IEEE TRANSACTIONS ON CYBERNETICS 1

A Many-Objective Evolutionary Algorithm Based

on a Two-Round Selection Strategy
Zhengping Liang , Kaifeng Hu, Xiaoliang Ma, and Zexuan Zhu , Member, IEEE

Abstract—Balancing population diversity and convergence is where x represents an n-dimensional decision vector in the
critical for evolutionary algorithms to solve many-objective decision space , fi (x) denotes the ith objective function, and
optimization problems (MaOPs). In this paper, a two-round envi- M indicates the number of objective functions. An MOP is
ronmental selection strategy is proposed to pursue good tradeoff
between population diversity and convergence for many-objective also called a many-objective optimization problem (MaOP)
evolutionary algorithms (MaOEAs). Particularly, in the first if M > 3 [2]–[7]. Given two different solutions x, y ∈ ,
round, the solutions with small neighborhood density are picked if ∀i = 1, 2, . . . , M, fi (x) = fi (y), and ∃j ∈ {1, 2, . . . , M},
out to form a candidate pool, where the neighborhood density fj (x) < fj (y), then x is said to dominate y (denoted as x ≺ y).
of a solution is calculated based on a novel adaptive position A solution x∗ ∈  not dominated by any other solutions is
transformation strategy. In the second round, the best solution
in terms of convergence is selected from the candidate pool and called a Pareto optimal solution. All Pareto optimal solutions
inserted into the next generation. The procedure is repeated until compose the Pareto optimal set (PS) and the corresponding
a new population is generated. The two-round selection strat- objective vectors of the PS in the objective space form the
egy is embedded into an MaOEA framework and the resulting Pareto front (PF) [7], [8].
algorithm, namely, 2REA, is compared with eight state-of-the- A large number of multiobjective evolutionary algo-
art MaOEAs on various benchmark MaOPs. The experimental
results show that 2REA is very competitive with the compared rithms (MOEAs) have been proposed to solve MOPs
MaOEAs and the two-round selection strategy works well on and shown promising performance [9]–[11]. However, these
balancing population diversity and convergence. MOEAs encounter great difficulties when dealing with
Index Terms—Adaptive position transformation (APT), many- MaOPs. The main reason is that the proportion of non-
objective evolutionary algorithm (MaOEA), many-objective dominated solutions in the evolution population explodes
optimization. as the number of objective functions increases, which
reduces the pressure of Pareto-based selection toward the
I. I NTRODUCTION true PF [12]–[14]. The most straightforward solution to this
issue is to increase the evolutionary pressure toward the
VER the last two decades, multiobjective optimization
O problems (MOPs) have received great attention in the
evolutionary computation community [1], [2]. Without loss of
PF using modified or relaxed Pareto dominance relations,
for example, α-dominance [12]; -dominance [15], [16]; and
L-dominance [17]. Modified or relaxed Pareto dominance rela-
generality, an MOP can be defined as
tions can accelerate the convergence of MOEAs by enlarging
min F(x) = (f1 (x), f2 (x), . . . , fM (x)) the dominating area of the individuals [18], [19]; however,
subject to: x ∈  ⊆ Rn (1) they tend to result in the loss of population diversity [20].
Balance of population convergence and diversity is required
Manuscript received November 2, 2018; revised January 28, 2019 and to better solve MaOPs.
March 27, 2019; accepted May 11, 2019. This work was supported in part
by the National Natural Science Foundation of China under Grant 61871272, A number of many-objective evolutionary algo-
Grant 61471246, and Grant 61672358, in part by the Project of Department rithms (MaOEAs) have been proposed in the literature
of Education of Guangdong Province under Grant 2016KTSCX121, in to solve MaOPs [3], [4]. Environmental selection is the key
part by the Guangdong Foundation of Outstanding Young Teachers in
Higher Education Institutions under Grant Yq2013141, in part by the operation in MaOEAs to achieve good tradeoff between
Guangdong Special Support Program of Top-Notch Young Professionals under population convergence and diversity. For example, the
Grant 2014TQ01X273, and in part by the Shenzhen Scientific Research and decomposition-based selection strategies maintain good
Development Funding Program under Grant JCYJ20170302154227954, Grant
JCGG20170414111229388, and Grant JCYJ20170302154328155. This paper population diversity via uniformly distributed reference
was recommended by Associate Editor K.-C. Tan. (Corresponding author: vectors and enhance the convergence through the aggregation
Zexuan Zhu.) functions. In these methods, a target MaOP is decomposed
Z. Liang, K. Hu, and X. Ma are with the College of Computer Science and
Software Engineering, Shenzhen University, Shenzhen 518060, China (e-mail: into multiple single-objective optimization subproblems
[email protected]; [email protected]; [email protected]). using a set of reference vectors, and the subproblems are
Z. Zhu is with the College of Computer Science and Software Engineering, optimized simultaneously in an evolutionary framework. An
Shenzhen University, Shenzhen 518060, China, and also with Shenzhen
Pengcheng Laboratory, Shenzhen 518000, China (e-mail: [email protected]). MOEA based on decomposition (MOEA/D) [21] is one of the
This paper has supplementary downloadable multimedia material available most typical decomposition-based MOEAs. Many MaOEAs
at http://ieeexplore.ieee.org provided by the authors. are implemented based on MOEA/D by introducing new
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. operators specifically designed for MaOPs [22]–[24]. The
Digital Object Identifier 10.1109/TCYB.2019.2918087 decomposition-based algorithms are computationally efficient,
2168-2267 c 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
2 IEEE TRANSACTIONS ON CYBERNETICS

