitat xi
are calculated as follows:
fmax − fi
λi = I . (8)
fmax − fmin
fi − fmin
μi = E . (9)
fmax − fmin
where fmax and fmin are the maximum and minimum fitness value of the popu-
lation, and I and E are the maximum possible immigration rate and emigration
rate which are typically set to 1. However, there are also many other nonlinear
migration models can be used [6].
At each generation of BBO, each SIV of a habitat xi has a probability of λi

to be immigrated, and the migrating SIV comes from an emigrating habitat xj ,
the selection probability of which is proportional to μj of the islands.
BBO also has a mutation operator for setting a SIV to a random value in
the search range, which is mainly for increasing solution diversity and hence
improving exploration.
In [7] Ma and Simon developed the blended BBO (B-BBO), which replaces
the clonal migration of BBO with the following blended migration:
Xid = αXid + (1 − α)Xjd . (10)
where α is a real number between 0 and 1.

Gong et al. [5] proposed a hybrid migration operator by combining BBO
migration and DE mutation, where each habitat component has a probability of
cr to be mutated by DE mutation, and a probability of (1 − cr ) to be changed by
BBO migration. Experiments show that the hybrid method outperforms both DE
and BBO on a set of benchmark problems. Boussaı̈d et al. [2] proposed another
approach combining BBO and DE, which evolves the population by alternately
applying BBO and DE iterations, and at each iteration always selects a fitter
one from the updated solution and its parent for the next iteration. In [3] the
authors extended the approach for constrained optimization, where the original
mutation operator of BBO is replaced by the DE mutation operator.
By default, BBO uses a global topology where any two habitats in the popu-
lation have a chance to communicate with each other. Zheng et al. [13] equipped
BBO with a local topology, and any habitat can only immigrate SIVs from its
neighboring habitats. In consequence, the flow of information can be moderately
pass through the neighborhood, and the algorithm can achieve a better balance
between exploration and exploitation. In [14] Zheng et al. enhanced BBO with
two new migration operators: global migration and local migration. The former
migrates features from a neighbor Xnb and a non-neighbor Xf ar , while the latter
only considers migration from a neighbor:
Xid = Xfdar + α(Xnb

d
− Xid ) . (11)
Xid = Xid + d
α(Xnb − Xid ) . (12)
4 A Hybrid DE and BBO with Self Adaptation
The HSDB method differentiates from the hybrid algorithms of Gong et al. [5]
and Boussaı̈d et al. [2] in that it does not combine DE mutation and BBO
migration into an integrated operator. Instead, HSDB prefers to make more use
of DE mutation in exploration in early search stage, and is more likely to adopt
BBO migration in exploitation in later search stage. The key to realize this
is a parameter named maturity index, denoted by ν, which increases with the
generation: The higher (lower) the ν, the more likely the algorithm is to conduct
migration (mutation). A simple approach is to linearly increase ν from an initial
value νmin to a final value νmax as follows (where g is the current generation
number and g max is the total generation number of the algorithm):
g
ν = ν min + (ν max − ν min ) . (13)
g max
At each generation, each habitat has a probability of ν to be modified by BBO
migration and a probability of (1 − ν) by DE mutation. Moreover, we embed
a set of DE mutation schemes (denoted by SDE ) and a set of BBO migration
schemes (denoted by SBBO ) in HSDB, for each scheme record its success and
failure numbers within a fixed number LP of previous generations, and calculate
its success rate in a similar way as SaDE [9]. However, in HSDB, we calculate
the success rates for DE mutation schemes and BBO migration schemes inde-
pendently. That is, at beginning every scheme is assigned with an equal selection
probability; at each generation G ≥ LP, the probability of choosing the kth DE
mutation scheme and that of the k th BBO migration scheme are respectively
updated as follows:
sk (G)
pk (G) = . (14)
k∈S sk (G)
DE
sk (G)
pk (G) = . (15)
k ∈S sk (G)
BBO
Currently, in HSDB we embed two DE mutation schemes including DE/rand

/1/bin (1) and DE/rand-to-best/2/bin (5), and two BBO migration schemes
including the original clonal migration and the local migration (12). This is
because the two DE mutation schemes exhibit good exploration ability and the
two BBO migration schemes have more exploitation ability than others.
To support local migration, HSDB also implements a local topology to estab-
lish the neighborhood structure for the population. Given an expected average
neighborhood size K, the local topology, represented by an adjacent 0-1 matrix
NB, is randomly set by the procedure described in Algorithm 1. As we can see,
for each solution in the population p, every other solution has a probability of
K/(|P | − 1) to be its neighbor. An advantage of this approach is that the neigh-
borhood size K does not need to be limited to an integer. However, isolated
solutions are not allowed. To increase diversity, HSDB will reset its local topol-
ogy if it fails to find a new better known solution after NP generations, where
NP is a control parameter typically ranges from 1 to 6.
For coherence, we replace the crossover rate cr of DE by the immigration rate
λi of each solution i. That is, the DE crossover operation (2) is replaced by the
following equation in HSDB:
% d
vi if rand(0, 1) < λi or d = ri
udi = . (16)
xdi else
We also use the same parameter scheme for coefficient F in Eq. (1) and (5)
and α in Eq. (12), all denoted by F in HSDB. We employ a Gaussian random
Algorithm 1. The procedure for setting the neighborhood topology.

1 Let NB be a |P | × |P | all-zeros matrix;
1 Let p = K/(|P | − 1);
2 for i = 1 to |P | do
3 for j = 1 to |P | do
4 if i = j ∧ rand(0, 1) < p then
5 NB(i, j) = NB(j, i) = 1;
6 if the ith solution has no neighbor then
7 Randomly choose a j other than i and set NB(i, j) = NB(j, i) = 1;
number N (Fμ , Fσ ) to approximate F , where Fμ and Fσ are typically set as

0.5 and 0.3 (as in SaDE [9]). Moreover, since the original BBO uses mutation
mainly for exploration purpose, while here we concentrate on the exploitation
ability of BBO, the HSDB method does not use the original BBO mutation
scheme. Algorithm 2 describes the framework of HSDB.
Algorithm 2. The HSDB algorithm.

1 Randomly initialize a population P of solutions to the problem;
2 Initialize the local topology of the population based on Algorithm 1;
3 while stop criterion is not satisfied do
4 for i = 1 to |P | do
5 Compute λi and μi for each solution xi ;
6 if rand(0, 1) < ν then
7 Select a DE mutation scheme k with probability ∝ sk (G);
8 else
9 Select a BBO migration scheme k with probability ∝ sk (G);
10 Create a mutant vector vi using the selected scheme;
11 Create a trial vector ui according to the crossover operation (16);
12 if ui is fitter than xi then
13 Replace xi with ui in P ;
14 Update the maturity index ν;
15 if G ≥ LP then
16 Update the success rates of the schemes;
17 if no new better solution has been found for NG generations then
18 Reset the local topology of the population;
19 return the best solution found so far.
5 Computational Experiment
5.1 Experiment Setup
We test the HSDB algorithm on the ICSI 2014 benchmark set, which consists
of 30 high-dimensional problems, denoted as f1 –f30 , all shifted and rotated.
Since all the test problem are considered as black box problems, we do not
tune parameters for each problem; instead we use a general parameter setting
of HSDB for the whole problem set as: the range of maturity index νmin = 0.05,
νmax = 0.95, the learning period LP = 30, the limit of non-improved generations
NG = 3, and the neighborhood size K = 3. The population size |P | is set to 50.
The experiments are conducted on a computer of Intel Core i5-2430M pro-
cessor and 4GB DDR3 memory. The HSDB algorithm has been implemented
using Matlab R2013a. As requested by the ICSI 2014 competition session, the
algorithm has been tested on each problem with dimensions 2, 10 and 30 re-
spectively, the search space is [−100, 100]D, the maximum number of function
evaluations is set as 10000D (where D denotes the problem dimension), and 51
simulation runs have been performed on each problem instance. Any function
value smaller than 2−52 ≈ 2.22e − 16 (the in Matlab) is considered as zero.
In our experimental environment, the mean time for executing the benchmark
program of ICSI 2014 competition over 5 runs is T 1 = 48.62s, and the mean time
of HSDB on function 9 and D = 30 over 5 runs is T 2 = 116.02s, and the time
complexity is evaluated by the ratio (T 2 − T 1)/T 1 = 1.386.
5.2 Experimental Results

Tables 1, 2 and 3 respectively present the computational results of HSDB on
2-D, 10-D and 30-D problems (where ‘Std’ denotes standard deviation). As we
can see, on the simple 2-D problems, HSDB achieves a function error value
less than on 21 problems. With the increase of dimension, the performance of
HSDB decreases on most problems. This phenomenon, known as the “curse of
dimensionality”, is serious on several problems such as f1 and f2 . However, an
interesting finding is that the solution accuracy of the algorithm increases on
some special problems including f12 and f26 .
5.3 Comparative Results

We have compared the performance of HSDB with DE/rand/1/bin scheme, SaDE,
and B-BBO on the benchmark set, and conducted nonparametric Wilcoxon rank
sum tests on the results of HSDB and that of the other three methods. Due to
the page limits, here we only present the comparison results on 30-D problems
in Table 4 (where an h value of 1 indicates that the performances of the HSDB
and DE/SaDE are statistically different with 95% confidence, h value of 0 implies
there is no statistical difference, the superscript + denotes HSDB has significant
performance improvement over DE/SaDE and − vice versa).
As we can see, HSDB significantly outperforms DE on 28 problems, DE only
outperforms HSDB on f26 , and there is no statistical difference between them on
f28 . In comparison with SaDE, HSDB has significant performance improvement
on 15 problems; SaDE outperforms HSDB on 6 problems, and there is no statis-
tical difference between them on 9 problems. HSDB also outperforms B-BBO on
29 problems, and there is no statistical difference only on f28 . The results show
that the overall performance of HSDB is better than the other three methods
on the benchmark set.
Table 1. The experimental results on 2-D problems
ID Max Min Mean Median Std

f1 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00
f2 2.98E−27 0.00E+00 1.13E−28 0.00E+00 4.18E−28
f3 1.67E+00 1.67E+00 1.67E+00 1.67E+00 1.12E−15
f4 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00
f5 1.52E−17 8.00E−20 2.45E−18 1.53E−18 2.44E−18
f6 1.97E−31 1.97E−31 1.97E−31 1.97E−31 0.00E+00
f7 1.21E−17 0.00E+00 6.44E−18 9.65E−18 5.26E−18
f8 8.88E−16 8.88E−16 8.88E−16 8.88E−16 9.96E−32
f9 -4.00E+00 -4.00E+00 -4.00E+00 -4.00E+00 2.24E−15
f10 -9.80E−01 -1.00E+00 -9.96E−01 -1.00E+00 5.86E−03
f11 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00
f12 5.20E+01 4.86E+01 4.90E+01 4.86E+01 8.74E−01
f13 1.94E−02 0.00E+00 7.00E−03 0.00E+00 9.15E−03
f14 6.99E−02 1.00E−11 5.90E−03 9.41E−04 1.06E−02
f15 3.60E−03 3.91E−08 7.92E−04 3.59E−04 9.40E−04
f16 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00
f17 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00
f18 -8.38E+02 -8.38E+02 -8.38E+02 -8.38E+02 4.59E−13
f19 1.35E−32 1.35E−32 1.35E−32 1.35E−32 1.11E−47
f20 5.77E−23 0.00E+00 2.58E−24 1.77E−33 8.18E−24
f21 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00
f22 0.00E+00 0.00E+00 0.00E+00 0.00E+00 0.00E+00
f23 8.97E+00 8.97E+00 8.97E+00 8.97E+00 7.64E−15
f24 -4.53E+00 -4.53E+00 -4.53E+00 -4.53E+00 6.28E−15
f25 -1.39E+00 -1.71E+00 -1.60E+00 -1.71E+00 1.54E−01
f26 6.67E−01 5.71E−01 5.87E−01 5.71E−01 2.79E−02
f27 -3.74E+07 -3.74E+07 -3.74E+07 -3.74E+07 4.01E+02
f28 -4.63E+00 -5.83E+00 -5.22E+00 -5.25E+00 2.63E−01
f29 2.00E+01 2.00E+01 2.00E+01 2.00E+01 8.92E−03
f30 1.01E+00 2.67E−01 4.00E−01 2.70E−01 2.71E−01

f1 1.35E+02 5.03E−02 8.23E+00 2.28E+00 1.87E+01
f2 2.64E+02 2.03E−01 3.26E+01 5.43E+00 5.33E+01
f3 1.70E+02 1.70E+02 1.70E+02 1.70E+02 1.77E−02
f4 2.96E+00 6.54E−04 3.45E−01 1.55E−01 4.67E−01
f5 1.27E−01 1.38E−05 1.14E−02 8.59E−04 2.33E−02
f6 9.57E+00 4.97E−01 5.02E+00 5.95E+00 2.40E+00
f7 1.15E−03 1.00E−12 1.46E−04 1.96E−05 2.74E−04
f8 1.16E+00 5.66E−13 1.71E−01 5.16E−03 3.71E−01
f9 -1.95E+01 -2.00E+01 -1.98E+01 -1.99E+01 1.35E−01
f10 -7.42E+00 -8.76E+00 -8.34E+00 -8.37E+00 2.85E−01
f11 4.56E+00 0.00E+00 1.09E+00 1.51E−02 1.82E+00
f12 5.40E−01 3.76E−01 4.52E−01 4.47E−01 4.17E−02
f13 6.35E−01 1.55E−01 3.99E−01 4.02E−01 1.37E−01
f14 9.91E−02 1.01E−02 5.17E−02 5.25E−02 2.19E−02
f15 9.83E−02 1.74E−02 4.89E−02 4.84E−02 1.76E−02
f16 4.01E−02 1.14E−08 6.05E−03 1.20E−03 9.83E−03
f17 4.72E−03 3.06E−06 8.07E−04 2.72E−04 1.03E−03
f18 -2.02E+03 -4.14E+03 -2.30E+03 -2.02E+03 6.89E+02
f19 1.62E−03 5.20E−21 6.90E−04 9.55E−04 6.29E−04
f20 3.68E−02 4.98E−06 5.61E−03 1.90E−03 8.75E−03
f21 9.99E−02 9.99E−02 9.99E−02 9.99E−02 7.70E−11
f22 1.08E−02 5.32E−31 8.48E−04 3.56E−05 1.90E−03
f23 2.37E+01 1.55E+01 1.94E+01 1.90E+01 2.01E+00
f24 -3.47E+01 -3.56E+01 -3.53E+01 -3.52E+01 2.40E−01
f25 4.30E+01 4.29E+01 4.29E+01 4.29E+01 2.80E−03
f26 7.59E−03 5.90E−03 6.72E−03 6.70E−03 4.87E−04
f27 -6.34E+07 -1.60E+08 -1.02E+08 -1.07E+08 1.70E+07
f28 -5.04E+00 -5.85E+00 -5.38E+00 -5.40E+00 1.72E−01
f29 2.00E+01 2.00E+01 2.00E+01 2.00E+01 9.31E−04
f30 1.01E+00 1.01E+00 1.01E+00 1.01E+00 1.86E−10

