This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

A Constrained Decomposition Approach with Grids

for Evolutionary Multiobjective Optimization
Xinye Cai, Member, IEEE, Zhiwei Mei, Zhun Fan, Senior Member, IEEE, Qingfu Zhang, Fellow, IEEE

Abstract—Decomposition based multiobjective evolutionary to tackle multiobjective optimization problems (MOPs). As

algorithms (MOEAs) decompose a multiobjective optimization a representative of such approaches, the multiobjective evo-
problem into a set of scalar objective subproblems and solve lutionary algorithm based on decomposition (MOEA/D) [42]
them in a collaborative way. Commonly used decomposition
approaches originate from mathematical programming and the has drawn a large amount of attention over the recent years.
direct use of them may not suit MOEAs due to their population- One critical difficulty for MOEA/D is on how to approximate
based property. For instance, these decomposition approaches a set of uniformly-distributed Pareto optimal solutions without
used in MOEAs may cause the loss of diversity and/or be very knowing the shape of the Pareto front (PF) [30], [9] a priori.
sensitive to the shapes of Pareto fronts (PFs). This paper proposes Commonly-used decomposition approaches in MOEA/D in-
a constrained decomposition with grids (CDG) that can better
address these two issues thus more suitable for MOEAs. In ad- cluding Weighted Sum (WS), Tchebycheff (TCH) and Penalty-
dition, different subproblems in CDG defined by the constrained based Boundary Intersection (PBI) [30] may fail to achieve
decomposition constitute a grid system. The grids have an such a goal due to the following two reasons [29], [37].
inherent property of reflecting the information of neighborhood First, WS, TCH and PBI tend to be very sensitive to the
structures among the solutions, which is a desirable property shapes of PFs [42]. An example of the Pareto optimal solutions
for restricted mating selection in MOEAs. Based on CDG,
a constrained decomposition MOEA with grid (CDG-MOEA) obtained by TCH on MOPs with an extremely convex or
is further proposed. Extensive experiments are conducted to concave PF is given in Fig. 1a. Although the Pareto optimal
compare CDG-MOEA with the domination-based, indicator- solutions obtained by TCH are well-distributed on the concave
based and state-of-the-art decomposition-based MOEAs. The PF, the distribution of the solutions on the extremely convex
experimental results show that CDG-MOEA outperforms the PF is not satisfactory. In Fig. 1b, another example shows the
compared algorithms in terms of both the convergence and
diversity. More importantly, it is robust to the shapes of PFs Pareto optimal solutions obtained by TCH on MOPs with
and can still be very effective on MOPs with complex PFs (e.g., disparately scaled objectives. It can be clearly seen that these
extremely convex, or with disparately scaled objectives). Pareto optimal solutions are very unevenly-distributed on PF,
Index Terms—Evolutionary multiobjective optimization, con- where almost half of the PF is not covered by any Pareto
strained decomposition, grids, robust to Pareto front optimal solution.
It is worth noting that there has already been some research
I. I NTRODUCTION in the literature to address either one of the above scenarios.
As far as we know, an inverted PBI has been proposed to tackle
Along with domination-based (e.g., [48], [12], [23], [24], MOPs with extremely convex PFs in [33]. However, the use of
[40], [7], [28]) and indicator-based (e.g., [47], [3], [20], inverted PBI to achieve well-distributed solution set still needs
[4], [2]) multiobjective evolutionary algorithms (MOEAs), to assume the convexity of PFs. A combination of normal
decomposition-based MOEAs (e.g., [31], [42], [32], [13], [34], boundary intersection and the Tchebycheff approach has been
[18], [19]) have been recognized as a major type of approaches proposed for MOPs with disparately scaled objectives in [43],
Xinye Cai and Zhiwei Mei are with the College of Computer Science and where a satisfactory distribution of solutions can be achieved
Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, in bi-objective optimization problems but fails to extend to
Jiangsu, 210016 P. R. China, and also with the Collaborative Innovation Center tri-objective optimization problems.
of Novel Software Technology and Industrialization, Nanjing, 210023, China
(e-mail: [email protected],[email protected]). Second, in those commonly-used decomposition methods,
Zhun Fan is with Guangdong Provincial Key Laboratory of Digital Sig- the same solution is very likely to be assigned to many differ-
nal and Image Processing and the Department of Electronic Engineering, ent subproblems, which may lead to the loss of diversity [37],
School of Engineering, Shantou University, Guangdong, P. R. China (e-mail:
[email protected]). [29]. The reason of such phenomenon can be explained as
Qingfu Zhang is with the Department of Computer Science, City follows. Let xi be the current solution for the i-th subproblem,
University of Hong Kong, Kowloon, Hong Kong SAR (e-mail: qing- then the improvement region of a solution xi can be defined
[email protected]); the School of Computer Science and Electronic
Engineering, University of Essex, Wivenhoe Park, Colchester, CO4 3SQ, UK, as the set {F (x)| x is better than xi for the i-th subproblem
(e-mail: [email protected]). } [37], as shown in Fig. 2a-2c. Any new solution in the
This work was supported in part by the National Natural Science Foundation improvement region for xi can improve and replace it for
of China (NSFC) under grant 61300159, 61732006, 61473241 and 61332002,
by the Natural Science Foundation of Jiangsu Province of China under the i-th subproblem. As it can be observed in Fig. 2a-2c, the
grant BK20130808, by China Postdoctoral Science Foundation under grant improvement regions for three commonly used decomposition
2015M571751. approaches (WS, TCH and PBI), may be too large for causing
Copyright (c) 2012 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be the loss of diversity.
obtained from the IEEE by sending a request to [email protected]. Over the recent years, some attempts have already been

1089-778X (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

O O O O
(a) WS (b) TCH (c) PBI (d) CDG

Fig. 2: An illustration of improvement regions for three commonly used decomposition approaches (Fig. 2a-2c). In each sub-
figure, the square point is solution xi of i-th subproblem with direction vector λi , the solid circle point is its optimal solution
and the dash line represents its contour. Therefore, the entire shaded region is the improvement region for xi .

݂ଶ ݂ଶ
ison approach has been proposed to balance the convergence
10 50
and diversity. An online geometrical metric has been proposed
to enhance the diversity of MOEA/D in [14], [6]. Designing
more suitable decomposition approaches for MOEAs is also
a possible way for maintaining better diversity in MOEA/D.
For instance, the use of different decomposition approaches
for different search phases has been studied in [21]. In a very
recent work [37], an angle θ is imposed on each subproblem
ܼ‫כ‬ 10 ݂ଵ ܼ‫כ‬ 10 ݂ଵ
as a constraint to improve the diversity in MOEA/D [37].
However, the appropriate setting of parameter θ for each
(a) The Pareto optimal solutions ob- (b) The Pareto optimal solutions ob-
tained by TCH on MOPs with con- tained by TCH on MOPs with dis-
subproblem is a very tedious task and different subproblems
cave and extremely convex PFs parately scaled objectives, where the on different evolutionary stages may have different θ values.
scale of f2 is five times as much as In this paper, a constrained decomposition with grids
that of f1
(CDG) is proposed to address the above two aspects for
݂ଶ ݂ଶ
ܼ௡௔ௗ ܼ௡௔ௗ decomposition-based MOEAs. In CDG, one objective function
10 50
is selected to be optimized while all the other objective
functions are converted into constraints by setting up both
upper and lower bounds. In a sense, CDG can be considered
as an extension of -constraint approach [17], [30]. If CDG
is applied to all the objectives, the volumes of the improve-
ment regions for a solution are appropriately reduced to the
narrowed regions, where the same solution can be assigned to
ܼ‫כ‬ 10 ݂ଵ ܼ‫כ‬ 10 ݂ଵ
at most m subproblems (m is the number of objectives), as
shown in Fig. 2d. This interesting characteristic gives CDG a
(c) The Pareto optimal solutions ob- (d) The Pareto optimal solutions ob-
tained by CDG on MOPs with con- tained by CDG on MOPs with dis- natural ability for maintaining better diversity for MOEAs. In
cave and extremely convex PFs parately scaled objectives, where the addition, as each objective is equally divided by constraints,
scale of f2 is five times as much as unlike the commonly-used decomposition methods (WS, TCH,
that of f1
PBI), CDG is very robust to the shapes of PFs. These can
Fig. 1: An illustration of Pareto optimal solutions obtained by be observed in Fig. 1c and Fig. 1d, where the Pareto optimal
TCH or CDG on MOPs with different shapes of PFs solutions obtained by CDG are well-distributed on MOPs with
concave or convex PFs, and/or disparately scaled objectives.
Also, another interesting observation in Fig. 1c and Fig. 1d
made to maintain better diversity in MOEA/D. Among them, is that the contour lines of different constrained subproblems
Li et al. limit the number of subproblems allowed to be updat- constitute a grid-coordinate-system. A grid has an inherent
ed by a single offspring in [26]. The offspring is only allowed property of reflecting the information of neighborhood struc-
to update the most suitable subproblem in [38]. Decomposition tures among the solutions [39]: each solution in the grid can
approaches have been hybridized with the R2 indicator in [36]. be located by grid coordinates and the grid coordinates can
An external archive is adopted to maintain the representative help the solutions to locate its neighboring solutions, which is
solutions and guide the search in the MOEA/D population essential for the restricted mating selection in an MOEA. More
in [5]. In [27], a global stable matching model (STM) is details of constrained decomposition with grids are specified
integrated into MOEA/D to find suitable matches between in Section III.
subproblems and solutions. In [1], an adaptive epsilon compar- The rest of this paper is organized as follows. Section II

