Multi-Objective Optimization For Football Team Member Selection
Multi-Objective Optimization For Football Team Member Selection
Multi-Objective Optimization For Football Team Member Selection
ABSTRACT Team composition is one of the most important and challenging directions in the recommen-
dation problem. Compared with a single person, the advantage of a team is mainly reflected in the synergy
of team members’ complementary collaboration. To build a high-efficiency team, how to choose the team
members has become a tricky problem. However, there is a lack of quantitative algorithms and validation
methods for team member selection. In this paper, we put forward three indicators to measure a team’s
ability and formulate the selection of football team members as a multi-objective optimization problem.
Subsequently, an evolutionary player selection algorithm based on the genetic algorithm is proposed to solve
the team composition problem. We verify the effectiveness of the team member recommendation algorithm
via data analysis, football game simulation under different budget constraints and provide comparisons with
existing methods.
definition of the objective function is subjective, and there and particle swarm optimization [16], [17]. Particularly,
is no reliable verification method to prove the effectiveness MOEA/D, an acronym for multi-objective evolutionary
of the algorithm. Grund [9] exploited the density of network algorithm based on decomposition, provides a new idea
data structure, taking the Premier League football team as for solving MOPs [18]. Inspired by MOEA/D, some
an example, to study the team performance. However, if the decomposition-based algorithms are also proposed [19]–[21].
network structure is used to connect players in the same team Furthermore, evolutionary algorithms also provide new ways
for player recommendation, there will be a problem that all for solving practical problems. For example, Li and Yin [22]
players cannot be connected by each other, i.e., the player used differential evolutionary algorithm to design a recon-
graph is not connected. Besides, a fuzzy inference system figurable antenna array with quantized phase excitations.
is also adopted into player selection [10], but this method In [23], a multi-search differential evolutionary algorithm
largely depends on the experiences and cognitive abilities with self-adaptive parameter control was proposed for solv-
of experts. Zeng et al. [11] hashed over a skill coverage ing global real-parameter optimization problems. Addition-
function by using the submodule optimization technique, ally, several heuristic optimization algorithms have also been
which selects players by maximizing the constructed skill explored to deal with the MOPs. For instance, Li et al. [24]
coverage function. Nevertheless, the constructed function still proposed a new heuristic optimization method, called animal
has some shortcomings, e.g., the skill coverage of a team will migration optimization algorithm. The algorithm is inspired
reach a maximum as long as one player has a very high score by animal migration behavior, which is a ubiquitous phe-
on a certain attribute, no matter how other players perform. nomenon that can be found in all major animal groups.
Besides, it is not suitable to predict players’ salaries based Besides, Li et al. [25] put forward a new multi-objective
on the fitted exponential function which only considers the forest algorithm that can identify protein-RNA interactions
performance-price ratio of players. from CLIP-seq data. This study provides a refreshing insight
Team composition is often more than a temporary consid- into the use of multi-objective optimization for genome
eration. For example, on the football field, player transactions informatics.
and contracts would not expire within four years. Therefore, In addition to the MOPs, many of the previous
the potentials of team members are of great importance. researches on dynamic multi-objective optimization prob-
Teams with greater potentials tend to achieve higher valua- lems (DMOPs) [26]–[28], which focuses on the multiple
tions and better long-term performance. There are a variety conflicting goals that change over time, have also been
of literature on the field of evaluating potentials of football explored. For example, Xu et al. [27] proposed a cooperative
players. E.g., Williams and Reilly [12] attempted to integrate co-evolutionary strategy based on environmental sensitivities
their main research findings with talent identification and for solving DMOPs. Furthermore, Rong et al. [26] put for-
development in soccer. Similarly, Unnithan et al. [13] did ward a multi-directional prediction strategy to enhance the
research on talent identification in youth soccer. In [14], performance of EAs for DMOPs.