f1 1.69E+05 4.08E+03 7.67E+04 7.80E+04 4.15E+04
f2 1.60E+04 1.34E+03 5.18E+03 3.71E+03 3.27E+03
f3 4.53E+03 4.51E+03 4.52E+03 4.52E+03 5.11E+00
f4 1.33E+01 1.59E+00 7.45E+00 7.31E+00 2.71E+00
f5 7.87E−02 8.39E−03 3.77E−02 3.66E−02 1.72E−02
f6 3.03E+01 2.82E+01 2.91E+01 2.90E+01 5.15E−01
f7 1.01E−02 3.23E−03 7.37E−03 7.84E−03 1.72E−03
f8 1.60E+00 1.86E−01 9.06E−01 9.91E−01 4.42E−01
f9 -5.76E+01 -5.91E+01 -5.83E+01 -5.82E+01 3.35E−01
f10 -1.99E+01 -2.79E+01 -2.48E+01 -2.54E+01 1.80E+00
f11 8.90E−02 4.55E−04 2.98E−02 2.85E−02 1.82E−02
f12 1.47E−02 1.22E−02 1.43E−02 1.44E−02 4.42E−04
f13 4.02E+00 1.40E+00 3.03E+00 3.14E+00 6.88E−01
f14 1.03E−01 2.53E−02 7.29E−02 7.43E−02 1.96E−02
f15 1.70E−01 2.04E−02 9.48E−02 9.36E−02 3.40E−02
f16 3.33E−01 8.08E−02 2.21E−01 2.24E−01 5.23E−02
f17 3.93E+00 4.31E−01 2.02E+00 2.10E+00 8.17E−01
f18 -3.69E+03 -6.06E+03 -5.79E+03 -6.06E+03 7.11E+02
f19 8.67E−03 2.00E−03 5.05E−03 4.81E−03 1.77E−03
f20 7.45E−01 9.08E−06 2.07E−01 1.11E−01 2.27E−01
f21 3.00E−01 9.99E−02 2.13E−01 2.00E−01 5.30E−02
f22 2.15E−01 1.01E−02 1.01E−01 9.68E−02 5.44E−02
f23 5.68E+01 3.09E+01 4.43E+01 4.32E+01 7.31E+00
f24 -1.13E+02 -1.16E+02 -1.15E+02 -1.15E+02 7.39E−01
f25 1.13E+03 1.13E+03 1.13E+03 1.13E+03 2.06E+00
f26 4.48E−04 4.12E−04 4.33E−04 4.35E−04 8.73E−06
f27 -2.99E+08 -5.25E+08 -4.76E+08 -5.21E+08 8.72E+07
f28 -5.24E+00 -5.81E+00 -5.48E+00 -5.48E+00 1.12E−01
f29 2.00E+01 2.00E+01 2.00E+01 2.00E+01 1.53E−04
f30 1.04E+00 1.01E+00 1.02E+00 1.01E+00 1.34E−02
Table 4. The Comparison of HSDB with DE and SaDE
432
DE SaDE B-BBO
ID Median Std p-value h Median Std p-value h Median Std p-value h
f1 1.73E+06 1.10E+06 3.30E−18 1+ 1.98E+04 2.95E+04 2.12E−08 1− 3.98E+06 9.84E+05 3.30E−18 1+
f2 8.28E+05 1.87E+05 3.30E−18 1+ 2.58E+04 8.72E+03 4.18E−18 1+ 5.72E+04 2.18E+04 3.30E−18 1+
f3 5.26E+03 7.97E+02 6.30E−17 1+ 4.51E+03 5.27E+00 1.24E−05 0 4.84E+03 1.30E+02 3.30E−18 1+
f4 3.72E+02 6.08E+01 3.30E−18 1+ 1.76E+02 3.06E+01 3.30E−18 1+ 1.36E+02 3.23E+01 3.30E−18 1+
f5 1.39E+00 9.06E−01 3.30E−18 1+ 2.31E−02 1.36E−02 2.41E−05 1− 1.64E+00 3.71E-01 3.30E−18 1+
f6 3.77E+01 9.47E+00 3.72E−18 1+ 2.84E+01 4.02E−01 3.77E−08 1− 6.91E+01 1.79E+01 3.30E−18 1+
f7 3.98E−02 3.87E−03 3.30E−18 1+ 8.74E−03 3.81E−03 2.06E−02 1+ 3.10E-02 1.38E-02 3.30E−18 1+
f8 2.29E+00 4.90E−01 1.24E−16 1+ 9.81E−01 6.12E−01 8.94E−01 0 3.37E+00 4.41E-01 3.30E−18 1+
f9 -5.61E+01 7.52E−01 3.30E−18 1+ -5.84E+01 3.60E−01 9.70E−02 0 -5.58E+01 5.06E-01 3.30E−18 1+
Y.-J. Zheng and X.-B. Wu
f10 -1.03E+01 1.12E+00 3.30E−18 1+ -1.84E+01 1.10E+00 4.70E−18 1+ -2.42E+01 2.43E+00 2.66E−03 1+
f11 2.82E−01 2.92E−01 5.29E−18 1+ 1.21E−02 1.14E−02 8.30E−06 1− 2.73E+00 8.30E-01 3.30E−18 1+
f12 1.53E−02 1.57E−04 3.30E−18 1+ 1.48E−02 1.87E−04 3.91E−12 1+ 1.46E-02 6.13E-04 1.44E−03 1+
f13 7.40E+00 4.03E−01 3.30E−18 1+ 5.16E+00 5.64E−01 3.50E−18 1+ 4.79E+00 5.59E-01 1.32E−16 1+
f14 1.40E−01 2.29E−02 1.79E−17 1+ 9.16E−02 1.97E−02 1.40E−05 1+ 8.37E-02 2.02E-02 4.94E−03 1+
f15 1.46E−01 4.37E−02 3.24E−08 1+ 1.19E−01 3.67E−02 4.42E−04 1+ 4.78E-01 8.24E-02 7.51E−18 1+
f16 5.63E−01 1.50E−01 3.72E−18 1+ 2.07E−01 8.51E−02 5.21E−01 0 1.01E+00 2.24E-01 3.30E−18 1+
f17 1.69E+02 2.91E+01 3.30E−18 1+ 1.19E+01 2.83E+00 3.30E−18 1+ 1.14E+01 3.15E+00 3.30E−18 1+
f18 -6.05E+03 1.66E+00 2.67E−12 1+ -6.06E+03 7.60E+02 8.63E−09 1+ -3.66E+03 1.91E+03 9.65E−16 1+
f19 2.85E−01 1.39E−01 3.30E−18 1+ 2.35E−02 1.53E−02 1.79E−17 1+ 1.90E-02 8.13E-03 3.30E−18 1+
f20 6.04E+02 2.07E+02 3.30E−18 1+ 7.82E+01 3.06E+01 3.30E−18 1+ 1.89E+01 7.50E+00 3.30E−18 1+
f21 7.78E−01 2.18E−01 5.38E−19 1+ 2.00E−01 6.16E−02 2.85E−01 0 8.00E-01 2.79E-01 9.80E−19 1+
f22 8.45E−01 6.23E−01 3.30E−18 1+ 1.05E−02 1.94E−02 4.61E−15 1− 7.75E+00 2.32E+00 3.30E−18 1+
f23 9.28E+01 5.00E+00 3.30E−18 1+ 6.98E+01 4.62E+00 3.30E−18 1+ 6.69E+01 4.92E+00 4.70E−18 1+
f24 -1.10E+02 1.49E+00 4.18E−18 1+ -1.14E+02 8.46E−01 3.63E−01 0 -1.10E+02 1.24E+00 3.30E−18 1+
f25 1.32E+03 7.89E+01 3.30E−18 1+ 1.13E+03 1.69E+00 4.64E−10 1− 1.21E+03 2.57E+01 3.30E−18 1+
f26 3.98E−04 1.19E−06 3.30E−18 1− 4.26E−04 1.58E−05 1.92E−01 0 6.26E-04 5.67E-05 3.30E−18 1+
f27 -4.64E+08 3.13E+07 5.36E−07 1+ -5.19E+08 4.89E+07 3.29E−01 0 -2.69E+08 5.82E+07 3.30E−18 1+
f28 -5.49E+00 1.26E−01 7.18E−01 0 -5.51E+00 1.44E−01 3.25E−01 0 -5.49E+00 1.14E-01 8.72E-01 0
f29 2.00E+01 1.26E−04 5.95E−03 1+ 2.00E+01 1.18E−04 2.77E−02 1+ 2.00E+01 1.57E-04 2.72E-02 1+
f30 1.04E+00 1.14E−02 8.40E−19 1+ 1.01E+00 1.34E−02 4.05E−04 1+ 1.04E+00 8.03E-03 6.07E-03 1+
6 Conclusion
The paper presents a new hybrid DE and BBO algorithm, named HSDB, which
uses a trial-and-error method to select among two DE mutation schemes and two
BBO migration schemes, and balances the exploration and exploitation based
on the maturity index parameter ν. Experiments show that HSDB outperforms
DE, SaDE, and B-BBO on the benchmark set for the ICSI 2014 Competition.
We are currently including more DE mutation and BBO migration schemes into
HSDB and testing more effective method for tuning ν.
References
1. Abbass, H.: The self-adaptive pareto differential evolution algorithm. In: Proceed-
ings of the 2002 Congress on Evolutionary Computation, vol. 1, pp. 831–836 (2002)
2. Boussaı̈d, I., Chatterjee, A., Siarry, P., Ahmed-Nacer, M.: Two-stage update
biogeography-based optimization using differential evolution algorithm (DBBO).
Comput. Oper. Res. 38(8), 1188–1198 (2011)
3. Boussaı̈d, I., Chatterjee, A., Siarry, P., Ahmed-Nacer, M.: Biogeography-based
optimization for constrained optimization problems. Comput. Oper. Res. 39(12),
3293–3304 (2012)
4. Chen, J., Xin, B., Peng, Z., Dou, L., Zhang, J.: Optimal contraction theorem for
exploration – exploitation tradeoff in search and optimization. IEEE Trans. Syst.
Man Cybern. Part A 39(3), 680–691 (2009)
5. Gong, W., Cai, Z., Ling, C.X.: DE/BBO: a hybrid differential evolution with
biogeography-based optimization for global numerical optimization. Soft Com-
put 15(4), 645–665 (2010)
6. Ma, H.: An analysis of the equilibrium of migration models for biogeography-based
optimization. Inform. Sci. 180(18), 3444–3464 (2010)
7. Ma, H., Simon, D.: Blended biogeography-based optimization for constrained op-
timization. Engin. Appl. Artif. Intell. 24(3), 517–525 (2011)
8. Omran, M.G.H., Salman, A., Engelbrecht, A.P.: Self-adaptive differential evolution.
In: Hao, Y., Liu, J., Wang, Y.-P., Cheung, Y.-m., Yin, H., Jiao, L., Ma, J., Jiao,
Y.-C. (eds.) CIS 2005. LNCS (LNAI), vol. 3801, pp. 192–199. Springer, Heidelberg
(2005)
9. Qin, A.K., Suganthan, P.N.: Self-adaptive differential evolution algorithm for nu-
merical optimization. In: 2005 IEEE Congress on Evolutionary Computation,
vol. 2, pp. 1785–1791 (2005)
10. Simon, D.: Biogeography-based optimization. IEEE Trans. Evol. Comput. 12(6),
702–713 (2008)
11. Storn, R., Price, K.: Differential evolution - a simple and efficient heuristic for global
optimization over continuous spaces. J. Global Optim. 11(4), 341–359 (1997)
12. Tan, Y., Li, J., Zheng, Z.: ICSI 2014 competition on single objective optimization.
Tech. rep., Peking University (2014),
http://www.ic-si.org/competition/ICSI.pdf
13. Zheng, Y.J., Ling, H.F., Wu, X.B., Xue, J.Y.: Localized biogeography-based opti-
mization. Soft Comput. (2014), doi:10.1007/s00500-013-1209-1
14. Zheng, Y.J., Ling, H.F., Xue, J.Y.: Ecogeography-based optimization: Enhancing
biogeography-based optimization with ecogeographic barriers and differentiations.
Comput. Oper. Res. 50, 115–127 (2014)
The Multiple Population Co-evolution PSO Algorithm
Xuan Xiao and Qianqian Zhang
Beijing Institute of Technology, Beijing, China

{1521494822,289314426}@qq.com
Abstract. In order to overcome the standard particle swarm optimization

algorithm which is easily trapped in local minima and optimize the shortcoming
of low precision, this paper proposed a way which can make multiple
information exchange between particles come true: the multiple population
co-evolution PSO algorithm. This paper proposes a multiple population
co-evolutionary algorithm to achieve communication among populations, and
then show the feasibility and effectiveness of this algorithm through
experiments.
Keywords: Particle swarm, co-evolution, PSO multiple population.
1 Introduction
Simulating the behaviors of biological populations has become a research hotspot

to solve calculating problems in the field of intelligence calculation. The theory t
hat the swarm intelligence is the core has formed, and it has made revolutionary
progress in a number of practical applications. Particle swarm optimization(PSO) i
s an intelligent optimization algorithm, which is used to handle the problem of co
ntinuous variables searching the search space, and it has been applied to many ar
eas, such as function optimization, constrained optimization and neural networks.
This paper proposes a multiple population co-evolutionary algorithm to achieve c
ommunication among populations, and then show the feasibility and effectiveness
of this algorithm through experiments.
2 The Description of Basic Particle Swarm Optimization
The basic concept of PSO comes from the study of the preying of birds developed by
Kennedy and Eberhart [1, 2]. Imaging a scenario like this: a flock of birds randomly
search food in an area, but no one knows where the food are and how far away their
current location is from the food. The strategy of flying and searching is to follow the
first bird in population. PSO get inspiration from this model and is used to solve
optimization problems. Every possible solution is a bird which is called "particles" in
search space, and all particles have been evaluated by fitness decided by fitness
function. Each particle is used to describe an alternative solution in the solution space,
The Multiple Population Co-evolution PSO Algorithm 435
and has a random velocity throughout the whole solution space. Each particle gets
heuristic information from each other and guides the movement of the entire group by
the exchange of information with other particles.
In the basic PSO algorithm, each particle represents a possible solution and all
the particles forms swarm. Particles depend on their own historical information an
d swarm information in the search space to determine the velocity and direction o
f flying to find the optimal solution. Assuming that solving problems in D-dimen
sion search space, swarm composed of m particles, Swarm = {x1(k ) , x2(k ) ,, xm(k ) } . At
time k+1, the position vector is xi( k +1) = ( xi(1k +1) , xi(2k +1) ,, xiD
( k +1)
) , i= 1,2, ...,m, whic
h is the location of individuals in the search space, and it is also a possible soluti
on of the problem. Corresponding to the individual position vector is its velocity
vector vi( k +1) = (vi(1k +1) , vi(2k +1) ,, viD
( k +1)
) , which describes the movement of particles o
f each dimension in search space. ( Note: The superscript of variable represents t
he iteration cycle, for example, xid(k +n) represents the k+n cycles; superscript of var
iable without parentheses represents power, asωn represents the n-th power of ω.)
The neighborhood function of PSO generates a new location status according t
o each individual's own position vector, velocity vector, individual historical infor
mation, group information, and disturbance. In standard PSO algorithm, function
calculating formulation of i-th particle at time k+1 in d-dimension is as follows:
vid(k+1) =ω⋅vid(k) +c1 ⋅ r1 ⋅(pid(k) − xid(k) ) +c2 ⋅ r2 ⋅( pld(k) − xid(k) )