1089-778X (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

introduces some preliminaries on MOPs and three popular m

decomposition approaches are also introduced in this section. minimize g ws (x|λi ) =
X
λij fj (x) ,
Section III details the proposed constrained decomposition j=1 (2)
with grids. Section IV describes the whole framework of CDG-
subject to x ∈ Ω .
MOEA. Section V introduces the benchmark test functions and
the performance indicators used in the experimental studies. Its search direction vector is defined as λi , as shown in
Experiments and discussions are presented in Section VI, Fig. 2a.
where CDG-MOEA is compared with five decomposition- 2) Tchebycheff (TCH) Approach: The i-th subproblem is
based, one domination-based and one indicator-based MOEAs defined as
on a set of well-known benchmark functions and a set of
benchmark MOPs with complex PFs. In addition, the sensi- minimize g te (x|λi , z ∗ ) = max {|fj (x) − zj∗ |/λij } ,
tivity analysis of parameters in CDG-MOEA is conducted in 1≤j≤m
Section VII. Finally, Section IX concludes this paper. subject to x ∈ Ω .
(3)
where Ω is the feasible region, but λj = 0 is replaced
II. P RELIMINARIES
by λj = 10−6 because λj = 0 is not allowed as a
A. Basic Definitions denominator in (3). Its search direction vector is defined
A multiobjective optimization problem (MOP) can be de- as λi , as shown in Fig. 2b.
fined as follows: 3) Penalty-based Boundary Intersection (PBI) Ap-
proach: This approach is a variant of Normal-Boundary
minimize F (x) = (f1 (x), . . . , fm (x))T , Intersection approach [8]. The i-th subproblem is defined
(1)
subject to x ∈ Ω . as
where Ω is the decision space and F : Ω → Rm consists
of m real-valued objective functions. {F (x)|x ∈ Ω} is the minimize g pbi (x|λi , z ∗ ) = di1 + βdi2 ,
attainable objective set. di1 = (F (x) − z ∗ )T λi /||λi || ,
Let u, v ∈ Rm , u is said to dominate v, denoted by (4)
di2 = ||F (x) − z ∗ − (di1 /||λi ||)λi || ,
u ≺ v, if and only if uj ≤ vj for every j ∈ {1, . . . , m}
and uk < vk for at least one index k ∈ {1, . . . , m}1 . Given subject to x ∈ Ω .
a set S in Rm , a solution x ∈ S is called nondominated where ||.|| denotes the L2 -norm and β is the penalty
if no other solution in S dominates it. A solution x∗ ∈ Ω parameter. Its search direction vector is defined as λi ,
is Pareto-optimal if F (x∗ ) is nondominated in the attainable as shown in Fig. 2c.
objective set. F (x∗ ) is then called a Pareto-optimal (objective)
vector. In other words, any improvement in one objective III. T HE C ONSTRAINED D ECOMPOSITION WITH G RIDS
of a Pareto optimal solution is bound to deteriorate at least
In this section, the setup of a grid system and the constrained
another objective. The set of all the Pareto-optimal solutions is
decomposition based on the grid system are introduced as
called the Pareto set (PS) and the set of all the Pareto-optimal
follows.
objective vectors is the Pareto front (PF) [30]. The ideal and
nadir objective vectors are another two important concepts
containing the information on the ranges of PFs as follows. A. The Setup of a Grid System
The ideal objective vector z ∗ is a vector z ∗ = {z1∗ , . . . , zm
∗ T
} , The constrained decomposition is based on a grid-system
∗
where zj = min fj (x), j ∈ {1, . . . , m}. The nadir objective and the setup of it is introduced in Algorithm 1, as follows.
x∈Ω
vector z nad is a vector z nad = (z1nad , . . . , zm
nad T
) , where Each objective is divided into K equal intervals within the
nad
zj = max fj (x), j ∈ {1, . . . , m}. approximations of the ideal and nadir points, where K is a
x∈P S preset parameter. The width of each interval is
dj = (zjnad − zj∗ + 2 × σ)/K . (5)
B. Decomposition Approaches
Popular decomposition methods [30] used in MOEAs in- Fig. 3a shows a grid division for a bi-objective problem
clude Weighted Sum, Tchebycheff and Penalty-based Bound- where K = 4.
ary Intersection, either one of which decomposes an MOP The grid location of x along the j-th objective gj (x) can
into a number of scalar optimization subproblems [42]. These be calculated as
widely used approaches can be defined as follows. gj (x) = d(fj (x) − zj∗ + σ)/dj e , (6)
Let λi = (λ1 , . . . , λm )T be a direction vector
Pmfor the i-th
subproblem, where λj ≥ 0, j ∈ 1, . . . , m and j=1 λj = 1. where d.e denotes the ceil function, gj (x) is the grid-
1) Weighted Sum (WS) Approach: The i-th subproblem coordinate of solution x and fj (x) is the value of j-th objective
is defined as function. A small positive number σ is introduced to ensure
that the value of gj is more than 0 but not more than
1 In the case of maximization, the inequality signs should be reversed. K. An example of the grid location along one objective is

1089-778X (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

demonstrated in Fig. 3b, where K = 4. In the example, the

grid-coordinates of the solutions (denoted by “•”) along f1 ܵଶሺʹሻ ‫ ݖ‬௡௔ௗ
߳ଶర
are assigned as follows: a = 1, b = 1, c = 2, d = 2, e = 3,
f = 3, g = 4 and h = 4.
߳ଶయ

Algorithm 1: Grid-system Setup (GS) ܵଵሺ͵ሻ

߳ଶమ
Input : P : the current population;
z ∗ : the approximation of the ideal point;
z nad : the approximation of the nadir point; ߳ଶభ
m: the number of objectives;
K: the number of the intervals on each ߳ଶబ
objective. œ ‫߳ כ‬ଵబ ߳ଵభ ߳ଵమ ߳ଵయ ߳ଵర
Output: the grid locations of P . (a) A grid system (K = 4) in a bi-objective
1 for j = 1 to m do optimization problem
2 dj = (zjnad − zj∗ + 2 × σ)/K; ‫ݖ‬௝ȗ ‫ݖ‬௝௡௔ௗ
3 end ݆
߳௝బ ܽ ܾ߳ ܿ ݀߳ ݁ ݂ ߳ ݃ ݄ ߳
4 foreach x ∈ P do ௝భ ௝మ ௝య ௝ర
5 for j = 1 to m do ݀௝
6 gj (x) = d(fj (x) − zj∗ + σ)/dj e; ߪ ߪ
7 end (b) Grid location along one objective (K = 4)
8 G(x) = (g1 (x), . . . , gm (x)); Fig. 3: An illustration of the grid system produced by con-
9 end
strained decomposition

To use the grid locations for the restricted mating selection is a (m − 1)-dimensional vector), which can be defined as:
in MOEAs, the Grid Distance and Grid Neighbor are defined
for convenience as follows. Sl (k) = {x|g1 (x) = k1 , . . . , gl−1 (x) = kl−1 ,
Definition 1 (Grid Distance): Let u, v ∈ Rm be two solu- gl+1 (x) = kl+1 , . . . , gm (x) = km } , (10)
tions, the grid distance GD(u, v) between u and v is defined subject to l ∈ {1, . . . , m} k ∈ {1, . . . , K} m−1
.
as
An example of such subproblems in a bi-objective opti-
GD(u, v) = max (|gj (u) − gj (v)|) . (7) mization is given in Fig. 3a, where the feasible regions of
j=1,...,m
two subproblems S1 (3) and S2 (2) are denoted in the shaded
Definition 2 (Grid Neighbors): The grid neighbors of a so- regions.
lution x within distance T is defined as
IV. CDG-MOEA
GN (x, T ) = {x∗ |GD(x, x∗ ) ≤ T x, x∗ ∈ Rm } . (8) A. The Main Framework of CDG-MOEA
In this section, the main framework of constrained de-
composition based multiobjective evolutionary algorithm with
B. Constrained Decomposition with Grids (CDG)
grids (CDG-MOEA) is presented in Algorithm 2, which
A constrained decomposition can be defined by adopting includes 6 steps: initialization, reproduction, update of the
the grid system specified in the last section. CDG can be ideal and nadir points, update of the grid system, rank-based
considered as an extension of -constraint approach [30]. The selection and termination. In the following sections, each step
constrained decomposition approach for the k-th subproblem is specified in details.
of the l-th objective can be defined as follows. Algorithm 2:
Input:
1) an MOP;
minimize fl (x) ,
2) a stopping criterion;
subject to gj (x) = kj for all j = 1, . . . , m, j 6= l , 3) N : the population size of P ;
kj ∈ {1, . . . , K} , 4) T : the maximum grid distance for neighborhood;
x∈Ω. 5) K: the number of the intervals in each objective.
(9) Output: A solution set P ;
where K is a division parameter which determines the number Step 1: Initialization:
of grids. Step 1.1 Generate an initial population P =
With K intervals on each objective, the grids decompose {x1 , . . . , xN } randomly;
an MOP into m × K m−1 subproblems. In general, the k-th Step 1.2 Approximate the ideal and nadir points: z ∗ =
subproblem of l-th objective contains a solution set Sl (k) (k U PDATE 1(P ), z nad = U PDATE 2(P );

1089-778X (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

Step 1.3 Initialize the grid system: GS(P ); Algorithm 4: Update the Nadir Point (U PDATE 2)
Step 1.4 Set gen = 0. Input : P : the combined population;
Step 2: Reproduction: z ∗ : the current ideal point;
Step 2.1 Generate an empty set Q = ∅; z nad : the current nadir point.
For each solution x ∈ P do Output: Updated nadir point z nad .
Step 2.2 Obtain the neighboring solutions as the mating /* To reduce computational cost, find a
pool of x: subset of P : SP */
1 SP = ∅;

 GN (x, T ), rand < δ and
2 foreach x ∈ P do
NS = |GN (x, T )| > 2 , (11)

