IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 26, NO.

2, APRIL 2022 263

An Evolutionary Multitasking Optimization

Framework for Constrained Multiobjective
Optimization Problems
Kangjia Qiao , Member, IEEE, Kunjie Yu , Boyang Qu , Jing Liang , Senior Member, IEEE,
Hui Song, and Caitong Yue , Member, IEEE

Abstract—When addressing constrained multiobjective I. I NTRODUCTION

optimization problems (CMOPs) via evolutionary algorithms,
ULTIOBJECTIVE optimization [1] is devoted to find-
various constraints and multiple objectives need to be satisfied
and optimized simultaneously, which causes difficulties for the
solver. In this article, an evolutionary multitasking (EMT)-based
M ing a set of nondominated solutions among multiple
conflicting objectives, which corresponds to the Pareto front
constrained multiobjective optimization (EMCMO) framework (PF) in the objective space. In general, there is a need to
is developed to solve CMOPs. In EMCMO, the optimization of find a well-distributed PF since more choices can be pro-
a CMOP is transformed into two related tasks: one task is for
the original CMOP, and the other task is only for the objectives
vided. During the past several decades, related studies have
by ignoring all constraints. The main purpose of the second developed toward many-objective optimization [2], large-scale
task is to continuously provide useful knowledge of objectives to multiobjective optimization [3], and multimodal multiobjective
the first task, thus facilitating solving the CMOP. Specially, the optimization [4]; and various multiobjective evolutionary algo-
genes carried by parent individuals or offspring individuals are rithms (MOEAs) have been proposed to solve them. Recently,
dynamically regarded as useful knowledge due to the different
complementarities of the two tasks. Moreover, the useful knowl-
some studies began to focus on constrained multiobjective
edge is found by the designed tentative method and transferred optimization [5] because many real-world applications contain
to improve the performance of the two tasks. To the best of both the satisfaction of various constraints and the optimization
our knowledge, this is the first attempt to use EMT to solve of multiple objectives [6], which propose significant chal-
CMOPs. To verify the performance of EMCMO, an instance of lenges for solvers. These problems are called constrained
EMCMO is obtained by employing a genetic algorithm as the
optimizer. Comprehensive experiments are conducted on four
multiobjective optimization problems (CMOPs). Generally, a
benchmark test suites to verify the effectiveness of knowledge CMOP can be formulated as
transfer. Furthermore, compared with other state-of-the-art
constrained multiobjective optimization algorithms, EMCMO
min f(x) = (f1 (x), f2 (x), . . . , fM (x)) (1)
can produce better or at least comparable performance. subject to
Index Terms—Constrained multiobjective optimization, evolu- ⎧
⎨x ∈ S
tionary multitasking (EMT), knowledge transfer, multitasking
gj (x) ≤ 0, j = 1, . . . , m (2)
optimization problem. ⎩
hj (x) = 0, j = m + 1, . . . , n
Manuscript received December 10, 2020; revised April 13, 2021, July 9, where x = (x1 , x2 , . . . , xD ) ∈ S represents a solution with D
2021, October 9, 2021, and December 20, 2021; accepted January 3, 2022. dimensions; S ⊆ RD indicates the decision space; f : S → RM
Date of publication January 25, 2022; date of current version March 31, 2022.
This work was supported in part by the National Natural Science Fund for denotes M objective functions; gj (x) and m, respectively, repre-
Outstanding Young Scholars of China under Grant 61922072; in part by sent the jth inequality constraint and the number of inequality
the National Natural Science Foundation of China under Grant 62176238, constraint; and hj (x) and (n − m), respectively, represent
Grant 61806179, Grant 61876169, and Grant 61976237; in part by the
China Postdoctoral Science Foundation under Grant 2020M682347; in part (j − m)th equality constraint and the number of equality con-
by the Training Program of Young Backbone teachers in Colleges and uni- straint. When a solution satisfies all constraints, it is defined
versities in Henan Province under Grant 2020GGJS006; and in part by the as a feasible solution; otherwise, it is an infeasible solu-
Henan Provincial Young Talents Lifting Project under Grant 2021HYTP007.
(Corresponding author: Jing Liang.) tion. Different from unconstrained multiobjective optimization,
Kangjia Qiao, Kunjie Yu, Jing Liang, and Caitong Yue are with constrained multiobjective optimization aims to find a set of
the School of Electrical Engineering, Zhengzhou University, Zhengzhou feasible nondominated solutions/Pareto-optimal solutions.
450001, China (e-mail: [email protected]; [email protected];
[email protected]; [email protected]). For CMOPs, due to the existence of constraints, the objec-
Boyang Qu is with the School of Electrical and Information Engineering, tive space contains two Pareto optimal fronts: 1) the uncon-
Zhongyuan University of Technology, Zhengzhou 450007, China (e-mail: strained PF (UPF), i.e., the PF obtained without considering
[email protected]).
Hui Song is with the School of Engineering, RMIT University, Melbourne, any constraint and 2) the constrained PF (CPF) corresponding
VIC 3000, Australia (e-mail: [email protected]). to the set of feasible nondominated solutions in the decision
This article has supplementary material provided by the space. Furthermore, based on the distribution of the UPF and
authors and color versions of one or more figures available at
https://doi.org/10.1109/TEVC.2022.3145582. the CPF in the objective space, literature [7] summarizes the
Digital Object Identifier 10.1109/TEVC.2022.3145582 relationships between them into four classes, as presented in
1089-778X 
c 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.
264 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 26, NO. 2, APRIL 2022

knowledge in this article. The feasibility and effectiveness of

EMT have been verified in various real-world problems, such
as the clustered shortest path tree problem [13], capacitated
vehicle routing problem [14], image classification [15], and
feature selection [16].
Inspired by the success of EMT and the similarities of
(a) (b) finding UPFs and CPFs, this article presents a novel EMT
framework to solve CMOPs. The main contributions of this
study are as follows.
1) The CMOP is modeled as a multitasking optimization
problem since finding two PFs can be regarded as two
related tasks. The first task considers both the con-
straints and objectives (i.e., the CMOP), and the second
task considers only the objectives extracted from the
(c) (d) first task (i.e., the MOP). Correspondingly, a novel
EMT-based constrained multiobjective optimization
Fig. 1. Four relationships between CPF and UPF, where the gray area is the (EMCMO) framework is proposed. EMCMO assigns
feasible region, blue line is the UPF, and red line is the CPF. (a) Type I: CPF
is same as UPF. (b) Type II: CPF is part of UPF. (c) Type III: CPF and UPF two populations to, respectively, optimize these two
overlap partially. (d) Type IV: CPF and UPF do not overlap. tasks in a parallel way. To the best of our knowledge,
this is the first study to use the EMT to solve CMOPs.
2) Based on the relationships between the CPF and UPF,
Fig. 1. They are: 1) type-I: the CPF is same as the UPF; EMCMO considers two types of useful knowledge
2) type-II: the CPF is part of the UPF; 3) type-III: the CPF sources: a) parent individuals and b) offspring individu-
and the UPF overlap partially; and 4) type-IV: the CPF and als. Moreover, for a given problem, a tentative method
UPF do not overlap. Although the ultimate purpose of solv- is devised to estimate the effectiveness of two types of
ing CMOPs is to find a well-distributed CPF, the UPF cannot knowledge sources to guide the knowledge transfer.
be ignored since it can provide some auxiliary information to 3) An instance of EMCMO is created by incorporating
find the CPF. For example, when the relationship between the the genetic algorithm as the optimizer. The created
CPF and UPF is type-I or type-II, the task of finding the CPF method is compared with five state-of-the-art con-
can be achieved by finding the UPF, since there is no con- strained multiobjective optimization algorithms on four
straint to prevent the evolution of the population. Conversely, benchmark test suites, and the results demonstrate the
if the population searches for the CPF directly, the popula- effectiveness and superiority of the proposed framework.
tion will fall into a poor feasible region when the CMOP has The remainder of this article is arranged as follows. The
multiple feasible regions. Therefore, it is promising to mea- related works are described in Section II. The proposed
sure and utilize the relationships between the UPF and CPF EMCMO is introduced in Section III. Section IV provides the
to solve CMOPs. During the past several decades, identify- experimental settings, and the results are shown in Section V.
ing and utilizing the relationships among different problems Finally, Section VI provides the detailed conclusion and future
have been widely studied, which has led to the development of work.
transfer learning [8] and multitasking learning [9]. Inspired by
these learning methods, in the evolutionary computation com-
II. R ELATED W ORK
munity, evolutionary multitasking (EMT) [10] has become a
new optimization pattern that aims to identify and utilize the A. Evolutionary Multitasking
relationships or similarities among multiple optimization tasks. The multifactorial evolutionary algorithm (MFEA) [10] is
Unlike the original evolutionary algorithm that can only the pioneering work on EMT. In the MFEA, a single pop-
solve one task in a run, EMT aims to transfer potential excel- ulation is created to accommodate the individuals belonging
lent solutions among different tasks to simultaneously solve to different tasks. The population can be regarded as a cross-
multiple relevant optimization tasks. In EMT, multiple deci- domain platform for multitasking optimization, and it makes
sion spaces corresponding to different tasks are searched by a full use of the power of the implicit parallelism of population-
population of individuals simultaneously, and intradomain and based search. In the population, each individual has a skill
interdomain knowledge are mined. Through various knowl- factor that indicates which task it belongs to. In the off-
edge transfer methods, such as intertask crossover [11] and spring generation process, if two parents have different skill
individual transfer [12], knowledge of other tasks represented factors, the generated offspring will randomly inherit the
by promising solutions will be utilized by the target task. skill factors of the parents. In this manner, the offspring can
Thus, the convergence performance of the population can be own the genes that come from another task, and knowledge
improved, and better solutions for each task can be found. It transfer is completed. The MFEA and its improved versions
is worth noting that the genes of individuals are the knowl- have achieved success in numerical optimization and various
edge of a task. Since the genes are carried by individuals, real-world applications. It should be noted that the MFEA
for the sake of convenience, the individuals are viewed as is an implicit EMT model. Conversely, the explicit EMT