Jimǹez and Pain explored the relative age effect in Spanish In this paper, we focus on the genetic algorithm, one of
Association football and provided a comprehensive elabora- the pivotal methods in EAs, which aims to balance different
tion on the influence of player ages. objective functions as well as find the solution set that makes
However, in the studies mentioned above, the indicators each objective function as optimal as possible. Among several
of selecting players are subjective and lacking quantitative multi-objective genetic algorithms, NSGA [29] is one of the
evaluation. The subjective player selection methods cannot most influential and widely used ones, and has been improved
be verified by numerical approaches. Besides, those methods by Deb et al. [30]. The improved one is called NSGA-II. Due
only focus on football players’ physical fitness, ignoring to the simplicity and effectiveness, the NSGA-II algorithm
the comprehensiveness of the composed group. Furthermore, has successfully been applied in various fields [31]–[33].
those researchers did not consider the potential of the com- Based on the relationship between football players and
posed team, i.e., the future performance of the team, which is teams, we propose three indicators to evaluate the ability of
of great importance. a team. Besides, we construct the team composition problem
as a novel multi-objective optimization function, which will
B. MULTI-OBJECTIVE OPTIMIZATION METHODS be resolved by a new modification of NSGA-II algorithm and
Multi-objective optimization methods have been widely we further verify the effectiveness of our proposed model and
used in many areas, including engineering and economics algorithm by simulating virtual video game.
where optimal decisions need to be taken in the pres-
ence of trade-offs between two or more conflicting objec- III. ESTABLISHMENT OF EVALUATION INDICATORS
tives. In the past decades, numerous studies have attempted In this section, we design three indicators to evaluate the per-
to solve multi-objective optimization problems (MOPs). formance of teams and football players, which contribute to a
Among them, evolutionary algorithm (EA) is one of the novel framework for team composition. The reason why we
mainstream algorithms. Many researches have been carried consider the three indicators will be discussed in Section V
out, including ant colony algorithm [15], genetic algorithm in detail.
A. OVERALL EVALUATION OF PLAYERS AND TEAMS as many as 26 positions on the pitch, to save space, readers
Given a football player Pi , the most important factor for team may refer to the website https://sofifa.com/ for more specific
composition is the player’s ability, which we define as the abbreviations.
overall evaluation φOverall (Pi ). It is a comprehensive property
based on the player’s general performance in a football game. C. PLAYER POTENTIAL
The overall evaluation of a player considers different football The two factors mentioned above are based on the current
skills, including the players’ physical attributes, football tech- situation, but if we want the result of team composition to
nology, and psychological quality. We integrate the overall be effective, workable, and sound, it must be grounded in
evaluations of all football players in a team to form the team the present and the future as well. For a football player,
ability, which is defined as φOverall (N ), where N is the total we describe future ability as the potential attribute. However,
number of football players in a team. The specific calculation how to measure the potential value of players in a team is a
method can be seen in Eq. (1). difficult problem. According to the analysis of the players’
N
data, we found that the characteristics associated with the
X potential value of players are players’ age, current perfor-
φOverall (N ) = φOverall (Pi ) (1)
mance, and so on. Given a player Pi , we use Po(Pi ) to rep-
i=1
resent his potential value, which consults the comprehensive
B. OFFENSIVE AND DEFENSIVE ABILITY OF PLAYERS evaluation results of football scouters in the dataset.
It is not enough to form a competitive football team only
based on a player’s overall evaluation, which will evoke some IV. TEAM COMPOSITION AS A MULTI-OBJECTIVE
shortcomings. For example, if the selected players are all for- OPTIMIZATION PROBLEM
wards, it will make the team’s defensive ability insufficient. As previously asserted, building a football team with a high
As we have analyzed above, there is a significant difference winning rate requires many considerations, including the
in soccer players’ abilities at different positions on the field. overall evaluation of the team, the offensive, and defensive
When measuring the ability of a forward, we require more abilities of the players, as well as the potential value. In this
offensive skills, such as control and speed of the ball and section, we model the team composition as a multi-objective
shooting skills. Similarly, when considering the ability of a optimization problem based on the three factors proposed in
guard, we desire higher defensive attributes such as physical Section III. We first formulate the football team composition
contact ability. Taking the famous football star Messi as an problem, and then elaborate on the algorithm for selecting
example, his ability in an offensive position is generally players.
higher than that in a defensive position, so it will be sensible
to put this player in the offensive position. A. MODEL FORMULATION
In this paper, we consider the football players’ abili- We formulate the team composition problem as a
ties for different positions, and divide the positions except multi-objective optimization problem based on a series of
for Goalkeeper into three parts: Attack position (e.g. Striker, optimization parameters. Given a football player Pi , let S(Pi )
Center Forward), Midfield position (e.g. Center Midfield) represent the player’s salary, and B be the team’s total budget,
and Defensive position (e.g. Center Back). Attack posi- the multi-objective optimization problem can be formulated
tion describes a player’s offensive ability, while Defensive as follows:
position measures the player’s defensive attribute. Besides,
X
N
φOverall (Pi )
we consider both offensive and defensive abilities for
Xi=1
N
the Midfield position because of its position specialty.