{x1 , . . . , xN }, otherwise . 3 for j = 1 to m do
z nad
k l 4 if fj (x) < zj∗ + j5 then
Step 2.3 Select two solutions x and x from N S
5 SP = SP ∪ x;
randomly; and generate an offspring solution y from
6 end
solution x, xk and xl by DE operators; then y is added
7 end
to Q.
8 end
End for
/* find all the nondominated solutions
Step 3: Update of the ideal and nadir points: in SP . */
Step 3.1 gen = gen + 1; 9 SP = N ONDOMINATED -S ELECTION(SP );
Step 3.2 P = P ∪ Q; 10 for j = 1 to m do
Step 3.3 Update the ideal point: z ∗ = U PDATE 1(P ); 11 zjnad = max {fj (x)};
Step 3.4 If gen is a multiplication of 50, update the x∈SP
nadir point: z nad = U PDATE 2(P ). 12 end
Step 4: Update of the grid system:
Step 4.1 P̄ = {x|x ∈ P ∧ ∃j ∈ {1, . . . , m}, fj (x) >
zjnad }; where δ is the probability that the mating solutions are selected
Step 4.2 P = P \P̄ ; from the grid neighbors.
Step 4.3 Update the grid system: GS(P ). In Step 2.3, two solutions xk and xl are selected from N S
Step 5: Rank-based selection: randomly. An offspring solution y is generated from solution
Step 5.1 If |P | < N , randomly select N − |P | solutions x, xk and xl by DE operators [26]; and then y is added to Q.
from P̄ and add them to P . Otherwise, P = RBS(P ).
Step 6: Termination: D. Update of the Ideal and Nadir Points
Step 6.1 If the stopping criterion is satisfied, terminate
the algorithm and output P . Otherwise, go to Step 2. In Step 3, the approximations of the ideal and nadir points
are updated by using the combined population P = P ∪ Q.
In Algorithm 3, the ideal point z ∗ is approximated by the
B. Initialization
minimum value of each objective in P .
In Step 1.1, a population P is initialized randomly. In Step In Algorithm 4, the nadir point is approximated by the
1.2, the ideal and nadir points are approximated based on P . maximum value of each objective in the nondominated so-
The update of the ideal point is presented in Algorithm 3 and lutions of P . To further lower the computational cost of
the update of the nadir point is presented in Algorithm 4. nondominated selection for approximating nadir point, only
a subset of solutions SP in P that are close to the corner
Algorithm 3: Update the Ideal Point (U PDATE 1) solutions is selected (line 1-8 of Algorithm 4), as follows:
Input : P : the current population.
Output: Updated ideal point z ∗ . zjnad
SP = {x|x ∈ P ∧ ∃j ∈ {1, . . . , m}, fj (x) < zj∗ + }.
1 for j = 1 to m do 5
(13)
2 zj∗ = min{fj (x)};
x∈P After that, the nondominated solutions are selected from
3 end SP and the nadir point is updated with the maximum value
of each objective in these nondominated solutions (line 9-12
of Algorithm 4).
C. Reproduction
In the Step 2, N offspring solutions are generated from P . E. Update of the Grid System
An empty set Q is generated for storing the offspring solutions. In Step 4.1-4.2, the solutions P̄ = {x|x ∈ P ∧ ∃j ∈
For each solution x ∈ P , its mating solutions are obtained by {1, . . . , m}, fj (x) > zjnad } ( the ones located outside the nadir

 GN (x, T ), rand < δ and point approximation) are eliminated from P first. Then, in Step
NS = |GN (x, T )| > 2 , (12) 4.3, the grid system is updated using P by calling Algorithm
{x1 , . . . , xN }, otherwise . 1, which was already specified in Section III.B.


1089-778X (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

F. Rank-Based Selection ሺͳǡʹሻ

In Step 5, if the size of P is less than N , then N − |P | ܼ௡௔ௗ
solutions are selected randomly from P̄ to fill in P ; otherwise, ሺʹǡͳሻ
the rank-based selection (RBS), presented in Algorithm 5, is
called. ሺʹǡʹሻ
In Algorithm 5, N solutions are selected from Q (|Q| > N ) ሺͳǡ͵ሻ
by rank-based selection, which include decomposition-based ሺͶǡͷሻ
ranking and lexicographic sorting. The details of them are ሺ͵ǡͳሻ ሺ͵ǡ͵ሻ
explained as follows. ሺͳǡͶሻ ሺͶǡʹሻ
In the decomposition-based ranking, each solution set Sl (k) ሺͷǡͶሻ
for the k-th subproblem of the l-th objective is ranked based ሺʹǡͳሻ
on constrained decomposition with grids defined in Eq. (9). ሺʹǡʹሻ
After ranking all the m objectives, each solution x has m ሺͳǡ͵ሻ ሺ͵ǡͳሻ
ranks, stored in a rank vector R(x) = (r1 (x), . . . , rm (x)). It ܼ‫כ‬
is worth noting that, with mK m−1 subproblems, at most mM
times rankings are needed as some subproblems may contain Fig. 4: The ranks of all the solutions in a grid system (K =
no solutions, where M = |Q|. 4) for a bi-objective optimization problem, where the final
In the lexicographic sorting, R(x) is sorted in an ascending selected solutions are denoted in .
order and stored in R0 (x). Then, each solution x ∈ Q is ranked
in lexicographic order based on R0 (x) and the first N solutions TABLE I: The procedures of decomposition-based ranking and
are selected from Q and assigned to P . lexicographic sorting for Fig. 4
Lexicographic
Algorithm 5: Rank-based Selection (RBS) R(x) R0 (x)
Sorting
Input : Q: the combined population. 1. (1,2) 1. (1,2) 1. (1,2) X
Output: A population P . 2. (2,1) 2. (1,2) 2. (1,2) X
/* Decomposition-based Ranking */ 3. (1,3) 3. (1,3) 8. (1,2) X
4. (2,2) 4. (2,2) 3. (1,3) X
1 foreach x ∈ Q do
5. (3,1) 5. (1,3) 5. (1,3) X
2 initialize R(x) = {r1 (x), . . . , rm (x)} = {0, . . . , 0} ; 6. (4,5) 6. (4,5) 12. (1,3) X
3 end 7. (1,4) 7. (1,4) 14. (1,3) X
8. (2,1) 8. (1,2) 7. (1,4) ×
4 for l = 1 to m do
9. (3,3) 9. (3,3) 4. (2,2) ×
5 forall subproblems Sl (k) do 10. (4,2) 10. (2,4) 13. (2,2) ×
6 if Sl (k) 6= ∅ then 11. (5,4) 11. (4,5) 10. (2,4) ×
12. (1,3) 12. (1,3) 9. (3,3) ×
7 [S 0 , I] = sortl (Sl (k)); // sort Sl (k) by 13. (2,2) 13. (2,2) 6. (4,5) ×
fl based on Eq. (9) and I stores 14. (3,1) 14. (1,3) 11. (4,5) ×
the ranks
8 foreach x ∈ Sl (k) do rl (x) = I(x);
9 end G. Computational Complexity
10 end
In Algorithm 1, the setup of a grid system requires O(mN )
11 end
computations, where m is the number of objectives and N
/* Lexicographic Sorting */ is the population size. In Algorithm 3, the update of the
12 foreach x ∈ Q do
ideal point needs O(mN ) computations. In Algorithm 4, the
13 R0 (x) = sort(R(x));
computational complexity for the update of the nadir point is
14 end
O(mL2 ), where L is the size of nondominated solutions in
/* sort all x ∈ Q based on R0 (x) in
SP .
lexicographic order */ In Algorithm 5, the rank-based selection is divided into two
15 Q = L EXICOGRAPHIC -S ORT (Q);
parts: the decomposition-based ranking and the lexicographic
16 P = Q(1 : N ); // select first N solutions
sorting. Among them, the decomposition-based ranking needs
to be done at most mM times, where M ≤ 2N is the size of
Fig. 4 and Table I show an example of rank-based selection solutions that are located inside the nadir point approximation.
in a grid system (K = 4) for a bi-objective optimization In the lexicographic sorting, the sorting in line 12-14 requires
problem, where |P | = 7 solutions are selected out of |Q| = 14 O(mM logm) computations and the second sorting in line
solutions in Algorithm 5, as follows. All the solutions are 15 requires O(mM logM ) computations. In summary, the
ranked as R(x) based on Eq. (9), as shown in Fig. 4 and computational cost of rank-based selection is O(mM logM ),
the first column of Table I . The ranks are sorted and stored where M ≤ 2N .
in R0 (x) (second column of Table I). After the lexicographic In the framework of CDG-MOEA (Algorithm 2), the com-
sorting, the first 7 solutions are selected (the last column of putational cost of both Step 1 and Step 2 is O(mN ). The com-
Table I). putational cost of Step 2-5 is O(mM logM ). In summary, the

1089-778X (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

computational complexity of CDG-MOEA is O(mM logM ). to z r ) is the volume of the region dominated by P and
bounded by z r , and it can be defined as
V. E XPERIMENTAL S ETTINGS [
A. Test Instances IH (P ) = volume( [f1 , z1r ] × . . . [fm , zm
r
]). (15)
f ∈P
UF suite [45] is considered as the benchmark functions
in our experimental studies. Among all the 10 benchmark Obviously, the higher the hypervolume, the better the
problems, UF1-7 are bi-objective and UF8-10 are tri-objective approximation is. In our experiments, z r is set to (2, 2)T
optimization problems. The number of decision variables for for UF1-7, (2, 2, 2)T for UF8-10, (1.2, 1.2)T for GLT1
all the test instances is set to 30. and GLT3, (1.2, 12)T for GLT2, (1.2, 2.2)T for GLT4
GLT suite [15], [41] is another set of the benchmark and (1.2, 1.2, 1.2)T for GLT5-6 when computing hyper-
problems used in our experimental studies to verify the ro- volume for the nondominated sets obtained by all the
bustness of CDG-MOEA on MOPs with complicated PFs (e.g., algorithms.
extremely convex or with disparately scaled objectives). The
number of decision variables for all the test instances is set to
10. VI. E XPERIMENTAL S TUDIES AND D ISCUSSIONS