 (k+1) (k) (k+1) . (1)
xid = xid +vid
Standard PSO has few parameters which need to be adjusted, and we usually s
elect them empirically:
Population size m is generally selected from 10 to 30, the number of particles is e
nough forthe general problem with small scale, it also reduces the complexity of calc
ulation. The constant limit number of particle velocity Vmax determines the maximu
m moving distance of particles in an iteration cycle. Achieving maximum iterations
or meeting the requirement of accuracy is generally selected as terminal conditions.
3 The PSO Algorithm Consensus Analysis and the Consensus

Region Boundary
In the fields of biological evolution, consensus problems are especially reflected in
self-organizing aggregation of biological systems, such as flocks of birds [3], schools
of fish [4], mammals [5]. Jiang and Jin [6, 7] mainly investigate the stochastic
convergence of PSO system from the perspective of stochastic process. In order to
solve the consensus problem of PSO, we think of a way to improve the velocity and
position updating formula of single particle according to the existing standard PSO
algorithm, this paper improved PSO by consensus Protocol U:
U =vi(k+1) =ω⋅vi(k) +ηφi(1k) ⋅(pi(k) −xi(k) )+ηφi(2k) ⋅(pl(k) −xi(k) ) . (2)
436 X. Xiao and Q. Zhang
The new PSO model is further described as:

U = ω ⋅ v(k ) +ηφ(k ) ⋅ ( p(k ) − x(k ) ) +ηφ(k ) ⋅ ( pl(k ) − x(k ) )
i i1 i i i2 i
 (k +1) ( k ) . (3)
 i
x = xi + U
and then,
xi(k+1) = (1+ ω −ηφi(k ) ) xi(k ) − ω ⋅ xi(k−1) +ηφi(k ) ⋅ pi(k ) . (4)

Particles in PSO achieve the optimization goal and consensus through mutual
cooperation. The cooperative behavior is that the particle and its connected
particles(neighbor) pass on information with each other, and change state according to
certain strategy and the received neighbor information, resulting in a corresponding
self-organizing behavior. Assuming that each particle in PSO population represents a
node, then the interactions between particles form a sensing figure, which decodes
relationship and interaction between particles and their neighbor.
In the improved consensus protocol U,
 0 η  0 .2 5

 η = 0 .2 5 . (5)
 η  0 .2 5

The corresponding region for 0 < η < 0.25 ， 0 ≤ φi (k) <

1
η
(1 + ω ) ，0 ≤φ i
(k)
< 4(1 + ω ) is in
1
complete consensus. The corresponding region for 4(1 + ω ) < φi (k) < (1 + ω ) is
η
inconsistent state area. When η =0.25 ，
φi (k) = 4(1 + ω ) is the critical value for
，
consensus area; when η > 0.25 the corresponding region for 0 ≤ φi (k) < 4(1 + ω ) is
complete consensus area.
Based on the research of PSO consistency theory, φi(k) = 4(1+ω) are the critical value
for consensus area. According to the boundary value, aiming at the shortcomings of the
particle swarm algorithm itself, this paper puts forward an improved particle swarm
optimization algorithm: particle swarm optimization algorithm of multiple population
co-evolution.
When w = 0.79, select a few particle swarms , the values of c within these particle
swarms are different from each other, and all values are in consensus area. Multiple
particle swarms search solution space independently, this way can enhance global
searching capability. The results of several test functions using the improved algorithm
and the standard particle swarm algorithm demonstrate the effectiveness of the
improved algorithm.
4 The Multiple Population Co-evolution PSO Algorithm

Experiment parameter:
Number of population: 3
Population size: 10
Particle neighborhood size: 2
、、
Dimension: 2 10 30
The scope of the search space: [-100 100]D

Maximum iterations: 10000*dimension
The number of goal: 1
Accuracy: 2.22e-16
Inertia weight: 0.8
Learning factor: for swarm 1, the learning factor is 1.5; for swarm 2 , the learning
factor is 1.2; for swarm 3, the learning factor is 0.9.
experimental environment: MATLAB 7.11.0 R2010b .（）
steps of algorithm:
Step1: Particle swarm initialization, including population number, population scale,
the inertiaweight, learning factor, the initial position and velocity of the particle, etc.
The particle swarm learning factor selection strategy: the inertia weight of each
particle group is 0.8, but learning factor is different, if the swarm is i, then the learning
factor of each swarm is 1.8 -0.3*i respectively.
Step2: Calculate the fitness of each particle of each group
Group update strategy:
Step3: Compare the fitness of each particle of each group with its previous
experience fitness at the best position, if good, then use its current fitness value as the
best fitness value of particles.
Step4: Compare the fitness of each particle in the particle group (fitness) with the
pbest, if good, it will be updated as the particle group best fitness value.
Step5: Compare the best fitness value of each particle group, select the smallest
value as the best fitness value for the particle swarm.
Step6: According to the speed and position update formula, update each particle's
velocity and position.
Step7: Algorithm set termination conditions (usually good enough to adapt to the
value or reaches the maximum iterations and precision) according to specific problems,
if not, then return to step 2; if do, stop the iteration, output the optimal solution.
5 Simulation Experiment and Result Analysis

In simulation experiments, we compare the multiple population co-evolution PSO
algorithm procedure with the result of the given example.m, both use 30-dimension in
the experiments, the test time is five, using function 9 Weierstrass Function as test
function. Run example.m and record the average time of five experiments as T1, run
the multiple population co-evolution PSO algorithm program and record the average
time of five experiments as T2.
T1=57.7936
T2=76.2999
(T2-T1)/T1=0.320214
、、
For dimension 2 10 30, run the multiple population co-evolution PSO algorithm
、
program 51 times respectively, results are saved in 20140418Bit_2d 20140418Bit_10d 、
20140418Bit_30d. Maximum, minimum, average, median, standard variance of fitness
、、
value are recorded in analysis_2d.csv analysis_10d.csv analysis_30d .csv respectively.
Table 1. For 2 dimension
function standard
max min mean median
number deviation
1 0.27622 7.15E-09 0.017526 0.000323 0.052023
2 381.39 0.006224 25.446 6.3034 60.777
3 1.6667 1.6667 1.6667 1.6667 2.24E-16
4 38.605 0.00205 3.9485 0.69603 7.1389
5 4.05E-05 2.18E-12 3.47E-06 2.37E-07 7.69E-06
6 1.73E-21 1.97E-31 5.66E-23 1.29E-28 2.65E-22
7 1.31E-05 1.09E-13 1.59E-06 1.01E-07 3.45E-06
8 4.44E-15 8.88E-16 1.03E-15 8.88E-16 6.96E-16
9 -3.9804 -4 -3.9992 -4 0.002984
10 -0.96952 -1 -0.99485 -0.99821 0.006531
11 2.58E-13 0 6.20E-15 0 3.64E-14
12 54.063 48.585 49.496 49.213 1.0739
13 0.019432 1.61E-08 0.010733 0.019432 0.009166
14 0.007925 5.40E-08 0.000566 0.000182 0.0012
15 0.001914 2.74E-09 0.000241 5.14E-05 0.000388
16 3.45E-16 0 6.76E-18 0 4.83E-17
17 0 0 0 0 0
18 -601.09 -837.97 -771.73 -837.68 102.84
19 1.67E-14 1.35E-32 4.44E-16 1.63E-28 2.39E-15
20 0 0 0 0 0
21 0 0 0 0 0
22 2.59E-31 0 5.08E-33 0 3.62E-32
23 8.9715 8.9715 8.9715 8.9715 0
24 -4.2842 -4.5265 -4.5063 -4.5237 0.043573
25 -1.3921 -1.7107 -1.6329 -1.7106 0.13568
26 0.67181 0.57074 0.60386 0.59685 0.02904
27 -37200000 -37400000 -37300000 -37400000 40169
28 -5.4038 -5.8966 -5.7391 -5.7558 0.11493
29 20.006 19.992 20.001 20.001 0.002386
30 1.0097 0.26706 0.31886 0.26966 0.17126
From analysis_2d.csv we can see that when dimension is 2, the maximum、minimum

、average、median and standard deviation of fitness value are shown in table 1.
From analysis_10d.csv we can see that when dimension is 10, the maximum 、
minimum、average、median and standard deviation of fitness value are shown in table 2.
Function Standard
max min mean median
number deviation
1 28464 10.482 4164.1 1115.3 6835.1
2 9820.6 206.77 3107.7 2103.8 2614.3
3 169.67 169.55 169.56 169.55 0.023359
4 255.7 10.773 78.632 74.071 51.557
5 0.10652 0.001738 0.025106 0.017069 0.024209
6 9.5293 0.090725 5.3674 6.1252 2.7106
7 0.027443 0.000625 0.008021 0.006973 0.005679
8 4.9899 0.002668 2.4819 2.5799 1.0874
9 -16.232 -19.288 -18.477 -18.623 0.6125
10 -6.3833 -8.7012 -7.9124 -7.9542 0.49787
11 14.881 9.27E-05 5.3959 4.6682 4.4417
12 0.55492 0.41675 0.4764 0.47755 0.029455
13 1.3251 0.077766 0.62293 0.57989 0.30117
14 0.12353 0.01847 0.058184 0.052282 0.025136
15 0.46897 0.015289 0.16643 0.12418 0.13413
16 0.49012 0.011708 0.21558 0.20158 0.10959
17 0.35762 0.000104 0.030355 0.015601 0.05335
18 -534.4 -3963.7 -1939.5 -2018.4 393.32
19 0.046597 0.001339 0.008284 0.005039 0.008649
20 0.020475 7.25E-17 0.001305 1.62E-06 0.004377
21 0.79987 0.099873 0.29399 0.29987 0.13916
22 0.047381 2.56E-06 0.003458 0.000649 0.007831
23 33.908 17.788 24.723 24.623 3.2122
24 -23.297 -34.026 -31.592 -31.87 1.9737
25 42.999 42.943 42.945 42.943 0.007724
26 0.010188 0.006475 0.008403 0.008328 0.000751
27 -101790000 -65600000.00 -28200000.00 -24800000.00 1.31E+07

28 -5.3948 -5.9232 -5.663 -5.642 0.13994
29 20.002 20 20 20 0.000393
30 1.0372 1.0097 1.0183 1.0097 0.012889
From analysis_30d.csv , we can see that when dimension is 30, the maximum 、
、、
minimum average median and standard deviation of fitness value are shown in table 3.
Function Standard
max min mean median
number deviation
1 1.65E+05 837.96 40120 28993 35732
2 35919 1041.7 9241.1 7925.3 6736.7
3 4554.3 4510.3 4521.8 4520.2 8.5227
4 122.78 22.076 67.24 64.817 21.807
5 0.12922 0.004461 0.040408 0.034891 0.025361
6 33.922 25.335 29.054 28.94 1.1685
7 0.034895 0.008832 0.018335 0.017374 0.00532
8 4.4911 1.5532 3.0353 3.0689 0.69854
9 -49.45 -55.663 -53.521 -53.673 1.1596
10 -20.794 -26.507 -23.939 -24.092 1.2552
11 29.852 0.001599 1.4679 0.03437 4.9221
12 0.014924 0.013084 0.013988 0.014076 0.000392
13 2.9674 0.61538 1.9113 1.8738 0.49718
14 0.036961 0.009863 0.021115 0.02106 0.005493
15 0.48666 0.044493 0.45524 0.4659 0.060491
16 1.4644 0.34813 0.76603 0.73933 0.2357
17 21.569 1.3109 7.1059 6.8476 3.7275
18 -1813.4 -6055.2 -2495 -1908.9 1145.1
19 0.13058 0.006985 0.025035 0.016779 0.023687
20 15.096 6.43E-06 0.64335 0.081668 2.2122
21 1.1999 0.29987 0.7234 0.69987 0.19554
22 0.13694 0.003823 0.048059 0.041579 0.035101
23 50.619 28.644 39.607 39.267 5.4406
24 -101.22 -110.21 -105.56 -105.43 2.1965
25 1151 1124.1 1128.2 1127.2 4.0849
26 0.000504 0.000441 0.000465 0.000459 1.67E-05
27 -47290000 -290950000 -108730000 -101790000 4.88E+07
28 -5.4465 -5.8975 -5.7366 -5.76 0.11652
29 20.005 20.004 20.005 20.005 0.00016
30 1.0782 1.0372 1.0404 1.0372 0.011123
6 Conclusion
This paper proposes a multiple population co-evolution particle swarm optimization

algorithm, through which the population information exchanges a lot than before and it
is more quickly and efficiently to find the optimal solution. What’s more, multiple
populations and changes of the learning factor can guarantee the diversity of
population, which can effectively improve the inherent defects of the particle swarm
algorithm. The experiment results show that the algorithm has good optimization
ability, and can get satisfactory results in less time.
References
1. Kennedy, J., Eberhart, R.C.: Particle swarm optimization. In: Proceedings of IEEE
International Conference on Neural Networks, vol. 4, pp. 1942–1948 (1995)
2. Eberhart, R.C., Kennedy, J.: A new optimizer using particle swarm theory. In: Proceedings of
the Sixth International Symposium on Micro Machine and Human Science, pp.39–43 (1995)
3. Emlen, J.T.: Flocking behavior in birds. The Auk 69(2), 160–170 (1952)
4. Barlow: Behaviour of teleost fishes. Reviews in Fish Biology and Fisheries 4(1), 126–128
(1994)
5. Gueron, S., Levin, S.A.: R. D. I., The dynamics of herds: from individuals to aggregations.
Journal of Theoretical Biology 182(1), 85–98 (1996)
6. Jiang, M., Luo, Y.P., Yang, S.Y.: Stochastic convergence analysis and parameter selection of
the standard particle swarm optimization algorithm. Information Processing Letters 102(1),
8–16 (2007)
7. Jin, X.L., Ma, L.H., Wu, T.J.: The analysis of pso convergence based on stochastic process.
Acta Automatica Sinica 33(12), 1263–1268 (2007)
Fireworks Algorithm and Its Variants for Solving
ICSI2014 Competition Problems
Shaoqiu Zheng, Lang Liu, Chao Yu, Junzhi Li, and Ying Tan
Department of Machine Intelligence, School of EECS, Peking University, China

Key Laboratory of Machine Perception (MOE), Peking University, China
{zhengshaoqiu,langliu,chaoyu,ljz,ytan}@pku.edu.cn
Abstract. Firework algorithm (FWA) is a newly proposed swarm intel-

ligence based optimization technique, which presents a different search
manner by simulating the explosion of fireworks to search within the po-
tential space till the terminal criterions are met. Since its introduction, a
lot of improved work have been conducted, including the enhanced fire-
works algorithm (EFWA), the dynamic search in FWA (dynFWA) and
adaptive fireworks algorithm (AFWA). This paper is to use the FWA and
its variants to take participate in the ICSI2014 competition, the perfor-
mance among them are compared, and results on 2-, 10-, 30-dimensional
benchmark functions are recorded.
Keywords: ICSI2014 competition, FWA, EFWA, dynFWA, AFWA.
1 Introduction
FWA is a population based swarm intelligence algorithm proposed by Tan and
Zhu [16] in 2010. It takes the inspiration from the phenomenon that the fireworks
explode and illuminate the local space around the fireworks in the night sky. Its
proposed explosion search manner for each firework and cooperative strategy
for allocating the resources among the fireworks swarm make it a novel and
promising algorithm.
Assume that objective function f is a minimization problem with the form
minx∈Ω f (x), and Ω is the feasible search region. The conventional FWA works
as follows: At first, a fixed number of fireworks (N ) are initialized within the
feasible search range, and the quality of the fireworks’ positions are evaluated,
based on which the explosion amplitudes and explosion sparks number are cal-
culated. Here, the principle idea for calculating them is that: the firework with
smaller fitness will have larger number of explosion sparks and smaller explo-
sion amplitude, while the firework with larger fitness will have smaller number
of explosion sparks and bigger explosion amplitude. In addition, to increase the
diversity of the population of the fireworks and explosion sparks, Gaussian mu-
tation sparks are also introduced. After these operations of generating explosion
and Gaussian mutation sparks, selection strategy is performed among the can-
didates set which includes fireworks, explosion sparks and Gaussian mutation