yet the population diversity of such algorithms is maintained detail. Section IV presents the experimental design and results.
by a set of predefined reference vectors, which might not Finally, Section V concludes this paper.
work well on MaOPs of irregular PFs [25].
The indicator-based environmental selection, which is eas- II. R ELATED W ORK
ier to implement, represents another commonly used strat- Environmental selection plays a key role in the performance
egy to handle MaOPs. Performance indicators are used of MaOEAs. The goal of environmental selection is to pursue
to select individuals during the environmental selection in the uniform distribution of individual solutions along the PF
MaOEAs [26]–[28]. The hypervolume (HV) [29] and the while maintaining good convergence. Generally, the existing
inverted generational distance (IGD) [30] are the two most environmental selection strategies can be classified into three
widely used indicators. They have achieved success in var- groups.
ious MaOPs. Nevertheless, using a single specific indicator The first group contains the decomposition-based selection
inevitably imposes bias in the selection that might dete-
strategies. To maintain good population diversity and conver-
riorate the generality of the algorithm. Some algorithms,
gence, many effective selection methods are integrated into
like SRA [31] and 1by1EA [32], consider multiple indicators
the MOEAs based on problem decomposition. For example,
simultaneously to balance the bias of the individual indicators
RVEA [33] devises a novel aggregation function (APD) of
and take advantage of each of them. Multiple indicators-based
decomposed subproblems to calculate the fitness of individ-
methods tend to obtain better performance than the counter-
uals. Based on the fitness evaluation and reference vectors,
part single indicator-based methods at the cost of involving
the population in RVEA is partitioned into N subpopulations,
extra parameters.
and the best individual in each subpopulation is selected into
In this paper, we propose a new two-round environmen-
the next generation. NSGA-III [34] associates each population
tal selection strategy to balance the population convergence
member with a reference vector based on its perpendicular
and diversity. The strategy is reference-vector free and multi-
distance to the reference vector. With this method, the popu-
indicator based. In the first round of the selection, solutions
lation members are decomposed into N subpopulations, and
with smaller neighborhood density are picked out to guar-
the Pareto nondominated individual with the shortest per-
antee the population diversity. In the second round, the
solution with the best convergence is selected from the out- pendicular distance from the reference vector is chosen to
put of the first round and survives into the next generation. be a member of the next-generation population. θ -DEA [5]
The selection procedure is repeated until the new popula- utilizes a clustering operator to decompose the population
tion is generated. The two-round selection is characterized members into a set of N clusters based on reference vec-
by a novel adaptive position transformation (APT) strategy tors, and the nondominated sorting based on θ -dominance
for distance calculation within the evaluation of individual is used to divide the population into multiple levels. The
neighborhood density. Particularly, in APT, the PF shape individuals in the first level are selected into next genera-
is estimated first and the distance between two individuals tion. The decomposition-based methods are computationally
is then calculated according to their mappings on the esti- efficient and exhibit good convergence capability. However,
mated PF rather than the original positions. The proposed the population diversity in such methods depends heavily on
environmental selection is integrated into an MaOEA frame- accurate generation of reference vectors, which might not be
work and the resulting algorithm is named 2REA. 2REA is available in advance. Even uniformly distributed reference vec-
compared with eight state-of-the-art MaOEAs on various tors might struggle to maintain evenly distributed solutions on
well-known benchmark MaOPs. The experimental results MOPs with a nonlinear PF (i.e., convex or concave) [25]. For
show that 2REA is capable of obtaining high-quality solu- example, based on uniformly distributed reference directions
tions and the two-round environmental selection works well {(f1 , f2 )|f1 + f2 = 1, f1 , f2 ≥ 0}, the obtained Pareto-optimal
on maintaining good balance of population diversity and solutions in biobjective optimization problems with different
convergence. PF shapes are shown in Fig. 1. It is observed that the distri-
The contributions of this paper are highlighted as follows. bution of Pareto-optimal solutions is uniform only when the
1) A two-round environmental selection strategy is PF shape is linear. On a concave (convex) PF, the distribu-
presented as an option to achieve good population diver- tion of the Pareto-optimal solutions on the boundaries is more
sity and convergence in solving MaOPs. The strategy is dense (sparser) than the other parts.
computationally efficient and can be easily incorporated The second group consists of indicator-based selection
into most exiting MaOEAs. strategies. In the environmental selection of indicator-based
2) An APT strategy is proposed to estimate the neigh- MaOEAs, some performance indicators (e.g., HV, IGD, and
borhood density of each individual. APT is simple yet Iε+ ) are employed to measure the quality of individuals,
efficient to handle various PF shapes, as it does not rely which helps to speed up the convergence to the true PF
on reference vectors. and simultaneously maintain good diversity. For example,
3) The resulting algorithm 2REA is extensively evaluated MaOEA/IGD [35] and AR-MOEA [36] apply an IGD indi-
and it can serve as a candidate solution for MaOPs. cator to select the solutions with favorable convergence and
The remainder of this paper is organized as follows. Section II diversity in each generation. Two_Arch2 [37] adopts a new
reviews the related environmental selection strategies of Lp -norm-based diversity maintenance scheme to enhance the
MaOEAs. Section III describes the proposed algorithm in diversity of an Iε+ indicator. SRA [31] utilizes a simple
LIANG et al.: MaOEA BASED ON TWO-ROUND SELECTION STRATEGY 3

Algorithm 1 General Framework of 2REA

Input: N (population size)
Output: S (final population)
1 S = Initialize(N)
2 While termination criterion is not fulfilled do
3 S = Mating_Selection(S)
4 S = Reproduction(S )
5 R=S∪S
6 R = Normalizatiom(R)
7 d = Distribution_Calculation(R)
8 c = Convergence_Calculation(R)
9 S = Environmental_Selection(R,c,d)
10 End While
11 Return S

sampled in the decision space. Second, the evolutionary opera-

Fig. 1. Distribution of uniform reference vectors in different PF shapes.
tors, including selection, crossover, and mutation, are executed
to generate offspring individuals. Third, the objective nor-
stochastic ranking technique to balance the ISDE [38] indi-
malization is performed to ensure a more reliable individual
cator and Iε+ indicator [39]. 1by1EA [32] uses a one-by-
distance measurement. Finally, the distribution indicator and
one selection strategy based on a convergence indicator and
convergence indicator are calculated and the two-round envi-
a distribution indicator for environmental selection. Once
ronmental selection is carried out to form the new population.
a solution is selected, its neighbors are de-emphasized by
The evolution procedure is repeated until some stopping crite-
using a niching technique to guarantee population diversity.
rion is satisfied. The key components of 2REA are described
The indicator-based selection strategies are easy to implement,
in detail as follows.
but they might suffer from the shortcomings: 1) the evalua-
tion cost of some indicators, for example, HV, is not trivial and
B. Mating Selection Strategy and Reproduction
2) some indicators (e.g., IGD and Iε+ ) are not strictly Pareto
compliant [40], which leads to performance degradation of the After population initialization, mating selection is carried
algorithms [41]. out in the evolution procedure. To obtain high-quality off-
The recently proposed selection strategies, which do not fall spring individuals, the selected parent individuals should have
within the first two groups, form the third group. For example, good convergence and diversity. A binary tournament selec-
the vector-angle-based evolutionary algorithm (VaEA) [42] uti- tion strategy based on the convergence indicator (described in
lizes the maximum-vector-angle-first strategy to select elite Section III-D) and neighborhood density estimation strategy
individuals. The selected solutions might converge poorly are used to select the promising parents for recombination.
in some cases; therefore, a worse-elimination principle is Binary tournament selection strategy can well maintain off-
adopted to replace the poor individuals with the better spring diversity and avoid getting trapped in local optima. It
ones. The coordinated selection strategy evolutionary algo- has been widely used to solve MaOPs.
rithm (MaOEA-CSS) [43] selects two individuals with the The neighborhood density of an individual is calculated as
smallest cosine value from the population, and then elim- follows:
inates the one more poorly converged. The two selection 1
strategies are in favor of diversity maintenance and they turn  (x) = k (2)
i=1 dAPT (x, i ) + 1
to other selection approaches to enhance the convergence
performance. Yet, it is usually difficult to tune the parameters where i is the ith nearest neighbor of x in population
of the convergence enhancement approaches. S, dAPT (x, i ) calculates the APT distance (described in
The existing environmental selection strategies of MaOEAs Section III-D) between x and i , and k defines the neigh-
have shown competitive performance in balancing population borhood size. The neighborhood density of an individual is
convergence and diversity. However, they also suffer from inversely proportional to the summarized distance
 of the indi-
some inherent disadvantages. In this paper, we try to solve vidual to its k-nearest neighbors, that is, ki=1 dAPT (x, i ).
the aforementioned issues of the existing selection strategies A smaller summarized distance indicates a greater neigh-
and propose a computationally efficient two-round-selection borhood density of an individual x in the objective space.
strategy, which does not rely on reference vectors. The details The neighborhood density estimation method is similar to
of the proposed strategy and the resulting algorithm 2REA are the kth nearest neighbor method used in [7], which relies
provided in the following section. on the Euclidean distance. However, the Euclidean distance
might not work well on MaOPs [29], so the APT distance
is proposed in this paper. Given two randomly selected indi-
III. P ROPOSED 2REA
viduals x and y from the population S, the one with smaller
A. General Framework neighborhood density and convergence indicator value is cho-
The framework of the proposed 2REA is outlined in sen for mating. If there is no winner with respect to both
Algorithm 1. First, a population S of N individuals is randomly metrics, then one is randomly selected. After the mating
4 IEEE TRANSACTIONS ON CYBERNETICS