Po(Pi )
For a soccer player Pi , we use λAtk (Pi ) and λDef (Pi ) to Xi=1
Nplayers
represent a player’s offensive ability for Attack position and max λAtk (Pj )
j=1
defensive score for Defensive position respectively, and λGK
X Nplayers
λDef (Pj )
refers to a player’s goalkeeping attribute. The calculation
j=1
N
GK
X
λGK (Pk )
method of a player’s offensive ability is the average score
k=1
of his abilities in different offensive positions (and similar N
applies to the measure of defensive ability). We adopt the
X
s.t. S(Pi ) 6 B (3)
following method to compute them: i=1
(
λAtk (Pi ) = Mean(µSt , · · · , µCam ) where Nplayers is the total number of football players in three
(2) positions (i.e. Attack position, Midfield position, and Defen-
λDef (Pi ) = Mean(µCb , · · · , µCdm )
sive position) and NGK is the number of goalkeepers. The
where µ represents the player’s performance in different parameters N , Nplayers and NGK are all constants. Particularly
positions. For example, µSt shows the player’s performance in our case, we have Nplayers = 10, NGK = 1, and N =
in the Striker (St). Similarly, µCam is the player’s perfor- Nplayers + NGK = 11. Indeed, the proposed model formulated
mance in the Center attack midfield (Cam). Since there are in Eq. (3) can be extend to a matchday team with reserves
B. ESP ALGORITHM
There exists multiple Pareto optimal solutions for multi-
objective optimization, and evolutionary algorithms funda-
mentally operate on a set of candidate solutions, thus we focus
on genetic algorithms, which is one of the major evolutionary
algorithm paradigms, for further solving the optimization
problem. In this section, we elaborate on the ESP algorithm,
which is a modification of NSGA-II. We first provide a brief
description of the NSGA-II procedure [30] and then illustrate
the ESP algorithm in detail.
Fig. 3 shows the general procedure of NSGA-II. The
basic flow of NSGA-II algorithm is similar to the tra-
ditional genetic algorithm, including critical components
such as coding, crossover, mutation, and selection. Besides,
NSGA-II explores three special characteristics (i.e., fast
non-dominated sorting approach, density estimation and
crowded-comparison operator). Specifically, according to the FIGURE 3. The flowchart of the NSGA-II algorithm.
criteria for the sorting process, NSGA-II first initializes a ran-
dom parent population. The population is sorted based on the TABLE 1. A toy example of the chromosome reordering process.
non-domination. Once the first sorting is completed, the usual
binary tournament selection, recombination, and mutation
operators are used to create an offspring population, which
is then combined with the current generation population.
Followed by the combination procedure, NSGA-II introduces
In addition, we also arrange the remaining 10 bits of
the elitism criterion to compare the current population with
the chromosome in ascending order according to the serial
the previous best solutions and selects the individuals of the
number, which represents the player’s salary (see Table 1).
next generation based on the crowded-comparison operator.
The smaller the value of the serial number, the higher the
Based on the multi-objective player selection optimization
player’s salary. During the cross recombination procedure,
model, the ESP algorithm absorbs the advantages of NSGA-II
rearranging the similar chromosome sequence makes a small
and some modification is done to make it better fit the need
difference in salary between the two players, and avoids
for the football player selection problem. Specific changes
meaningless updating operations. Furthermore, it can better
are explained from the following two aspects.
control the total cost of the team and reduce the probability
1) CODING METHOD
of non-feasible solutions.
In the genetic algorithm NSGA-II, the individual variables of
2) FAST NON-DOMINATED SORTING PROCESS BASED ON
each generation are usually continuous. However, we make
BUDGET CONSTRAINTS
each variable of individual as an integer representing the
serial number of each player. Besides, the length of a chro- It is not an easy task to solve the team composition problem
mosome is set to 11, which equals the number of football with budget constraints. In the ESP algorithm, we introduce
players. Considering there is only one goalkeeper in a team, a variable Constraint Violation (CV) in Eq. (4) to make the
we choose the first bit of the chromosome limited to the goal- final solution satisfy the budget constraint.
keeper selection and the remaining 10 bits of a chromosome TC − B
CV = max 0, (4)
represent non-goalkeeper players. B
VOLUME 9, 2021 90479
H. Zhao et al.: Multi-Objective Optimization for Football Team Member Selection
where TC is the total cost of the team. Based on the above TABLE 2. An example of football players’ athletic abilities.
formula, the constraint violation value of a feasible solution is
always 0, while the one for the non-feasible solution is greater
than 0. The larger the violation value, the greater the deviation
of the non-feasible solution.