Fireworks Algorithm and Its Variants for Solving ICSI2014 443
sparks, and a fixed number of (N ) fireworks are selected for the next iteration.
The algorithm continues the search until the termination criterions are reached.
Since its first presentation in [16], FWA has attracted a number of researchers
to develop the conventional algorithm and apply the algorithm for optimization
of real world problems. For the algorithm developments, it includes the single ob-
jective algorithm developments [13] [12] [18] [14] [11], multi-objective algorithm
developments [21], hybrid version with other algorithms [20] [2] [4] and paral-
lel implementation versions [3]. For the application, FWA has been applied for
FIR and IIR digital filters design [4], the initialization of Non-negative Matrix
Factorization (NMF) and iterative optimization of NMF [10], [8], [9], spam de-
tection [5], finger-vein identification [19] and power system reconfiguration [7] [6].
Experimental results suggest that FWA is a promising swarm intelligence algo-
rithm, which needs further research and developments.
Motivation and Synopsis: The original motivation of this paper is to let
FWA and its variants to participate the competition in ICSI2014 competition,
and the performance among some typical improved work are compared. The
remainder of this paper is organized as follows: Section 2 briefly introduces the
framework of conventional fireworks algorithm, and the FWA variants are pre-
sented in Section 3, Experiments are given in Section 4 and finally conclusions
are drawn in Section 5.
2 The Conventional FWA
In FWA, it works with a population of fireworks which can generate the explosion
sparks and Gaussian mutation sparks thus to maintain the fireworks swarm
with global and local search abilities. After generating two kinds of sparks, the
selection strategy is performed for the selection of fireworks to the next iteration.
Algorithm 1 gives the framework of conventional FWA.
In FWA, to make a contract among the fireworks and balance between the ex-
ploration and exploitation capacities, the fireworks are designed to take different
explosion amplitudes and explosion sparks number. Assume that the fireworks
number is N , then for each firework, the explosion sparks number si and explo-
sion amplitude Ai are calculated as following:
f (Xi ) − ymin + ε
Ai = Â · N , (1)
i=1 (f (Xi ) − ymin ) + ε
ymax − f (Xi ) + ε
si = M e · N , (2)
i=1 (ymax − f (Xi )) + ε
where, ymax = max(f (Xi )), ymin = min(f (Xi )), and Â and Me are two con-
stants to control the explosion amplitude and the number of explosion sparks,
respectively, and ε is the machine epsilon. In addition, to avoid the overwhelming
effects of fireworks at good/bad locations, the max/min number of sparks are
444 S. Zheng et al.
Algorithm 1. General structure of conventional FWA

1: Initialize N fireworks Xi
2: repeat
3: Explosion operator
4: (i) Calculate explosion amplitude Ai and explosion sparks number si
5: (ii) Generate the explosion sparks
6: for each firework Xi , perform si times do
7: Initialize the location of the “explosion sparks”: X̂i = Xi
8: Calculate offset displacement:
X = Ai × rand(−1, 1)
9: z = round(D ∗ rand(0, 1))
10: Randomly select z dimensions of X̂i
11: for each select dimension of X̂ik do
12: X̂ik = X̂ik +
X
13: if X̂ik out of bounds then
14: X̂ik = Xmin
k
+ |X̂ik | % (Xmax
k
− Xmin
k
)
15: end if
16: end for
17: (iii) Evaluate fitness of newly created explosion sparks
18: end for
19: Gaussian mutation operator
20: (i) Generate the Gaussian sparks
21: for perform Mg times do
22: Randomly initialize the location of the “Gaussian sparks”: X̃i = Xi
23: Calculate offset displacement: e = Gaussian(1, 1)
24: Set z k = round(rand(0, 1)), k = 1, 2, ..., d
25: for each dimension of X̂ik , where z k == 1 do
26: X̃ik = X̃ik × e
27: if X̃ik out of bounds then
28: X̃ik = Xmin
k
+ |X̃ik | % (Xmax
k
− Xmin
k
)
29: end if
30: end for
31: end for
32: (ii) Evaluate fitness of newly created Gaussian sparks
33: Selection strategy
34: (i) Select fireworks for next iteration
35: until termination is met.
bounded by
⎧
⎪
⎨round(aMe ) if si < aMe ,
si = round(bMe ) if si > bMe , (3)
⎪
⎩
round(si ) otherwise.
where, a and b are constant parameters which confine minimal/maximal sparks

number (the range of the sparks number). Then for each firework, the explosion
sparks are generated according to Algorithm 1.
To increase the diversity, Gaussian mutation sparks are generated based on a
Gaussian mutation operator (cf. Algorithm 1).
To retain the information to the next iteration, selection strategy is performed

as most of the swarm intelligence algorithms and evolutionary algorithms. In the
candidates set, the individual with minimal fitness is always selected while for
the rest xi in candidates set, the selection probability pi is calculated as
R(xi )
p(xi ) = (4)
j∈KR(xj )

R(xi ) = d(xi , xj ) = ||xi − xj || (5)
j∈K j∈K
where K is the set of all current locations including original fireworks and both
types of sparks.
3 The Selected Typical Improvement Work
3.1 Enhanced Fireworks Algorithm
Although FWA has shown its great performance when dealing with function
optimization in [16], which outperforms SPSO [1] and CPSO [15] in the selected
benchmark functions, in [18], Zheng et al presented a comprehensive study of
operators in conventional FWA and proposed the enhanced FWA (EFWA). Some
details of the EFWA are as following.
Amplitude of Explosion: In FWA, the explosion amplitude of the best fire-

work is usually very close to 0. In EFWA, a lower bound Amin is introduced to
avoid this problem:
Ai = max(Ai , Amin ), (6)
and Amin is non-linearly decreased with the evaluation times going on:
Ainit − Af inal
Akmin (t) = Ainit − (2 ∗ evalsmax − t)t, (7)
evalsmax
where Ainit and Af inal are the initial and final minimum explosion amplitude,
evalsmax is the maximum evaluation times and t is the current evaluation times.
Generating Sparks: In FWA, the number of to-be-mutated dimensions is

calculated first and the displacement is calculated just once then used for all
the selected dimensions. While in EFWA, for each dimension, an independent
displacement ΔXik = Ai × U (−1, 1) is calculated with the selection probabil-
ity U (0, 1), (U (a, b) denotes the generated random number is under the mean
distribution between a and b).
Xîk = Xik + ΔXik . (8)

446 S. Zheng et al.
Moreover, the way to generate Gaussian sparks makes use of the currently
best location XB to avoid the concentrated search on origin region.
X̃ik = Xik + e ∗ (XB

k
− Xik ), (9)
where e ∼ N (0, 1).

A new-generated spark will be mapped into a random place in the variable
space with uniform distribution if the generated location exceeds the bounds.
Xîk = U (Xkmin , Xkmax). (10)
Selection Strategy: To decrease the computational cost of selection strategy

in FWA (Eq. 4), in EFWA, the best of the set will be selected first while the
rest are randomly selected.
3.2 The dynFWA and AFWA

In FWA and EFWA, the explosion amplitude for fireworks is one of the most key
features relevant to the performance. For each firework, its explosion amplitude
is calculated by Eq. 1. In fact, the fitness of firework’s position is only one kind of
information to characterize the local information around Xi , good position needs
further local search. However, for an optimization problem, the optimization
process is dynamic, the previous static explosion amplitude calculation strategy
only suggests that positions of fireworks are good or bad, not the local region
within the positions of fireworks. So the explosion amplitude calculation method
will lead to a bad local search ability and the experimental results in [18] are
consistent with this idea.
For simplicity, the firework with minimal fitness in the fireworks swarm is
defined as the core firework (CF, XCF ), which has the property that (i) its
fitness is best among the fireworks, (ii) it is always selected to the next itera-
tion. To overcome the limitations presented above, the dynamic search in FWA
(dynFWA) and adaptive fireworks algorithm (AFWA) in [14] [11] are proposed
respectively.
The dynFWA – Dynamic Explosion Amplitude Strategy: Assume that

f (X̂best ) is the minimal fitness among the explosion sparks and f (XCF ) denotes
the fitness of CF. Here, in dynFWA, it concerns whether the generated explosion
sparks can get better fitness than the CF, i.e. the Δf = f (X̂best ) − f (XCF ) .
1) Δf < 0
It means at least one of the newly generated sparks has smaller fitness than CF’s
fitness. If so, the X̂best is probably created by the CF or the rest of fireworks
other than CF. If X̂best is created by CF, in order to speedup the search for
the global optimum, the explosion amplitude of CF will become a bigger one
compared with the current value. If X̂best is created by one firework (Xi ) other
than CF, it has a high chance that the Xi is close to XCF . If Xi is close to XCF ,
the same explosion amplitude strategy for the CF in the next iteration is taken.
If Xi is not close to XCF , then the current explosion amplitude is in fact not
effective for the newly generated CF for search any more. However, as it is hard
to define the closeness and it is believed that the dynamic explosion amplitude
strategy has its ability to adjust the explosion amplitude itself in the following
iterations, so dynFWA just sets the explosion amplitude of newly selected CF
with a increasing amplitude.
2) Δf ≥ 0
It means that none of the explosion sparks has found a position with better
fitness compared to the CF. The reason for this situation is that the explosion
amplitude of firework is too big for CF to search a better position. The CF needs
to narrow down the search range. That is to reduce the explosion amplitude thus
increasing the probability that the fireworks swarm can find a better position.
In fact, if the CF is far away from the global optimal position, increasing
the explosion amplitude is one of the most effective methods to speedup the
convergence. The reduction of the explosion amplitude makes it move towards
the global optimal position, i.e. the CF finding a better solution.
In FWA and EFWA, to increase the diversity of the fireworks swarm, Gaus-
sian mutation sparks are introduced. However, due to the selection method, the
Gaussian mutation sparks do not work effectively as they are designed to, thus
in dynFWA [14], they are eliminated.
The AFWA – Adaptive Fireworks Algorithm: The motivation of AFWA

is to guarantee the progress made in current iteration is bigger than in the
previous iteration [11].
AFWA tries to find the spark whose fitness is the minimal among the can-
didates whose fitness is worse than CF, and whose Infinite Norm is closest to
the best candidate (i.e., Core firework or best spark), then the Infinite Norm
between the found candidate with the firework will be taken as the explosion
amplitude for the next iteration.
Under this explosion amplitude updating strategy, there are two cases. The
first one is that the CF does not generate any good sparks whose fitness is smaller
than the firework, then the fitness of all the sparks are larger than the firework,
and the explosion amplitude will take the Infinite norm between the firework
and the selected candidate. The explosion amplitude in the next iteration will be
reduced. For the second case, it has two situations, and the explosion amplitude
will be amplified according to the simulation with high chance. Moreover, as the
Infinite Norm between the calculated candidate and the firework may change
radically, in AFWA, it introduces the smoothing strategy.
4 Experiments
4.1 Experimental Setup
For the implementation of FWA, EFWA, dynFWA and AFWA in this paper, all
the parameters are taken from [14] without any modifications. The experimental
448 S. Zheng et al.
Table 1. Run time results on f9
T1(s) T2(s) (T2-T1)/T1

dynFWA 28.0029 28.8269 0.0294
AFWA 28.0029 28.4186 0.0148
platform used in the experiments is MATLAB 2011b (Windows 7; Intel Core i7-
2600 CPU @ 3.7 GHZ; 8 GB RAM) while ICSI-2014 competition problems are
used as benchmark functions to compare the performance.
The description of the ICSI-2014 competition benchmark functions is as fol-
lows: It contains 30 functions, and for each function, the feasible range is set to
[−100, 100]. Moreover, to make a comprehensive comparison, in the competition,
three groups of experiments with dimension set as D = 2, 10, 30, and maximum
evaluation times D ∗ 10000 are designed. For each function, the max, min, mean,
median value and standard deviation of 51 times results are recorded.
4.2 Experimental Results
The experimental results can be found in Table 2, Table 3 and Table 4. For the
run time consuming, the experimental runs on f9 of dynFWA and AFWA are
given in Table 1 according to [17].
Table 2. Results for 2D functions