B. Parameter Settings In this section, the following experiments are conducted to

test the performance of CDG-MOEA:
To make a fair comparison, the following parameters of all
the compared algorithms are set in a similar manner as in • comparisons on UF test suite, to verify the general
MOEA/D-DE [26]: performance of CDG-MOEA on balancing between con-
• δ = 0.9;
vergence and diversity.
• In DE: CR = 1.0, F = 0.5, η = 20 and pm = 1/n;
• comparisons on GLT test suite with complicated PFs (ex-
• Function evaluations: 300,000 for each test instance;
tremely convex and/or with disparately scaled objectives),
to verify the robustness of CDG-MOEA on the shapes of
The population size of all the compared algorithms is
PFs.
set to 300 for UF1-7 and GLT1-6; and 600 for UF8-10.
For MOEA/D-DE (WS, TCH and PBI), the parameter of
neighborhood size is set to 20, the same as the one in [26].
For CDG-MOEA, the grid division parameter K is set to 180 A. Comparisons on UF test suite
for bi-objective and 30 for tri-objective problems; and the grid In this section, CDG-MOEA is compared with four
neighborhood distance is set to 5 for bi-objective and 1 for tri- classical decomposition-based MOEAs (MOEA/D-DE (WS,
objective problems. The settings of the grid division parameter TCH or PBI) [26] and MSOPS-II [19]), one state-of-the-
and population size are based on the sensitivity analysis in art decomposition-based MOEA (MOEA/D-ACD [37]), one
Section VII. Each of all the compared algorithms is run 30 domination-based MOEA (NSGA-II [12]), one indicator-based
times independently for all the benchmark problems. MOEA (IBEA [47]) and three grid-based MOEAs (Epsilon
MOEA [11], Borg MOEA [16] and OMOPSO [35]) on UF
C. Performance Metrics test problems.
1) Inverted Generational Distance (IGD) [46], [44]: It The performance of eleven compared algorithms in terms of
measures the average distance from a set of reference IGD or IH is presented in Table II, where the performance
points P ∗ in the PF to the approximation set P . It can of the algorithm with the best mean IGD or IH value is
be formulated as follows. highlighted in boldface. It can be observed that, in terms of
1 X IGD, CDG-MOEA significantly outperforms other compared
IGD(P, P ∗ ) = ∗ dist(v, P ) (14) algorithms on all the test problems, except for UF3, UF4, UF8
|P | ∗ v∈P and UF10. MOEA/D-ACD achieves the best performance on
where dist(v, P ) is the Euclidean distance between the UF3, Borg has the best performance on UF4, MOEA/D-DE
solution v and its nearest point in P , and |P ∗ | is the car- (PBI) has the best performance on UF8, and MSOPS-II has
dinality of P ∗ . If |P ∗ | is large enough to represent the PF the best performance on UF10.
very well, IGD(P, P ∗ ) could measure both the diversity Similar performance can be observed on the comparisons
and convergence of P in a sense. In our experiments, of eight algorithms in terms of IH , where CDG-MOEA is
626 reference points are used on GLT1, 1000 reference significantly better than other compared algorithms on six out
points on GLT2-4, 2600 reference points on GLT5, and of ten test problems.
1377 reference points on GLT6, for calculating IGD. The comparisons of CDG-MOEA with other ten algorithms
2) Hypervolume Indicator (IH ) [49]: Let z r = on UF test suite verify that CDG in CDG-MOEA is able to
(z1r , . . . , zm
r T
) be a reference point in the objective space achieve better diversity while maintaining good convergence
that is dominated by all Pareto-optimal objective vectors. in most test problems, compared with other commonly-used
Let P be the obtained approximation to the PF in the decomposition methods thus more suitable for MOEA frame-
objective space. Then, the IH value of P (with regard work.

1089-778X (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

TABLE II: Performance comparisons of CDG-MOEA with ten MOEAs on UF problems in terms of the mean and standard
deviation values of IGD and IH .
MOEA/D-DE
CDG-MOEA MOEA/D-ACD MSOPS-II NSGA-II IBEA -MOEA Borg OMOPSO
instance
WS TCH PBI
IGD
† † † † † †
mean 2.072E-03 6.113E-02 2.439E-03 2.394E-03 2.096E-03 7.023E-02 8.390E-02 1.084E-01 1.206E-01† 8.217E-02† 6.790E-02†
UF1
std 5.063E-05 1.399E-02 5.026E-04 7.929E-04 9.026E-05 5.731E-03 1.187E-02 1.715E-02 3.709E-02 2.426E-02 1.325E-02
† † † † † † †
mean 5.191E-03 5.442E-02 1.118E-02 3.532E-02 7.970E-03 5.707E-02 3.272E-02 4.877E-02 4.930E-02† 1.871E-02† 2.047E-02†
UF2
std 1.276E-03 1.102E-02 3.263E-03 3.582E-02 1.918E-03 2.137E-02 2.355E-03 2.148E-02 2.007E-02 4.799E-03 1.888E-03
† † ‡ † † †
mean 1.832E-02 1.252E-01 2.539E-02 4.670E-02 6.120E-03 3.141E-01 7.031E-02 3.136E-01 2.916E-01† 9.785E-02† 6.579E-02†
UF3
std 1.367E-02 3.208E-02 2.157E-02 3.615E-02 3.894E-03 1.784E-02 1.163E-02 1.145E-02 4.240E-02 4.687E-02 2.771E-02
† † † † † † †
mean 4.077E-02 3.431E-01 6.767E-02 6.177E-02 6.241E-02 5.393E-02 7.606E-02 4.389E-02 4.205E-02† 3.230E-02‡ 4.787E-02†
UF4
std 7.424E-04 4.382E-03 2.849E-03 3.588E-03 4.784E-03 2.545E-03 1.370E-02 2.477E-03 1.776E-03 7.427E-04 1.943E-03
† † † † † † †
mean 1.446E-01 3.026E-01 2.901E-01 3.674E-01 3.236E-01 3.429E-01 6.793E-01 2.851E-01 3.161E-01† 1.842E-01† 6.398E-01†
UF5
std 2.713E-02 3.446E-02 4.636E-02 1.476E-01 9.074E-02 1.004E-01 1.006E-01 9.974E-02 1.060E-01 3.768E-02 3.972E-01
† † † † † † †
mean 6.140E-02 2.731E-01 1.868E-01 3.792E-01 1.108E-01 2.960E-01 3.207E-01 2.993E-01 4.145E-01† 2.851E-01† 3.204E-01†
UF6
std 3.085E-02 1.451E-01 1.361E-01 2.329E-01 7.349E-02 2.346E-01 7.719E-02 1.567E-01 1.796E-01 1.018E-01 1.926E-01
† † † † † † †
mean 2.634E-03 3.514E-01 4.067E-03 6.088E-03 2.722E-03 3.858E-02 3.504E-01 2.279E-01 2.678E-01† 7.735E-02† 4.009E-02†
UF7
std 1.441E-04 4.038E-03 9.467E-04 4.618E-03 4.302E-04 6.747E-03 8.797E-03 1.549E-01 1.724E-01 1.276E-01 4.461E-02
‡ † ‡ † † † †
mean 5.846E-02 5.359E-01 6.213E-02 2.401E-02 6.593E-02 1.902E-01 2.671E-01 4.420E-01 3.143E-01† 3.706E-01† 1.693E-01†
UF8
std 1.259E-02 5.586E-02 7.577E-03 7.142E-04 6.511E-03 4.524E-03 5.537E-02 2.371E-04 5.935E-02 1.186E-01 2.531E-02
† † † † † † †
mean 4.530E-02 3.653E-01 6.111E-02 9.091E-02 1.068E-01 2.344E-01 1.840E-01 2.000E-01 1.608E-01† 2.200E-01† 3.567E-01†
UF9
std 3.002E-02 3.942E-02 3.914E-02 5.620E-02 5.281E-02 3.507E-02 7.033E-02 6.355E-02 5.736E-02 6.541E-02 4.588E-02
‡ ‡ ‡ ‡ ‡ ‡ ‡
mean 9.121E-01 4.066E-01 4.971E-01 5.632E-01 7.763E-01 2.438E-01 6.630E-01 5.189E-01 3.310E-01‡ 4.080E-01‡ 1.518E+00†
UF10
std 1.542E-01 6.183E-02 4.517E-02 1.081E-01 1.240E-01 1.205E-01 6.928E-02 6.460E-02 1.512E-01 1.087E-01 3.549E-01
IH
† † † † † † †
mean 3.663E+00 3.594E+00 3.650E+00 3.643E+00 3.656E+00 3.442E+00 3.386E+00 3.311E+00 3.317E+00† 3.440E+00† 3.493E+00†
UF1
std 1.000E-04 1.711E-02 7.779E-03 1.090E-02 2.100E-03 9.338E-02 1.784E-02 1.024E-01 1.089E-01 1.389E-01 9.189E-02
† † † † † † †
mean 3.650E+00 3.583E+00 3.607E+00 3.553E+00 3.639E+00 3.462E+00 3.609E+00 3.510E+00 3.500E+00† 3.601E+00† 3.622E+00†
UF2
std 1.500E-02 2.500E-02 4.059E-02 1.046E-01 1.512E-02 8.993E-02 7.740E-03 6.899E-02 6.053E-02 3.516E-02 1.397E-02
† † † † †
mean 3.630E+00 3.450E+00 3.549E+00 3.357E+00 3.654E+00 2.571E+00 3.519E+00 2.559E+00 2.590E+00† 3.137E+00† 3.559E+00†
UF3
std 3.900E-02 6.599E-02 1.111E-01 1.929E-01 7.162E-03 4.653E-02 4.087E-02 3.747E-02 1.052E-01 1.815E-01 4.195E-02
† † † † † † †
mean 3.231E+00 2.877E+00 3.135E+00 3.124E+00 3.152E+00 3.188E+00 3.079E+00 3.217E+00 3.213E+00† 3.241E+00 3.201E+00†
UF4
std 4.870E-03 1.433E-02 1.674E-02 2.155E-02 1.978E-02 6.537E-03 1.368E-01 6.390E-03 3.752E-03 3.564E-03 5.123E-03
† † † † † † †
mean 3.165E+00 2.834E+00 2.581E+00 2.362E+00 2.566E+00 2.587E+00 1.590E+00 2.514E+00 2.397E+00† 2.827E+00† 1.935E+00†
UF5
std 1.000E-01 1.197E-01 1.649E-01 2.461E-01 3.017E-01 3.366E-01 2.378E-01 2.156E-01 2.895E-01 2.009E-01 8.799E-01
† † † † † † †
mean 3.231E+00 2.761E+00 2.906E+00 2.526E+00 3.067E+00 2.665E+00 2.623E+00 2.702E+00 2.390E+00† 2.617E+00† 2.534E+00†
UF6
std 7.290E-02 2.919E-01 2.469E-01 4.142E-01 2.546E-01 4.582E-01 1.522E-01 3.278E-01 3.686E-01 3.202E-01 5.677E-01
† † † † † † †
mean 3.494E+00 3.001E+00 3.480E+00 3.455E+00 3.484E+00 3.403E+00 2.545E+00 2.854E+00 2.784E+00 3.305E+00 3.403E+00
UF7
std 2.580E-03 2.074E-03 8.107E-03 5.177E-02 8.424E-03 6.379E-02 9.270E-03 4.184E-01 4.475E-01 3.632E-01 1.626E-01
† † ‡ † † † †
mean 7.323E+00 6.338E+00 7.273E+00 7.363E+00 7.297E+00 6.399E+00 6.373E+00 6.424E+00 6.476E+00 6.527E+00 6.657E+00
UF8
std 3.373E-02 4.857E-01 2.209E-02 8.799E-03 1.372E-02 1.597E-02 2.263E-01 3.303E-04 1.456E-01 2.338E-01 2.619E-01
† † † † † † †
mean 7.653E+00 7.061E+00 7.419E+00 7.309E+00 7.312E+00 5.819E+00 7.085E+00 6.494E+00 6.761E+00† 6.989E+00† 5.797E+00†
UF9
std 1.417E-01 2.444E-01 1.513E-01 2.459E-01 2.389E-01 2.444E-01 2.235E-01 3.205E-01 2.346E-01 3.053E-01 3.945E-01
‡ ‡ ‡ ‡ ‡ ‡ ‡
mean 2.185E+00 4.144E+00 3.347E+00 3.406E+00 2.425E+00 5.874E+00 2.510E+00 5.904E+00 5.216E+00‡ 4.099E+00‡ 6.961E-01†
UF10
std 6.270E-01 4.462E-01 2.021E-01 3.266E-01 4.125E-01 9.841E-01 3.098E-01 7.103E-01 1.223E+00 1.013E+00 7.549E-01