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.
QIAO et al.: EMT OPTIMIZATION FRAMEWORK FOR CMOPs 265

TABLE I
S UMMARY OF E XISTING CMOEA S

model [17] optimizes multiple tasks using multiple popula- In this article, the proposed EMCMO is an explicit EMT
tions. The advantages of the explicit EMT model are that it model and focuses on what to transfer.
can improve the concentration of the population on the target
task to reduce the harm of negative transfer [18]. In general, B. Existing Constrained Multiobjective Optimization
the success of knowledge transfer is twofold: 1) knowledge Methods
transfer improves the diversity of the population to jump out This section will briefly describe the existing related
of local optimal regions and 2) the solutions generated by the CMOEAs. Table I summarizes these algorithms.
simple task help the population of complex tasks quickly find When addressing CMOPs, constraint-handling techniques
the global optimal solutions or regions. Therefore, if knowl- (CHTs) and optimizers are two important components.
edge transfer can help a task jump out local regions or find the Therefore, some CMOEAs have been proposed by combining
global optimal solutions or regions, it will be a positive trans- MOEAs and CHTs. For example, Ning et al. [25] proposed
fer. Otherwise, if the knowledge transfer makes the population a novel algorithm that combined a constrained nondominated
fall into local areas, it will be a negative transfer. sorting (CNS) scheme and an improved hybrid MOEA. The
Inspired by the MFEA, numerous researchers have proposed CNS assigned each individual a rank value for environmen-
different EMT algorithms. As the core of EMT, knowledge tal selection. The rank value was obtained by summing the
transfer naturally receives more attention. These algorithms ranking of nondominated sorting and the degree of constraint
mainly solve the problems of how and what to transfer. How to violation. Similarly, Ma et al. [26] proposed a new fitness
transfer refers to the form of knowledge transfer. For example, function with the two rankings method (ToR) in which the new
some algorithms embed the knowledge transfer into the muta- fitness of individuals was obtained by combining two rankings.
tion strategy of differential evolution or the velocity update These two rankings were obtained by sorting the popula-
strategy of particle swarm optimization [19]–[21]. In addi- tion based on the constrained dominance principle (CDP) and
tion, some methods realize knowledge transfer by replacing Pareto dominance. However, it is difficult for CNS and ToR
some individuals with individuals from other tasks [12], [22] to maintain the traderoff between constraints and objectives
directly. In addition, what to transfer refers to which solutions since the weights of these two rankings are problem dependent.
should be transferred. There is no doubt that useful knowledge Liu et al. [27] first extended the indicator-based MOEAs to
should be transferred since random knowledge transfer would solve CMOPs. In the study, different indicator-based MOEAs
lead to negative transfer that destroys the evolutionary direc- and different CHTs were combined to form various algo-
tion of the target task. However, finding useful knowledge is rithms. The results show that indicator-based MOEAs can
a difficult issue. To address this problem, Zheng et al. [23] produce excellent performance and should be further stud-
proposed a new self-regulated EMTO (SREMTO) algorithm. ied. Fan et al. [28] devised a novel angle-based CDP (ACDP)
In SREMTO, the individuals who performed well on two tasks for CMOPs. The ACDP measured the angle between the par-
could be regarded as useful solutions. Lin et al. [24] proposed ent individual and the corresponding offspring individual to
a new EMTO algorithm based on incremental learning to select guide the environmental selection, which could retain and uti-
valuable transferred solutions. Accurately identifying useful lize excellent infeasible solutions to find more feasible regions.
knowledge is still a problem that needs further study. Although these algorithms have produced good results, they

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.
266 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 26, NO. 2, APRIL 2022

cannot significantly balance objectives and constraints. The cooperation of two populations could balance constraints and
main reason is that each constraint-handling technique has a objectives. However, for the two tasks with low similarity,
specific preference for objectives or constraints. the helper problem still consumed half of the computing
Very recently, some researchers developed two-stage resources.
CMOEAs to balance constraints and objectives. For example, These multipopulation-based methods exchange information
Fan et al. [29] employed the push and pull search (PPS) to between two populations. There is no doubt that the exchange
solve CMOPs. In the push stage, the population was updated of high-quality information can effectively improve the
by an MOEA to approach the UPF, which may help the pop- performance of the two populations. Unfortunately, these algo-
ulation rapidly find feasible regions. Then, in the pull stage, rithms do not attempt to achieve a high-quality information
constraints were considered to help the population converge exchange, which motivates us to design a more effective
to the true CPF. However, when the UPF corresponds to one framework to solve CMOPs by implementing a high-quality
small area in the decision space, it is difficult for the pop- information exchange.
ulation to spread from the UPF to the CPF. Tian et al. [30]
proposed a two-stage CMOEA (CMOEA-MS), which adjusted
III. P ROPOSED F RAMEWORK
the search behaviors of the population based on the ratio
of feasible solutions in the combined population, including A. Motivation
the parent population and offspring population. CMOEA-MS To solve complex optimization problems, EMT has been
implemented a deep search in infeasible regions to utilize used to improve the search abilities of the working population
infeasible solutions; however, using only the ratio of feasible by creating helper tasks. For example, Ma et al. [36] used
solutions may reduce the convergence speed of the population. prior-knowledge-based multiobjectivization via decomposition
Liu and Wang [31] devised a two-phase framework (ToP) to to create meme helper tasks for each component task in a mul-
solve CMOPs. The ToP focused on finding feasible areas in the titasking optimization environment. The results revealed that
first phase and Pareto-optimal solutions in the second phase. the search bias provided by helper tasks could not only reduce
However, it is not easy to set a suitable condition to transform the number of local regions but also improve the population
the first phase into the second phase. diversity. Chen et al. [21] proposed an EMT method to solve
To avoid introducing new parameters to control the phase high-dimensional classification feature selection problems, in
transition, some dual-population CMOEAs have also been which the helper task was created to speed up the convergence
proposed. Wang et al. [32] developed a cooperative dif- by selecting the feature subset from the available features. The
ferential framework. In this framework, the problem was results demonstrated that the created helper task can reduce the
transformed into M constrained single-objective problems and search difficulties by searching in a low-dimensional space.
one constrained M-objective problem. In addition, M small Inspired by EMT, this article attempts to use EMT pattern to
populations and an archive population are created to optimize solve CMOPs, which is related to the relatedness between two
them, respectively. The knowledge transfer among M + 1 tasks. If two tasks share more of the same optimal solutions
populations helps the archive population utilize the objective and have complementary evolutionary trials, the similarities
information. However, optimizing M + 1 populations at the between them will be higher. For a CMOP, the main difficul-
same time will inevitably cause serious computational over- ties come from the constraints that may cause several small and
head. Qian et al. [33] adopted the immune system model to disconnected feasible regions, which put serious requirements
solve CMOPs. In the proposed algorithm, the feasible non- on population diversity. Therefore, we consider creating helper
dominated individuals constituted the first population and were tasks with simple constraints to utilize some solutions that are
used to explore feasible regions, and infeasible/dominated feasible for the helper tasks while are infeasible for the main
individuals constituted the second population and focused task (the CMOP). In general, the fewer the number of con-
on searching the entire space. Then, the offspring of the straints is, the fewer difficulties the population will encounter.
two populations were combined to perform a transformation Therefore, a constraint subset that is randomly sampled from
technique to enhance the convergence performance. Although all constraints can be used as the constraints of a helper
this algorithm maintained diverse search abilities, it ignored task. Different constraint subsets may form different feasible
the diversity in feasible regions. Li et al. [34] proposed a regions; thus, the created helper tasks can provide some search
two-archive EA for CMOPs (CTAEA), which maintained a biases for the main task. In extreme cases, the helper task con-
convergence-oriented archive (CA) and a diversity-oriented tains zero constraint; thus, it becomes an unconstrained MOP.
archive (DA). The CA and DA were responsible for con- In Fig. 2, a general framework is provided, in which arbitrary
vergence and diversity, respectively. Moreover, the mating numbers of helper tasks with different constraint subsets can
parent was selected from these two archives to generate off- be created and the knowledge transfer is implemented among
spring for balancing the convergence, diversity, and feasibility. tasks to improve the overall search performance. In addition,
However, the strong cooperation between the two archives it is worth noting that the prior knowledge about the CMOP
interfered with the evolutionary direction of each archive. can be used to guide the creation of helper tasks to guarantee
Therefore, Tian et al. [35] proposed a co-evolutionary frame- a high similarity between the main task and the helper task.
work (CCMO) with a weak connection. CCMO evolved two Based on Fig. 2, a simplified “special case” of only two
populations that were responsible for the original CMOP and tasks, which contains the main task and a helper task, is cre-
a helper problem derived from the original problem. The ated. Moreover, an unconstrained MOP with zero constraints

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.
QIAO et al.: EMT OPTIMIZATION FRAMEWORK FOR CMOPs 267

Fig. 3. Illustration of what is the useful knowledge in the early generations,

where red circles and red triangles are, respectively, the parent individuals and
Fig. 2. Illustration of transforming a CMOP into a multitasking optimization offspring individuals of main task, and the blue square and blue hexagons are,
task by creating arbitrary numbers of helper tasks. respectively, the parent individuals and offspring individuals of helper task.