By introducing the concept of the constraint violation,
we change the rules of the fast non-dominated sorting pro-
cess, which is the key component of the NSGA-II algorithm.
Given any two solutions (i.e. individuals in the population) xa TABLE 3. An example of football players’ personal attributes.
and xb , of which decision values are a , CV (xa ) and
b , CV (xb ) respectively, the new dominant relationship is
obtained as follows:
• If CV (xa ) = 0 and CV (xb ) > 0, we have a dominates b,
or otherwise, b dominates a.
• If CV (xa ) > 0 and CV (xb ) > 0, we have the smaller CV
dominates the larger CV . We first perform data preprocessing by filtering out some
• If CV (xa ) = 0 and CV (xb ) = 0, the dominant relation- irrelevant attributes and sort the data (in descending order)
ship is determined according to the rules mentioned in according to players’ salaries. Table 2 and Table 3 list some
Pareto dominance. players’ basic properties after data preprocessing. We then
Algorithm 1 presents the details of the player selection pro- elaborate on the three phenomena discovered from the exper-
cedure for the ESP algorithm. We initialize a set of candidate imental data, which consist of our assumptions.
players P, a budget constraint B, and the hyperparameters
such as mutation probability pm , crossover probability pc , Algorithm 1 Finding a Best Team Based on ESP Algorithm
and polynomial mutation distribution index ηm . After sorting Input: crossover probability(pc ), mutation probability(pm ),
players based on their salaries, the initial population of a given polynomial mutaion parameter(ηm ), player dataset(P),
size is randomly generated by integer coding (Line 1 - Line 7). player number(N ), budget(B).
The first generation of offspring population is obtained by Output: optimal solution set(P̂).
basic operations of crossover and mutation of genetic algo- 1: pop = ∅.
rithm (Line 8 - Line 11). Starting from the second genera- 2: // Individual coding and population initialization
tion, the parent population and the offspring population are 3: for i = 1 to PopSize do
merged, followed by calculating the CV values and perform- 4: RandomSequence = randomly choose the index of N
ing the fast non-dominated sorting process with constraints. players from P.
At the same time, we compute the crowding degree of the 5: individual = RealNumberCoding(RandomSequence,
individuals in each Pareto front. Note that among the two budget).
solutions with different Pareto fronts, we prefer the solution 6: pop.add(individual).
with a better dominance ranking. Otherwise, if the two solu- 7: end for
tions belong to the same front, we prefer the solution in the 8: for i = 1 to Max_Generation do
relatively less crowded region. According to the crowding 9: newpop = CrossOver(pop, pc ).
degree of individuals, the appropriate individuals are selected 10: newpop = Mutation(newpop, pm , ηm ).
to form a new parent population (Line 12 - Line 17). The algo- 11: newpop = newpop + pop.
rithm loops until the conditions for the end of the program 12: // Calculate constraint violation for each individ-
are met. ual
13: CV = ConstraintViolation(newpop, budget).
V. DATA ANALYSIS AND EXPERIMENTS
14: newpop = FastNondominantSort(newpop, CV ).
We implement the algorithms in Python 3.85 and conduct
15: crowding = CrowdingCompare(newpop).
all the numerical computations on a Windows PC with a
16: offspring = Selection(newpop, crowding).
4-core Intel i5-1135g7 2.40GHz CPU and 16GB memory.
17: pop = offspring.
All the experimental data is collected from the website
18: end for
https://sofifa.com/ and all the games are simulated in a quick
19: return P̂.
game of PES2021.
A. DATA ANALYSIS
1) WAGES ARE NOT DIRECTLY PROPORTIONAL TO
In this section, we analyze the experimental data and present
PLAYERS’ ABILITIES
some facts related to our proposed multi-objective function.
We first analyze the relationship between the player’s ability
5 Part of the code and dataset are available from: https://github.com/haoyu- and wages. Taking the player’s overall rating as the X-axis,
zhao/Multi-Objective-Optimization-for-Football-Team-Member-Selection the corresponding weekly salary as the Y-axis, we plot the
FIGURE 7. Pareto solutions for ESP algorithm without budget constraints. FIGURE 9. The influence of budgets on Pareto solutions.
