10 1109@iceaa 2019 8879110
10 1109@iceaa 2019 8879110
10 1109@iceaa 2019 8879110
Abstract—We review machine learning and its applications in II. DESIGN AND ANALYSIS
a wide range of electromagnetic problems, including radar,
communication, imaging and sensing. We extensively discuss some Typically, in order to have a fully trained NN, one needs to
recent progress in development and use of intelligent algorithms provide a sufficiently large set of data points, i.e., inputs and
for antenna design, synthesis, and characterization. We also their corresponding outputs. For antenna designs, this data set
provide some perspectives for future research directions in this can be built by intensively performing numerical simulations
emerging field of study. for different inputs and/or their combinations. In [2], design
goals and objectives (e.g., resonant frequency, directivity, and
Keywords—Electromagnetics, Artificial Neural Network,
efficiency) as functions of geometric parameters of a microstrip
Machine Learning, Deep Learning
antenna (e.g., slot and gap) have been investigated using the
I. INTRODUCTION intelligent method. Microstrip antennas with different slot sizes
and air gaps were first simulated using a commercial software,
In the past few years, there has been growing interest in
machine learning (ML), more specifically artificial neural IE3D, and the corresponding antenna characteristics were
networks (ANN) and one part of its broader family - deep recorded to construct the data set for training (and also testing)
learning (DL). To date, ML and DL have been demonstrated to the NNs. The trained NN took a fraction of second to calculate
be effective in image classification, speech processing, and other radiation properties of an antenna with a set of randomly
information processing tasks. Very recently, ML and DL have generated design parameters. In [3], the date set was
been further extended to complex electromagnetic problems, constructed using empirical formulas for a single-feed
such as structural design and optimal parameter extraction for circularly polarized square microstrip antenna. The results
antennas, beamforming algorithms for adaptive antenna arrays, showed that Levenberg-Marquardt (LM) algorithm may help
and data interpretation for radar and MIMO systems. Although designing a practical antenna with desired operating frequency
many ML and DL algorithms have been proposed, choosing the and bandwidth (or Q-factor). In [4], the a contoured beam
most appropriate one for a specific electromagnetic problem is reflect array composed of antenna elements was designed using
still an open-ended question. A comprehensive review of NNs. The method of moments (MoM) was used to generate the
applications of neural network (NN) in smart antennas has been training data points. It was shown that the ANN can be faster
conducted in [1]. The advantages of NN-assisted antenna than the MoM by a factor of 100, when both techniques are used
systems reside in the reduced computational complexity and to compute radiation characteristics of an unknown antenna. In
time, lower cost compared to alternatives, and highly accurate [5], an ANN has been used to predict aperture antenna shape,
data analysis. In that review paper, results predicted by ML and demonstrating capabilities of dynamically producing the
DL models agree quite well with theory, numerical simulation,
desired radiation pattern. In contrast, currently existing
and measurements.
numerical approaches take days to analyze the design. This
1378
speed and when the BSs deploy large antenna arrays. With VII. CONCLUSION AND PERSPECTIVES FOR FUTURE RESEARCH
sufficient learning time, the DL model efficiently adapts to Based on the above discussions, ML approaches could be
changing environment, yielding a robust beamforming system. successfully applied to different kinds of electromagnetic
problems, with aims of reducing computational complexity,
VI. DIRECTION OF ARRIVAL (DOA) ESTIMATION allowing faster real-time operation, and helping to deal with
problems of complicated structures when accurate simulation is
DoA estimation is another problem that has been either impossible or very challenging. Though the performance
solved using ML. So far, many papers devoted to this problem of the ML-based system depends on the quality of the training
have demonstrated that ML may improve the DoA estimation. data, once the training phase is done, the model can work very
Particularly, the deep neural network (DNN) was applied to the fast and accurate.
case of ULA in [19], making possible the detection of signals One main drawback of ML is the large amount of data
coming within a 120 degree sector and showing better needed for the training [18]. In turn, this may require a large
performance than multiple signal classification (MUSIC). This time to generate the data [9]. Efforts are then required to reduce
methodology was designed for a switched-beam system (SBS) the number of samples for the training, which may be critical
for direct sequence code division multiple access (DS-CDMA) for the use of ML methods for many real-time applications.