selection, the selected parent individuals are employed to gen-

erate N offspring individuals with simulated binary crossover
(SBX) and polynomial mutation [44]. The offspring individ-
uals and the parent population are combined to form a new
population R.

C. Objective Normalization
After mating selection and reproduction, the objective val-
ues of all individuals in the combined population R are (a) (b)
normalized into a uniform scale. In a target MaOP, the
scales of the objectives could largely vary, and the dis-
tance measurement defined in the objective space to maintain
population diversity could therefore be dominated by some
objectives without normalization [45]. In this regard, objective
normalization is necessary to ensure a more reliable distance
measurement. Given a solution x ∈ R, each of its objectives
fi (x), i = 1, 2, . . . , M, is normalized as follows:
fi (x) − z∗i
fi (x) = , i = 1, 2, . . . , M (3) (c) (d)
znad
i − z∗i
Fig. 2. Simple biobjective minimization example to illustrate the process of
where z∗ and znad are the ideal point and nadir point, respec- PF shape estimation.
tively. Since the PF is unknown, z∗ and znad are usually
estimated based on the current population. Each element of
z∗ = (z∗1 , z∗2 , . . . , z∗M ) can be estimated as the minimum PF geometric shape. Mathematically, the Lp-norm distance is
value of the corresponding objective found in the nondomi- defined as follows:
nated individuals. In the proposed 2REA, znad is approximated M 1/p
by M corner individuals. A corner individual, denoted as xci ,  p
g(x|p) = fi (x) ,p > 0 (6)
is the individual closest to a corresponding objective axis i in
i=1
the objective space, that is
    where M is the number of objective functions.
xci = x|x = arg min dist± x, ei , i = 1, 2, . . . , M (4) The PF shape of an MOP can be approximated using
the following method. First, the Lp -norm distances Gp =
where dist± (x, ei ) indicates the perpendicular Euclidean dis-
{g(x1 |p), g(x2 |p), g(x3 |p), . . . , g(xN |p)} of all individuals to
tance between a vector x and a direction vector of the ith
the origin point with different p-values are calculated. Second,
axis ei . The nadir point znad can be defined as
the standard deviation Vp is computed, that is
znad = xci,i , i = 1, . . . , M (5)  
i
Vp = std Gp . (7)
where xci,i is the ith objective value of the corner individual
A smaller Vp value indicates a better distribution consistence
associated with the ith axis.
of the individuals in terms of the LP -norm distance. The PF
shape is approximated by the contour curve of the Lpmin -
D. Calculation of Distribution and Convergence Indicators norm method, where pmin is the p-value corresponding to the
After normalization, distribution and convergence indica- smallest Vp .
tors are calculated, based on which good tradeoff between To alleviate the impact of noise in the estimation of the
population diversity and convergence could be achieved. PF shape, the box plot [48] is used to detect the outlier val-
The pivotal issue of maintaining good population diversity ues in Gp . Specifically, values greater than Q3 + δ ∗ (Q3 − Q1 )
is to estimate the neighborhood density of the individuals. are considered noise, where Q1 and Q3 are the lower and
The majority of the existing MaOEAs addresses this issue upper quartiles of Gp , respectively, and the parameter δ is set
by applying the Euclidean distance, the cosine value of two to 1.5 following [48]. The noise values are deleted from Gp
objective vectors, or the shift-based density estimation [38]. before calculating the standard deviation Vp .
Nevertheless, most of these methods are not suitable for prob- An example is provided in Fig. 2 to illustrate the process
lems with a large number of objectives and/or incompetent of PF shape estimation. Given five solutions {A, B, C, D, E}
to handle different shapes of PFs. To address these problems, of a biobjective problem, the contour lines of the LP -norm
this paper proposes the APT strategy to estimate the popu- distance with different p-values, where p = 0.5, 1, 1.5, and 2,
lation diversity. Particularly, the distribution indictor evaluates in the objective space are plotted in Fig. 2(a)–(d), respectively.
the difference between two individuals in the population based It is shown that with p = 0.5, the standard deviation, that
on the Lp -norm distance [46], [47] and the estimation of the is, V 0.5 , is the smallest and the PF geometric shape is best
LIANG et al.: MaOEA BASED ON TWO-ROUND SELECTION STRATEGY 5

Algorithm 2 Shape_Estimate (R) distance, that is,

Input: R (combined population), candidate CP = {0.1, 0.2, 0.3, . . . , 10} M
Output: p value  2
1 For i = 1 to length(Cp ) dAPT (x, y) =  f̃i (x) − f̃i (y) , x, y ∈ R. (9)
2 p = Cp [i]
i=1
3 For j = 1 to length(R)
4 x = R[j]  For example, as shown in Fig. 3, suppose the PF surface is
5 Gp [j] = ( M p 1/p
u=1 fu (x) ) best estimated with p = 0.5, and the individuals A, B, C, D,
6 End For
7 Gp is denoised using box plot and E in the population are mapped to A , B , C , D , and
8 Vp[i] = Std(Gp ) E on the estimation PF, respectively. The Euclidean distances
9 End for
10 index = arg min (Vp)
between every pair of A , B , C , D , and E are calculated as
11 Return Cp [index] the distribution indicator of the population.
Besides population diversity, population convergence is
another major issue considered in the design of an
Algorithm 3 Distribution_Calculation (R) MaOEA. After distribution estimation, a convergence indicator
Input: R (combined population) of each solution is measured using the sum of each objective
Output: DAPT (the distribution indictor)
1 p = Shape_Estimate (R) value as
2 Foreach x∈R
f (x)
M
3 f̃i (x) = M i
(
, i = 1, . . . , M c(x) = fi (x). (10)
u=1 fu (x) )
p 1/p
4 End For i=1
5 Foreach x,y ∈ R
M The sum of objective values is a simple aggregation function,
6 dAPT (x, y) = i=1 (f̃i (x) − f̃i (y))
2
which can reflect the convergence information of a solution
7 End For
8 Return DAPT = {dAPT (x, y)|x, y ∈ R} and is widely used in MaOEAs, such as VaEA [42] and
ISDE+ [49]. A smaller c(x) indicates that x converges closer
to the PF.