F 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Max 1.63E+03 1.89E+03 1.67E+00 1.13E+03 2.09E+00 1.68E-01 2.12E-03 2.58E+00 -3.74E+00 -4.14E-01 9.37E-03 6.62E+01 7.45E-02 8.05E-02 5.83E-02
Min 8.03E-02 6.43E-01 1.67E+00 5.86E+00 1.39E-03 9.00E-05 2.00E-05 2.00E-05 -4.00E+00 -9.99E-01 1.90E-05 4.88E+01 1.90E-05 1.37E-03 3.40E-05
Mean 1.97E+02 4.23E+02 1.67E+00 1.38E+02 3.39E-01 4.40E-02 6.91E-04 1.01E-01 -3.92E+00 -8.77E-01 1.04E-03 5.54E+01 2.10E-02 9.88E-03 1.03E-02
Median 8.13E+01 2.10E+02 1.67E+00 7.99E+01 2.28E-01 2.37E-02 5.77E-04 4.16E-02 -3.93E+00 -9.10E-01 6.47E-04 5.48E+01 1.94E-02 4.46E-03 4.66E-03
Std 2.96E+02 4.65E+02 2.26E-04 1.83E+02 3.94E-01 4.37E-02 4.89E-04 3.59E-01 6.26E-02 1.17E-01 1.44E-03 4.73E+00 1.14E-02 1.51E-02 1.30E-02
FWA
F 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Max 4.79E-02 5.94E-01 -6.01E+02 8.60E-03 0.00E+00 2.01E-01 1.62E-01 1.27E+01 -4.01E+00 -1.37E+00 1.03E+00 -3.62E+07 -4.82E+00 2.00E+01 1.01E+00
Min 1.13E-03 2.32E-04 -8.38E+02 9.00E-06 0.00E+00 2.60E-05 2.00E-06 8.97E+00 -4.52E+00 -1.71E+00 5.74E-01 -3.74E+07 -5.86E+00 2.00E+01 2.68E-01
Mean 1.54E-02 1.27E-01 -7.89E+02 2.71E-03 0.00E+00 4.35E-02 1.95E-02 9.36E+00 -4.31E+00 -1.50E+00 7.15E-01 -3.72E+07 -5.45E+00 2.00E+01 4.18E-01
Median 1.34E-02 7.02E-02 -8.38E+02 2.81E-03 0.00E+00 7.93E-03 7.86E-03 8.98E+00 -4.31E+00 -1.39E+00 6.85E-01 -3.73E+07 -5.45E+00 2.00E+01 2.96E-01
Std 1.10E-02 1.38E-01 9.06E+01 2.19E-03 0.00E+00 5.57E-02 3.19E-02 8.58E-01 1.16E-01 1.41E-01 9.67E-02 2.22E+05 2.54E-01 6.07E-03 2.17E-01
F 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Max 7.31E+02 4.06E+03 1.67E+00 1.27E+03 7.12E-04 7.14E-01 2.28E-02 2.58E+00 -3.99E+00 -6.65E-01 3.13E-06 8.33E+01 3.56E-01 3.16E-01 4.78E-01
Min 1.16E-04 1.27E-02 1.67E+00 7.69E-01 2.32E-06 2.47E-09 7.02E-07 3.44E-04 -4.00E+00 -1.00E+00 4.02E-10 4.86E+01 6.98E-08 2.30E-02 2.11E-03
Std 1.92E+02 1.13E+03 2.93E-09 3.33E+02 1.14E-04 1.00E-01 9.23E-03 3.61E-01 2.63E-03 7.60E-02 6.92E-07 8.41E+00 6.61E-02 7.72E-02 6.60E-02
EFWA
F 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Max 1.88E-04 1.59E-05 -6.01E+02 4.86E-03 6.14E-32 1.08E-07 1.04E-05 1.07E+01 -2.11E+00 7.38E-01 1.31E+00 -1.46E+07 -4.51E+00 2.00E+01 1.12E+00
Min 1.21E-05 2.68E-08 -8.38E+02 4.91E-10 1.59E-58 3.07E-11 4.61E-09 8.97E+00 -4.52E+00 -1.71E+00 5.72E-01 -3.74E+07 -5.74E+00 1.99E+01 2.67E-01
Mean 8.94E-05 3.09E-06 -6.11E+02 5.68E-04 1.92E-33 1.46E-08 1.54E-06 9.01E+00 -4.09E+00 -1.32E+00 7.65E-01 -3.01E+07 -5.08E+00 2.00E+01 6.71E-01
Median 8.51E-05 1.95E-06 -6.01E+02 1.16E-06 6.74E-39 4.82E-09 5.32E-07 8.97E+00 -4.34E+00 -1.39E+00 7.30E-01 -2.36E+07 -5.07E+00 2.00E+01 1.01E+00
Std 4.64E-05 3.62E-06 4.64E+01 1.57E-03 9.39E-33 2.15E-08 2.04E-06 2.41E-01 6.85E-01 4.69E-01 1.69E-01 7.19E+06 2.68E-01 7.74E-03 3.76E-01
F 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Max 1.07E+03 5.40E+03 1.67E+00 1.45E+03 3.29E-02 1.81E-02 2.36E-02 1.60E-01 -3.87E+00 -9.13E-01 1.47E-03 5.33E+01 1.94E-02 2.85E-01 6.37E-02
Min 3.37E-04 6.11E-02 1.67E+00 1.61E-02 5.16E-05 2.33E-07 2.39E-05 2.45E-04 -4.00E+00 -1.00E+00 1.42E-08 4.86E+01 1.90E-06 2.84E-03 7.42E-05
Std 2.44E+02 9.25E+02 1.16E-06 3.18E+02 6.32E-03 2.85E-03 3.30E-03 3.95E-02 2.45E-02 2.08E-02 2.09E-04 7.73E-01 7.75E-03 6.18E-02 1.55E-02
dynFWA
F 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Max 1.04E-02 7.81E-02 -8.38E+02 5.10E-03 0.00E+00 9.99E-02 1.59E-02 9.13E+00 -4.22E+00 -1.39E+00 9.18E-01 -3.73E+07 -4.73E+00 2.00E+01 1.01E+00
Min 1.19E-04 7.38E-07 -8.38E+02 4.42E-07 0.00E+00 1.97E-08 1.07E-06 8.97E+00 -4.53E+00 -1.71E+00 5.76E-01 -3.74E+07 -5.89E+00 2.00E+01 2.67E-01
Mean 2.31E-03 3.63E-03 -8.38E+02 1.14E-03 0.00E+00 1.98E-03 2.27E-03 8.99E+00 -4.42E+00 -1.57E+00 6.68E-01 -3.74E+07 -5.24E+00 2.00E+01 3.39E-01
Median 1.83E-03 1.54E-04 -8.38E+02 4.24E-04 0.00E+00 4.57E-06 5.31E-04 8.98E+00 -4.46E+00 -1.71E+00 6.45E-01 -3.74E+07 -5.23E+00 2.00E+01 2.72E-01
Std 2.16E-03 1.24E-02 3.38E-04 1.63E-03 0.00E+00 1.40E-02 4.05E-03 3.21E-02 9.62E-02 1.59E-01 7.80E-02 9.24E+02 2.50E-01 5.86E-03 1.45E-01
F 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Max 1.97E+03 4.52E+03 1.67E+00 1.31E+03 3.04E-03 1.27E+01 2.38E-03 4.44E-15 -3.86E+00 -8.82E-01 2.74E-04 5.82E+01 1.94E-02 2.84E-01 7.07E-02
Min 5.52E-02 5.52E-02 1.67E+00 5.15E-04 5.37E-07 1.42E-10 0.00E+00 8.88E-16 -4.00E+00 -1.00E+00 1.07E-14 4.86E+01 0.00E+00 5.11E-04 1.89E-05
Std 4.12E+02 1.23E+03 4.50E-09 3.53E+02 5.87E-04 2.07E+00 4.24E-04 8.44E-16 2.08E-02 2.56E-02 4.87E-05 2.76E+00 4.62E-03 7.09E-02 2.05E-02
AFWA
F 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Max 0.00E+00 0.00E+00 -6.01E+02 4.85E-03 0.00E+00 0.00E+00 0.00E+00 1.07E+01 -3.12E+00 -6.50E-01 1.31E+00 -3.74E+07 -5.04E+00 2.00E+01 5.57E-01
Min 0.00E+00 0.00E+00 -8.38E+02 0.00E+00 0.00E+00 0.00E+00 0.00E+00 8.97E+00 -4.53E+00 -1.71E+00 5.71E-01 -3.74E+07 -5.93E+00 2.00E+01 2.67E-01
Mean 0.00E+00 0.00E+00 -8.29E+02 1.22E-03 0.00E+00 0.00E+00 0.00E+00 9.01E+00 -4.35E+00 -1.49E+00 7.05E-01 -3.74E+07 -5.53E+00 2.00E+01 3.11E-01
Median 0.00E+00 0.00E+00 -8.38E+02 0.00E+00 0.00E+00 0.00E+00 0.00E+00 8.97E+00 -4.43E+00 -1.39E+00 6.67E-01 -3.74E+07 -5.54E+00 2.00E+01 2.70E-01
Std 0.00E+00 0.00E+00 4.45E+01 2.03E-03 0.00E+00 0.00E+00 0.00E+00 2.41E-01 2.61E-01 1.94E-01 1.36E-01 8.99E+01 2.46E-01 8.51E-03 7.74E-02

F 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Max 7.42E+06 1.35E+05 1.82E+02 3.52E+02 4.59E+00 1.63E+02 6.03E-02 7.97E+00 -1.51E+01 -3.20E+00 2.57E+01 6.15E-01 1.76E+00 2.06E-01 2.01E-01
Min 2.25E+05 2.49E+03 1.71E+02 2.28E+01 9.06E-02 5.90E+00 9.08E-03 2.86E+00 -1.88E+01 -8.05E+00 2.36E-01 4.35E-01 2.19E-01 1.28E-02 1.12E-02
Mean 2.78E+06 4.15E+04 1.74E+02 1.19E+02 1.95E+00 3.83E+01 3.39E-02 5.01E+00 -1.70E+01 -5.67E+00 8.25E+00 5.35E-01 8.91E-01 9.38E-02 7.14E-02
Median 2.32E+06 3.35E+04 1.74E+02 1.01E+02 1.57E+00 2.84E+01 3.36E-02 4.83E+00 -1.72E+01 -5.77E+00 7.47E+00 5.36E-01 8.52E-01 8.77E-02 5.92E-02
Std 1.81E+06 2.95E+04 2.57E+00 6.01E+01 1.23E+00 3.35E+01 1.47E-02 1.30E+00 9.77E-01 1.04E+00 6.28E+00 3.99E-02 3.02E-01 4.60E-02 4.34E-02
FWA
F 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Max 1.05E+00 2.54E+01 -1.66E+03 9.41E-01 4.65E+00 2.20E+00 1.71E+01 3.22E+01 -2.10E+01 4.54E+01 1.53E-02 -1.47E+07 -5.16E+00 2.00E+01 1.08E+00
Min 9.41E-02 1.65E+00 -4.12E+03 3.36E-02 1.92E-05 9.99E-02 2.61E-01 2.09E+01 -3.27E+01 4.31E+01 8.55E-03 -9.88E+07 -5.93E+00 2.00E+01 1.01E+00
Mean 4.22E-01 1.14E+01 -2.92E+03 2.65E-01 5.86E-01 7.65E-01 4.39E+00 2.63E+01 -2.78E+01 4.36E+01 1.12E-02 -5.30E+07 -5.53E+00 2.00E+01 1.03E+00
Median 3.76E-01 1.16E+01 -2.95E+03 1.95E-01 5.78E-02 7.00E-01 3.30E+00 2.70E+01 -2.81E+01 4.35E+01 1.12E-02 -4.67E+07 -5.53E+00 2.00E+01 1.04E+00
Std 2.07E-01 6.18E+00 1.04E+03 2.09E-01 1.02E+00 4.65E-01 3.29E+00 2.58E+00 2.66E+00 4.20E-01 1.77E-03 2.44E+07 1.97E-01 1.74E-03 1.50E-02
F 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Max 1.37E+05 2.00E+04 1.70E+02 6.60E+01 2.53E-01 1.00E+01 3.07E+00 2.01E+01 -1.22E+01 8.07E+00 4.25E+01 5.84E-01 3.26E+00 2.02E+00 1.67E+00
Min 1.73E+04 7.47E+02 1.70E+02 3.62E+00 1.07E-02 2.94E+00 1.98E-03 2.99E+00 -1.57E+01 -7.25E+00 3.65E+00 3.90E-01 6.40E-01 8.65E-02 8.09E-02
Mean 5.83E+04 7.52E+03 1.70E+02 1.89E+01 1.12E-01 7.81E+00 8.51E-01 1.66E+01 -1.48E+01 -2.16E+00 2.04E+01 4.96E-01 2.11E+00 7.24E-01 3.02E-01
Median 5.47E+04 6.49E+03 1.70E+02 1.71E+01 1.24E-01 8.46E+00 7.29E-01 2.00E+01 -1.51E+01 -2.53E+00 1.79E+01 5.03E-01 2.12E+00 3.40E-01 3.45E-01
Std 2.81E+04 4.46E+03 4.11E-02 1.37E+01 6.73E-02 1.74E+00 9.63E-01 5.97E+00 8.49E-01 3.63E+00 1.05E+01 4.51E-02 5.64E-01 7.60E-01 2.40E-01
EFWA
F 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Max 1.46E+00 5.94E-02 -3.00E+03 4.82E+00 4.84E+02 6.00E-01 2.28E-01 5.51E+01 -8.86E+00 5.29E+01 1.37E-02 -3.68E+07 -5.13E+00 2.00E+01 1.08E+00
Min 3.79E-02 7.01E-03 -3.00E+03 1.51E-02 0.00E+00 9.99E-02 2.64E-03 2.36E+01 -3.26E+01 4.29E+01 7.05E-03 -1.07E+08 -5.66E+00 2.00E+01 1.01E+00
Mean 2.45E-01 2.44E-02 -3.00E+03 8.23E-01 9.57E+00 2.29E-01 9.15E-02 3.88E+01 -2.18E+01 4.37E+01 9.84E-03 -9.75E+07 -5.37E+00 2.00E+01 1.04E+00
Median 2.11E-01 2.13E-02 -3.00E+03 4.68E-01 3.90E-03 2.00E-01 9.34E-02 3.95E+01 -2.21E+01 4.29E+01 9.81E-03 -1.06E+08 -5.36E+00 2.00E+01 1.04E+00
Std 2.31E-01 1.26E-02 3.33E-02 9.42E-01 6.77E+01 1.25E-01 6.71E-02 7.04E+00 5.33E+00 2.62E+00 1.57E-03 1.87E+07 1.29E-01 4.34E-04 1.68E-02
F 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Max 1.32E+05 5.20E+04 1.70E+02 1.86E+02 2.00E-01 9.71E+00 3.32E-01 4.56E+00 -1.69E+01 -5.32E+00 1.52E+01 5.67E-01 1.66E+00 2.03E-01 1.89E-01
Min 1.80E+03 6.97E+02 1.70E+02 4.78E+00 4.23E-03 8.88E-01 5.74E-04 2.19E-02 -1.94E+01 -8.73E+00 7.35E-04 3.89E-01 2.08E-01 2.83E-02 7.27E-03
Mean 3.18E+04 9.49E+03 1.70E+02 6.48E+01 6.31E-02 6.56E+00 1.64E-02 2.03E+00 -1.84E+01 -7.24E+00 6.54E+00 4.70E-01 9.80E-01 8.83E-02 1.03E-01
Median 2.49E+04 6.61E+03 1.70E+02 5.74E+01 5.46E-02 7.04E+00 4.19E-03 2.04E+00 -1.85E+01 -7.34E+00 5.86E+00 4.63E-01 9.96E-01 8.04E-02 1.04E-01
Std 2.90E+04 9.87E+03 7.52E-02 4.78E+01 5.23E-02 2.51E+00 4.85E-02 1.23E+00 6.30E-01 7.17E-01 4.17E+00 4.42E-02 3.85E-01 4.11E-02 4.01E-02
dynFWA
F 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Max 8.67E-01 1.68E-02 -1.68E+03 1.66E-01 8.76E+00 8.00E-01 2.33E-01 4.20E+01 -2.70E+01 5.33E+01 9.77E-03 -1.91E+07 -5.23E+00 2.00E+01 1.04E+00
Min 3.89E-02 7.06E-05 -4.14E+03 1.32E-03 6.79E-12 9.99E-02 5.62E-05 1.96E+01 -3.51E+01 4.29E+01 6.68E-03 -1.09E+08 -5.94E+00 2.00E+01 1.01E+00
Mean 2.67E-01 2.57E-03 -2.31E+03 2.42E-02 3.95E-01 2.88E-01 4.80E-02 2.90E+01 -3.29E+01 4.36E+01 7.92E-03 -7.83E+07 -5.77E+00 2.00E+01 1.02E+00
Std 1.89E-01 2.70E-03 7.54E+02 3.26E-02 1.44E+00 1.48E-01 5.98E-02 5.18E+00 1.70E+00 2.41E+00 6.37E-04 2.89E+07 1.37E-01 1.28E-04 1.39E-02
F 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Min 2.07E+03 8.97E+02 1.70E+02 1.04E+01 1.12E-02 1.93E+00 1.69E-03 1.18E+00 -1.95E+01 -8.54E+00 1.85E-03 3.89E-01 3.24E-01 1.16E-02 1.48E-02
Std 2.84E+05 1.16E+04 3.96E-01 7.98E+01 2.59E-01 1.24E+01 1.87E-01 1.47E+00 8.05E-01 9.47E-01 6.82E+00 4.17E-02 3.85E-01 4.52E-02 5.26E-02
AFWA
F 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Max 1.23E+00 1.66E+00 -1.23E+03 2.82E-01 2.09E+01 1.90E+00 1.27E+00 4.19E+01 -2.52E+01 5.34E+01 1.15E-02 -1.73E+07 -4.86E+00 2.00E+01 1.04E+00
Min 5.55E-02 7.31E-02 -4.14E+03 1.69E-03 0.00E+00 9.99E-02 0.00E+00 2.00E+01 -3.50E+01 4.29E+01 6.74E-03 -1.06E+08 -5.92E+00 2.00E+01 1.01E+00
Mean 3.41E-01 4.24E-01 -2.22E+03 5.90E-02 7.22E-01 5.19E-01 2.78E-01 3.16E+01 -3.10E+01 4.39E+01 8.69E-03 -5.37E+07 -5.70E+00 2.00E+01 1.03E+00
Std 2.47E-01 3.28E-01 7.49E+02 6.69E-02 3.20E+00 3.38E-01 3.17E-01 5.21E+00 2.02E+00 2.99E+00 9.22E-04 2.86E+07 2.15E-01 2.38E-04 1.33E-02