TABLE II
C HALLENGES OF F OUR D IFFERENT R ELATIONSHIPS
is created as the helper task, which considers some practical B ETWEEN UPF AND CPF
issues as follows.
1) Creating helper tasks is expected to reduce the search
difficulties; however, without prior knowledge, it is dif-
ficult to know the difficulties of the constraint subset
contained in a helper task. In this case, a helper task may
not have a complementary evolutionary trial with the
main task. However, when all constraints are removed,
there is no infeasible region that will block the evolution
of the population. In this case, the created unconstrained
MOP can help the main task cross infeasible regions via
knowledge transfer.
2) When generating many helper tasks, optimizing them
would consume more computing resources. In addition, a locally feasible region due to the large infeasible regions.
optimizing these tasks requires more storage space and Since the constraints are not considered by the helper task
computing time. However, unlike generic multitasking (denoted as T2 ), its parent population (denoted as P2 ) is located
optimization problems, in which computing resources in the infeasible region. Since the evolution of the popula-
are proportional to the number of tasks, the total amount tion depends on the better offspring individuals, in this phase,
of computing resources is fixed no matter how many some parent individuals may be dominated by partial offspring
tasks CMOP is decomposed into. When creating only individuals. Therefore, P1 and P2 generate some offspring
one task, most of the computing resources can be used individuals with better objective values. However, it is not
to solve the main task. easy for P1 to escape from the local feasible region without
In the simplified “special case” of only two tasks, from the knowledge transfer since its offspring individuals with better
viewpoint of objective space, the main task is to find a well- objective values are infeasible. Furthermore, when P2 is trans-
distributed CPF, and the helper task is to find a well-distributed ferred into T1 , P1 still cannot jump out of the local feasible
UPF. As stated in Section II-A, what to transfer is an impor- region. Conversely, the offspring individuals of P2 are located
tant issue for EMT since only useful knowledge is helpful. in a better feasible region; thus, when they are transferred
If the quality of transferred knowledge is not distinguished, into T1 , P1 can cross the large infeasible region. Therefore, in
negative transfer will occur and destroy the evolutionary direc- the early phase, the offspring population is considered useful
tion. Therefore, it is necessary and significant to find useful knowledge.
knowledge. Next, we will first answer the problem of what to 2) Later Phase: During the evolution, two populations may
transfer. be located in or draw near to their PFs in the later phase.
Next, based on the four types of relationships between the CPF
B. What to Transfer and UPF, which population can be regarded as useful knowl-
During the evolution, the knowledge of a task contains two edge is explained. In addition, the challenges of the above
types: 1) parent individuals and 2) offspring individuals. Parent four relationships between the CPF and UPF are described in
individuals and offspring individuals form the parent popula- Table II.
tion and the offspring population, respectively. Then, we will 1) Type-I: When the UPF and CPF are the same, the parent
introduce which population is useful knowledge in different population is more effective as the transferred knowl-
evolutionary phases. edge. Although the two PFs are the same, the parent
1) Early Phase: As shown in Fig. 3, the parent popula- individuals of the two populations may be located in
tion (denoted P1 ) of the main task (denoted T1 ) falls into different locations. This is because two populations of

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.
268 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 26, NO. 2, APRIL 2022

(a) (b)

Fig. 5. Updating process of EMCMO in the early phase.

(c) (d)

Fig. 4. Illustration of what is the useful knowledge in the later generations,

where red circles and red triangles are, respectively, the parent individuals and
offspring individuals of main task, and the blue square and blue hexagons are,
respectively, the parent individuals and offspring individuals of helper task.
(a) Type-I: CPF is same as UPF (b) Type-II: CPF is part of UPF (c) Type-III:
CPF and UPF overlap partially (d) Type-IV: CPF and UPF do not overlap

individuals are not exactly the same and two popula- Fig. 6. Updating process of EMCMO in the later phase.
tions have different environmental selection methods. As
shown in Fig. 4(a), when the parent individuals of T2 are
transferred into T1 , T1 can obtain more Pareto-optimal may be near or located on the CPF, even if this proba-
solutions since the parent individuals of two tasks are bility is small. Similarly, partial offspring individuals of
nondominated. However, the partial offspring individu- T1 may be near or located on the UPF. In Fig. S-VII of
als of T2 are dominated by the parent individuals of T1 . Section S-II in the supplementary material, the Pareto
Therefore, the benefit of transferring offspring individu- fronts of four test functions corresponding to the above
als is lower than that of transferring parent individuals. four types are plotted to observe the real difference
In the same way, the parent individuals of T1 are non- between the CPF and UPF.
dominated solutions for T2 , and the parent population of
T1 may be viewed as useful knowledge for T2 . C. Procedure of EMCMO
2) Type-II: When the CPF is part of the UPF, the par-
ent population is more effective as the transferred Inspired by the observations above, the evolutionary pro-
knowledge. As shown in Fig. 4(b), partial parent individ- cess in the proposed EMCMO framework is divided into two
uals of T2 are the Pareto-optimal solutions of T1 while all phases. Fig. 5 plots the framework of the EMCMO in the early
parent individuals of T1 are the Pareto-optimal solutions phase, and Fig. 6 plots the framework of the EMCMO in the
of T2 . later phase.
3) Type-III: When the CPF and UPF partially overlap, As shown in Algorithm 1, EMCMO begins with two popula-
which population is more effective depends on the tions P1 and P2 that focus on two tasks. Then, two populations
degree of overlap of the UPF and CPF. In fact, under this are evaluated on the corresponding tasks. During the main
relationship, as shown in Fig. 4(c), two PFs can be split loop, the evolutionary process is divided into two phases since
into two parts: a) the overlap part and b) the nonoverlap the useful knowledge is different in different phases. In the
part. For the overlap part, the parent population is more early stage, two populations evolve according to Algorithm 2.
effective, which is the same as the situation in Fig. 4(a). Then, in the later phase, they are updated by Algorithm 3. It
For the nonoverlap part, the offspring population is more should be noted that the CDP method and fast nondominated
effective, which is the same as the situation in Fig. 4(d). sorting method are used as the environmental selection meth-
4) Type-IV: When the CPF and UPF do not overlap, the ods of the two tasks. Here, the parameter β is set to switch
offspring population is more effective as the transferred the evolutionary phase from the early phase to the later phase.
knowledge. As shown in Fig. 4(d), the parent individ- The impact of β will be investigated in Section V.
uals of one task are not the Pareto-optimal solutions
of another task; thus, the parent population cannot be D. Early Phase
considered useful knowledge. Although some offspring The updating process of EMCMO in the early phase is
individuals of T2 are inferior to their parent individu- described in Algorithm 2. Note that when one task is repre-
als, these offspring individuals may be effective for T1 . sented by Tj (j = (1, 2)), another task is represented by T2/j .
From Fig. 4(d), the partial offspring individuals of T2 In each generation, for each task Tj , its mating parents are

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.
QIAO et al.: EMT OPTIMIZATION FRAMEWORK FOR CMOPs 269