Besides, we analyze the Pareto fronts under different bud- by the ESP algorithm has certain numerical advantages over
get constraints, as shown in Fig. 9. The green dot set is other teams without budget constraints.
the Pareto solutions without budget constraints. In Fig. 9, In the t-test settings, given a team attribute (i.e. overall
we see that with the increase of the budget, the Pareto front evaluation, attacking rating, defence rating, or goalkeeping),
shifts to the lower-left corner, that is, the higher the budget, we assume that the team generated from the ESP algorithm
the stronger the abilities of attack and defense of the team, and the random team share the same average value. We first
which is consistent with our assumptions. take out 200 teams selected by the ESP algorithm, as well
As can be seen in Fig. 9, there are many candidate solu- as 200 random teams to simulate the distribution of team
tions, which form a Pareto front under the same budget. attribute values with large samples. For each type of teams,
By descending sorting the crowding degree of the solutions we then draw a histogram by calculating the frequency of
in Pareto fronts, we obtain the optimal solution (i.e., the best attribute scores in Fig. 10, and the corresponding P-values
team). Unless otherwise indicated, we use ESP Dream Team are shown in Table 5. Based on the P-value of all attributes,
to represent the best team in our simulation experiments. we can reject the original hypothesis with confidence that the
average values of any team attribute of both the ESP team
1) BUDGET UNCONSTRAINED CASE and random team are the same, which in turn demonstrates
In this section, we analyze the selection results of the ESP the effectiveness of our algorithm.
algorithm without budget constraints. We set a large num-
ber, which equals 770 million euros to simulate a sufficient 2) BUDGET CONSTRAINED CASE
budget. Table 4 provides a comparison of players’ average We compare the numerical performance of our team with a
athletic abilities between the ESP Dream Team and a random randomly selected team under the constraints of the budget
team under a similar total budget level. The players’ average in this section. We set the budget at 150 million euros, which
athletic abilities include the average of the overall evaluation, is a representative budget of a football team. The comparison
attack rating and defense rating, as well as the goalkeeper’s results are shown in Table 6. It can be seen that all the
goalkeeping ability. In Table 4, we see that the team selected numerical results of our proposed method are better than
FIGURE 12. Results of players’ average athletic abilities of the ESP Dream
Team and Random Team.
FIGURE 11. Histogram of the distribution of four attributes with budget
constraints.
football team. Table 8 shows the optimal solution selected
that of the random values. From Table 6, it is clear that the by the ESP algorithm when the budget is 100 million euros.
ESP algorithm delivers the best performance in all aspects It can be seen that with a similar budget, the ESP algorithm
while using fewer salaries. Similar to the t-test settings in considering the players’ potentials can select a team with
Section V-B1, we show the distributions of all attributes’ excellent potential value while keeping the four better men-
average values and P-values based on the budget constrained tioned indicators. To better understand the proposed method,
case in Fig. 11 and Table 7 respectively, and the correspond- Fig. 12 gives a preview of two team’s properties, where each
ing results also demonstrate the effectiveness of our proposed dimension shows a kind of average ability. From Fig. 12,
algorithm. we can also see that the team selected by the ESP algorithm
is better than the team generated from the random algorithm
3) POTENTIAL IMPACT in all aspects. Likewise, we also show the distributions of all
It is hard to verify the influence of the potential attribute attributes’ average values and P-values when considering the
for the football player selection in reality because we can potential effect. The corresponding results (see Fig. 13 and
only observe the current performance of players. Therefore, Table 9) demonstrate the effectiveness and rationality of the
we provide the numerical result of the potential value for a proposed factors in our modeling.
TABLE 10. Game simulation results of the ESP Dream Team v.s. Random Teams.
TABLE 11. Game simulation results of the ESP Dream Team v.s. Ten
Representative Real Teams from different football leagues.
TABLE 12. Game simulation results of the ESP Dream Team v.s. Real Teams.
TABLE 13. Game simulation results of the ESP Dream Team (with budget constraints) v.s. Random Teams.
TABLE 14. Budget level. Dream team, which indicates the stability and reliability of
the proposed algorithm.
TABLE 15. Game simulation results of the ESP Dream Team (with budget constraints) v.s. Real Teams.
TABLE 16. Game simulation results of the ESP Dream Team v.s. CEFG Dream Team.