F 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Min 2.79E+06 6.02E+04 4.79E+03 7.04E+01 1.07E+00 7.51E+01 3.25E-02 4.36E+00 -5.23E+01 -2.21E+01 1.59E+00 1.39E-02 1.48E+00 4.96E-03 8.57E-03
Mean 6.27E+06 2.38E+05 5.16E+03 1.57E+02 2.61E+00 1.36E+02 6.65E-02 6.12E+00 -4.85E+01 -1.80E+01 4.38E+00 1.48E-02 3.68E+00 2.98E-02 3.30E-02
Median 5.37E+06 1.98E+05 5.13E+03 1.55E+02 2.64E+00 1.31E+02 6.47E-02 6.19E+00 -4.86E+01 -1.81E+01 4.42E+00 1.48E-02 3.71E+00 2.85E-02 2.68E-02
Std 2.36E+06 1.11E+05 2.24E+02 3.50E+01 8.88E-01 3.41E+01 2.08E-02 7.97E-01 1.37E+00 2.51E+00 1.37E+00 3.57E-04 1.19E+00 1.73E-02 2.32E-02
FWA
F 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Max 2.59E+00 1.07E+02 -8.94E+03 1.34E+00 3.87E+02 3.60E+00 1.91E+01 7.52E+01 -8.74E+01 1.40E+03 9.67E-04 -3.48E+08 -5.29E+00 2.00E+01 1.18E+00
Min 8.97E-01 1.84E+01 -1.25E+04 3.11E-01 1.29E+00 1.10E+00 4.05E+00 3.67E+01 -1.02E+02 1.18E+03 4.77E-04 -7.54E+08 -5.94E+00 2.00E+01 1.08E+00
Mean 1.68E+00 5.81E+01 -9.31E+03 7.27E-01 4.80E+01 2.08E+00 1.09E+01 5.72E+01 -9.50E+01 1.28E+03 6.19E-04 -5.17E+08 -5.61E+00 2.00E+01 1.10E+00
Median 1.71E+00 5.63E+01 -8.97E+03 7.16E-01 3.36E+01 2.10E+00 1.09E+01 5.59E+01 -9.48E+01 1.28E+03 6.05E-04 -4.37E+08 -5.61E+00 2.00E+01 1.08E+00
Std 3.89E-01 1.73E+01 1.04E+03 2.59E-01 6.31E+01 5.53E-01 3.37E+00 9.05E+00 3.35E+00 4.77E+01 8.01E-05 1.38E+08 1.63E-01 2.22E-04 2.66E-02
F 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Max 3.62E+05 2.71E+04 4.53E+03 2.26E+01 1.62E-01 9.03E+01 3.79E+00 2.02E+01 -4.87E+00 -3.50E-01 1.53E+02 1.51E-02 1.20E+01 2.01E+00 1.92E+00
Min 4.11E+03 3.27E+03 4.51E+03 2.97E+00 2.26E-02 2.70E+01 9.18E-03 6.94E+00 -4.75E+01 -1.98E+01 8.66E+00 1.29E-02 1.86E+00 5.92E-02 4.26E-02
Mean 6.63E+04 1.23E+04 4.52E+03 8.91E+00 7.12E-02 3.03E+01 2.12E+00 1.92E+01 -1.07E+01 -1.27E+01 7.82E+01 1.41E-02 7.32E+00 1.94E+00 8.91E-01
Median 4.60E+04 1.16E+04 4.52E+03 8.13E+00 6.62E-02 2.91E+01 1.99E+00 2.01E+01 -6.34E+00 -1.40E+01 7.43E+01 1.41E-02 7.52E+00 1.98E+00 4.62E-01
Std 7.26E+04 5.59E+03 5.27E+00 4.43E+00 2.96E-02 8.60E+00 8.40E-01 2.93E+00 1.23E+01 4.22E+00 4.07E+01 4.78E-04 2.70E+00 2.69E-01 6.66E-01
EFWA
F 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Max 1.35E+00 3.12E+00 -9.01E+03 4.99E+00 4.80E-05 1.20E+00 2.84E-01 1.26E+02 -7.58E+01 1.13E+03 4.73E-04 -5.08E+08 -5.31E+00 2.00E+01 1.18E+00
Mean 4.92E-01 1.07E+00 -9.01E+03 1.27E+00 1.20E-05 6.86E-01 6.72E-02 9.27E+01 -8.55E+01 1.13E+03 4.31E-04 -6.11E+08 -5.52E+00 2.00E+01 1.09E+00
Std 2.65E-01 6.63E-01 1.80E-01 1.26E+00 9.00E-06 2.10E-01 6.24E-02 2.04E+01 4.91E+00 1.30E+00 1.70E-05 1.50E+08 1.24E-01 2.31E-04 3.09E-02
F 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Min 2.61E+03 4.01E+03 4.51E+03 1.04E+01 8.11E-03 2.74E+01 4.82E-03 9.59E-01 -5.63E+01 -2.61E+01 6.59E-04 1.31E-02 1.52E+00 2.39E-02 3.06E-02
Std 5.21E+04 6.72E+03 6.35E+00 1.14E+01 5.14E-02 2.01E+01 4.62E-03 8.24E-01 1.76E+00 2.02E+00 1.45E+01 5.07E-04 7.50E-01 1.31E-02 1.88E-01
dynFWA
F 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Max 1.21E+00 5.49E+00 -3.68E+03 9.42E-01 4.44E+01 2.00E+00 3.70E-01 7.09E+01 -9.33E+01 1.14E+03 4.72E-04 -3.95E+07 -5.38E+00 2.00E+01 1.13E+00
Mean 6.31E-01 2.17E+00 -5.03E+03 1.95E-01 2.76E+00 8.87E-01 7.58E-02 4.98E+01 -1.04E+02 1.13E+03 4.23E-04 -2.64E+08 -5.80E+00 2.00E+01 1.06E+00
Median 5.94E-01 1.98E+00 -6.05E+03 1.29E-01 5.70E-01 9.00E-01 3.94E-02 4.97E+01 -1.05E+02 1.13E+03 4.19E-04 -2.90E+08 -5.84E+00 2.00E+01 1.04E+00
Std 2.65E-01 1.17E+00 1.18E+03 2.15E-01 7.05E+00 3.07E-01 8.57E-02 9.16E+00 3.57E+00 3.47E+00 1.66E-05 1.02E+08 1.40E-01 2.14E-04 2.46E-02
F 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Min 7.49E+02 2.83E+03 4.51E+03 1.43E+01 1.42E-02 2.69E+01 6.80E-03 2.02E+00 -5.62E+01 -2.59E+01 6.37E-03 1.31E-02 1.34E+00 1.80E-02 1.95E-02
Median 1.97E+04 1.02E+04 4.52E+03 2.69E+01 5.79E-02 2.92E+01 1.40E-02 3.03E+00 -5.28E+01 -2.20E+01 1.35E-01 1.42E-02 2.97E+00 3.95E-02 4.22E-01
Std 6.60E+04 5.84E+03 7.28E+00 1.18E+01 4.48E-02 1.48E+01 6.38E-03 7.93E-01 1.87E+00 2.03E+00 8.29E+00 3.87E-04 9.45E-01 1.20E-02 1.87E-01
AFWA
F 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Max 1.44E+00 6.34E+00 -3.69E+03 6.64E-01 5.80E+01 1.80E+00 1.87E-01 7.59E+01 -9.54E+01 1.14E+03 4.51E-04 -7.18E+07 -5.37E+00 2.00E+01 1.13E+00
Mean 7.35E-01 2.40E+00 -5.39E+03 1.33E-01 3.70E+00 9.78E-01 4.91E-02 5.11E+01 -1.04E+02 1.13E+03 4.23E-04 -3.07E+08 -5.81E+00 2.00E+01 1.07E+00
Median 7.15E-01 2.21E+00 -6.05E+03 8.82E-02 3.90E-01 1.00E+00 3.38E-02 5.02E+01 -1.04E+02 1.13E+03 4.21E-04 -2.91E+08 -5.85E+00 2.00E+01 1.08E+00
Std 2.83E-01 1.34E+00 1.45E+03 1.55E-01 9.30E+00 3.48E-01 4.71E-02 8.84E+00 3.47E+00 3.17E+00 1.20E-05 1.53E+08 1.36E-01 2.07E-04 2.76E-02
450 S. Zheng et al.
From the results in different dimension, it can be seen that with the increasing
of the dimension, the results optimized by all the algorithms get worsen, which
is usually called “dimension of curse”. From the run time results in Table 1, it
can be seen that AFWA achieve smaller (T 2 − T 1)/T 1 than dynFWA. Here we
also need to point out that the implementation of the code is one of the core
factors to influence the run time.
From the results of 2D functions in Table 2, it can be seen that AFWA achieves
better results than FWA, EFWA and dynFWA. Especially on f16 , f17 , f20 , f21 ,
f22 , AFWA gets the optimum of these functions. Table 3 gives the results of
10D functions. The dynFWA and AFWA still outperform EFWA and FWA. For
the comparison between dynFWA and AFWA, dynFWA achieves smaller mean
fitness. Table 4 shows the results on 30D functions. None of the algorithms
works well, since all the maximum and minimum are different for each function.
The dynFWA and AFWA still outperform EFWA and FWA due to their great
local search ability, while the performances of dynFWA and AFWA do not differ
much.
5 Conclusion
In this paper, the FWA and its variants are used to take the ICSI2014 competi-
tion for solving competition problems which contains 30 functions, and the three
groups of experimental results with the dimensions set to 2, 10, 30 are recorded.
In the competition, the error smaller than 2−52 ≈ 2.22e−16 is set to 0. It can be
seen that for some functions, the most recent work dynFWA and AFWA still can
not get the optimum, thus further research needs to be taken and it is believed
that there is a long way to go for fireworks algorithm in the future.
Acknowledgements. This work was supported by National Natural Science

Foundation of China (NSFC), Grant No. 61375119, No. 61170057 and No.
60875080.
References
1. Bratton, D., Kennedy, J.: Defining a standard for particle swarm optimization. In:
Swarm Intelligence Symposium, SIS 2007, pp. 120–127. IEEE (2007)
2. Yu, C., Kelley, L., Zheng, S.: Fireworks algorithm with differential mutation for
solving the cec 2014 competition problems. In: 2014 IEEE Congress on Evolution-
ary Computation (CEC). IEEE (2014)
3. Ding, K., Zheng, S., Tan, Y.: A gpu-based parallel fireworks algorithm for op-
timization. In: Proceeding of the Fifteenth Annual Conference on Genetic and
Evolutionary Computation Conference, GECCO 2013, pp. 9–16. ACM, New York
(2013), http://doi.acm.org/10.1145/2463372.2463377
4. Gao, H., Diao, M.: Cultural firework algorithm and its application for digital filters
design. International Journal of Modelling, Identification and Control 14(4), 324–
331 (2011)
5. He, W., Mi, G., Tan, Y.: Parameter optimization of local-concentration model for
spam detection by using fireworks algorithm. In: Tan, Y., Shi, Y., Mo, H. (eds.)
ICSI 2013, Part I. LNCS, vol. 7928, pp. 439–450. Springer, Heidelberg (2013)
6. Imran, A.M., Kowsalya, M.: A new power system reconfiguration scheme for power
loss minimization and voltage profile enhancement using fireworks algorithm. In-
ternational Journal of Electrical Power & Energy Systems 62, 312–322 (2014)
7. Imran, A.M., Kowsalya, M., Kothari, D.: A novel integration technique for optimal
network reconfiguration and distributed generation placement in power distribution
networks. International Journal of Electrical Power & Energy Systems 63, 461–472
(2014)
8. Janecek, A., Tan, Y.: Iterative improvement of the multiplicative update nmf al-
gorithm using nature-inspired optimization. In: 2011 Seventh International Con-
ference on, Natural Computation (ICNC), vol. 3, pp. 1668–1672. IEEE (2011)
9. Janecek, A., Tan, Y.: Swarm intelligence for non-negative matrix factorization.
International Journal of Swarm Intelligence Research (IJSIR) 2(4), 12–34 (2011)
10. Janecek, A., Tan, Y.: Using population based algorithms for initializing nonnega-
tive matrix factorization. In: Tan, Y., Shi, Y., Chai, Y., Wang, G. (eds.) ICSI 2011,
Part II. LNCS, vol. 6729, pp. 307–316. Springer, Heidelberg (2011)
11. Junzhi Li, S.Z., Tan, Y.: Adaptive fireworks algorithm. In: 2014 IEEE Congress on
Evolutionary Computation (CEC). IEEE (2014)
12. Liu, J., Zheng, S., Tan, Y.: The improvement on controlling exploration and ex-
ploitation of firework algorithm. In: Tan, Y., Shi, Y., Mo, H. (eds.) ICSI 2013, Part
I. LNCS, vol. 7928, pp. 11–23. Springer, Heidelberg (2013)
13. Pei, Y., Zheng, S., Tan, Y., Hideyuki, T.: An empirical study on influence of ap-
proximation approaches on enhancing fireworks algorithm. In: Proceedings of the
2012 IEEE Congress on System, Man and Cybernetics, pp. 1322–1327. IEEE (2012)
14. Zheng, S., Andreas, J., Li, J., Tan, Y.: Dynamic search in fireworks algorithm. In:
2014 IEEE Congress on Evolutionary Computation (CEC). IEEE (2014)
15. Tan, Y., Xiao, Z.: Clonal particle swarm optimization and its applications. In:
IEEE Congress on Evolutionary Computation, CEC 2007, pp. 2303–2309. IEEE
(2007)
16. Tan, Y., Zhu, Y.: Fireworks algorithm for optimization. In: Tan, Y., Shi, Y., Tan,
K.C. (eds.) ICSI 2010, Part I. LNCS, vol. 6145, pp. 355–364. Springer, Heidelberg
(2010)
17. Tan, Y., Li, J., Zheng, Z.: Icsi 2014 competition on single objective optimization
(2014)
18. Zheng, S., Andreas, J., Tan, Y.: Enhanced fireworks algorithm. In: 2013 IEEE
Congress on Evolutionary Computation (CEC), pp. 2069–2077. IEEE (2013)
19. Zheng, S., Tan, Y.: A unified distance measure scheme for orientation coding in
identification. In: 2013 IEEE Congress on Information Science and Technology, pp.
979–985. IEEE (2013)
20. Zheng, Y., Xu, X., Ling, H.: A hybrid fireworks optimization method with differ-
ential evolution. Neurocomputing (2012)
21. Zheng, Y.J., Song, Q., Chen, S.Y.: Multiobjective fireworks optimization for
variable-rate fertilization in oil crop production. Applied Soft Computing 13(11),
4253–4263 (2013)
Performance of Migrating Birds Optimization
Algorithm on Continuous Functions
Ali Fuat Alkaya1 , Ramazan Algin1 , Yusuf Sahin2 ,