Algorithm 1 Total Procedure of EMCMO offspring population is regarded as useful knowledge due to
Input: NP (population size), β (the parameter control when its higher diversity and better quality.
to switch the evolutionary phase) P1 (the population for
CMOP), P2 (the population for MOP), MaxFES (the total E. Later Phase
number of function evaluations). Compared with the early phase, the main difference in the
Output: P1 later phase is that the transferred population needs to be deter-
1: Initialize P1 and P2 of size NP
mined. Since it is difficult to know the relationship between
2: FES = 0
the CPF and UPF without prior knowledge, we propose a ten-
3: Evaluate P1 on the CMOP
tative method to evaluate which population is more effective.
4: Evaluate P2 on the MOP
First, Pj with NP individuals and OPj with NP/2 individuals
5: FES = FES + 2 · NP;
are combined to obtain a new population TOPj with (3 · NP/2)
6: while termination criterion not satisfied do
individuals. Then, the best NP individuals in TOPj are selected
7: if FES < β · MaxFES then based on the environmental selection rule of T2/j . Next, the
8: Perform the early phase → Algorithm 2 success rates of Pj and OPj on T2/j are computed as follows:
9: else
Perform the later phase → Algorithm 3 num− Pj
10: αPj = (3)
11: end if NP
FES = FES + NP; num− OPj
12: αOPj = NP
(4)
13: end while 2
where αPj ∈ [0, 1] and αOPj ∈ [0, 1], respectively, indicate
Algorithm 2 Early Phase of EMCMO the success rate of Pj and OPj ; num− Pj and num− OPj are,
Input: NP (population size), P1 (the population for CMOP), respectively, the number of parent individuals and offspring
P2 (the population for MOP), individuals in the best NP individuals. When αPj < αOPj ,
1: MatingParent1 ← Randomly select NP/2 individuals from
it means OPj is more suitable for T2/j . In this case, OPj is
P1 regarded as the transferred population TrPj of Tj . Otherwise,
2: OP1 ← Generate NP/2 offspring based on the evolution-
Pj would be a better choice for T2/j . In this case, to maintain
ary operators by the MatingParent1 the diversity of T2/j , NP/2 individuals selected randomly from
3: MatingParent2 ← Randomly select NP/2 individuals from
Pj are regarded as the transferred population TrPj of Tj .
P2 Algorithm 3 gives the updating process of EMCMO in the
4: OP2 ← Generate NP/2 offspring based on the evolution-
later phase. In each generation, for each task Tj , its mating par-
ary operators by the MatingParent2 ents are selected from Pj based on their quality to improve the
5: TP1 ← P1 ∪ OP1 ∪ OP2
convergence performance. These mating parents are used to
6: TP2 ← P2 ∪ OP2 ∪ OP1
generate NP/2 offspring individuals. According to (3) and (4),
7: Evaluate TP1 on the CMOP
the transfer population TrPj of Tj is determined. Then, Pj ,
8: Evaluate TP2 on the MOP
OPj , and TrP2/j are combined into a temporary population TPj .
9: P1 ← Select NP individuals from TP1 based on the
Finally, the environmental selection is performed to select NP
environmental selection individuals from TPj to obtain a new Pj .
10: P2 ← Select NP individuals from TP2 based on the
environmental selection F. Computational Complexity
The developed EMCMO framework does not introduce
any new selection operator; thus, its computational complex-
ity mainly depends on the adopted MOEA (NSGA-II). As
randomly selected from the corresponding parent population
suggested in NSGA-II, the main operators are mating selec-
Pj to improve population diversity. High population diversity
tion, genetic operators, and environmental selection, and their
can prevent premature convergence. Then, the mating parents
worst complexities are O(NP), O(NP · D), and O(M · NP2 ),
are used to generate NP/2 offspring individuals, where NP is
respectively. Since each population generates NP/2 individ-
the parent population size. The offspring population is OPj .
uals, the worst complexities are 2 · O(NP/2) = O(NP),
The temporary population TPj is obtained by combining Pj ,
2·O(NP·D/2) = O(NP·D), and 2·O(M·NP2 ) = O(M·NP2 ). In
OPj , and OP2/j . Next, TPj is evaluated on the correspond-
summary, the worst computational complexity of EMCMO is
ing task. Finally, NP individuals are selected from TPj based
the same as that of the employed MOEA. However, because all
on environmental selection to form a new Pj . It should be
operators are performed twice a generation, EMCMO would
noted that when evaluating one individual on T2 , not only
consume more CPU time.
objective function values but also the values of the degree of
constraint violation are obtained. Therefore, evaluating OP2
IV. E XPERIMENT S ETTINGS
on T1 will not consume extra function evaluations, and only
newly generated offspring will consume function evaluations. A. Test Functions and Performance Indicators
In the early phase, population diversity is considered in To test the performance of EMCMO, we adopt four bench-
the selection process of the mating parent. Furthermore, the mark test suites: 1) the MW test suite [7] includes 14 functions

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.
270 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 26, NO. 2, APRIL 2022

Algorithm 3 Later Phase of EMCMO 3) The feasible solutions ratio (FSR) indicates the average
Input: NP (population size), P1 (the population for CMOP), proportion of feasible solutions in the last generation
P2 (the population for MOP), found by one algorithm on 30 runs.
1: MatingParent1 ← Select NP/2 individuals from P1 based It is worth noting that there are EMT benchmark prob-
on their quality using binary tournament selection method lems and performance metrics in the community of EMT [42].
2: OP1 ← Generate NP/2 offspring based on the evolution- However, the constrained multitask multiobjective test prob-
ary operators by the MatingParent1 lems and performance metrics have not attracted attention.
3: MatingParent2 ← Select NP/2 individuals from P2 based Therefore, the related studies are necessary and significant.
on their quality using binary tournament selection method
4: OP2 ← Generate NP/2 offspring based on the evolution- B. Compared Algorithms and Their Parameter Settings
ary operators by the MatingParent2
Five state-of-the-art algorithms, including CCMO [35],
5: TOP1 ← P1 ∪ OP1
CTAEA [34], NSGAII-ToR [26], MOEADDAE [6], and
6: TOP2 ← P2 ∪ OP2
MSCMO [39], are employed. Among these algorithms,
7: Evaluate TOP1 on the MOP (the second task)
CCMO and CTAEA are dual-population-based CMOEAs, and
8: Evaluate TOP2 on the CMOP (the first task)
they adopt the CDP method to tackle constraints. NSGAII-
9: TrP1 ← Based on Eqs. (3) - (4)
ToR employs a ToR to balance constraints and objectives.
10: TrP2 ← Based on Eqs. (3) - (4)
MOEADDAE uses the improved  method, which regards
11: TP1 ← P1 ∪ OP1 ∪ TrP2
some infeasible solutions as feasible solutions. MSCMO grad-
12: TP2 ← P2 ∪ OP2 ∪ TrP1
ually adds constraints during the evolution and employs a
13: Evaluate TP1 on the CMOP
multiobjective-based method and CDP to tackle constraints.
14: Evaluate TP2 on the MOP
These five compared methods employ different CHTs and dif-
15: P1 ← Select NP individuals from TP1 based on the
ferent evolutionary mechanisms; thus, comparison experiments
environmental selection
among the methods can provide fair and sufficient proof for
16: P2 ← Select NP individuals from TP2 based on the
the superiority of EMCMO. The parameter settings of these
environmental selection
five algorithms are set as provided in their original literature,
as shown in Table IV. To make a fair comparison, NP is
set to 100, and MaxFES is set to 100 000 for all algorithms
TABLE III
S UMMARY OF B ENCHMARK T EST S UITES
on all test functions. Moreover, Wilcoxon’s test at the 0.05
level is performed to estimate the significance of the differ-
ence between the two methods. “+” or “−” indicates that
the compared algorithm has a better or worse performance
than EMCMO, respectively. “=” shows that EMCMO and the
compared algorithm have a similar performance.
Since all compared methods adopt a genetic algorithm with
simulated binary crossover [43] and polynomial mutation [44]
as the optimizer, EMCMO uses the same evolutionary oper-
ators. In addition, a fitness method similar to SPEA2 [45] is
adopted in EMCMO. The fitness of one solution x is equal
to the sum of two parts. The first part is the number of solu-
tions dominated by the solutions dominating x. The second
with D = 15; 2) the DTLZ test suite [37] has ten func- part indicates the crowding distance of x, which equals the
tions, including four functions with D = 7 (C1_DTLZ1, √
distance from x to its 2NP nearest neighbor. Furthermore,
DC1_DTLZ1, DC2_DTLZ1, and DC3_DTLZ1) and other six to maintain diversity, EMCMO employs the truncation strat-
functions with D = 12; 3) the LIRCMOP test suite [38] con- egy proposed in SPEA2, which is activated when the number
sists of 14 functions with D = 10; and 4) the DOC test of nondominated solutions is larger than NP. In addition, the
suite [31] is constituted by nine functions with different dimen- differential operators adopted by the MOEADDAE for the
sions, and their dimension settings are the same as those in LIRCMOP and DOC test suites are replaced by the same
the literature [31]. The features of these four test suites are genetic operators. Note that all experiments are implemented
provided in Table III. on PlatEMO [46] since it contains the codes of all compared
In addition, three performance indicators are as follows. algorithms.
1) The inverted generational distance (IGD) [40] is a pop-
ular indicator that can reflect the convergence of the
V. E XPERIMENT R ESULTS
obtained PF and the coverage ratio between the found
PF and true PF. A smaller IGD is desired. A. Compared Experiment Results
2) The hypervolume (HV) [41] is also widely used in the Table V gives the mean IGDs and standard deviations of
multiobjective optimization. The HV mainly reflects the each method on all four benchmark test suites, and the best
convergence and diversity of the obtained PF, and a results are marked in bold. The corresponding HV results are
larger HV is better. listed in Table S-I in the supplementary material.