E. Two-Round Environmental Selection Strategy

In the two-round environmental selection, the individuals in
population R are selected into the next generation population S
one by one based on the distribution indicator and convergence
indicator.
First of all, S is initialized to contain the best converged
individual in R. Afterward, a two-round selection strategy is
applied to select individuals one by one from R to S. In the
first round selection, the minimum APT distance of each non-
dominated individual x in R to the existing individuals in S
is calculated, that is, dmin = miny∈S dAPT (x, y). A larger dmin
value suggests a smaller neighborhood density (with k = 1) of
the individual x. The corner individuals are the most represen-
Fig. 3. Illustration of position transformation strategy.
tative individuals on the PF and their dmin values are set to
+∞, such that they are certainly selected into S. All nondomi-
nated individuals are sorted in descending order based on their
approximated by the corresponding contour lines as shown in dmin values. The top-ranked N − |S|(| · | is the cardinality func-
Fig. 2(a). The PF is, therefore, estimated by the contour curve tion) nondominated individuals are moved to a temporary set
of the L0.5 -norm method. The pseudocode of the PF geometric RT . If the number of nondominated individuals in R is smaller
shape estimation is provided in Algorithm 2. than N − |S|, all nondominated individuals move to RT . In
Once the PF is estimated, the APT distances between every the second-round selection, the best converged individual r
two individuals in R are calculated as the distribution indicator in RT , that is, r = arg min(c(RT )) is moved to S. The above
of the population. The corresponding pseudocode is provided two-round selection procedure is repeated until the size of
in Algorithm 3. First, the mapping objective of each individual population S reaches N. The pseudocode of the environmental
in R is calculated by selection procedure is provided in Algorithm 4.

fi (x) IV. E XPERIMENTAL S TUDY

f̃i (x) = , i = 1, . . . , M (8)
M 1/p The performance of the proposed algorithm 2REA is eval-
u=1 fu (x)
p
uated and compared to the other state-of-the-art MaOEAs on
benchmark MaOPs. The effects of the APT strategy are also
where p is estimated in Algorithm 2. Then, the Euclidean dis- investigated. The experimental design and the analysis of the
tance between two mapping positions is calculated as the APT experimental results are presented as follows.
6 IEEE TRANSACTIONS ON CYBERNETICS

TABLE I
Algorithm 4 Environmental_Selection(R, c, d) S ETTING OF THE P OPULATION S IZE
Input: R (combined population), c (the convergence indicator), d (the
distribution indictor)
Output: S (next generation population)
1 Move the individual r = arg min(c(R)) to S
2 While |S| < N do
3 Let RN be the set of all non-dominated individuals in R
4 Foreach x∈RN
5 dmin = miny∈S (dAPT (x,y))
6 End For
7 Set the dmin value of each corner individual to +∞ TABLE II
8 Sort the individuals in RN in descending order in terms of dmin PARAMETER S ETTINGS FOR C ROSSOVER AND M UTATION
9 If |RN | > N − |S|
10 Let RT be the set of the top ranked N −|S| individuals in RN
11 Else
12 RT = RN // add all non-dominated individuals to RT
13 End If
14 r = arg min(c(RT ))
15 S =S∪r // add the selected individual r to S
16 Remove r from R
17 End While
18 Return S

set to y∗ = (1.5, 1.5, . . . , 1.5) in these problems. To evalu-

ate the diversity of the final solution set of each algorithm,
A. Experimental Design the SPREAD indicator is adopted to measure the extent of
1) Test Problems: In this paper, the test suites, includ- spread by the final solution set. The details of SPREAD can
ing DTLZ [50], WFG [51], and MaF [52] are used to assess be found in [42]. A smaller SPREAD value indicates a more
the performance of the proposed algorithm. DTLZ and WFG diverse final solution set.
are the two most commonly used MaOP test sets. They are 3) Compared Algorithms and Parameter Settings:
characterized by various PF shapes, including concave, lin- 2REA is compared with eight representative state-of-the-art
ear, disconnected, degenerate, multimodal, scaled, and biased MaOEAs, namely, NSGA-III [34], 1by1EA [32], VaEA [42],
PFs. In DTLZ problems, only DTLZ1–4 and DTLZ7 are con- RVEA [33], MaOEA-CSS [43], SRA [31], AR-MOEA [36],
sidered, since the PFs of DTLZ5 and DTZL6 are not well and SPEA2+SDE [38]. The compared algorithms, falling
defined when the number of objectives exceeds three [40]. within different categories, have been shown to be effec-
MaF is a newly proposed MaOP test set, which poses more tive in dealing with various MaOPs. Particularly, NSGA-III
challenges to the existing MaOEAs. In MaF problems, only and RVEA are decomposition-based algorithms. SRA,
MaF1–6, MaF8, MaF9, MaF13, MaF14, and MaF15 are SPEA2+SDE, and AR-MOEA are indicator-based algorithms.
considered because the remaining problems, that is, MaF7, VaEA, 1by1EA, and MaOEA-CSS are characterized by novel
MaF10, MaF11, and MaF12, have been involved in DTLZ selection strategies. The corresponding brief descriptions of
and WGF sets. The parameters of DTLZ and WFG prob- these algorithms are provided in Sections I and II.
lems are set according to [20] and [33], and the corresponding For the compared algorithms, the population size and
parameters of MaF problems are configured following [52]. the parameters used to control the number of reference
2) Performance Metrics: HV and SPREAD [53] indicators vectors are summarized in Table I. Other parameter set-
are used to evaluate the performance of the compared algo- tings of the compared algorithms are set according to the
rithms in this paper. Since the IGD metric requires a large corresponding references. The SBX and polynomial muta-
number of sampling points to obtain accurate estimation in tion (PM) [44] are employed to generate the offspring solutions
MaOPs [32], [54], [55], it might not be practical to apply in 2REA. The parameters of SBX and PM are presented in
IGD in MaOPs. The HV metric has a strict Pareto-compliant Table II. The parameter δ is set to 1.5 following [48] and
nature and nice theoretical qualities [40], [56], [57] that is k is set to 0.1∗ N following [29]. The range of p is set
used to confirm the effectiveness of the proposed algorithm to {0.1, 0.2, 0.3, . . . , 10.0}. In principle, the range should be
here. The HV value can be obtained by calculating the vol- (0, +∞). However, since p cannot be zero and the con-
ume of the objective space between the obtained solutions set tour lines vary within small limits when p > 10, the range
and a specified reference point in the objective space. The [0.1, 10] is a reasonable selection. The interval 0.1 enables
final solution set is normalized if the PFs of the problems a fine-grained search of the range. The sensitivity analyses
have differently scaled objective values. To improve the com- of k and p are provided in Section A of the supplemen-
putational efficiency, the Monte Carlo sampling method [58] tary material. The fluctuations of p-values as the evolutionary
is used to compute HV values, and the required reference search proceeds are also presented in the supplementary mate-
point is set to y∗ = (1.1, 1.1, . . . , 1.1). However, for MaF1, rial. The termination criterion of all algorithms is defined
MaF4, MaF8, MaF9, and MaF15 problems, the region deter- in terms of the maximum number of generations (Gens).
mined by the final solution set and the reference point is On DTLZ1, DTLZ3, DTLZ4, WFG1, WFG2, WFG3, MaF1,
very small, that is, the sampling points in the hypercube MaF3, MaF13, MaF14, and MaF15, all algorithms are termi-
hardly fall into the region. Therefore, the reference point is nated when a maximum of 1000 Gens is reached. For the other
LIANG et al.: MaOEA BASED ON TWO-ROUND SELECTION STRATEGY 7