applications. The main advantages over MUSIC are speed and Intelligent approaches may be developed to augment machine
simplicity, whereas a drawback could be the operation in a 120 learning methods by making use of prior knowledge stemming
degree instead of a 180 degrees domain. Later on, in [20], a from the understanding of electromagnetic problems. Ray-
similar problem was solved for 180-degrees sector by tracing methods [24] provide a way to introduce prior
modifying the DNN configuration, which also make estimation knowledge, while preserving a simple understanding of the
less insensitive to array imperfections, such as non-ideal sensor basic physical phenomena that are involved. Ray-tracing
design and manufacture error, array installation and inter sensor methods are generally fast and can be applied to electrically
mutual interference, and background radiation. The proposed large problems. Ray-tracing methods have been applied to find
method showed better robustness compared to MUSIC. the ray trajectories that most contribute to a specific
For electronically steerable parasitic array radiator antenna phenomenon. For example, they have been applied to radar
(ESPAR), the DoA estimation problem was solved by means of problems in [25]-[27] for identifying the most likely
SVC in [21]. The one-versus-all classification was done for propagation channels in a beamforming application.
received signal strength (RSS) values recorded at the antenna’s
output port. The antenna radiation patterns measured precisely
in an anechoic room were used as the training data. The VIII. REFERENCES
performance of the NN-based method was higher than that of
the conventional power pattern cross correlation (PPCC)
method. [1] A. Rawat, R. N. Yadav and S. C. Shrivastava, "Neural
For the purpose of multiple source tracking, the DoA is network applications in smart antenna arrays: a review."
done by the family of RBFNN [22], which makes the system AEU-International Journal of Electronics and
capable of tracking several sources with any angular separation, Communications," AEU-International Journal of
without any prior knowledge of the number of sources. The Electronics and Communications, pp. 903-912, 2012.
problem was tackled by dividing the field of view of the [2] T. Khan, A. De and M. Uddin, "Prediction of slot-size
antenna into angular spatial sectors, training each network in and inserted air-gap for improving the performance of
the first stage to detect signals emanating in that sector, and rectangular microstrip antennas using artificial neural
later activating of one or more networks of the second stage to networks," IEEE Antennas and Wireless Propagation
estimate the exact location of the sources. This method allows Letters, pp. 1367-1371, 2013.
dramatical reduction in the size of the training set and locating [3] Z. Wang, S. Fang, Q. Wang and H. Liu, "An ANN-based
sources greater than the number of antenna elements. synthesis model for the single-feed circularly-polarized
DNN were applied to MIMO system in [23] to perform square microstrip antenna with truncated corners," IEEE
DoA and channel estimation. The DNN was first trained using Transactions on Antennas and Propagation, vol. 60, no.
simulation data obtained under different channel conditions 12, pp. 5989-5992, 2012.
(with the aid of the offline learning), and then output data were
[4] P. Robustillo, J. Zapata, J. A. Encinar and M. Arrebola,
obtained based on the current input data during online learning
"Design of a contoured-beam reflectarray for a
process. The mean square error (MSE) of DoA estimation with
EuTELSAT European coverage using a stacked-patch
different training length sequences, learning rate, and batch size
element characterized by an artificial neural network,"
was examined and compared with the one when other methods
IEEE Antennas and Wireless Propagation Letters, vol.
(SBEM Scheme, PA channel tracking scheme, ADMA user
11, pp. 977-980, 2012.
scheduling scheme, and compressed sensing scheme) were
used. The proposed DL based method had lower error that the [5] G. Washington, "Aperture antenna shape prediction by
others. feedforward neural networks," IEEE Transactions on
Antennas and Propagation, vol. 45, pp. 683-688, 1997.
1379
[6] C. Gianfagna, H. Yu, M. Swaminathan, R. Pulugurtha, R. highly-mobile millimeter wave systems," IEEE Access 6,
Tummala and G. Antonini, "Machine-learning approach pp. 37328-37348, 2018.
for design of nanomagnetic-based antennas," Journal of [19] K. A. Gotsis, K. Siakavara and J. N. Sahalos, "On the
Electronic Materials, vol. 48, no. 6, pp. 4963-4975, 2017. direction of arrival (DoA) estimation for a switched-
[7] B. K. Singh, "Design of rectangular microstrip patch beam antenna system using neural networks," IEEE
antenna based on Artificial Neural Network algorithm," Transactions on Antennas and Propagation, vol. 57, no.
in 2015 2nd International Conference on Signal 5, pp. 1399-1411, 2009.