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.
QIAO et al.: EMT OPTIMIZATION FRAMEWORK FOR CMOPs 271

TABLE IV
C OMPARED A LGORITHMS AND R ELATED PARAMETER S ETTINGS

TABLE V
IGD R ESULTS O BTAINED BY EMCMO AND F IVE C OMPARED A LGORITHMS ON A LL F OUR B ENCHMARK T EST S USITES

For type-I problems, the key is to utilize the nondominated outperforms CCMO, CTAEA, NSGAII-ToR, MOEADDAE,
solutions on the UPF to help find the CPF. In this case, trans- and MSCMO on 8, 11, 11, 11, and 10 functions. However,
ferring parent individuals is more significant than transferring only CCMO presents a better performance than EMCMO on
offspring individuals. Since EMCMO can identify the relation- LIRCMOP6. Similarly, EMCMO obtains excellent results on
ship using the tentative approach and transfer correct knowl- the HV indicator, and no algorithm outperforms EMCMO
edge, it obtains the best IGD results on 10 out of 11 functions. on any function. They are, respectively, outperformed by
Wilcoxon’s test results indicate that EMCMO, respectively, EMCMO on 9, 11, 11, 10, and 10 functions. These results

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.
272 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 26, NO. 2, APRIL 2022

TABLE VI
verify the superiority of EMCMO. It should be noted that W ILCOXON ’ S T EST R ESULTS ON A LL F OUR B ENCHMARK T EST S UITES
only the second population can converge to the true UPF, and I NCLUDING 47 F UNCTIONS
knowledge transfer can maximize value. For example, the CPF
of MW2 is constituted by several CPFs with large feasible
regions and several CPFs with small feasible regions. Although
the second population can converge to the UPF, partial par-
ent individuals that draw near to small feasible regions are
still infeasible. The reasons are that the convergence accuracy
of the second population is not very good. In this case, the
infeasible parent individuals would be deleted by the first task;
thus, the small feasible regions cannot be found by knowledge
transfer. algorithms is not good due to the complex characteristics of
Similar to type-I problems, the key for type-II problems is DOC. In summary, compared with the compared algorithms,
to find a complete UPF and transfer parent individuals. The EMCMO is a promising method for DOC.
IGD results show that EMCMO ranks first on 6 of 9 func- In summary, compared with all five compared methods,
tions. Moreover, EMCMO produces better performance than EMCMO presents outstanding optimization results on all
CCMO, CTAEA, NSGAII-ToR, MOEADDAE, and MSCMO four benchmark test suites regarding both the IGD and HV
on 6, 7, 9, 8, and 7 functions, respectively. Additionally, only indicators.
MOEADDAE outperforms EMCMO on one function. Based
on the HV results, EMCMO is still better than the other meth-
ods on 4, 8, 9, 7, and 8 functions, respectively. Moreover, B. Statistical Results on All Four Benchmark Test Suites
MOEADDAE produces better performance on two LIRCMOP It is necessary to observe the FSR values of EMCMO and
functions. This is because MOEADDAE uses the  method to the other five algorithms on all test functions. As shown in
utilize diverse solutions rather than those near the UPF. In Table S-II in the supplementary material, the FSRs of only
conclusion, EMCMO is the best method for type-II problems. the proposed EMCMO are 1 on MW, DTLZ, and LIRCMOP
For type-III problems, the IGD results show that EMCMO test suites. In addition, for DOC3 and DOC7, the FSR values
does not rank first on any function. However, no algo- of EMCMO are larger than those of the other five algorithms.
rithm outperforms EMCMO on any function. Wilcoxon’s test Based on the statistical results, EMCMO achieves better FSR
results indicate that EMCMO is better than CCMO, CTAEA, results than CCMO, CTAEA, NSGAII-ToR, MOEADDAE,
NSGAII-ToR, MOEADDAE, and MSCMO on 1, 4, 6, 5, and 2 and MSCMO on 3, 10, 6, 24, and 24 test functions, respec-
functions, respectively. In addition, the HV results indicate that tively. Therefore, these results reasonably prove the superiority
CCMO is better than and inferior to EMCMO on LIRCMOP12 and convergence ability of EMCMO.
and MW3, respectively. However, other peers are still inferior Table VI shows Wilcoxon’s test results on four benchmark
to EMCMO on most functions. It should be noted that the test suites. Obviously, for each compared algorithm, the num-
degree of overlap between the CPF and UPC determines the ber of “+”s is much larger than the number of “−”s on both
importance of parent or offspring individuals. For example, the IGD and HV indicators, which fully demonstrates the
the degree of overlap in MW13 is higher than other func- superiority of EMCMO on different types of test problems.
tions; thus, EMCMO outperforms CCMO due to the transfer Furthermore, the multiproblem Wilcoxon’s test and
of parent individuals. Friedman’s test in terms of IGD and HV are performed using
For type-IV problems, CCMO and EMCMO produce bet- the keel software [47] on a total of 47 test functions to observe
ter results than other peers in terms of both IGD and HV. the significance of the difference between EMCMO and each
However, CCMO is still inferior to EMCMO on LIRCMOP1 compared method. Table VII lists the multiproblem Wilcoxon
and LIRCMOP2. This is because EMCMO employs the ran- test results, in which the R+ value and the R− value indi-
dom mating operator to improve the diversity that is required cate the performance of EMCMO and the compared algorithm,
for the functions with small feasible regions. In addition, com- respectively. The results show R+ is much greater than R− for
pared with the other four compared algorithms, EMCMO is all compared algorithms. Moreover, all P values are smaller
still the best. It must be admitted that the similarity between than 0.05, which indicates that EMCMO is significantly super
the CPF and UPF for type-IV problems is very low, which than all compared algorithms when considering all four bench-
reduces the effect of knowledge transfer. Therefore, it is sig- mark test suites. In addition, the ranking of each algorithm
nificant to develop new strategies rather than the individual obtained by Friedman’s test is presented in Fig. 7. EMCMO
transfer method to utilize the knowledge of the UPF. achieves the best rankings for both the IGD and HV, which
For DOC functions, in which the types are unknown, the demonstrates the superiority of EMCMO.
IGD results show that EMCMO, respectively, outperforms The nonparametric Vargha–Delaney ( A12 ) test [48] is also
CCMO, CTAEA, NSGAII-ToR, MOEADDAE, and MSCMO used in this article to measure the effect size. When two algo-
on 3, 7, 4, 2, and 5 functions. Among the six methods, CTAEA rithms have similar performances,  A12 = 0.5. When  A12 >
is the worst algorithm since it cannot solve six functions. For 0.56, > 0.64, and > 0.71, it is believed that the two algorithms,
all nine functions, the other four compared algorithms are very respectively, have small, median, and large differences [49].
competitive with EMCMO. In fact, the performance of all six The results are provided in Table S-III in the supplementary

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.
QIAO et al.: EMT OPTIMIZATION FRAMEWORK FOR CMOPs 273

TABLE VII
M ULTIPROBLEM W ILCOXON ’ S T EST R ESULTS ON A LL F OUR C. Parameter Analysis
B ENCHMARK T EST S UITES I NCLUDING 47 F UNCTIONS In EMCMO, the parameter β is set to control the conversion
of the evolutionary phase. It should be noted that β is related
to MaxFES. In this article, under the premise that MaxFES
is set to 10 000, different β values are used by EMCMO to
investigate its impact. The IGD results obtained by EMCMO
with β = 0, 0.2, 0.4, 0.6, 0.8, and 1 on all 47 functions are
provided in Table S-IV in the supplementary material. First,
EMCMO with β = 0.2 outperforms EMCMO with β = 0, 0.4,
0.6, 0.8, and 1 on 17, 4, 9, 9, and 12 functions, respectively.
When β = 0 or β = 1, EMCMO obtains very poor results,
which reflects that both the early phase and later phase are
significant and important. In addition, as β increases, the num-
ber of “−”s also increases, which might imply that the later
stage is more important than the early phase. Moreover, when
β = 0.2, EMCMO obtains the best results on most functions.
Therefore, 0.2 is recommended as the value of β.