TABLE III
P ERFORMANCE C OMPARISON OF 2REA TO OTHER A LGORITHMS IN T ERMS OF AVERAGE HV ON DTLZ P ROBLEMS .
T HE B EST AVERAGE VALUE A MONG THE A LGORITHMS FOR E ACH I NSTANCE I S H IGHLIGHTED IN B OLD

Fig. 4. Average performance score over all objective dimensions for different test problems, namely, DTLZ (Dx), WFG (Wx), and MaF (Mx). The smaller
the score, the better the PF approximation in terms of HV. The values of 2REA are connected by a solid line to assess the score more easily.

test problems, the algorithms are stopped when a maximum study of 2REA with the other state-of-the-art algorithms on the
of 500 Gens is reached. benchmark problems. The aim of this part is to demonstrate
The Mann–Whither–Wilcoxon rank-sum test [59] is used the effectiveness of 2REA in achieving the desired conver-
to quantify whether one algorithm is better than the others gence and diversity. The second part verifies the effectiveness
in a statistically meaningful sense. 2REA is compared with of the APT strategy.
the other eight MaOEAs over 20 independent runs in all 1) Comparison With Other State-of-the-Art Algorithms:
selected test problems. Symbols “+,” “≈,” and “−” are used 2REA is compared with the other state-of-the-art MaOEAs
to denote that 2REA is worse than, equivalent to, and better on DTLZ, WFG, and MaF test problems with M set to 5,
than a corresponding competing algorithm in a significance 10, and 15, respectively. Tables III–V show the mean val-
level of 5%, respectively. ues of HV on DTLZ, WFG, and MaF problems, respectively.
The best result on each test problem is highlighted in bold-
face. To enable a comparison of the overall performance, the
B. Experimental Results performance score [58] is also introduced to rank the algo-
In this section, the analyses of experimental results are rithms. Fig. 4 shows the ranks of the algorithms on all test
divided into two parts. The first one presents the comparison problems in terms of HV. Fig. 5(a) and (b) shows the average
8 IEEE TRANSACTIONS ON CYBERNETICS

TABLE IV
P ERFORMANCE C OMPARISON OF 2REA TO OTHER A LGORITHMS IN T ERMS OF AVERAGE HV ON WFG P ROBLEMS .
T HE B EST AVERAGE VALUE A MONG THE A LGORITHMS FOR E ACH I NSTANCE I S H IGHLIGHTED IN B OLD

performance score over all 75 test instances for all algorithms MaF14 is not as competitive as that for the other problems.
in terms of HV and SPREAD, respectively. The smaller the The high dimensionality of MaF14 calls for a more computa-
score is, the better the algorithm performance is. tional budget to reach convergence, whereas 2REA tends to put
According to the empirical results, 2REA achieves the best more effort in maintaining population diversity, that is, ineffi-
overall performance in terms of HV and SPREAD values. cient computational resource is allocated for pursuing conver-
Thanks to the two-round selection strategy, 2REA attains bet- gence, which could deteriorate the convergence performance
ter balance of population convergence and diversity than the on MaF14.
other algorithms. The superiority of 2REA is more obvious RVEA and NSGA-III are decomposition-based algorithms.
in dealing problems with concave or convex PFs, for exam- As shown in Table III, RVEA outperforms the other algorithms
ple, DTLZ2–4, WFG4–9, and MaF3, as shown in Fig. 4. in most instances of DTLZ1, because of the well-predefined
The APT strategy used to estimate the diversity of popula- reference vectors that are evenly distributed on the surface of
tion does help to handle the concave and convex problems the hyperplane. RVEA and NSGA-III maintain better popu-
more efficiently. The PFs of some test problems are difficult lation diversity than the other algorithms. Since most PFs of
to estimate, for example, the PFs of DTLZ7 and WFG2 are the DTLZ and WFG test problems are regular, RVEA and
disconnected, and the PF of WFG1 contains both convex and NSGA-III show competitive performance on these problems
concave segments. It is difficult for the other algorithms to as expected. However, on MaF problems, some of the PFs
accurately estimate such PFs, yet the APT strategy works are irregular. For example, MaF1 and MaF4 have inverted
well on such PFs, leading to superior performance on the PFs, and MaF6 is featured by a degenerate PF. In these cases,
corresponding test problems. The performance of 2REA on most reference vectors could become invalid; therefore, it is
LIANG et al.: MaOEA BASED ON TWO-ROUND SELECTION STRATEGY 9

TABLE V
P ERFORMANCE C OMPARISON OF 2REA TO OTHER A LGORITHMS W ITH R ESPECT TO THE AVERAGE HV ON MAF P ROBLEMS .
T HE B EST AVERAGE VALUE A MONG THE A LGORITHMS FOR E ACH I NSTANCE I S H IGHLIGHTED IN B OLD

difficult for decomposition-based algorithms to maintain good terms of average performance, as shown in Fig. 5(a). AR-
distribution. MOEA adopts an adaptive reference points adjustment strat-
VaEA and MaOEA-CSS adopt the cosine value of two egy, yet it is still difficult to adjust the reference points for the
objective vectors to estimate the difference of two individ- test problems with irregular PFs. SPEA2+SDE achieves good
uals. Thanks to their selection strategies, both algorithms population convergence to the true PF, but it is impotent to
manage to obtain good population diversity, however, there distribute the solution set uniformly on concave PFs, that is,
is still room for the improvement of convergence. As the it tends to miss solutions near the boundaries. 1by1EA shows
results show in Table III, VaEA and MaOEA-CSS achieve rela- good population convergences on the test problems, but the
tively good performance on DTLZ2 and DTLZ4 problems, but diversity of the obtained final population might not be well
their performance on DTLZ3 deteriorates sharply. DTLZ3 is maintained.
a highly multimodal problem and it is more difficult for VaEA To show the distribution of the final population more intu-
and MaOEA-CSS to converge to the PF on DTLZ3 than on itively, Fig. 6 plots the final population of all algorithms on
DTLZ2 and DTLZ4. a representative problem MaF3 with M = 15 by parallel
AR-MOEA, SRA, SPEA2+SDE, and 1by1EA are four coordinates [58]. It can be seen that the solution set obtained
indicator-based algorithms. AR-MOEA and SPEA2+SDE are by 2REA is the best in terms of convergence and diversity. The
ranked as the second and third positions, respectively, in solutions obtained by VaEA, NSGA-III, and MaOEA-CSS are
10 IEEE TRANSACTIONS ON CYBERNETICS

(a) (b)

Fig. 5. (a) Ranking in the average performance score over all test problem instances for the selected ten algorithms. The smaller the score, the better the
overall performance in terms of HV. (b) Ranking in the average performance score over all test problem instances for the selected eight algorithms. The
smaller the score, the better the overall performance in terms of SPREAD.