Mustafa Agaoglu1 , and Vural Aksakalli3
1
Marmara University, Department of Computer Engineering, Istanbul, Turkey
2
Marmara University, Department of Electrical and Electronics Engineering,
Istanbul, Turkey
3
Istanbul Sehir University, Department of Industrial Engineering, Istanbul, Turkey
{falkaya,ysahin,agaoglu}@marmara.edu.tr,
algin.ramazan@gmail.com, aksakalli@sehir.edu.tr
Abstract. In this study, we evaluate the performance of a recently pro-

posed metaheuristic on several well-known functions. The objective of
this evaluation is to participate in a competition where several meta-
heuristics are compared. The metaheuristic we exploit is the recently
proposed migrating birds optimization (MBO) algorithm. Our contribu-
tion in this study is to develop a novel neighbor generating function for
MBO to be used in multidimensional continuous spaces. After a set of
preliminary tests presenting the best performing values of the parame-
ters, the results of computational experiments are given in 2, 10 and 30
dimensions.
Keywords: migrating birds optimization, continuous functions, single

objective optimization.
1 Introduction
The MBO algorithm is a newly proposed, population-based neighborhood search

technique inspired from the V formation flight of the migrating birds which is
proven to be an effective formation in energy minimization. In the analogy,
initial solutions correspond to a flock of birds. Likewise the leader bird in the
flock, a leader solution is chosen and the rest of the solutions is divided into two
parts. Each solution generates a number of neighbor solutions. This number is
a determiner value on exploration and it corresponds to the speed of the flock.
The higher this value, the more detailed the flock explores its surroundings.
The algorithm starts with a number of initial solutions corresponding to birds
in a V formation. Starting with the first solution (corresponding to the leader
bird) and progressing on the lines towards the tales, each solution is tried to be
improved by its neighbor solutions. If any of the neighbor solutions is better,
the current solution is replaced by that one. There is also a benefit mechanism
for the solutions (birds) from the solutions in front of them. Here we define the

Performance of Migrating Birds Optimization Algorithm 453
benefit mechanism as sharing the best unused neighbors with the solutions that
follow. In other words, a solution evaluates a number of its own neighbors and
a number of best neighbors of the previous solution and is replaced by the best
of them. Once all solutions are improved (or tried to be improved) by neighbor
solutions, this procedure is repeated a number of times (tours) after which the
first solution becomes the last, and one of the second solutions becomes the
first and another loop starts. The algorithm is terminated after a predetermined
number of neighbors are generated. Pseudocode of our MBO is given in Figure 1.
1. Generate n initial solutions in a random manner and place them on an hypothetical

V formation arbitrarily
2. while termination condition is not satisfied
3. for m times
4. Try to improve the leading solution by generating and evaluating k
neighbors of it
5. for each solution si in the flock (except leader)
6. Try to improve si by evaluating (k-x) neighbors of it and x unused
best neighbors from the solution in the front
7. endfor
8. endfor
9. Move the leader solution to the end and forward one of the solutions
following it to the leader position
10. endwhile 11. return the best solution in the flock
Fig. 1. Pseudocode of the MBO
MBO algorithm has four parameters: number of solutions (n), number of

tours (m), number of neighbor solutions to be generated from a solution (k) and
number of solutions to be shared with the following solution (x). However, due
to the inherent design of the algorithm n value has to be equal to or greater
than 2 ∗ x + 1.
This new metaheuristic was proposed by Duman et al. [1]. They applied it to
solve quadratic assignment problem instances arising from printed circuit board
assembly workshops. Its performance was compared with those of metaheuris-
tics implemented and compared in two previous studies. These metaheuristics
are simulated annealing, tabu search, genetic algorithm, scatter search, parti-
cle swarm optimization, differential evolution and guided evolutionary simulated
annealing. In this comparison, the MBO outperformed the best performed meta-
heuristic (simulated annealing) in the previous studies by approximately three
percent on the average. In addition, MBO was tested with some benchmark
problem instances obtained from QAPLIB and in most of the instances it ob-
tained the best known solutions. As a result of these tests, it is concluded that
the MBO is a promising metaheuristic and it is a candidate to become one of
the highly competitive metaheuristics. Duman and Elikucuk [2] applied MBO to
454 A.F. Alkaya et al.
solve fraud detection problem. They also proposed a new version of MBO where
a different benefit mechanism is used. They tested the original MBO algorithm
and its new version on real data and compared their performance with that of
genetic algorithm hybridized with scatter search (GASS). Test results showed
that the MBO algorithm and its new version performed significantly better than
the GASS algorithm.
In this study, we exploit MBO to solve problems in continuous environments.
The set of functions used are given in [3] which are tried to be minimized on
2, 10 and 30 dimensional spaces. The search space is [−100, 100]D where D
is the dimension. We believe that defining an effective neighboring function is
much more important than any other modifications on the MBO. In line with
this observation, our contribution in this study is to develop a novel neighbor
generating function for MBO to be used in multidimensional continuous spaces.
In the next section we present an effective neighbor generating function de-
signed for MBO. Section three presents experimental setup which includes pa-
rameter analysis of the MBO algorithm. Section four gives the results where
MBO is run on 30 different functions and various dimensions. Section five gives
the conclusive remarks together with some future work.
2 A Novel Neighbor Generating Function
In order to design a well performing MBO algorithm, an effective neighbor gen-

erating function is essential. To have a more effective exploration plan in the D
dimensional solution space, we used D dimensional spheres (D-spheres for short
throughout this paper). A neighbor of a solution can be obtained only within
the D-sphere around it. A neighbor of a solution can be at most r units away
from the original solution where r is the radius of the D-sphere that surrounds
it. To find the radius of a D-sphere, we firstly calculate the volume allocated to
it using the following formula.
VD = T V /n (1)
where VD is the volume of a D-sphere and T V is the total volume of the
solution space. In order to calculate the radii for the D-spheres in a D dimen-
sional space, the volume of the solution space is divided by n. In this way, we
try to make an effective exploration and fair distribution of volume for all birds
(solutions) to fly around. When the volume of a D-sphere is calculated, we need
to find the radius of the D-sphere. The following inductive formulas give the
volumes of D-spheres.
V1 = 2 ∗ r (2)
V2 = π ∗ r2 (3)
VD = VD−2 ∗ 2 ∗ π/r f orD > 2
D
(4)
Once the volume for each sphere is calculated, the radii of each sphere can be
easily calculated using Equations(2-4). After calculating the radius of D-sphere,
we can develop a method to find a neighbor solution (point) within the sphere
using some trigonometry. The distance that how far will the new solution be
away from the original solution will be a random number in [0, r] where r is the
radius of the sphere.
Additionally, we also need to determine the location (coordinate in each axis)
of the point in the D dimensional space. For this, we used the following set of
trigonometric formula.
xD = l ∗ cos(θD−1 )
xD−1 = l ∗ sin(θD−1 ) ∗ cos(θD−2 )
xD−2 = l ∗ sin(θD−1 ) ∗ sin(θD−2 ) ∗ cos(θD−3 )
...
x2 = l ∗ sin(θD−1 ) ∗ sin(θD−2 ) ∗ ... ∗ sin(θ2 ) ∗ cos(θ1 )
x1 = l ∗ sin(θD−1 ) ∗ sin(θD−2 ) ∗ ... ∗ sin(θ2 ) ∗ sin(θ1 )
where l is the distance that how far will the new solution be away from the
original solution, xi is the coordinate of the point in the ith axis and θi is the
angle between ith and (i + 1)th axis. Before using this set of formula θi ’s must
be obtained randomly such that θ1 ∈ [0, 2π] and θi ∈ [0, π] for i = 2, . . . , D − 1.
An example for the formulas given above is presented in Figure 2 for D = 3.
Fig. 2. Representation of a point (solution) and the vectors constituting it in three

dimensions
From this setting, one can easily observe that if the number of birds (solutions)
is small, then the volume that they are going to explore will be large whereas
if the number of birds is large, the volume that they are going to explore will
be small. Since we are limited by the number of function evaluation (neighbor
generations) due to the competition rules, with a large number of solutions we
will be able to explore small number of neighbors in smaller regions whereas with
small number of solutions we will be able to explore large number of neighbors
in larger regions. Hence, an efficient value for the n parameter must be found
for the best performance of the algorithm.
3 Experimental Setup
The experiments are run on an HP Z820 workstation with Intel Xeon E5 proces-
sor at 3.0 GHz with 128 GB RAM running Windows 7. The MBO algorithm is
implemented in Java language. The stopping criterion for the MBO algorithm is
a given number of function evaluations which correspond to number of neighbors
generated. Specifically the allowed number of function evaluations is 10000*D.
Table 1. Statistics of 51 runs on 30 different functions when D=2

Function ID min max avg med std
1 40.82 19721.55 1951.86 1113.82 2964.70
2 2.48 4532.26 408.61 98.97 751.68
3 1.67 1.67 1.67 1.67 0.00
4 15.22 4166.54 488.89 227.25 704.77
5 0.17 1.20 0.57 0.59 0.24
6 0.03 3.58 0.95 0.82 0.80
7 0.00 0.01 0.00 0.00 0.00
8 0.91 6.44 3.99 4.17 1.17
9 0.07 0.43 0.26 0.26 0.10
10 0.02 0.39 0.20 0.21 0.10
11 0.01 1.10 0.35 0.23 0.36
12 0.08 1.28 0.55 0.55 0.30
13 0.02 0.08 0.02 0.02 0.01
14 0.23 1.21 0.73 0.75 0.21
15 0.28 2.62 0.98 0.81 0.54
16 0.01 0.15 0.07 0.06 0.04
17 0.01 0.82 0.16 0.12 0.15
18 -861.55 0.00 -840.06 -837.83 5.82
19 0.01 0.28 0.09 0.09 0.06
20 0.39 7.27 2.85 2.97 1.45
21 0.02 1.38 0.51 0.47 0.32
22 0.03 1.44 0.34 0.21 0.34
23 0.32 10.86 7.77 8.34 2.01
24 17.73 7069.62 551.74 261.12 1056.58
25 2.13 2390.53 355.57 138.30 535.51
26 1.33 12.67 2.75 2.03 1.92
27 -41721.46 0.00 -38311.70 -38587.87 2274.27
28 -1.57 0.00 -0.88 -0.87 0.26
29 8.47 17.97 12.75 12.35 2.77
30 0.05 0.84 0.33 0.33 0.16
In order to reveal the best performing parameter values of the MBO on the con-
tinuous functions, we run a set of extensive computational experiments. These pre-
liminary tests are conducted on six functions selected out of 30 given in [3]. Best
performing values for the parameters are as follows: n = 5001, k = 3, m = 1, x = 1.
4 Results
In this section we provide the results of the MBO algorithm on the aforementioned
continuous benchmark functions. One of the results to be delivered as a rule of the
competition is the T 1 and T 2 values. T 1 is the average run time of five runs of the
following piece of MATLAB code in our environment.

1 20323203.08 109661239.94 61390076.47 63430416.62 18862528.62
2 240.72 1998.77 801.74 755.39 358.08
3 192.16 304.70 247.67 248.40 21.37
4 354.30 2204.80 1001.13 934.47 468.90
5 1.09 1.64 1.35 1.36 0.12
6 477.57 5023.73 1975.59 1786.44 1048.17
7 0.17 1.53 0.84 0.83 0.25
8 7.04 11.55 9.73 9.74 1.01
9 3.22 5.43 4.75 4.91 0.51
10 1.25 2.12 1.72 1.75 0.22
11 11.01 35.33 27.20 28.07 5.30
12 0.66 2.42 1.54 1.60 0.37
13 1.54 2.58 2.13 2.13 0.25
14 5.30 13.38 8.59 8.56 2.05
15 33.46 214.90 120.04 122.06 42.16
16 1.69 3.46 2.71 2.75 0.43
17 578.73 4028.84 1831.87 1830.40 681.23
18 -3470.64 0.00 -2717.47 -2617.14 376.96
19 1.97 8.05 4.56 4.36 1.30
20 5.90 13.83 9.96 9.91 1.73
21 5.10 19.54 13.41 13.80 3.44
22 38.72 213.23 122.00 126.23 38.49
23 29.19 43.87 38.97 39.58 2.96
24 395.74 1997.54 980.50 935.94 390.40
25 217.89 3921.61 1169.83 1048.91 596.34
26 838.83 6476.15 2421.21 2310.05 1154.72
27 -19655.46 0.00 -13727.18 -13385.63 2203.31
28 14.54 15.72 15.14 15.17 0.29
29 21.37 21.38 21.37 21.37 0.00
30 1.08 1.35 1.22 1.23 0.07

1 131323505.00 371642428.18 272954269.37 278704299.47 49176717.62
2 370.11 1232.52 730.82 729.31 191.71
3 19702.97 36266.85 28845.16 29096.35 4050.11
4 397.31 2204.13 920.60 891.43 351.70
5 1.04 1.23 1.10 1.10 0.04
6 1607.06 12377.82 7458.20 7420.84 2176.52
7 1.80 3.42 2.80 2.80 0.39
8 10.20 11.89 11.05 11.11 0.46
9 13.70 17.27 15.64 15.67 0.88
10 2.97 4.82 3.96 4.06 0.42
11 81.94 140.23 114.50 117.65 13.35
12 2.77 4.47 3.74 3.74 0.39
13 8.00 10.33 9.18 9.27 0.58
14 6.30 11.09 8.99 9.18 1.14
15 194.15 371.56 298.56 311.88 42.88
16 5.71 8.75 7.47 7.60 0.72
17 35345.91 77582.40 54386.92 55674.38 10061.42
18 -5215.18 0.00 -4565.75 -4527.92 252.22
19 8.43 17.05 11.84 11.86 1.89
20 5.71 9.16 7.86 8.05 0.70
21 20.15 38.43 30.42 30.40 4.19
22 189.36 418.69 316.94 322.05 59.04
23 100.05 116.60 108.54 108.75 3.99
24 603.02 1419.04 951.32 938.95 199.66
25 11240.74 29044.32 19293.43 19268.09 4740.83
26 1980.60 11722.33 7516.65 8055.23 2296.92
27 -6756.60 0.00 -5080.52 -5009.40 753.36
28 54.70 55.72 55.18 55.18 0.28
29 21.61 21.62 21.61 21.61 0.00
30 1.38 1.47 1.43 1.44 0.02
for i=1:300000
evaluate(9,rand(30,1)*200-100);
end
T 2 is the average run time of five runs of the function 9 on D=30 in our environ-
ment.
According to our experimental work T 1, T 2 and (T 2-T 1)/T 1 are as follows:
– T 1=29.965
– T 2=73.369
– (T 2-T 1)/T 1=1.448
Table 1, 2 and 3 present the statistics when D=2, 10 and 30, respectively. The
major observation among the tables is that in higher dimensions the performance
of the MBO algorithm gets worse. This is an expected result because the search
space grows much faster than the allowed number of function evaluations.
5 Conclusion
In this study we applied migrating birds optimization algorithm to 30 different
functions on continuous domain. Our contribution in this study is to develop an
effective novel neighbor generation function for MBO. The tests are conducted on
2, 10 and 30 dimensions. Results present that even though MBO is a recently pro-
posed algorithm it is also promising for problems in continuous domain.
References
1. Duman, E., Uysal, M., Alkaya, A.F.: Migrating Birds Optimization: A New Meta-
heuristic Approach and Its Performance on Quadratic Assignment Problem. Infor-
mation Sciences 217, 65–77 (2012)
2. Duman, E., Elikucuk, I.: Solving Credit Card Fraud Detection Problem by the New
Metaheuristics Migrating Birds Optimization. In: Rojas, I., Joya, G., Cabestany, J.
(eds.) IWANN 2013, Part II. LNCS, vol. 7903, pp. 62–71. Springer, Heidelberg (2013)
3. Website of Fifth International Conference on Swarm Intelligence,
http://www.ic-si.org/competition
Author Index
Abdul-Kareem, Sameem I-284 Chiroma, Haruna I-284