D. Verification of the Effectiveness of the Main Strategies

1) Verification of Knowledge Transfer: The main contri-
bution of this article is the two-phase knowledge transfer
strategy. Thus, we design two EMCMO variants (EMCMO-s
and EMCMO-m) to verify the effectiveness of knowledge
transfer at two phases for different CMOPs. In EMCMO-s,
only the first population evolves, and knowledge transfer
is not implemented during the entire evolutionary process.
In EMCMO-m, two populations evolve by helping each
other in the early phase, and the helper task is deleted in
Fig. 7. Average rankings of all six algorithms obtained by Friedman’s test the later phase. The compared results are given in Table
on total 47 test functions.
S-V in the supplementary material. For all test functions,
the two variants are inferior to EMCMO on 16 and 28
functions, respectively. To be specific, for type-I and type-II
material. Regarding the IGD and HV, most  A12 values are problems, in which the CPF and UPF have high similarities,
larger than 0.71, which means that there are large differences EMCMO-s is outperformed by EMCMO-m on most func-
between EMCMO and the compared methods. tions, which demonstrates the effectiveness of transferring
In addition, to observe the results of the algorithms more offspring individuals in the early phase. For type-III and
intuitively, the Pareto-optimal solutions with the median IGDs type-IV problems, EMCMO-s is inferior to EMCMO on
on 30 runs obtained by them on three functions are plotted almost all functions; however, EMCMO-m can outperform
in Figs. S-I–S-III in the supplementary material, respectively. EMCMO on several functions. The reasons may be that
For C1_DTLZ1, the distribution of the feasible Pareto-optimal the similarities of the two tasks for these functions are
solutions of EMCMO is the best since C1_DTLZ1 belongs to very low in both the decision space and objective space,
type-I CMOPs and EMCMO transfers the parent populations resulting in the waste of computing resources. In addi-
between the two tasks. CCMO, MOEADDAE, and MSCMO tion, the improvement effect (IE) of knowledge transfer
obtain relatively good distributions. However, NSGAII-ToR regarding the IGD value is calculated: IEEMCMO-m =
and CTAEA search for only partial feasible Pareto-optimal (IGDEMCMO-m − IGDEMCMO )/IGDEMCMO-m and
solutions. LIRCMOP3 is a type-IV CMOP. The results show IEEMCMO−s = (IGDEMCMO-s − IGDEMCMO )/IGDEMCMO-s .
that only a very small number of feasible Pareto-optimal solu- From Table S-V in the supplementary material, it can be seen
tions are found by CTAEA, MOEADDAE, and MSCMO. The that the IEEMCMO-m and IEEMCMO-s values on most of the
feasible Pareto-optimal solutions of NSGAII-ToR are located functions are larger than 0. Meanwhile, the Vargha–Delaney
in the central regions of the CPF. The feasible Pareto-optimal (
A12 ) test results show that EMCMO-m and EMCMO-s are
solutions of CCMO are located in the upper left corner of the significantly worse than EMCMO on most of functions, which
CPF. The feasible Pareto-optimal solutions of EMCMO are demonstrates the effectiveness of the knowledge transfer.
more widely distributed. MW7 is a type-III CMOP. Obviously, Moreover, to observe the effect of knowledge transfer more
the convergence of CTAEA, NSGAII-ToR, and MOEADDAE clearly, we plot the average TSR from the helper task to the
is poor. The distribution and convergence of the feasible main task on MW4 and MW9 on 30 runs. MW4 is a three-
Pareto-optimal solutions of CCMO, MSCMO, and EMCMO objective type-I problem, and MW9 is a two-objective type-IV
are similar. problem. The results are, respectively, provided in Figs. 8

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.
274 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 26, NO. 2, APRIL 2022

Since EMCMO-1 transfers only the offspring population, it

performs poorly on type-I and type-II functions, such as MW1,
MW2, MW4, C1_DTLZ1, and C1_DTLZ3. EMCMO-2 has a
similar performance with EMCMO on most functions. The
reason is that the offspring population has been transferred
in the early phase, resulting that EMCMO-2 can solve most
simple type-IV functions. However, for complex type-IV func-
tions, such as MW9, LIRCMOP1, and LIRCMOP2, EMCMO-
2 cannot solve them well. The uncertainty of transferring
Fig. 8. Average transfer successful rate (TSR) from the helper task to the useful knowledge in EMCMO-3 leads to worse performance
main task on MW4 on 30 runs. on 12 functions, which can demonstrate the effectiveness of the
designed tentative method. Although the offspring population
and NP/2 parent individuals are transferred in EMCMO-4, its
performance is not satisfied on type-I and type-II problems.
The main reason is that two populations may own similar
evolutionary trials, which reduces the effect of knowledge
transfer. In addition, Table S-VII in the supplementary mate-
rial records the Vargha–Delaney ( A12 ) test results, which show
that EMCMO has significant differences from EMCMO-1,
EMCMO-3, and EMCMO-4 on type-I and type-II problems;
and EMCMO has a significant difference from EMCMO-2 on
type-IV problems.
Fig. 9. Average TSR from the helper task to the main task on MW9 on 30
runs. Moreover, the distribution of feasible Pareto-optimal solu-
tions corresponding to median IGDs over 30 runs obtained
and 9. Fig. 8 shows that the TSR is maintained at low and by EMCMO and its four variants on MW4 (type-I),
high levels in the former 600 generations and the latter 400 C2_DTLZ2 (type-II), and LIRCMOP2 (type-IV) is plotted in
generations, respectively. Since P2 may be located in infeasible Figs. S-IV–S-VI in the supplementary material, respectively.
regions in the former generations, the generated infeasible The figures show that EMCMO and EMCMO-2 obtain better
solutions are deleted by P1 . However, after P2 converges to the distributions of feasible Pareto-optimal solutions among the
feasible regions containing feasible nondominated solutions, five algorithms on MW4. This is because the parent population
the knowledge transfer can lead P1 to these regions. Then, since is more effective for type-I problems, which can be verified in
two tasks have the same optimal solutions, TSR is maintained at Table S-VI supplementary material, where only EMCMO-2 has
a high level. For MW9, the TSR is maintained at a very low level a similar performance to EMCMO on MW4. For C2_DTLZ2
in the later evolutionary process; however, it is always larger (type-II), the distributions of feasible Pareto-optimal solutions
than 0. Even if the auxiliary efficiency is very low, the main obtained by EMCMO-1, EMCMO-3, and EMCMO-4 are sig-
task can utilize the successfully transferred feasible offspring nificantly worse than that of EMCMO, and only EMCMO-2
solutions to generate more nondominated solutions. However, has a similar performance to EMCMO, which demonstrates
we cannot directly cancel the knowledge transfer operator for that the parent population is more effective for type-II CMOPs.
type-IV problems because these transferred feasible solutions For LIRCMOP2 (type-IV), the population of EMCMO-4 finds
may be hard to generate by P1 . solutions closer to the boundary point, although its distri-
2) Verification of Two Knowledge Sources: Two knowledge butional uniformity is poor. Compared with the other three
resources, including the parent population and offspring pop- variants, the population of EMCMO is widely distributed on
ulation, are used in EMCMO. To verify their effectiveness, the CPF. In fact, EMCMO-4 produces a similar performance to
several EMCMO variants are designed. EMCMO-1 indicates ECMCMO on some type-IV problems, such as MW9, MW11-
that only the offspring population is transferred in the later 12, and LIRCMOP1-3. For type-IV CMOPs, the UPF and CPF
phase, EMCMO-2 indicates that only the parent population is do not overlap; thus, when the parent and offspring populations
transferred in the later phase, EMCMO-3 indicates that the of T2 are transferred to T1 , the population diversity of T1 is
parent population and offspring population have a 50% trans- not significantly affected. However, for type-I CMOPs, when
fer probability in the later phase, and EMCMO-4 indicates the parent and offspring populations of T2 are transferred into
that both the parent population and offspring population are T1 , a large number of the same solutions or crowded solutions
transferred in the later phase. The IGD results obtained by will not only increase the selection pressure but also affect
EMCMO and its four variants on four benchmark test suites the calculations of the dominance relationship and crowding
are recorded in Table S-VI in the supplementary material. distance, which lead to a inferior performance.
According to Wilcoxon’s test results, the performance of
EMCMO is significantly better than those of EMCMO-1, E. Further Discussion
EMCMO-2, EMCMO-3, and EMCMO-4 on 11, 4, 12, and 10 Due to space limitations, further discussions, including
test functions, respectively. EMCMO ranks first on 16 func- knowledge transfer at the population level and the individ-
tions, followed by EMCMO-2 that ranks first on 9 functions. ual level, two knowledge sources, and two-phase knowledge

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.
QIAO et al.: EMT OPTIMIZATION FRAMEWORK FOR CMOPs 275