TABLE VI
P ERFORMANCE C OMPARISON B ETWEEN 2REA, 2REA-SDE, 2REA-ED, 2REA-A NGLE , AND 2REA-PCD W ITH
R ESPECT TO THE AVERAGE HV ON S IX R EPRESENTATIVE T EST P ROBLEMS

not able to converge to the PF. 1by1EA, AR-MOEA, RVEA, methods. The corresponding variants are denoted as 2REA-Ed,
SRA, and SPEA2+SDE obtain relatively well-converged solu- 2REA-Angle, 2REA-PCD, and 2REA-SDE, respectively. Six
tion sets on MaF3, but the corresponding population diversities representative test problems with different PF shapes, that is,
are unsatisfactory. DTLZ1, DTLZ2, WFG1, WFG2, MaF1, and MaF3 are chosen
2) Effectiveness of the APT Strategy: The APT strategy is to investigate the performances of the five variant algorithms.
the key component of 2REA to maintain population diver- The mean HV values of the five algorithms on the test prob-
sity. To assess the effects of this strategy, we test 2REA by lems are shown in Table VI. It can be seen that 2REA achieves
replacing it with other diversity estimation strategies, includ- better performance than the other variants on most of the test
ing Euclidean distance, angle, PCD [60], and SDE-based problems. The PFs of WFG1 and WFG2 are concave–convex
LIANG et al.: MaOEA BASED ON TWO-ROUND SELECTION STRATEGY 11

Fig. 6. Visualization of PF approximation by nine algorithms on the 15-objective MaF3.

mixed and nonlinear, respectively, which can only be roughly first round selection characterized by a novel APT strat-
or partially estimated by the APT strategy, yet the strategy egy is adopted to maintain the diversity of the population,
still manages to obtain good performance in these problems. and the second-round selection is utilized to enhance the
The observation suggests that the APT strategy is more robust convergence. The two-round selection strategy is integrated
than the other diversity estimation strategies in maintaining into an MaOEA framework to form the resulting algorithm
diversity. 2REA. The proposed 2REA is compared with eight state-
It is noted that real-world problems could be much more of-the-art MaOEAs, including NSGA-III, 1by1EA, VaEA,
complicated than the benchmark problems, and the method RVEA, MaOEA-CSS, SRA, AR-MOEA, and SPEA2+SDE
proposed in this paper may not be able to accurately esti- on DTLZ, WFG, and MaF problems. The experimental results
mate the true PF shapes in such problems. However, as long show that 2REA is very competitive with the compared
as the PF shape can be roughly or partially estimated, the algorithms. The proposed two-round environmental selection,
performance of the algorithm can benefit from the estima- together with the APT, succeeds in achieving good tradeoff
tion. In this regard, the algorithm proposed in this paper between population convergence and diversity.
does not depend too much on the problem features. As the Although 2REA has shown superior performance on var-
complexity analysis shown in Section C of the supplemen- ious test problems, there are still some unresolved issues.
tary material, the proposed algorithm 2REA also demonstrates For example, 2REA should be further improved to better
good time efficiency. The experimental results presented in handle irregular PFs, where the geometric shape estimation
Section D of the supplementary material on five represen- is more challenging. More complicated problems should be
tative constraint MaOPs, namely, C1-DTLZ1, C2-DTLZ2, considered in future work to verify the scalability of 2REA.
C3-DTLZ4, DTLZ8, and DTLZ9 demonstrate the capability 2REA is implemented in MATLAB based on the platform
of the proposed algorithm to handle constraint problems. of PlatEMO [61]. The source code is publically available at
https://github.com/CIA-SZU/HKF.

V. C ONCLUSION R EFERENCES
This paper proposed a new environmental selection method, [1] A. Zhou et al., “A survey of multiobjective evolutionary algorithms,”
namely, the two-round selection strategy, for MaOEAs. The Swarm Evol. Comput., vol. 1, no. 1, pp. 32–49, Mar. 2014.
12 IEEE TRANSACTIONS ON CYBERNETICS