[5] V. Di Salvo, R. Baron, H. Tschan, F. C. Montero, N. Bachl, and F. Pigozzi, [30] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, ‘‘A fast and elitist
‘‘Performance characteristics according to playing position in elite soccer,’’ multiobjective genetic algorithm: NSGA-II,’’ IEEE Trans. Evol. Comput.,
Int. J. Sports Med., vol. 28, no. 3, pp. 222–227, Mar. 2007. vol. 6, no. 2, pp. 182–197, Apr. 2002.
[6] E. Ö. Ozceylan, ‘‘A mathematical model using AHP priorities for soccer [31] J. Jemai, M. Zekri, and K. Mellouli, ‘‘An NSGA-II algorithm for the green
player selection: A case study,’’ South Afr. J. Ind. Eng., vol. 27, no. 2, vehicle routing problem,’’ in Evolutionary Computation in Combinatorial
pp. 190–205, Aug. 2016. Optimization. Berlin, Germany: Springer, 2012, pp. 37–48.
[7] A. Kamble, R. Rao, A. Kale, and S. Samant, ‘‘Selection of cricket play- [32] S. Bandyopadhyay and R. Bhattacharya, ‘‘Solving multi-objective paral-
ers using analytical hierarchy process,’’ Int. J. Sports Sci. Eng., vol. 5, lel machine scheduling problem by a modified NSGA-II,’’ Appl. Math.
pp. 207–212, Sep. 2011. Model., vol. 37, nos. 10–11, pp. 6718–6729, Jun. 2013.
[8] F. Ahmed, K. Deb, and A. Jindal, ‘‘Multi-objective optimization and [33] J. Sadeghi, S. Sadeghi, and S. T. A. Niaki, ‘‘A hybrid vendor managed
decision making approaches to cricket team selection,’’ Appl. Soft Comput., inventory and redundancy allocation optimization problem in supply chain
vol. 13, no. 1, pp. 402–414, Jan. 2013. management: An NSGA-II with tuned parameters,’’ Comput. Oper. Res.,
[9] T. U. Grund, ‘‘Network structure and team performance: The case of vol. 41, pp. 53–64, Jan. 2014.
English Premier League soccer teams,’’ Social Netw., vol. 34, no. 4, [34] D. A. Van Veldhuizen and G. B. Lamont, ‘‘Evolutionary computation and
pp. 682–690, Oct. 2012. convergence to a Pareto front,’’ in Proc. Late Breaking Genetic Program.
[10] M. Tavana, F. Azizi, F. Azizi, and M. Behzadian, ‘‘A fuzzy inference Conf., 1998, pp. 221–228.
system with application to player selection and team formation in multi- [35] H. Yetgin, K. T. K. Cheung, and L. Hanzo, ‘‘Multi-objective routing opti-
player sports,’’ Sport Manage. Rev., vol. 16, no. 1, pp. 97–110, Feb. 2013. mization using evolutionary algorithms,’’ in Proc. IEEE Wireless Commun.
[11] Y. Zeng, G. Shen, B. Chen, and J. Tang, ‘‘Team composition in Netw. Conf. (WCNC), Apr. 2012, pp. 3030–3034.
PES2018 using submodular function optimization,’’ IEEE Access, vol. 7,
pp. 76194–76202, 2019.
[12] A. M. Williams and T. Reilly, ‘‘Talent identification and development in
soccer,’’ J. Sports Sci., vol. 18, no. 9, pp. 657–667, Jan. 2000.
[13] V. Unnithan, J. White, A. Georgiou, J. Iga, and B. Drust, ‘‘Talent identi-
fication in youth soccer,’’ J. Sports Sci., vol. 30, no. 15, pp. 1719–1726, HAOYU ZHAO received the B.S. degree in
Nov. 2012. automation from Xiamen University. His research
[14] I. P. Jiménez and M. T. G. Pain, ‘‘Relative age effect in Spanish association interests include artificial intelligence, data min-
football: Its extent and implications for wasted potential,’’ J. Sports Sci.,
ing, and recommendation systems.
vol. 26, no. 10, pp. 995–1003, Aug. 2008.
[15] M. Dorigo, M. Birattari, and T. Stutzle, ‘‘Ant colony optimization,’’ IEEE
Comput. Intell. Mag., vol. 1, no. 4, pp. 28–39, Nov. 2006.