Wilcoxon’s rank sum test at a 0.05 significance level is performed between CDG-MOEA and each of the other competing algorithms. † and ‡ denotes that the performance of the
corresponding algorithm is significantly worse than or better than that of CDG-MOEA, respectively. The best mean is highlighted in boldface.

B. Comparisons on GLT test suite rithms is very similar, except for MOEA/D-DE (WS), NSGA-
II and IBEA. The nondominated set obtained by MOEA/D-DE
To verify the robustness of CDG-MOEA on MOPs with
(WS) is degenerated to two small regions on the corners of
different shapes of PFs, CDG-MOEA is compared with ten
PF. This is because the weighted sum approach is unable to
algorithms on GLT test suite in this section.
approximate the nonconvex parts of PF. For GLT2 whose PF
To visualize the performance of all the compared algorithms has disparately scaled objectives and/or GLT6 whose PFs are
on MOPs with different characteristics, the final nondominated extremely convex, only CDG-MOEA is able to approximate
sets obtained by all the compared algorithms in the run the widely and uniformly distributed nondominated solution
with the median IGD value on GLT1, GLT2 and GLT6 are set. These results verify that, unlike other algorithms, CDG-
presented in Fig. 5-7. For GLT1 whose PF is two segments MOEA is able to remain effective on MOPs with complex PFs
of disconnected straight lines, the performance of all the algo-

1089-778X (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

TABLE III: Performance comparisons of CDG-MOEA with ten MOEAs on GLT problems in terms of the mean and standard
deviation values of IGD and IH .
MOEA/D-DE
CDG-MOEA MOEA/D-ACD MSOPS-II NSGA-II IBEA -MOEA Borg OMOPSO
instance
WS TCH PBI
IGD
† ‡ ‡ ‡ † † †
mean 1.748E-03 1.771E-01 1.253E-03 1.430E-03 1.216E-03 1.780E-02 1.121E-01 1.012E-01 1.570E-01† 9.084E-02† 4.831E-02†
GLT1
std 1.629E-04 9.065E-14 2.941E-06 3.824E-06 9.914E-06 6.084E-03 3.738E-02 2.479E-02 2.494E-02 5.252E-02 3.463E-02
† † † † † † †
mean 1.158E-02 4.522E-02 1.205E-01 2.429E+00 1.354E-01 6.773E-02 1.087E-01 3.040E-01 4.712E-01† 1.926E-01† 8.558E-02†
GLT2
std 1.492E-03 1.350E-03 1.947E-02 4.209E-04 1.980E-02 1.886E-02 1.359E-01 8.396E-02 3.242E-01 1.600E-01 2.885E-03
† † † † † † †
mean 1.834E-03 2.374E-01 6.537E-03 4.535E-01 6.536E-03 4.743E-02 6.639E-02 1.506E-01 8.166E-02† 1.047E-01† 5.554E-02†
GLT3
std 4.287E-05 1.383E-06 1.535E-04 1.932E-04 4.719E-04 1.598E-02 2.846E-02 6.907E-02 2.382E-02 2.081E-02 1.833E-03
† † † † † † †
mean 2.923E-03 2.724E-01 3.383E-03 9.696E-02 3.246E-03 1.276E-02 1.444E-01 2.316E-01 3.251E-01† 2.515E-01† 1.229E-01†
GLT4
std 6.339E-05 6.933E-02 8.270E-06 2.594E-02 6.589E-05 6.010E-03 4.927E-02 4.598E-03 1.371E-01 1.121E-01 6.068E-02
† † † † † † †
mean 2.016E-02 2.904E-02 6.839E-02 1.246E-01 3.887E-02 1.306E-01 3.115E-02 6.964E-02 5.025E-02† 3.112E-02† 2.677E-02†
GLT5
std 1.524E-03 5.505E-04 9.744E-04 4.637E-03 4.485E-03 7.923E-02 3.193E-03 7.197E-03 2.588E-02 7.382E-03 2.548E-03
† † † † † † †
mean 1.882E-02 6.799E-02 4.115E-02 1.785E-01 2.398E-02 1.148E-01 3.207E-02 6.005E-02 3.927E-02† 3.724E-02† 3.059E-02†
GLT6
std 5.105E-03 6.283E-02 8.534E-04 2.456E-02 2.688E-03 5.480E-02 4.172E-03 7.718E-03 7.995E-03 4.213E-03 4.620E-03
IH
† ‡ ‡ ‡ † † †
mean 8.128E-01 4.400E-01 8.141E-01 8.140E-01 8.133E-01 6.599E-01 5.663E-01 5.542E-01 4.777E-01† 5.895E-01† 6.834E-01†
GLT1
std 3.679E-03 4.168E-15 1.142E-05 7.386E-06 7.031E-04 4.850E-02 5.928E-02 4.832E-02 5.546E-02 7.306E-02 4.206E-02
† † † † † † †
mean 1.223E+01 1.218E+01 1.214E+01 8.635E+00 1.213E+01 1.201E+01 1.197E+01 1.162E+01 1.105E+01† 1.199E+01† 1.212E+01†
GLT2
std 2.245E-04 1.510E-03 1.779E-02 1.061E-03 1.458E-02 7.229E-02 4.109E-01 2.146E-01 8.633E-01 3.071E-01 4.250E-03
† † † † † † †
mean 1.389E+00 1.342E+00 1.388E+00 1.325E+00 1.388E+00 1.374E+00 1.376E+00 1.359E+00 1.373E+00† 1.367E+00† 1.376E+00†
GLT3
std 6.441E-04 2.772E-07 1.242E-04 2.342E-05 2.815E-04 3.875E-03 5.631E-03 7.815E-03 4.673E-03 4.209E-03 4.490E-04
† ‡ † †
mean 1.635E+00 1.370E+00 1.635E+00 1.551E+00 1.413E+00 1.490E+00 1.532E+00 † 1.475E+00 †
1.137E+00† 1.297E+00† 1.517E+00†
GLT4
std 9.133E-05 2.235E-01 2.908E-05 2.803E-02 6.374E-02 6.773E-02 3.496E-02 /; 5.266E-03 3.643E-01 3.153E-01 1.129E-01
† † † † †
mean 1.690E+00 1.691E+00 1.686E+00 1.590E+00 1.692E+00 1.667E+00 1.688E+00 1.648E+00 1.685E+00 1.680E+00† 1.688E+00†
GLT5
std 2.481E-03 1.445E-04 2.140E-04 2.471E-03 2.442E-04 3.610E-02 2.065E-03 6.947E-03 1.080E-02 2.248E-03 1.372E-03
† † † † † †
mean 1.686E+00 1.672E+00 1.680E+00 1.579E+00 1.684E+00 1.671E+00 1.675E+00 1.621E+00 1.678E+00† 1.664E+00† 1.673E+00†
GLT6
std 1.281E-03 3.448E-02 1.167E-03 6.630E-02 1.166E-03 2.059E-02 1.064E-03 1.140E-02 3.506E-03 1.585E-03 1.826E-03

Wilcoxon’s rank sum test at a 0.05 significance level is performed between CDG-MOEA and the other competing algorithms. † and ‡ denotes that the performance of the
corresponding algorithm is significantly worse than or better than that of CDG-MOEA, respectively. The best mean is highlighted in boldface.

(extremely convex PFs or with disparately scaled objectives). MSOPS-II, NSGA-II, IBEA, -MOEA, Borg and OMOPSO)
Table III shows the performance of the eleven compared during the evolutionary process, in terms of the average IGD
algorithms in terms of IGD and IH on GLT1-6. It can values over 30 runs, is illustrated in Fig. 8 on UF and GLT
be observed that CDG-MOEA performs significantly better problems. It can be observed that CDG-MOEA has the best
than other compared algorithms on all test problems except performance in terms of both the convergence speed and the
for GLT1 in terms of IGD. MOEA/D-ACD achieves the quality of the final nondominated sets on UF2, GLT2-3 and
best performance on GLT1, although the final nondominated GLT5. On UF6, CDG-MOEA converges more slowly at the
solutions obtained by MOEA/D-ACD and CDG-MOEA are early stage, but it performs increasingly better and outperforms
both uniformly distributed on the PF of GLT1, as shown in all the compared algorithms at the final stage. It is worth
Fig. 5. noting that the performance of MOEA/D-DE (PBI), in terms
In terms of IH , CDG-MOEA has the best performance on of IGD values, becomes even increasingly worse during the
GLT2-3 and MOEA/D-DE (TCH) has the best performance on optimization process on GLT2-3. This phenomenon can be
GLT1 with statistical significance. The performance of CDG- verified by the final nondominated sets obtain by MOEA/D-
MOEA is very similar to that of MOEA/D-DE (TCH) on DE (PBI) on GLT2 (Fig. 6 (d)), where only half of the PF can
GLT4; and the performance of CDG-MOEA is very similar to be approximated by MOEA/D-DE (PBI) on GLT2.
that of MOEA/D-ACD on GLT5-6, although it can be observed
in Fig. 7 that the solution set obtained by CDG-MOEA is
more uniformly distributed than that obtained by MOEA/D- VII. S ENSITIVITY A NALYSIS
DE (TCH) or MOEA/D-ACD on GLT6. This can be explained
by the fact that the solutions on the extremely convex regions In this section, we investigate the sensitivity of the control
of PFs have very little contribution on the value of IH . parameters in CDG-MOEA, including the population size N ,
the division parameter K, the grid neighborhood distance T
C. Convergence Plots and the probability δ for the mating restriction, and the control
The performance of eleven compared algorithms (CDG- parameters F and CR in the DE reproduction operator, on
MOEA, MOEA/D-DE (WS, TCH, PBI), MOEA/D-ACD, both bi- and tri-objective UF and GLT test problems.