[2] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist [26] D. A. V. Veldhuizen and G. B. Lamon, “Multiobjective evolution-
multiobjective genetic algorithm: NSGA-II,” IEEE Trans. Evol. Comput., ary algorithm research: Analyzing the state-of-the-art,” Evol. Comput.,
vol. 6, no. 2, pp. 182–197, Apr. 2002. vol. 8, no. 2, pp. 125–147, Feb. 2000.
[3] V. A. Shim, K. C. Tan, and H. J. Tang, “Adaptive memetic computing for [27] N. Beume, B. Naujoks, and M. Emmerich, “SMS-EMOA: Multiobjective
evolutionary multiobjective optimization,” IEEE Trans. Cybern., vol. 45, selection based on dominated hypervolume,” Eur. J. Oper. Res., vol. 181,
no. 4, pp. 610–621, Apr. 2015. no. 3, pp. 1653–1669, 2007.
[4] M. Li, S. Yang, and X. Liu, “Diversity comparison of Pareto front [28] D. Brockhoff, T. Wagner, and H. Trautmann, “On the properties of the
approximations in many-objective optimization,” IEEE Trans. Cybern., R2 indicator,” in Proc. Annu. Conf. Genet. Evol. Comput., Philadelphia,
vol. 44, no. 12, pp. 2568–2584, Dec. 2014. PA, USA, 2012, pp. 465–472.
[5] Y. Yuan, H. Xu, B. Wang, and X. Yao, “A new dominance relation- [29] E. Zitzler and L. Thiele, “Multiobjective evolutionary algorithms: A
based evolutionary algorithm for many-objective optimization,” IEEE comparative case study and the strength Pareto approach,” IEEE Trans.
Trans. Evol. Comput., vol. 20, no. 1, pp. 16–37, Feb. 2016. Evol. Comput., vol. 3, no. 4, pp. 257–271, Nov. 1999.
[6] X. Zhang, Y. Tian, and Y. Jin, “A knee point-driven evolutionary algo- [30] A. Zhou, Y. Jin, Q. Zhang, B. Sendhoff, and E. Tsang, “Combining
rithm for many-objective optimization,” IEEE Trans. Evol. Comput., model-based and genetics-based offspring generation for multi-objective
vol. 19, no. 6, pp. 761–776, Dec. 2015. optimization using a convergence criterion,” in Proc. IEEE Int. Conf.
[7] E. Zitzler, M. Laumanns, and L. Thiele, “SPEA2: Improving the Evol. Comput., Vancouver, BC, Canada, Jul. 2006, pp. 892–899.
performance of the strength Pareto evolutionary algorithm,” Comput. [31] B. Li, K. Tang, J. Li, and X. Yao, “Stochastic ranking algorithm for
Eng. Commun. Netw. Lab., Swiss Federal Inst. Technol. (ETH) Zurich, many-objective optimization based on multiple indicators,” IEEE Trans.
Zürich, Switzerland, Rep. TIK-Rep. 103, pp. 95–100, Jan. 2001. Evol. Comput., vol. 20, no. 6, pp. 924–938, Dec. 2016.
[8] X. Ma, Q. Zhang, G. Tian, J. Yang, and Z. Zhu, “On Tchebycheff decom- [32] Y. Liu, D. Gong, J. Sun, and Y. Jin, “A many-objective evolutionary
position approaches for multiobjective evolutionary optimization,” IEEE algorithm using a one-by-one selection strategy,” IEEE Trans. Cybern.,
Trans. Evol. Comput., vol. 22, no. 2, pp. 226–244, Apr. 2018. vol. 47, no. 9, pp. 2689–2702, Sep. 2017.
[9] W. Fang, L. Zhang, S. Yang, J. Sun, and X. Wu, “A multiobjective [33] R. Cheng, Y. Jin, M. Olhofer, and S. Bernhard, “A reference vector
evolutionary algorithm based on coordinate transformation,” IEEE Trans. guided evolutionary algorithm for many-objective optimization,” IEEE
Cybern., vol. 49, no. 7, pp. 2732–2743, Jul. 2019. Trans. Evol. Comput., vol. 20, no. 5, pp. 773–791, Oct. 2016.
[10] Y. Hua, Y. Jin, and K. Hao, “A clustering-based adaptive evolutionary [34] K. Deb and H. Jain, “An evolutionary many-objective optimization
algorithm for multiobjective optimization with irregular Pareto fronts,” algorithm using reference-point-based nondominated sorting approach,
IEEE Trans. Cybern., vol. 49, no. 7, pp. 2758–2770, Jul. 2019. Part I: Solving problems with box constraints,” IEEE Trans. Evol.
[11] H. Han, W. Lu, L. Zhang, and J. Qiao, “Adaptive gradient multiobjective Comput., vol. 18, no. 4, pp. 577–601, Aug. 2014.
particle swarm optimization,” IEEE Trans. Cybern., vol. 48, no. 11, [35] Y. Sun, G. G. Yen, and Z. Yi, “IGD indicator-based evolutionary algo-
pp. 3067–3079, Nov. 2018. rithm for many-objective optimization problems,” IEEE Trans. Evol.
[12] K. Ikeda, H. Kita, and S. Kobayashi, “Failure of Pareto-based MOEAs: Comput., vol. 23, no. 2, pp. 247–257, Apr. 2019.
Does non-dominated really mean near to optimal?” in Proc. IEEE Congr. [36] Y. Tian, R. Cheng, X. Zhang, F. Cheng, and Y. Jin, “An indicator-based
Evol. Comput., vol. 2. Seoul, South Korea, May 2001, pp. 957–962. multiobjective evolutionary algorithm with reference point adaptation
[13] V. Khare, X. Yao, and K. Deb, “Performance scaling of multi-objective for better versatility,” IEEE Trans. Evol. Comput., vol. 22, no. 4,
evolutionary algorithms,” in Evolutionary Multi-Criterion Optimization. pp. 609–622, Feb. 2018.
Heidelberg, Germany: Springer, Apr. 2003, pp. 376–390.
[37] H. Wang, L. Jiao, and X. Yao, “Two_Arch2: An improved two-archive
[14] R. C. Purshouse and P. J. Fleming, “Evolutionary many-objective opti- algorithm for many-objective optimization,” IEEE Trans. Evol. Comput.,
misation: An exploratory analysis,” in Proc. IEEE Congr. Evol. Comput., vol. 19, no. 4, pp. 524–541, Aug. 2015.
Canberra, ACT, Australia, Dec. 2003, pp. 2066–2073.
[38] M. Li, S. Yang, and X. Liu, “Shift-based density estimation for Pareto-
[15] M. Laumanns, L. Thiele, K. Deb, and E. Zitzler, “Combining conver-
based algorithms in many-objective optimization,” IEEE Trans. Evol.
gence and diversity in evolutionary multiobjective optimization,” Evol.
Comput., vol. 18, no. 3, pp. 348–365, Jun. 2014.
Comput., vol. 10, no. 3, pp. 263–282, Sep. 2002.
[39] E. Zitzler and S. Künzli, “Indicator-based selection in multiobjective
[16] K. Deb, M. Mohan, and S. Mishra, “Evaluating the ε-domination
search,” in Parallel Problem Solving From Nature-PPSN VIII.
based multi-objective evolutionary algorithm for a quick computation of
Heidelberg, Germany: Springer, Sep. 2004, pp. 832–842.
Pareto-optimal solutions,” Evol. Comput., vol. 13, no. 4, pp. 501–525,
Feb. 2005. [40] E. Zitzler, L. Thiele, M. Laumanns, C. M. Fonseca, and
[17] X. Zou, Y. Chen, M. Liu, and L. Kang, “A new evolutionary algorithm V. G. Da Fonseca, “Performance assessment of multiobjective
for solving many-objective optimization problems,” IEEE Trans. Syst., optimizers: An analysis and review,” IEEE Trans. Evol. Comput., vol. 7,
Man, Cybern. B, Cybern., vol. 38, no. 5, pp. 1402–1412, Oct. 2008. no. 2, pp. 117–132, Apr. 2003.
[18] M. Garza-Fabre, G. T. Pulido, and C. A. C. Coello, “Ranking meth- [41] B. Li, J. Li, K. Tang, and X. Yao, “Many-objective evolutionary algo-
ods for many-objective optimization,” in Proc. Mexican Int. Conf. Artif. rithms: A survey,” ACM Comput. Surveys, vol. 48, no. 1, pp. 1–35,
Intell., Guanajuato, Mexico, Nov. 2009, pp. 633–645. Sep. 2015.
[19] A. L. Jaimes and C. A. C. Coello, “Study of preference relations in [42] Y. Xiang, Y. Zhou, M. Li, and Z. Chen, “A vector angle-based evolu-
many-objective optimization,” in Proc. Conf. Genet. Evol. Comput., tionary algorithm for unconstrained many-objective optimization,” IEEE
Montreal, QC, Canada, Jul. 2009, pp. 611–618. Trans. Evol. Comput., vol. 21, no. 1, pp. 131–152, Feb. 2017.
[20] J. Cheng, G. G. Yen, and G. Zhang, “A many-objective evolutionary [43] Z. He and G. G. Yen, “Many-objective evolutionary algorithms based
algorithm with enhanced mating and environmental selections,” IEEE on coordinated selection strategy,” IEEE Trans. Evol. Comput., vol. 21,
Trans. Evol. Comput., vol. 19, no. 4, pp. 592–605, Aug. 2015. no. 2, pp. 220–233, Apr. 2017.
[21] Q. Zhang and H. Li, “MOEA/D: A multiobjective evolutionary algorithm [44] K. Deb and R. B. Agrawal, “Simulated binary crossover for continuous
based on decomposition,” IEEE Trans. Evol. Comput., vol. 11, no. 6, search space,” Complex Syst., vol. 9, no. 2, pp. 115–148, 1995.
pp. 712–731, Dec. 2007. [45] K. Deb, K. Miettinen, and S. Chaudhuri, “Toward an estimation of
[22] L. Paquete and T. Stützle, “A two-phase local search for the biob- nadir objective vector using a hybrid of evolutionary and local search
jective traveling salesman problem,” in Evolutionary Multi-Criterion approaches,” IEEE Trans. Evol. Comput., vol. 14, no. 6, pp. 821–841,
Optimization. Heidelberg, Germany: Springer, Apr. 2003, pp. 479–493. Dec. 2010.
[23] Z. Zhou, Y. Xiang, Z. Chen, J. He, and J. Wang, “A scalar projection [46] R. Wang, Q. Zhang, and T. Zhang, “Decomposition-based algorithms
and angle-based evolutionary algorithm for many-objective optimization using Pareto adaptive scalarizing methods,” IEEE Trans. Evol. Comput.,
problems,” IEEE Trans. Cybern., vol. 49, no. 6, pp. 2073–2084, vol. 20, no. 6, pp. 821–837, Dec. 2016.
Jun. 2019. [47] E. C. Jara, “Multi-objective optimization by using evolutionary algo-
[24] Y.-H. Zhang et al., “DECAL: Decomposition-based coevolutionary algo- rithms: The p-optimality criteria,” IEEE Trans. Evol. Comput., vol. 18,
rithm for many-objective optimization,” IEEE Trans. Cybern., vol. 49, no. 2, pp. 167–179, Apr. 2014.
no. 1, pp. 27–41, Jan. 2019. [48] W. John, “Mathematics and the picturing of data,” in Proc. Int. Congr.
[25] H. Ishibuchi, Y. Setoguchi, H. Masuda, and Y. Nojima, “Performance Mathematicians., 1975, pp. 523–531.
of decomposition-based many-objective algorithms strongly depends [49] T. Pamulapati, R. Mallipeddi, and P. Suganthan, “ISDE +—An indicator
on Pareto front shapes,” IEEE Trans. Evol. Comput., vol. 21, no. 2, for multi and many-objective optimization,” IEEE Trans. Evol. Comput.,
pp. 169–190, Apr. 2017. vol. 23, no. 2, pp. 346–352, Apr. 2019.
LIANG et al.: MaOEA BASED ON TWO-ROUND SELECTION STRATEGY 13