[16] J. Kennedy and R. Eberhart, ‘‘Particle swarm optimization,’’ in Proc. IEEE
ICNN, vol. 4, Nov./Dec. 1995, pp. 1942–1948.
[17] Z. Yong, Y. Li-Juan, Z. Qian, and S. Xiao-Yan, ‘‘Multi-objective opti-
mization of building energy performance using a particle swarm opti-
mizer with less control parameters,’’ J. Building Eng., vol. 32, Nov. 2020,
Art. no. 101505.
[18] Q. Zhang and H. Li, ‘‘MOEA/D: A multiobjective evolutionary algorithm HAIHUI CHEN received the B.S. degree in
based on decomposition,’’ IEEE Trans. Evol. Comput., vol. 11, no. 6, automation from Xiamen University. His research
pp. 712–731, Dec. 2007. interest includes recommendation systems.
[19] X. Li and S. Ma, ‘‘Multi-objective memetic search algorithm for multi-
objective permutation flow shop scheduling problem,’’ IEEE Access, vol. 4,
pp. 2154–2165, 2016.
[20] X. Li and K.-C. Wong, ‘‘Evolutionary multiobjective clustering and its
applications to patient stratification,’’ IEEE Trans. Cybern., vol. 49, no. 5,
pp. 1680–1693, May 2019.
[21] Y. Qi, X. Ma, F. Liu, L. Jiao, J. Sun, and J. Wu, ‘‘MOEA/D with adaptive
weight adjustment,’’ Evol. Comput., vol. 22, no. 2, pp. 231–264, Jun. 2014.
[22] X. Li and M. Yin, ‘‘Design of a reconfigurable antenna array with discrete
phase shifters using differential evolution algorithm,’’ Prog. Electromagn. SHENBAO YU received the M.S. degree in
Res. B, vol. 31, pp. 29–43, May 2011. automation from Xiamen University, in 2017,
[23] X. Li, S. Ma, and J. Hu, ‘‘Multi-search differential evolution algorithm,’’ where he is currently pursuing the Ph.D. degree
Appl. Intell., vol. 47, no. 1, pp. 231–256, Jul. 2017. in artificial intelligence and system engineering.
[24] X. Li, J. Zhang, and M. Yin, ‘‘Animal migration optimization: An optimiza- His research interests include probabilistic graph-
tion algorithm inspired by animal migration behavior,’’ Neural Comput. ical models, data mining, and recommendation
Appl., vol. 24, nos. 7–8, pp. 1867–1877, Jun. 2014. systems.
[25] X. Li, S. Zhang, and K.-C. Wong, ‘‘Multiobjective genome-wide RNA-
binding event identification from CLIP-seq data,’’ IEEE Trans. Cybern.,
early access, Jan. 10, 2020, doi: 10.1109/TCYB.2019.2960515.
[26] M. Rong, D. Gong, Y. Zhang, Y. Jin, and W. Pedrycz, ‘‘Multidirectional
prediction approach for dynamic multiobjective optimization problems,’’
IEEE Trans. Cybern., vol. 49, no. 9, pp. 3362–3374, Sep. 2019.
BILIAN CHEN received the Ph.D. degree
[27] B. Xu, Y. Zhang, D. Gong, Y. Guo, and M. Rong, ‘‘Environment sensitivity-
from The Chinese University of Hong Kong,
based cooperative co-evolutionary algorithms for dynamic multi-objective
optimization,’’ IEEE/ACM Trans. Comput. Biol. Bioinf., vol. 15, no. 6, in 2012. She is currently an Associate Profes-
pp. 1877–1890, Nov. 2018. sor with Xiamen University. Her publications
[28] D. Gong, B. Xu, Y. Zhang, Y. Guo, and S. Yang, ‘‘A similarity-based appeared in SIAM Journal on Optimization, IEEE
cooperative co-evolutionary algorithm for dynamic interval multiobjec- TRANSACTIONS ON NEURAL NETWORKS AND LEARNING
tive optimization problems,’’ IEEE Trans. Evol. Comput., vol. 24, no. 1, SYSTEMS, Journal of Global Optimization, and
pp. 142–156, Feb. 2020. Information Sciences. Her research interests
[29] N. Srinivas and K. Deb, ‘‘Muiltiobjective optimization using nondom- include machine learning, optimization theory, and
inated sorting in genetic algorithms,’’ Evol. Comput., vol. 2, no. 3, recommendation systems.
pp. 221–248, Sep. 1994.