1089-778X (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

10

GLT1 GLT1 GLT1 GLT1

1 1 1 1

0.8 0.8 0.8 0.8

f2

f2

f2

f2
0.6 0.6 0.6 0.6

0.4 0.4 0.4 0.4

0.2 0.2 0.2 0.2

0 0 0 0
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2

f1 f1 f1 f1

(a) CDG-MOEA (b) MOEA/D-DE (WS) (c) MOEA/D-DE (TCH) (d) MOEA/D-DE (PBI)

GLT1 GLT1 GLT1 GLT1

1 1 1 1

0.8 0.8 0.8 0.8

f2

f2

f2
f2
0.6 0.6 0.6 0.6

0.4 0.4 0.4 0.4

0.2 0.2 0.2 0.2

0 0 0 0
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2

f1 f1 f1 f1

(e) MOEA/D-ACD (f) MSOPS-II (g) NSGA-II (h) IBEA

GLT1 GLT1 GLT1

1 1 1

0.8 0.8 0.8

f2

f2

f2
0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2

f1 f1 f1

(i) -MOEA (j) the Borg MOEA (k) OMOPSO

Fig. 5: The final nondominated solution set obtained by eleven algorithms in the run with the median IGD value on GLT1.

A. Sensitivity to Population Size N and Division Parameter positively correlated for achieving the best performance for
K CDG-MOEA on both bi- and tri-objective problems. When N
For CDG, as described in Section III, there are a total = θK m−1 , CDG-MOEA can achieve the best performance.
number of mK m−1 subproblems and each solution can be
at most the optimal solution for m subproblems, where m is
the number of objectives. Although there is a complex inter- B. Sensitivity to Division Parameter K and Grid Neighbor-
dependence between the population size N and the division hood Distance T
parameter K, their relation for a MOP with a non-degenerate As the division parameter K and the grid neighborhood dis-
PF, can be analyzed as follows. tance T for the mating restriction also play an interdependent
role on the performance of CDG-MOEA, we set combinations
αmK m−1
N= = θK m−1 (16) of different values of K (100-260 for bi-objective UF1 and
β GLT3; 10-50 for tri-objective UF9 and GLT5) and T (1-10
where α is the optimal average number of solutions for a for bi-objective UF1 and GLT3; 1-5 for tri-objective UF9 and
subproblem; β is a coefficient depending on the shape of GLT5) in the experiments. The population size is set to 600
PF (e.g. the convexity or disconnection) and θ is the final for UF1, 9 and 300 for GLT3, 5. The other parameters are
coefficient after the simplification. the same as those in Section V-B. Fig. 10 shows the mean
As the population size N and the division parameter K play IGD values of the populations obtained by CDG-MOEA
an interdependent role on the performance of CDG-MOEA, with different division parameters K and grid neighborhood
we set combinations of different values of N (200-500 for bi- distance T on UF1, UF9, GLT3 and GLT5.
objective UF1 and GLT3; 300-1000 for tri-objective UF9 and Fig. 10 shows that CDG-MOEA with different K and T
GLT5) and K (100-260 for bi-objective UF1 and GLT3; 10- values may have the different performance. For UF problems,
50 for tri-objective UF9 and GLT5) in the experiments. The it can be observed that when the value of K increases, the
other parameters are the same as those in Section V-B. Fig. 9 value of T also needs to increase for achieving better perfor-
shows the mean IGD values of the populations obtained by mance. For GLT problems, CDG-MOEA is very sensitive to K
CDG-MOEA with different population sizes N and division and a suitable K value is needed for optimal performance, as
parameters K on UF1, UF9, GLT3 and GLT5. discussed in the last section. However, given a fixed K value,
Eq. 16 can be verified from Fig. 9 that N and K are CDG-MOEA is robust with regard to T on GLT problems.

1089-778X (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

11

GLT2 GLT2 GLT2 GLT2

12 12 12 12

10 10 10 10

8 8 8 8
f2

f2

f2
f2
6 6 6 6

4 4 4 4

2 2 2 2

0 0 0 0
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2

f1 f1 f1 f1

(a) CDG-MOEA (b) MOEA/D-DE (WS) (c) MOEA/D-DE (TCH) (d) MOEA/D-DE (PBI)

GLT2 GLT2 GLT2 GLT2

12 12 12 12

10 10 10 10

8 8 8 8
f2

f2

f2

f2
6 6 6 6

4 4 4 4

2 2 2 2

0 0 0 0
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2

f1 f1 f1 f1

(e) MOEA/D-ACD (f) MSOPS-II (g) NSGA-II (h) IBEA

GLT2 GLT2 GLT2

12 12 12

10 10 10

8 8 8
f2

f2
f2

6 6 6

4 4 4

2 2 2

0 0 0
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2

f1 f1 f1

(i) -MOEA (j) the Borg MOEA (k) OMOPSO

Fig. 6: The final nondominated solution set obtained by eleven algorithms in the run with the median IGD value on GLT2.
GLT6 GLT6 GLT6 GLT6

1.2 1.2 1.2 1.2

1 1 1 1

0.8 0.8 0.8 0.8

f3

f3

f3
f3

0.6 0.6 0.6 0.6

0.4 0.4 0.4 0.4

0 0 0 0
0.2 0.2 0.2 0.2
0.2 0.2 0.2 0.2
0 0.4 0 0.4 0 0.4 0 0.4
0 0 0 0
0.6 0.6 0.6 0.6
0.2 0.2 0.2 0.2
0.4 0.8 0.4 0.8 0.4 0.8 0.4 0.8
0.6 0.6 0.6 0.6
0.8 1 0.8 1 0.8 1 0.8 1
1 1 1 1
1.2 1.2 1.2 1.2
1.2
f1 1.2
f1 1.2
f1 1.2
f1
f2 f2 f2 f2

(a) CDG-MOEA (b) MOEA/D-DE (WS) (c) MOEA/D-DE (TCH) (d) MOEA/D-DE (PBI)

GLT6 GLT6 GLT6 GLT6

1.2 1.2 1.2 1.2

1 1 1 1

0.8 0.8 0.8 0.8

f3

f3

f3
f3

0.6 0.6 0.6 0.6

0.4 0.4 0.4 0.4

0 0 0 0
0.2 0.2 0.2 0.2
0.2 0.2 0.2 0.2
0 0.4 0 0.4 0 0.4 0 0.4
0 0 0 0
0.6 0.6 0.6 0.6
0.2 0.2 0.2 0.2
0.4 0.8 0.4 0.8 0.4 0.8 0.4 0.8
0.6 0.6 0.6 0.6
0.8 1 0.8 1 0.8 1 0.8 1
1 1 1 1
1.2 1.2 1.2 1.2
1.2
f1 1.2
f1 1.2
f1 1.2
f1
f2 f2 f2 f2

(e) MOEA/D-ACD (f) MSOPS-II (g) NSGA-II (h) IBEA

GLT6 GLT6 GLT6

1.2 1.2 1.2

1 1 1

0.8 0.8 0.8

f3

f3
f3

0.6 0.6 0.6

0.4 0.4 0.4

0 0 0
0.2 0.2 0.2
0.2 0.2 0.2
0 0.4 0 0.4 0 0.4
0 0 0
0.6 0.6 0.6
0.2 0.2 0.2
0.4 0.8 0.4 0.8 0.4 0.8
0.6 0.6 0.6
0.8 1 0.8 1 0.8 1
1 1 1
1.2 1.2 1.2
1.2
f1 1.2
f1 1.2
f1
f2 f2 f2

(i) -MOEA (j) the Borg MOEA (k) OMOPSO

Fig. 7: The final nondominated solution set obtained by eleven algorithms in the run with the median IGD value on GLT6.

1089-778X (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

12

0
UF2 1
UF6 0.015
10 10

0.2

IGD Metric Values

IGD Metric Values

0.01

IGD

IGD
−1 0
0.005 0.1
10 10
50
0 40
260 0
220 30
2 180
4 1 2 20
−2
10
−1
10 6 140 3
8 100 4 5 10 K
10 K
T T

−3
10
0 0.5 1 1.5 2 2.5 3
−2
10
0 0.5 1 1.5 2 2.5 3
(a) UF1 (b) UF9
Number of Function Evaluations (×105) Number of Function Evaluations (×105)

(a) UF2 (b) UF6

−3
x 10

1
GLT2 0
GLT3 6
10 10

4 0.05
IGD Metric Values

IGD Metric Values

IGD

IGD
0 −1
2
10 10
50
0 40
260 0
220 30
2 180
4 1 2 20
−1
10
−2
10 6 140 3
8 100 4 5 10 K
10 K
T T

−2
10
0 0.5 1 1.5 2 2.5 3
−3
10
0 0.5 1 1.5 2 2.5 3
(c) GLT3 (d) GLT5
Number of Function Evaluations (×105) Number of Function Evaluations (×105)

(c) GLT2 (d) GLT3

Fig. 10: The mean IGD values obtained by CDG-MOEA with
different combinations of K and T on UF1, UF9, GLT3 and
0
10
GLT5 GLT5 over 30 runs.
CDG−MOEA
IGD Metric Values

MOEA/D−DE (WS)
MOEA/D−DE (TCH)
MOEA/D−DE (PBI)
−1
MOEA/D−ACD 1 1
10 10 10
MSOPS−II δ=0.5 δ=0.5
NSGA−II δ=0.6 δ=0.6
IBEA 0 δ=0.7 δ=0.7
10
OMOPSO δ=0.8 0 δ=0.8
10
EMOEA δ=0.9 δ=0.9
−2
10
IGD