[50] K. Deb, L. Thiele, M. Laumanns, and E. Zitzler, “Scalable Kaifeng Hu received the B.S. degree in network
test problems for evolutionary multiobjective optimization,” in engineering from Jiangxi Agricultural University,
Evolutionary Multiobjective Optimization. London, U.K.: Springer, Nanchang, China, in 2016. He is currently pursuing
2005, pp. 105–145. the M.S. degree in computer science and engineer-
[51] S. Huband, P. Hingston, L. Barone, and L. While, “A review of ing from Shenzhen University, Shenzhen, China.
multiobjective test problems and a scalable test problem toolkit,” IEEE His current research interests include evolutionary
Trans. Evol. Comput., vol. 10, no. 5, pp. 477–506, Oct. 2006. computation, many-objective optimization, and their
[52] R. Cheng et al., “A benchmark test suite for evolutionary many-objective applications.
optimization,” Complex Intell. Syst., vol. 3, no. 1, pp. 67–81, Mar. 2017.
[53] H. Jain and K. Deb, “An evolutionary many-objective optimization algo-
rithm using reference-point based nondominated sorting approach, part
II: Handling constraints and extending to an adaptive approach,” IEEE
Trans. Evol. Comput., vol. 18, no. 4, pp. 602–622, Aug. 2014.
[54] H. Ishibuchi, H. Masuda, and Y. Nojima, “A study on performance
evaluation ability of a modified inverted generational distance indica-
tor,” in Proc. Conf. Genet. Evol. Comput., Madrid, Spain, Jul. 2015,
pp. 695–702.
[55] Y. Yuan, H. Xu, B. Wang, B. Zhang, and X. Yao, “Balancing conver-
gence and diversity in decomposition-based many-objective optimizers,” Xiaoliang Ma received the B.S. degree in comput-
IEEE Trans. Evol. Comput., vol. 20, no. 2, pp. 180–198, Apr. 2016. ing computer science and technology from Zhejiang
[56] I. Yevseyeva, A. P. Guerreiro, M. T. Emmerich, and C. M. Fonseca, “A Normal University, Jinhua, China, in 2006 and the
portfolio optimization approach to selection in multiobjective evolution- Ph.D. degree in computing from Xidian University,
ary algorithms,” in Parallel Problem Solving From Nature—PPSN XIII. Xi’an, China, in 2014.
Cham, Switzerland: Springer, Sep. 2014, pp. 672–681. He is currently an Assistant Professor with
[57] A. P. Guerreiro and C. M. Fonseca, “Hypervolume sharpe-ratio indicator: the College of Computer Science and Software
Formalization and first theoretical results,” in Parallel Problem Solving Engineering, Shenzhen University, Shenzhen, China.
From Nature—PPSN XIV. Cham, Switzerland: Springer, Sep. 2016, His current research interests include evolutionary
pp. 814–823. computation, multiobjective optimization, coopera-
[58] J. Bader and E. Zitzler, “HypE: An algorithm for fast hypervolume-based tive coevolution, and bioinformatics.
many-objective optimization,” Evol. Comput., vol. 19, no. 1, pp. 45–76,
Mar. 2011.
[59] L. M. Antonio and C. A. C. Coello, “Use of cooperative coevolution for
solving large scale multiobjective optimization problems,” in Proc. IEEE
Congr. Evol. Comput., Cancún, Mexico, Jun. 2013, pp. 2758–2765.
[60] W. Hu, G. Yen, and G. Luo, “Many-objective particle swarm
optimization using two-stage strategy and parallel cell coordinate
system,” IEEE Trans. Cybern., vol. 47, no. 6, pp. 1446–1459, Jun. 2017.
[61] Y. Tian, R. Cheng, X. Zhang, and Y. Jin, “PlatEMO: A MATLAB plat-
form for evolutionary multi-objective optimization [educational forum],” Zexuan Zhu (M’12) received the B.S. degree
IEEE Comput. Intell. Mag., vol. 12, no. 4, pp. 73–87, Nov. 2017. in computer science and technology from Fudan
University, Shanghai, China, in 2003 and the
Ph.D. degree in computer engineering from
Nanyang Technological University, Singapore, in
2008.
Zhengping Liang received the B.S. degree in com- He is currently a Professor with the College
puter science and technology from Hunan Normal of Computer Science and Software Engineering,
University, Changsha, China, in 2001 and the Shenzhen University, Shenzhen, China. His current
Ph.D. degree in computer science and technology research interests include computational intelli-
from Wuhan University, Wuhan, China, in 2006. gence, machine learning, and bioinformatics.
He is currently an Associate Professor with Dr. Zhu is an Associate Editor of the IEEE T RANSACTIONS ON
the College of Computer Science and Software E VOLUTIONARY C OMPUTATION and the IEEE T RANSACTIONS ON
Engineering, Shenzhen University, Shenzhen, China. E MERGING T OPICS IN C OMPUTATIONAL I NTELLIGENCE. He serves as the
His current research interests include computational Editorial Board Member for Memetic Computing and Soft Computing. He
intelligence, multiobjective optimization, and big is also the Chair of the IEEE CIS, Emergent Technologies Task Force on
data analysis. Memetic Computing.