Processing and Integrated Networks (SPIN). IEEE, 2015. [20] Z. M. Liu, C. Zhang and P. S. Yu, "Direction-of-arrival
[8] N. Dandanov, H. Al-Shatri, A. Klein and V. Poulkov, estimation based on deep neural networks with
"Dynamic self-optimization of the antenna tilt for best robustness to array imperfections," IEEE Transactions on
trade-off between coverage and capacity in mobile Antennas and Propagation, vol. 66, no. 12, pp. 7315-
networks," Wireless Personal Communications, vol. 92, 7327, 2018.
no. 1, pp. 251-278, 2017. [21] M. Tarkowski and L. Kulas, "RSS-based DoA Estimation
[9] B. Liu, H. Aliakbarian, Z. Ma, G. A. Vandenbosch, G. for ESPAR Antennas Using Support Vector Machine,"
Gielen and P. Excell, "An efficient method for antenna IEEE Antennas and Wireless Propagation Letters, 2019.
design optimization based on evolutionary computation [22] A. H. E. Zooghby, C. G. Christodoulou and M.
and machine learning techniques," IEEE Transactions on Georgiopoulos, "A neural network-based smart antenna
Antennas and Propagation, vol. 62, no. 1, pp. 7-18, 2014. for multiple source tracking," IEEE Transactions on
[10] V. Sathi, S. Taherizadeh, A. Lotfi and C. Ghobadi, Antennas and Propagation, vol. 48, no. 5, pp. 768-776,
"Optimisation of multi-frequency microstrip antenna 2000.
using genetic algorithm coupled with method of [23] H. Huang, J. Yang, H. Huang, Y. Song and G. Gui, "Deep
moments," IET microwaves, Antennas & Propagation, learning for super-resolution channel estimation and
vol. 4, no. 4, pp. 477-483, 2010. DOA estimation based massive MIMO system," IEEE
[11] A. Patnaik, B. Choudhury, P. Pradhan, R. Mishra and C. Transactions on Vehicular Technology, vol. 67, no. 9, pp.
Christodoulou, "An ANN application for fault finding in 8549-8560, 2018.
antenna arrays," IEEE Transactions on Antennas and [24] P. H. Pathak, "Techniques for High-Frequency
Propagation, vol. 55, pp. 775-777, 2007. Problems," in Antenna Handbook, New York, Van
[12] C. G. Christodoulou, J. A. Rohwer and C. T. Abdallah, Nostrand, 1993.
"The use of machine learning in smart antennas," in IEEE [25] H. T. Hayvaci, A. De Maio and D. Erricolo, "Improved
Antennas and Propagation Society Symposium, 2004. Detection Probability of a Radar Target in the Presence
[13] P. R. Chang, W. H. Yang and K. K. Chan, "A neural of Multipath with Prior Knowledge of the Environment,"
network approach to MVDR beamforming problem," IET Radar, Sonar & Navigation, vol. 7, no. 1, pp. 36-46,
IEEE Transactions on Antennas and Propagation, vol. 2013.
40, no. 3, pp. 313-322, 1992. [26] P. Setlur, T. Negishi, N. Devroye and D. Erricolo,
[14] A. H. E. Zooghby, C. G. Christodoulou and M. "Multipath Exploitation in Non-LOS Urban Synthetic
Georgiopoulos, "Neural network-based adaptive Aperture Radar," IEEE Journal of Selected Topics in
beamforming for one-and two-dimensional antenna Signal Processing, vol. 8, no. 1, pp. 137-152, Feb. 2014.
arrays," IEEE Transactions on Antennas and [27] V. Va, J. Choi, T. Shimizu, G. Bansal and R. W. Heath,
Propagation, vol. 46, no. 12, pp. 1891-1893, 1998. "Inverse Multipath Fingerprinting for Millimeter Wave,"
[15] A. H. Sallomi and S. Ahmed, "Elman Recurrent Neural IEEE Transactions on Vehicular Technology, vol. 67, no.
Network Application in Adaptive Beamforming of Smart 5, pp. 4042-4058, 2018.
Antenna System," International Journal of Computer
Applications, p. 8887, 2015.
[16] C. Antón-Haro and X. Mestre, "Learning and Data-
Driven Beam Selection for mmWave Communications:
An Angle of Arrival-Based Approach," IEEE Access 7,
pp. 20404-20415, 2019.
[17] H. Huang, W. Xia, J. Xiong, J. Yang, G. Zheng and X.
Zhu, "Unsupervised learning-based fast beamforming
design for downlink MIMO," IEEE Access 7, pp. 7599-
7605, 2019.
[18] A. Alkhateeb, S. Alex, P. Varkey, Y. Li, Q. Qu and D.
Tujkovic, "Deep learning coordinated beamforming for
1380