IGD
0 0.5 1 1.5 2 2.5 3 BORG 10
−1

Number of Function Evaluations (×105)

−1
10
(e) GLT5 (f) legends of eleven compared al- 10
−2

gorithms
−3 −2
10 10
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Fig. 8: The mean IGD values vs. the number of function Number of Function Evaluations (×105) Number of Function Evaluations (×105)

evaluations obtained by eleven algorithms over 30 runs. (a) UF1 (b) UF9
0 0
10 10
δ=0.5 δ=0.5
δ=0.6 δ=0.6
δ=0.7 δ=0.7
−1 δ=0.8 δ=0.8
10
δ=0.9 δ=0.9
IGD

IGD

−1
10

−2
10

−3
x 10
−3 −2
6 0.4 10 10
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
0.3 Number of Function Evaluations (×105) Number of Function Evaluations (×105)
4
IGD
IGD

0.2 (c) GLT3 (d) GLT5

2
0.1

0 100 0 10
Fig. 11: The mean IGD values vs. the numbers of function
200 140 400 20
180
300
400
500 260
220 600
800
1000 50
40
30
evaluations obtained by CDG-MOEA with different δ over 30
K K
N N
runs on UF1, UF9, GLT3 and GLT5.
(a) UF1 (b) UF9

0.015 0.06
C. Sensitivity to Probability of Mating Restriction
0.01 0.04
To investigate the sensitivity of the probability for the
IGD
IGD

0.005 0.02 mating restriction δ, CDG-MOEA with δ = 0.5, 0.6, 0.7, 0.8
0
200 140
100 0
400 20
10 and 0.9 is separately tested. The remaining parameters are the
300 180 600 30
400
500 260
220
K
800
1000 50
40
K
same as those in Section V-B. Fig. 11 shows the evolution of
N N
the mean IGD values of the populations obtained by CDG-
(c) GLT3 (d) GLT5
MOEA with different δ values on bi-objective UF1, GLT3 and
Fig. 9: The mean IGD values obtained by CDG-MOEA with tri-objective UF9, GLT5. It can be observed that CDG-MOEA
different combinations of N and K on UF1, UF9, GLT3 and with different δ achieves very similar performance, in terms
GLT5 over 30 runs. of the final IGD values, which indicates that CDG-MOEA is
not sensitive to δ. However, on UF1 and UF9, CDG-MOEA
converges faster with larger δ values (δ = 0.8, 0.9).

1089-778X (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

13

IX. C ONCLUSION

0.015 0.2 This paper proposed a novel constrained decomposition

0.01
0.15 approach with grids (CDG) to better fit the decomposition-
IGD

IGD
0.005
0.1
based MOEA framework. The proposed CDG-MOEA is com-
0.05

0 0
pared with seven classical or state-of-the-art MOEAs on two
0.2 0.1 0.2 0.1
0.3
0.4
0.6
0.8
1.0 0.9
0.7
0.5 0.4
0.6
0.8 0.9
0.7
0.5
0.3
sets of test suites. The experimental results show that CDG-
F 1.0
F
CR CR MOEA outperforms the compared algorithms in most test
(a) UF1 (b) UF9 problems. More importantly, CDG-MOEA is very robust with
the shapes of PFs and can remain effective on MOPs with
x 10
−3 complex PFs (e.g., extremely convex or with disparately scaled
3 0.08 objectives). The sensitivity analysis of the parameters are also
0.06
2 conducted in this paper. Further studies include the extension
IGD

IGD

0.04
1
0.02
of CDG-MOEA for combinatorial multiobjective optimization
0
0.2 0.1
0
0.2 0.1
problems and many-objective optimization problems.
0.4 0.3 0.4 0.3
0.6 0.5 0.6 0.5
0.8 0.7 0.8 0.7
0.9 1.0 0.9
1.0
F F
CR CR

(c) GLT3 (d) GLT5 R EFERENCES

Fig. 12: The mean IGD values obtained by CDG-MOEA with [1] M. Asafuddoula, T. Ray, and R. A. Sarker, “A decomposition-based evo-
different combinations of F and CR on UF1, UF9, GLT3 and lutionary algorithm for many objective optimization,” IEEE Transaction
on Evolutionary Computation, vol. 19, no. 3, pp. 445–460, 2015.
GLT5 over 30 runs. [2] J. Bader and E. Zitzler, “HypE: An algorithm for fast hypervolume-based
many-objective optimization,” Evolutionary Computation, vol. 19, no. 1,
pp. 45–76, 2011.
[3] M. Basseur and E. Zitzler, “Handling uncertainty in indicator-based
D. Sensitivity to F and CR multiobjective optimization,” International Journal of Computational
The sensitivity of the control parameters F and CR in the Intelligence Research, vol. 2, no. 3, pp. 255–272, 2006.
DE reproduction operator are tested in this section. We set a [4] N. Beume, B. Naujoks, and M. Emmerich, “SMS-EMOA: Multiobjec-
tive selection based on dominated hypervolume,” European Journal of
total of 25 combinations of five F values (0.1, 0.3, 0.5, 0.7 Operational Research, vol. 181, no. 3, pp. 1653–1669, 2007.
and 0.9) and five CR values (0.2, 0.4, 0.6, 0.8 and 1) on [5] X. Cai, Y. Li, Z. Fan, and Q. Zhang, “An external archive guided mul-
UF1, UF9, GLT3 and GLT5. All the other parameters remain tiobjective evolutionary algorithm based on decomposition for combi-
natorial optimization,” IEEE Transaction on Evolutionary Computation,
the same as those in Section V-B. The mean IGD values vol. 19, no. 4, pp. 508–523, 2015.
of final populations obtained by CDG-MOEA with different [6] X. Cai, Z. Yang, Z. Fan, and Q. Zhang, “Decomposition-based-sorting
combinations of F and CR on UF1, UF9 and GLT5 are shown and angle-based-selection for evolutionary multiobjective and many-
objective optimization,” IEEE Transactions on Cybernetics, 2016.
in Fig. 12. It can be observed that when both F and CR have [7] C. A. C. Coello, G. T. Pulido, and M. S. Lechuga, “Handling multiple
relatively large values (e.g., F = 0.5, 0.7 and CR = 0.8, 1), objectives with particle swarm optimization,” IEEE Transactions on
CDG-MOEA has better performance. Evolutionary Computation, vol. 8, no. 3, pp. 256–279, 2004.
[8] I. Das and J. E. Dennis, “Normal-boundary intersection: A new method
for generating the pareto surface in nonlinear multicriteria optimization
VIII. M ORE DISCUSSIONS OF CDG-MOEA ON problems,” SIAM Journal on Optimization, vol. 8, no. 3, pp. 631–657,
MANY- OBJECTIVE OPTIMIZATION 1998.
[9] K. Deb, Multi-Objective Optimization using Evolutionary Algorithms.
CDG-MOEA can be further extended to many-objective Chichester, UK: John Wiley & Sons, 2001.
optimization problems if the following two issues can be well- [10] K. Deb and H. Jain, “An evolutionary many-objective optimization algo-
rithm using reference-point-based nondominated sorting approach, part
addressed. i: Solving problems with box constraints,” Evolutionary Computation,
First, the setup of the grid-system in CDG is based on the IEEE Transactions on, vol. 18, no. 4, pp. 577–601, 2014.
approximation of both ideal and nadir point. In the current [11] K. Deb, M. Mohan, and S. Mishra, “A fast multi-objective evolutionary
algorithm for finding well-spread pareto-optimal solutions,” KanGAL
CDG-MOEA, Pareto-domination is used to approximate the report, vol. 2003002, pp. 1–18, 2003.
nadir point. However, it is well-known that the selection pres- [12] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist
sure of Pareto-domination becomes weaker with the increasing multiobjective genetic algorithm: NSGA-II,” IEEE Transactions on
Evolutionary Computation, vol. 6, no. 2, pp. 182–197, 2002.
number of objectives [22], [10], [25]. Therefore, to maintain [13] J. J. Durillo, A. J. Nebro, C. A. C. Coello, F. Luna, and E. Alba,
the convergence for CDG-MOEA/D, other methods are needed “A comparative study of the effect of parameter scalability in multi-
for the better approximation of the nadir point. objective metaheuristics,” in Evolutionary Computation, 2008. CEC
Second, as discussed in Section VII-A, when N = θK m−1 , 2008. IEEE Congress on. IEEE, 2008, pp. 1893–1900.
[14] S. B. Gee, K. C. Tan, V. A. Shim, and N. R. Pal, “Online diversity
where θ is a coefficient, N is the population size, K is the grid assessment in evolutionary multiobjective optimization: A geometrical
division parameter and m is the number of objectives, CDG- perspective,” IEEE Transaction on Evolutionary Computation, vol. 19,
MOEA can achieve the best performance. This indicates that no. 4, pp. 542–559, 2015.
[15] F. Gu, H.-L. Liu, and K. C. Tan, “A multiobjective evolutionary
the value of K (at least 2) would be far away from its optimal algorithm using dynamic weight design method,” International Journal
value (less than 1) when N is a small number with a large of Innovative Computing, Information and Control, vol. 8, no. 5B, pp.
m value. In other words, a large K value and a very large 3677–3688, 2012.
[16] D. Hadka and P. Reed, “Borg: An auto-adaptive many-objective evo-
N value are needed for CDG-MOEA on a many-objective lutionary computing framework,” Evolutionary Computation, vol. 21,
optimization problem, to achieve good performance. no. 2, pp. 231–259, 2013.