transfer, are provided in Section S-III in the supplementary because two successive environments may have high
material. similarity. However, only a few studies have focused on
DMOPs with dynamic constraints [53]. Therefore, we
will study how to extend the proposed method to solve
VI. C ONCLUSION these complex problems.
In this article, the CMOP is transformed into a multitasking 2) Measuring the similarity between tasks in the deci-
optimization problem that involves helper tasks with differ- sion space is significant, and it is generally accepted
ent constraint subsets. According to some practical issues, by related researchers. However, it is difficult to effec-
a simplified “special case” of only two tasks is created. tively measure the similarity in the decision space; thus,
The first task is the original CMOP, and the second task most of the existing works [11], [54], [55] define and
is the MOP extracted from the original CMOP. The pur- use the objective space similarity. In future work, we
poses of the CMOP and MOP are to find the well-distributed will study how to define the decision space similar-
CPF and well-distributed UPF, respectively. In the proposed ity to better improve the algorithm proposed in this
EMCMO framework, by analyzing four types of relation- article.
ships between the CPF and UPF, specific useful knowledge
is identified and transferred between the two tasks. Moreover,
the genetic algorithm is adopted as the optimizer of two R EFERENCES
populations to form an instance of EMCMO. The instance [1] A. Trivedi, D. Srinivasan, K. Sanyal, and A. Ghosh, “A survey of
is tested on four constrained multiobjective benchmark test multiobjective evolutionary algorithms based on decomposition,” IEEE
Trans. Evol. Comput., vol. 21, no. 3, pp. 440–462, Jun. 2017.
suites including a total of 47 different functions. Compared
[2] J. Cheng, G. G. Yen, and G. Zhang, “A many-objective evolutionary
with five state-of-the-art CMOEAs, the results demonstrate the algorithm with enhanced mating and environmental selections,” IEEE
superiority of EMCMO. Furthermore, the effectiveness of the Trans. Evol. Comput., vol. 19, no. 4, pp. 592–605, Aug. 2015.
parameters, knowledge transfer, knowledge sources, and the [3] W. Hong, K. Tang, A. Zhou, H. Ishibuchi, and X. Yao, “A scalable
indicator-based evolutionary algorithm for large-scale multiobjective
tentative method is verified. Note that the method proposed in optimization,” IEEE Trans. Evol. Comput., vol. 23, no. 3, pp. 525–537,
this article can obtain the CPF with a better distribution and Jun. 2019.
convergence, which can help decision makers use preference- [4] R. Tanabe and H. Ishibuchi, “A review of evolutionary multimodal
multiobjective optimization,” IEEE Trans. Evol. Comput., vol. 24, no. 1,
based solution selection methods [50] or knee point-based pp. 193–200, Feb. 2020.
solution selection methods [51] to obtain better final results. [5] J. Chen, J. Li, and B. Xin, “Dmoea-εC: Decomposition-based
Although EMCMO presents a excellent performance, it can multiobjective evolutionary algorithm with the ε -constraint framework,”
IEEE Trans. Evol. Comput., vol. 21, no. 5, pp. 714–730, Oct. 2017.
be further improved from the following aspects.
[6] Q. Zhu, Q. Zhang, and Q. Lin, “A constrained multiobjective evolu-
1) In EMCMO, the tentative method considers only the tionary algorithm with detect-and-escape strategy,” IEEE Trans. Evol.
objective space. A more effective similarity computing Comput., vol. 24, no. 5, pp. 938–947, Oct. 2020.
method should be studied to measure the relationship [7] Z. Ma and Y. Wang, “Evolutionary constrained multiobjective
optimization: Test suite construction and performance comparisons,”
between two tasks in both decision and objective spaces. IEEE Trans. Evol. Comput., vol. 23, no. 6, pp. 972–986, Dec. 2019.
2) Although transferring offspring is more effective than [8] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE Trans.
transferring parents, it may be ineffective for CMOPs Knowl. Data Eng., vol. 22, no. 10, pp. 1345–1359, Oct. 2010.
[9] R. Zhang, H. Zhang, and X. Li, “Robust multi-task learning with flexible
with extremely low similarity between the CPF and manifold constraint,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 43,
UPF. Thus, other knowledge transfer methods should be no. 6, pp. 2150–2157, Jun. 2021.
studied to solve type-III and type-IV CMOPs. [10] A. Gupta, Y. Ong, and L. Feng, “Multifactorial evolution: Toward evo-
lutionary multitasking,” IEEE Trans. Evol. Comput., vol. 20, no. 3,
3) For type-I and type-II CMOPs, in which feasible pp. 343–357, Jun. 2016.
nondominated solutions are located in small feasible [11] H. Xu, A. K. Qin, and S. Xia, “Evolutionary multi-task optimization
regions, the poor convergence of the second popula- with adaptive knowledge transfer,” IEEE Trans. Evol. Comput., early
tion makes it unable to provide high-quality transferred access, Aug. 24, 2021, doi: 10.1109/TEVC.2021.3107435.
[12] R. Hashimoto, H. Ishibuchi, N. Masuyama, and Y. Nojima, “Analysis
knowledge for the first task. Moreover, different CMOPs of evolutionary multi-tasking as an island model,” in Proc. Genet. Evol.
have different features and landscapes; thus, different Comput. Conf. Companion, 2018, pp. 1894–1897.
search engineers should be studied for the two tasks. [13] P. T. H. Hanh, P. D. Thanh, and H. T. T. Binh, “Evolutionary algorithm
and multifactorial evolutionary algorithm on clustered shortest-path tree
4) More helper tasks can be created to provide more com- problem,” Inf. Sci., vol. 553, pp. 280–304, Apr. 2021.
plementary evolutionary trials for the main task so that [14] L. Feng et al., “Explicit evolutionary multitasking for combinatorial
the population of the main task can improve the diversity optimization: A case study on capacitated vehicle routing problem,”
IEEE Trans. Cybern., vol. 51, no. 6, pp. 3143–3156, Jun. 2021.
and break through the blockade of complex constraints. [15] Y. Bi, B. Xue, and M. Zhang, “A divide-and-conquer genetic pro-
The difficulty measurement of constraints and the simi- gramming algorithm with ensembles for image classification,” IEEE
larity measurement mechanism can be studied to create Trans. Evol. Comput., vol. 25, no. 6, pp. 1148–1162, Dec. 2021,
doi: 10.1109/TEVC.2021.3082112.
useful helper tasks and self-adaptively select a few
[16] N. Zhang, A. Gupta, Z. Chen, and Y.-S. Ong, “Evolutionary
helper tasks to evolve. machine learning with minions: A case study in feature selec-
In future work, we will continue to study the above issues. tion,” IEEE Trans. Evol. Comput., early access, Jul. 26, 2021,
In addition, several promising studies can also be studied. doi: 10.1109/TEVC.2021.3099289.
[17] G. Li, Q. Lin, and W. Gao, “Multifactorial optimization via explicit
1) In [52], EMT is also used to solve dynamic multipopulation evolutionary framework,” Inf. Sci., vol. 512, no. 1,
multiobjective optimization problems (DMOPs). This is pp. 1555–1570, 2020.

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.
276 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 26, NO. 2, APRIL 2022