Abdullahi Muaz, Sanah I-284 Chu, Hua I-442
Agaoglu, Mustafa II-452 Crawford, Broderick I-189
Agrawal, Puja II-212 Cui, Yu I-350
Akhmedova, Shakhnaz I-499
Aksakalli, Vural II-452 Diao, Liang I-442
Al-Betar, Mohammed Azmi II-87 Diez, Matteo I-126
Alejo, Roberto II-17 Ding, Ke II-66
Algin, Ramazan II-452 Ding, Sheng II-221
Alkaya, Ali Fuat II-452 Ding, Shuyu II-228
An, Xueli II-146 Djenouri, Youcef II-50
Anto, P. Babu II-275 Drias, Habiba II-50
Anwar, Khairul II-87 Du, Huimin II-114, II-125
Arvin, Farshad I-1 Du, Mingyu I-74
Awadallah, Mohammed A. II-87 Duan, Haibin II-96
Ba-Karait, Nasser Omer II-352

Emre Turgut, Ali I-1
Bao, Aorigele I-246
Batouche, Mohamed I-450
Beegom, A.S. Ajeena II-79 Fasano, Giovanni I-126
Bellotto, Nicola I-1 Feng, Qianqian I-267
Beltaief, Olfa I-9 Feng, Tao I-374
Benmounah, Zakaria I-450 Folly, Komla A. II-135
Bian, Zijiang II-34 Fu, Xiaowei II-221
Boulesnane, Abdennour II-412 Fu, Yao-Wei I-342, II-42
Cai, Hongfei II-204 Gao, Chao I-27, I-173, I-424

Cai, Zhen-Nao I-342 Gao, Jie II-188
Campana, Emilio F. I-126 Gao, Shangce I-246
Cao, Lianlian II-221 Gao, Xiaozhi I-86
Chen, Beibei I-246 Gao, Yang I-223
Chen, Hanwu I-357 Garro, Beatriz Aurora I-207
Chen, Hua I-95 Geng, Huang II-34
Chen, Hui-Ling I-342, II-42 Geng, Mengjiao I-103, I-115
Chen, Junfeng I-95 Ghazali, Rozaida I-197
Chen, Li II-221 Ghedira, Khaled I-9
Chen, Min-You I-394 Giove, Silvio I-126
Chen, Qinglan II-236 Gnaneswar, A.V. II-8
Chen, Su-Jie I-342, II-42 Gong, Dunwei I-386
Chen, Xianjun II-58 Gong, Zhaoxuan II-34
Chen, Zhigang I-318 Gu, Jiangshao I-460
Cheng, Shan I-394 Guo, Jian I-142
Cheng, Shi II-319 Guo, Xiaoping II-340
Chenggang, Cui II-309 Guo, Yejun I-294
462 Author Index
Hadouaj, Sameh El I-9 Liang, Xiao-lei I-134

Han, Fei I-350 Liang, XiaoLong I-36
Hao, Junling I-64 Liang, Zhi-feng I-302
He, Jieyue II-180 Lin, Heng II-204
He, Nana I-302 Lin, Sheng-Min II-267
He, Ping II-1 Lin, Yuan-Lung I-158
Herawan, Tutut I-197, I-284 Liu, Bingyu I-404
Hu, Gang I-394 Liu, Cong I-431
Huang, Huali I-150 Liu, Ju II-155
Huang, Shan-Shan I-342 Liu, Lang II-442
Huang, Yantai II-292 Liu, Lili I-103, I-115
Huang, Yin-Fu II-267 Liu, Wei II-384
Liu, Weifeng II-228
Iemma, Umberto I-126 Liu, Xiyu I-267, I-470
Liu, Yabin I-431
Jain, Aruna I-165 Liu, Yu I-86
Janghel, R.R. II-8 Liu, Yuxin I-173, I-424
Jiang, He I-44 Liu, Zhaozheng I-374
Jiang, Yue II-163 Liu, Zhihao I-357
Jithesh, K. II-275 Loo, Chu Kiong I-332
Johnson, Franklin I-189 López-González, Erika II-17
Lu, Bingbing II-155
Kang, Qi I-294, II-163, II-401 Lu, Lin II-1
Kang, Zhenhua I-470 Lu, Mingli II-236, II-244, II-253
Ke, Liangjun II-301 Lu, Yuxiao I-173, I-424
Khader, Ahamad Tajudin II-87 Lu, Zhigang I-374
Khan, Abdullah I-284 Luo, Wenjie II-170
Khurshid, Aleefia II-212 Lv, Qing II-114, II-125
Kobayashi, Kunikazu I-324 Lv, Yawei II-196
Kuremoto, Takashi I-324
Ma, Chuncheng II-259
Lai, Xiaochen I-44 Ma, Chunsen I-275
Lei, Xiujuan I-74, I-479 Ma, Yinghong I-267
Leotardi, Cecilia I-126 Mabu, Shingo I-324
Li, Bin I-134 Meng, Xianbing I-86
Li, Changhe I-181 Meshoul, Souham I-450, II-412
Li, Fang II-196 Mi, Guyue I-223
Li, Fenglin I-142 Mo, Hongwei I-103, I-115, I-234
Li, Jian-Jun II-106
Li, Jinlong I-27, I-365 Naseem, Rashid I-197
Li, Junzhi II-442 Ni, Qingjian II-114, II-125
Li, Kanwen II-1 Niu, Ben I-150
Li, Li II-319
Li, Lian II-24 Obayashi, Masanao I-324
Li, Li-Juan I-342 Olguı́n, Eduardo I-189
Li, Qingshan I-442
Li, QiuQuan II-42 Pacheco-Sánchez, J. Horacio II-17
Li, Shuai II-284 Palma, Wenceslao I-189
Li, Yiguo I-215 Pan, Luoping II-146
Liang, Jane Jing I-150, II-384 Pan, Qianqian II-114, II-125
Author Index 463
Paredes, Fernando I-189 Tian, Hongjun II-401

Peng, Pengfei I-318 Tian, Xuedong II-170
Peri, Daniele I-126 Tingyu, Gao II-309
Phoa, Frederick Kin Hing I-158
Piao, Yong I-44 Valdovinos, Rosa Marı́a II-17
Pradeepkumar, Dadabada II-363 Vazquez, Roberto Antonio I-207
Qi, Feng I-267 Wan, Shuzhen II-392

Qian, Heng II-1 Wang, Chao II-155
Qin, Alex Kai II-384 Wang, Chaoxue I-275
Qin, Quande II-319 Wang, Cong I-404
Qu, Boyang II-376, II-384 Wang, Cuirong I-404
Wang, Dongyun II-376
Rajasree, M.S. II-79 Wang, Fei II-24, II-253
Ravi, Vadlamani II-363 Wang, Jiabao II-401
Ren, Xiongwei I-318 Wang, Lei I-294, II-163, II-292, II-401
Ren, Yayun II-236 Wang, Lin I-302
Ren, Zhilei I-44 Wang, Qing II-155
Rodrı́guez, Katya I-207 Wang, Shuaiqun I-246
Wang, Tai-Chi I-158
Sahana, Sudip Kumar I-165 Wang, Wanping II-259
Sahin, Yusuf II-452 Wang, YaLi I-36
Sánchez-Crisostomo, Juan Pablo II-17 Wang, Yanmei II-58
Sari, Eka Novita I-284 Weise, Thomas I-27
Semenkin, Eugene I-310, I-499 Wen, Chenglin II-228
Semenkina, Maria I-310 Wen, Kunmei I-460
Serani, Andrea I-126 Wu, Qidi I-294, II-163, II-292, II-401
Shah, Habib I-197 Wu, Xiao-Bei II-422
Shamsuddin, Siti Mariyam II-352 Wu, Yali II-328, II-340
Shang, Zhigang II-376 Wu, Ya-Qi I-350
Sharma, Sanjeev II-8 Wu, Yuheng I-424
Shen, Jiong I-215 Wu, Zhenqiang I-479
Shen, LiMing II-42
Shi, Guangda II-196 Xia, Changhong I-386
Shi, Jian II-244 Xia, Yong I-181
Shukla, Anupam II-8 Xiao, Xuan II-434
Song, Bo II-188 Xiaofei, Yang II-309
Song, Hui II-376, II-384 Xie, A.S. I-19
Soto, Ricardo I-189 Xie, Lixia II-328, II-340
Su, Yanhui II-284 Xie, Yingjuan I-95
Sudirman, Rubita II-352 Xing, Huanlai I-414
Sun, Qiang I-36 Xu, Benlian II-236, II-244, II-253
Sun, Xiaoyan I-386 Xu, Hongji II-155
Xu, Jiao II-24
Taherzadeh, Ghazaleh I-332 Xu, Jin I-74
Tan, Lijing I-150 Xu, Lifang I-234
Tan, Wenjun II-34 Xu, Xiaohua II-1
Tan, Ying I-74, I-223, I-479, I-489, Xu, Yanping II-376
II-66, II-442 Xu, Ying I-414
Tao, Li I-173, I-424 Xue, Jie I-470
464 Author Index
Xue, Puheng II-259 Zhang, Kai I-275

Xue, Xilin I-357 Zhang, Lei I-302
Zhang, Lihang I-442
Yan, Mingying I-365 Zhang, Minxia I-53
Yan, Yiwen II-204 Zhang, Qianqian II-434
Yang, Bo I-302 Zhang, Ru-Bo II-106
Yang, Jinzhu II-34 Zhang, Shuwei I-44
Yang, Lin I-134 Zhang, Wei II-1, II-204
Yang, Xiang I-489
Zhang, Xiaoqian I-302
Yang, Yu II-106
Zhang, Xing I-275
Yao, Shengqiang II-180
Zhang, Yong I-386
Yao, Yiyun II-114, II-125
Zhang, Zili I-173, I-424
Ye, Bin I-150
Zhao, Dazhe II-34
Yin, ZhongHai I-36
Zhao, Jie I-479
Yu, Botao II-170
Zhao, Xinchao I-64, I-258
Yu, Chao II-442
Yue, Shigang I-1 Zheng, Shaoqiu II-442
Zheng, Yu-Jun I-53, II-422
Zeng, Sanyou I-181 Zhou, Qiang I-431
Zhai, Laipeng II-301 Zhou, Xia I-215
Zhan, Yongsong II-58, II-284 Zhu, Peiyi II-236, II-244
Zhang, Bei I-53 Zhu, Wanning I-357
Zhang, Bo II-96 Zhu, Wumei I-275
Zhang, Hengzhen I-86 Zuo, Xingquan I-258

Advances in Swarm Intelligence 2014

Uploaded by

Copyright:

Available Formats

Advances in Swarm Intelligence 2014

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Advances in Swarm Intelligence 2014

Uploaded by

Copyright:

Available Formats

Ying Tan Yuhui Shi

Carlos A. Coello Coello (Eds.)

ISSN 0302-9743 e-ISSN 1611-3349

Library of Congress Control Number: 2014949260

LNCS Sublibrary: SL 1 – Theoretical Computer Science and General Issues

friends, and supporters who helped us in immeasurable ways; we express our

July 2014 Ying Tan

Programme Committee Chairs

Advisory Committee Chairs

Technical Committee Chairs

Special Sessions Chairs

Competition Session Chairs

Finance and Registration Chairs

Local Arrangement Chairs

Mu-Song Chen Da-Yeh University, Taiwan

Franziska Klügl Örebro University, Sweden

Ke Tang University of Science and Technology of China,

Alves, Felipe Rakitianskaia, Anna

Scheduling and Path Planning

Development on Harmony Search Hyper-heuristic Framework for

Predator-Prey Pigeon-Inspired Optimization for UAV

Research on Route Obstacle Avoidance Task Planning Based on

Wireless Sensor Network

An Improved Energy-Aware Cluster Heads Selection Method for

Power System Optimization

Vibration Adaptive Anomaly Detection of Hydropower Unit in Variable

Capacity and Power Optimization for Collaborative Beamforming with

Solving Power Economic Dispatch Problem Subject to DG Uncertainty

An Eﬃcient OLAP Query Algorithm Based on Dimension Hierarchical

The Enhancement and Application of Collaborative Filtering in

A Method to Construct a Chinese-Swedish Dictionary via English

The Design and Implementation of the Random HTML Tags and

Special Session on Swarm Intelligence in Image and

Application and Comparison of Three Intelligent Algorithms in 2D

A Shape Target Detection and Tracking Algorithm Based on the Target

Multi-cell Contour Estimate Based on Ant Pheromone Intensity

A Cluster Based Method for Cell Segmentation . . . . . . . . . . . . . . . . . . . . . . 253

Research on Lane Departure Decision Warning Methods Based on

Searching Images in a Textile Image Database . . . . . . . . . . . . . . . . . . . . . . . 267

IIS: Implicit Image Steganography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275

Humanized Game Design Based on Augmented Reality . . . . . . . . . . . . . . . 284

Special Session on Applications of Swarm Intelligence

A Multiobjective Large Neighborhood Search for a Vehicle Routing

A Self-adaptive Interior Penalty Based Diﬀerential Evolution Algorithm

A Novel Hybrid Algorithm for Mean-CVaR Portfolio Selection with

A Modiﬁed Multi-Objective Optimization Based on Brain Storm

Modiﬁed Brain Storm Optimization Algorithm for Multimodal

Special Session on Swarm Intelligence for Real-World

FOREX Rate Prediction Using Chaos, Neural Network and Particle

Path Planning Using Neighborhood Based Crowding Diﬀerential

Neural Network Based on Dynamic Multi-swarm Particle Swarm

Dynamic Diﬀerential Evolution for Emergency Evacuation

Centralized Charging Strategies of Plug-in Electric Vehicles on Spot

A New Multi-region Modiﬁed Wind Driven Optimization Algorithm

Special Session on ICSI 2014 Competition on Single

The Multiple Population Co-evolution PSO Algorithm . . . . . . . . . . . . . . . . 434

Fireworks Algorithm and Its Variants for Solving ICSI2014 Competition

Performance of Migrating Birds Optimization Algorithm on Continuous

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461

Novel Swarm-Based Search Methods