1089-778X (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

14

[17] Y. Y. Haimes, L. S. Lasdon, and D. A. Wismer, “On a bicriterion for- [39] S. Yang, M. Li, X. Liu, and J. Zheng, “A grid-based evolutionary
mulation of the problems of integrated system identifcation and system algorithm for many-objective optimization,” IEEE Transactions on Evo-
optimization,” IEEE Transactions on Systems, Man, and Cybernetics, lutionary Computation, vol. 17, no. 5, pp. 721–736, 2013.
vol. 1, pp. 296–297, 1971. [40] G. G. Yen and H. Lu, “Dynamic multiobjective evolutionary algorithm:
[18] E. J. Hughes, “Multiple single objective pareto sampling,” in Evolution- adaptive cell-based rank and density estimation,” IEEE Transactions on
ary Computation, 2003. CEC’03. The 2003 Congress on, vol. 4. IEEE, Evolutionary Computation, vol. 7, no. 3, pp. 253–274, 2003.
2003, pp. 2678–2684. [41] H. Zhang, A. Zhou, S. Song, Q. Zhang, X.-Z. Gao, and J. Zhang, “A self-
[19] ——, “MSOPS-II: A general-purpose many-objective optimiser,” in organizing multiobjective evolutionary algorithm,” IEEE Transactions
Evolutionary Computation, 2007. CEC 2007. IEEE Congress on. IEEE, on Evolutionary Computation, vol. 20, no. 5, pp. 792–806, 2016.
2007, pp. 3944–3951. [42] Q. Zhang and H. Li, “MOEA/D: A multiobjective evolutionary algorithm
[20] C. Igel, N. Hansen, and S. Roth, “Covariance matrix adaptation for based on decomposition,” IEEE Transactions on Evolutionary Compu-
multi-objective optimization,” Evolutionary Computation, vol. 15, no. 1, tation, vol. 11, no. 6, pp. 712–731, 2007.
pp. 1–28, 2007. [43] Q. Zhang, H. Li, D. Maringer, and E. Tsang, “MOEA/D with NBI-
[21] H. Ishibuchi, Y. Sakane, N. Tsukamoto, and Y. Nojima, “Adaptation style Tchebycheff approach for portfolio management,” in Evolutionary
of scalarizing functions in MOEA/D: An adaptive scalarizing function- Computation (CEC), 2010 IEEE Congress on. IEEE, 2010, pp. 1–8.
based multiobjective evolutionary algorithm,” in International Confer- [44] Q. Zhang, A. Zhou, and Y. Jin, “RM-MEDA: A regularity model-based
ence on Evolutionary Multi-Criterion Optimization. Springer, 2009, multiobjective estimation of distribution algorithm,” IEEE Transactions
pp. 438–452. on Evolutionary Computation, vol. 12, no. 1, pp. 41–63, 2008.
[45] Q. Zhang, A. Zhou, S. Zhao, P. N. Suganthan, W. Liu, and S. Tiwari,
[22] H. Ishibuchi, N. Tsukamoto, and Y. Nojima, “Evolutionary many-
“Multiobjective optimization test instances for the CEC 2009 special ses-
objective optimization: A short review,” in 2008 Congress on Evolu-
sion and competition,” University of Essex, Colchester, UK and Nanyang
tionary Computation (CEC’2008). Hong Kong: IEEE Service Center,
technological University, Singapore, special session on performance
June 2008, pp. 2424–2431.
assessment of multi-objective optimization algorithms, technical report,
[23] J. D. Knowles and D. W. Corne, “Approximating the nondominated front vol. 264, 2008.
using the Pareto archived evolution strategy,” Evolutionary Computation, [46] A. Zhou, Q. Zhang, Y. Jin, and E. Tsang, “A model-based evolutionary
vol. 8, no. 2, pp. 149–172, 2000. algorithm for bi-objective optimization,” in Evolutionary Computation,
[24] M. Laumanns, L. Thiele, K. Deb, and E. Zitzler, “Combining con- 2005. The 2005 IEEE Congress on, 2005, pp. 2568–2575 Vol. 3.
vergence and diversity in evolutionary multiobjective optimization,” [47] E. Zitzler and S. Künzli, “Indicator-based selection in multiobjective
Evolutionary Computation, vol. 10, no. 3, pp. 263–282, 2002. search,” in International Conference on Parallel Problem Solving from
[25] B. Li, J. Li, K. Tang, and X. Yao, “Many-objective evolutionary Nature. Springer, 2004, pp. 832–842.
algorithms: A survey,” ACM Computing Surveys (CSUR), vol. 48, no. 1, [48] E. Zitzler, M. Laumanns, and L. Thiele, “SPEA2: Improving the strength
p. 13, 2015. Pareto evolutionary algorithm,” 2001.
[26] H. Li and Q. Zhang, “Multiobjective optimization problems with com- [49] E. Zitzler and L. Thiele, “Multiobjective evolutionary algorithms: a com-
plicated Pareto sets, MOEA/D and NSGA-II,” IEEE Transactions on parative case study and the strength Pareto approach,” IEEE Transactions
Evolutionary Computation, vol. 13, no. 2, pp. 284–302, 2009. on Evolutionary Computation, vol. 3, no. 4, pp. 257–271, 1999.
[27] K. Li, Q. Zhang, S. Kwong, M. Li, and R. Wang, “Stable matching based
selection in evolutionary multiobjective optimization,” IEEE Transaction
on Evolutionary Computation, vol. 18, pp. 909–923, 2014.
[28] M. Li, S. Yang, and X. Liu, “Shift-based density estimation for pareto-
based algorithms in many-objective optimization,” IEEE Transactions
on Evolutionary Computation, vol. 18, no. 3, pp. 348–365, 2014.
[29] H. Liu, F. Gu, and Q. Zhang, “Decomposition of a multiobjective opti-
mization problem into a number of simple multiobjective subproblems,” Xinye Cai received the BSc in information engi-
IEEE Transaction Evolutionary Computation, vol. 18, no. 3, pp. 450– neering from Huazhong University of Science and
455, 2014. Technology, China in 2004, the MSc in electronic
[30] K. Miettinen, Nonlinear Multiobjective Optimization. Boston: Kluwer engineering from University of York, UK, in 2006
Academic Publishers, 1999. and the PhD in electrical engineering from Kansas
[31] T. Murata, H. Ishibuchi, and M. Gen, “Specification of genetic search State University, U.S, in 2009, respectively.
directions in cellular multi-objective genetic algorithms,” in Internation- He is an Associate Professor at the Department
al Conference on Evolutionary Multi-Criterion Optimization. Springer, of Computer Science and Technology, Nanjing U-
2001, pp. 82–95. niversity of Aeronautics and Astronautics (NUAA),
[32] A. J. Nebro, J. J. Durillo, F. Luna, B. Dorronsoro, and E. Alba, China. His current research interests include evolu-
“Mocell: A cellular genetic algorithm for multiobjective optimization,” tionary computation, optimization, data mining and
International Journal of Intelligent Systems, vol. 24, no. 7, pp. 726–746, their applications. He is currently leading an Intelligent Optimization Research
2009. Group in NUAA, who has won the Evolutionary Many-Objective Optimization
[33] H. Sato, “Inverted PBI in MOEA/D and its impact on the search per- Competition at the Congress of Evolutionary Computation 2017.
formance on multi and many-objective optimization,” in Proceedings of
the 2014 Annual Conference on Genetic and Evolutionary Computation.
ACM, 2014, pp. 645–652.
[34] J. D. Schaffer and J. J. Grefenstette, “Multi-objective learning via genetic
algorithms.” in Ijcai, vol. 85, 1985, pp. 593–595.
[35] M. R. Sierra and C. A. C. Coello, “Improving pso-based multi-
objective optimization using crowding, mutation and ε-dominance,” in
International Conference on Evolutionary Multi-Criterion Optimization. Zhiwei Mei received his Master degree in Computer
Springer, 2005, pp. 505–519. Science at the College of Computer Science and
[36] H. Trautmann, T. Wagner, and D. Brockhoff, “R2-EMOA: Focused mul- Technology, Nanjing University of Aeronautics and
tiobjective search using R2-indicator-based selection,” in International Astronautics, China, in 2017. Currently, he is work-
Conference on Learning and Intelligent Optimization. Springer, 2013, ing as an engineer in MeiTuan. His research interests
pp. 70–74. include evolutionary computation and multiobjective
[37] L. Wang, Q. Zhang, A. Zhou, M. Gong, and L. Jiao, “Constrained optimization.
subproblems in a decomposition-based multiobjective evolutionary algo-
rithm,” IEEE Transactions on Evolutionary Computation, vol. 20, no. 3,
pp. 475–480, 2016.
[38] Z. Wang, Q. Zhang, A. Zhou, M. Gong, and L. Jiao, “Adaptive re-
placement strategies for MOEA/D,” IEEE Transactions on Cybernetics,
vol. 46, no. 2, pp. 474–486, 2016.

1089-778X (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TEVC.2017.2744674, IEEE
Transactions on Evolutionary Computation

15

Zhun Fan is the head of the Department of Elec-

tronic Engineering and Full Professor at Shantou
University. Prior to joining Shantou University, he
was an Assistant Professor, and then an Associate
Professor at the Technical University of Denmark.
He has also been working for the BEACON Center
for Study of Evolution in Action at Michigan State
University.
Dr. Zhun Fan received his Ph.D. (Electrical and
Computer Engineering) in 2004 from the Michigan
State University. He received the B.S. degree (Con-
trol Engineering) in 1995 and M.S degree (Control Engineering) in 2000
from Huazhong University of Science and Technology, China. His major
research interests include intelligent control and robotic systems, robot vision
and cognition, MEMS, computational intelligence, design automation and
optimization of mechatronic systems etc.

Qingfu Zhang (M01-SM06) received the BSc in

mathematics from Shanxi University, China in 1984,
the MSc in applied mathematics and the PhD in in-
formation engineering from Xidian University, Chi-
na, in 1991 and 1994, respectively.
He is a Professor at the Department of Computer
Science, City University of Hong Kong, Hong Kong,
a Professor on leave from the School of Computer
Science and Electronic Engineering, University of
Essex, UK, and a Changjiang Visiting Chair Pro-
fessor in Xidian University, China. He holds two
patents and is the author of many research publications. His main research
interests include evolutionary computation, optimization, neural networks,
data analysis, and their applications. He is currently leading the Metaheuristic
Optimization Research (MOP) Group in City University of Hong Kong.
Dr. Zhang is an Associate Editor of the IEEE Transactions on Evolutionary
Computation and the IEEE Transactions on Systems, Man, and Cybernetic-
sCPart B. He is also an Editorial Board Member of three other international
journals. MOEA/D, a multobjevitve optimization algorithm developed in his
group, won the Unconstrained Multiobjective Optimization Algorithm Com-
petition at the Congress of Evolutionary Computation 2009, and was awarded
the 2010 IEEE Transactions on Evolutionary Computation Outstanding Paper
Award.

1089-778X (c) 2017 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.