[18] Z. Tang, M. Gong, F. Jiang, H. Li, and Y. Wu, “Multipopulation [41] E. Zitzler and L. Thiele, “Multiobjective evolutionary algorithms: A
optimization for multitask optimization,” in Proc. IEEE Congr. Evol. comparative case study and the strength pareto approach,” IEEE Trans.
Comput. (CEC), 2019, pp. 1906–1913. Evol. Comput., vol. 3, no. 4, pp. 257–271, Nov. 1999.
[19] K. Chen, B. Xue, M. Zhang, and F. Zhou, “Evolutionary multitasking for [42] Y. Yuan et al., “Evolutionary multitasking for multiobjective continuous
feature selection in high-dimensional classification via particle swarm optimization: Benchmark problems, performance metrics and baseline
optimisation,” IEEE Trans. Evol. Comput., early access, Jul. 26, 2021, results,” 2017, arXiv:1706.02766.
doi: 10.1109/TEVC.2021.3100056. [43] K. Deb and R. B. Agrawal, “Simulated binary crossover for continuous
[20] H. Song, A. K. Qin, P.-W. Tsai, and J. J. Liang, “Multitasking multi- search space,” Complex Syst., vol. 9, no. 2, pp. 115–148, 1995.
swarm optimization,” in Proc. IEEE Congr. Evol. Comput. (CEC), 2019, [44] K. Deb and M. Goyal, “A combined genetic adaptive search (GeneAS)
pp. 1937–1944. for engineering design,” Comput. Sci. Informat., vol. 26, no. 4,
[21] K. Chen, B. Xue, M. Zhang, and F. Zhou, “An evolutionary pp. 30–45, 1996.
multitasking-based feature selection method for high-dimensional [45] E. Zitzler, M. Laumanns, and L. Thiele, “SPEA2: Improving the
classification,” IEEE Trans. Cybern., early access, Dec. 31, 2020, strength pareto evolutionary algorithm,” Eidgenössische Technische
doi: 10.1109/TCYB.2020.3042243. Hochschule Zürich (ETH), Institut für Technische Informatik und
[22] J. Liang et al., “Evolutionary multi-task optimization for parameters Kommunikationsnetze (TIK), Zürich, Switzerland, TIK-Rep. 103, 2001.
extraction of photovoltaic models,” Energy Convers. Manag., vol. 207, [46] Y. Tian, R. Cheng, X. Zhang, and Y. Jin, “PlatEMO: A MATLAB
pp. 112509–112514, Mar. 2020. platform for evolutionary multi-objective optimization,” IEEE Comput.
[23] X. Zheng, A. K. Qin, M. Gong, and D. Zhou, “Self-regulated evolution- Intell. Mag., vol. 12, no. 4, pp. 73–87, Nov. 2017.
ary multitask optimization,” IEEE Trans. Evol. Comput., vol. 24, no. 1, [47] J. Alcalá-fdez et al., “KEEL: A software tool to assess evolutionary
pp. 16–28, Feb. 2020. algorithms for data mining problems,” Soft Comput., vol. 13, no. 3,
[24] J. Lin, H.-L. Liu, B. Xue, M. Zhang, and F. Gu, “Multiobjective multi- pp. 307–318, 2008.
tasking optimization based on incremental learning,” IEEE Trans. Evol. [48] A. Vargha and H. D. Delaney, “A critique and improvement of the CL
Comput., vol. 24, no. 5, pp. 824–838, Oct. 2020. common language effect size statistics of mcgraw and wong,” J. Educ.
[25] W. Ning, B. Guo, Y. Yan, X. Wu, J. Wu, and D. Zhao, “Constrained Behav. Stat., vol. 25, no. 2, pp. 101–132, 2000.
multi-objective optimization using constrained non-dominated sorting [49] G. Guizzo, F. Sarro, J. Krinke, and S. R. Vergilio, “Sentinel:
combined with an improved hybrid multi-objective evolutionary algo- A hyper-heuristic for the generation of mutant reduction strate-
rithm,” Eng. Optim., vol. 49, no. 10, pp. 1645–1664, 2017. gies,” IEEE Trans. Softw. Eng., early access, Jun. 15, 2020,
[26] Z. Ma, Y. Wang, and W. Song, “A new fitness function with two rank- doi: 10.1109/TSE.2020.3002496.
ings for evolutionary constrained multiobjective optimization,” IEEE [50] J.-H. Kim, J.-H. Han, Y.-H. Kim, S.-H. Choi, and E.-S. Kim,
Trans. Syst., Man, Cybern., Syst., vol. 51, no. 8, pp. 5005–5016, “Preference-based solution selection algorithm for evolutionary
Aug. 2021. multiobjective optimization,” IEEE Trans. Evol. Comput., vol. 16,
[27] Z.-Z. Liu, Y. Wang, and B.-C. Wang, “Indicator-based constrained no. 1, pp. 20–34, Feb. 2012.
multiobjective evolutionary algorithms,” IEEE Trans. Syst., Man, [51] J. J. Liang, X. P. Zhu, C. T. Yue, Z. Li, and B.-Y. Qu, “Performance
Cybern., Syst., vol. 51, no. 9, pp. 5414–5426, Sep. 2021. analysis on knee point selection methods for multi-objective sparse
[28] Z. Fan, Y. Fang, W. Li, X. Cai, C. Wei, and E. Goodman, “MOEA/D optimization problems,” in Proc. IEEE Congr. Evol. Comput. (CEC),
with angle-based constrained dominance principle for constrained 2018, pp. 1–8.
multi-objective optimization problems,” Appl. Soft Comput., vol. 74, [52] L. Feng, W. Zhou, W. Liu, Y.-S. Ong, and K. C. Tan,
pp. 621–633, Jan. 2019. “Solving dynamic multiobjective problem via autoencoding evolu-
[29] Z. Fan et al., “Push and pull search for solving constrained multi- tionary search,” IEEE Trans. Cybern., early access, Oct. 1, 2020,
objective optimization problems,” Swarm Evol. Comput., vol. 44, doi: 10.1109/TCYB.2020.3017017.
pp. 665–679, Feb. 2019. [53] R. Azzouz, S. Bechikh, and L. B. Said, “Multi-objective optimization
[30] Y. Tian, Y. Zhang, Y. Su, X. Zhang, K. C. Tan, and Y. Jin, “Balancing with dynamic constraints and objectives: New challenges for evolu-
objective optimization and constraint satisfaction in constrained evolu- tionary algorithms,” in Proc. Annu. Conf. Genet. Evol. Comput., 2015,
tionary multiobjective optimization,” IEEE Trans. Cybern., early access, pp. 615–622.
Mar. 17, 2021, doi: 10.1109/TCYB.2020.3021138. [54] Z. Liang, X. Xu, L. Liu, Y. Tu, and Z. Zhu, “Evolutionary
[31] Z.-Z. Liu and Y. Wang, “Handling constrained multiobjective many-task optimization based on multi-source knowledge trans-
optimization problems with constraints in both the decision and objec- fer,” IEEE Trans. Evol. Comput., early access, Aug. 2, 2021,
tive spaces,” IEEE Trans. Evol. Comput., vol. 23, no. 5, pp. 870–884, doi: 10.1109/TEVC.2021.3101697.
Oct. 2019. [55] J. Lin, H.-L. Liu, K. C. Tan, and F. Gu, “An effective knowledge trans-
[32] J. Wang, G. Liang, and J. Zhang, “Cooperative differential evolution fer approach for multiobjective multitasking optimization,” IEEE Trans.
framework for constrained multiobjective optimization,” IEEE Trans. Cybern., vol. 51, no. 6, pp. 3238–3248, Jun. 2021.
Cybern., vol. 49, no. 6, pp. 2060–2072, Jun. 2019.
[33] S. Qian, Y. Ye, B. Jiang, and J. Wang, “Constrained multiobjective
optimization algorithm based on immune system model,” IEEE Trans.
Cybern., vol. 46, no. 9, pp. 2056–2069, Sep. 2016.
[34] K. Li, R. Chen, G. Fu, and X. Yao, “Two-archive evolutionary algo-
rithm for constrained multiobjective optimization,” IEEE Trans. Evol.
Comput., vol. 23, no. 2, pp. 303–315, Apr. 2019.
[35] Y. Tian, T. Zhang, J. Xiao, X. Zhang, and Y. Jin, “A coevolutionary
framework for constrained multiobjective optimization problems,” IEEE
Trans. Evol. Comput., vol. 25, no. 1, pp. 102–116, Feb. 2021.
[36] X. Ma et al., “Enhanced multifactorial evolutionary algorithm with
meme helper-tasks,” IEEE Trans. Cybern., early access, Feb. 10, 2021,
doi: 10.1109/TCYB.2021.3050516.
[37] 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. Kangjia Qiao (Member, IEEE) received the B.E.
[38] Z. Fan et al., “An improved epsilon constraint-handling method in degree from Jianghan University, Wuhan, China,
MOEA/D for CMOPs with large infeasible regions,” Soft Comput., in 2018, and the M.S. degree from Zhengzhou
vol. 23, no. 23, pp. 12491–12510, 2019. University, Zhengzhou, China, in 2021, where he is
[39] H. Ma, H. Wei, Y. Tian, R. Cheng, and X. Zhang, “A multi-stage currently pursuing the Ph.D. degree.
evolutionary algorithm for multi-objective optimization with complex His current research interests include differ-
constraints,” Inf. Sci., vol. 560, pp. 68–91, Jun. 2021. ential evolution, multitasking optimization, con-
[40] P. A. N. Bosman and D. Thierens, “The balance between proximity and strained optimization, and multimodal multiobjective
diversity in multiobjective evolutionary algorithms,” IEEE Trans. Evol. optimization.
Comput., vol. 7, no. 2, pp. 174–188, Apr. 2003.

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.
QIAO et al.: EMT OPTIMIZATION FRAMEWORK FOR CMOPs 277

Kunjie Yu received the Ph.D. degree in control sci- Hui Song received the M.S. degree from Zhengzhou
ence and engineering from the East China University University, Zhengzhou, China, in 2014, and the
of Science and Technology, Shanghai, China, in Ph.D. degree in computer science and IT from the
2017. School of Science, RMIT University, Melbourne,
He is currently an Associate Professor with VIC, Australia in 2019.
the School of Electrical Engineering, Zhengzhou She is currently a Research Fellow with the
University, Zhengzhou, China. His current research School of Engineering, RMIT University. Her
interests include evolutional computation, con- research interests mainly include evolutionary com-
strained optimization, multiobjective optimization, putation, machine learning, time-series analytics,
and their applications in chemical process, photo- evolutionary multitasking learning, and EV charg-
voltaic system, and energy system. ing/discharging.

Boyang Qu received the B.E. and Ph.D. degrees

from the School of Electrical and Electronic
Engineering, Nanyang Technological University,
Singapore.
He is a Professor with the School of Electric
and Information Engineering, Zhongyuan University
of Technology, Zhengzhou, China. His research
interests include machine learning, neural network,
genetic and evolutionary algorithms, swarm intelli-
gence, and multiobjective optimization.
Dr. Qu is an Associate Editor for the Swarm and
Evolutionary Computation.

Caitong Yue (Member, IEEE) received the

Jing Liang (Senior Member, IEEE) received the B.E. and Ph.D. degrees from the School of
B.E. degree from the Harbin Institute of Technology, Electrical Engineering from Zhengzhou University,
Harbin, China, and the Ph.D. degree from the School Zhengzhou, China. He studied from the School
of Electrical and Electronic Engineering, Nanyang of Electrical and Electronic Engineering, Nanyang
Technological University, Singapore. Technological University, Singapore, from June
She is currently a Professor with the School 2019 to June 2020.
of Electrical Engineering, Zhengzhou University, He is currently a Lecturer with the School of
Zhengzhou, China. Her main research interests Electrical Engineering, Zhengzhou University. His
are evolutionary computation, swarm intelligence, research interests include multimodal multiobjective
multiobjective optimization, and neural network. optimization, pattern recognition, neural network,
and particle swarm optimization.

Authorized licensed use limited to: Universite De Sherbrooke. Downloaded on June 06,2022 at 15:09:58 UTC from IEEE Xplore. Restrictions apply.