IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 23, NO.

2, APRIL 2019 173

IGD Indicator-Based Evolutionary Algorithm for

Many-Objective Optimization Problems
Yanan Sun, Member, IEEE, Gary G. Yen , Fellow, IEEE, and Zhang Yi, Fellow, IEEE

Abstract—Inverted generational distance (IGD) has been I. I NTRODUCTION

widely considered as a reliable performance indicator to concur-
ANY-OBJECTIVE optimization problems (MaOPs)
rently quantify the convergence and diversity of multiobjective
and many-objective evolutionary algorithms. In this paper, an
IGD indicator-based evolutionary algorithm for solving many-
M refer to the optimization tasks involving m (i.e.,
m > 3) conflicting objectives to be optimized concurrently [1].
objective optimization problems (MaOPs) has been proposed. Generally, an MaOP is with the mathematic form repre-
Specifically, the IGD indicator is employed in each generation sented as
to select the solutions with favorable convergence and diversity. 
In addition, a computationally efficient dominance comparison y = f(x) = (f1 (x), . . . , fm (x))
(1)
method is designed to assign the rank values of solutions along s.t. x ∈ 
with three newly proposed proximity distance assignments. Based
on these two designs, the solutions are selected from a global view where  ⊆ Rn is the feasible search space for the deci-
by linear assignment mechanism to concern the convergence and sion variables x = (x1 , . . . , xn )T , and f :  →  ⊆ Rm
diversity simultaneously. In order to facilitate the accuracy of the is the corresponding objective vector including m objectives
sampled reference points for the calculation of IGD indicator, we which maps the n-dimensional decision space  to the m-
also propose an efficient decomposition-based nadir point estima-
tion method for constructing the Utopian Pareto front (PF) which dimensional objective space . Without the loss of generality,
is regarded as the best approximate PF for real-world MaOPs at f(x) is assumed to be minimized since the maximization prob-
the early stage of the evolution. To evaluate the performance, lems can be transformed into the minimization problems due
a series of experiments is performed on the proposed algo- to the duality principle. Because of the conflicting nature in
rithm against a group of selected state-of-the-art many-objective the objective functions, there is no single perfect solution for
optimization algorithms over optimization problems with 8-, 15-,
and 20-objective. Experimental results measured by the chosen f(x), but a set of tradeoff solutions which form the Pareto set
performance metrics indicate that the proposed algorithm is very (PS) in the decision space and the corresponding Pareto front
competitive in addressing MaOPs. (PF) in the objective space.
Index Terms—Inverted generational distance (IGD), lin- Optimization algorithms for addressing an MaOP aim at
ear assignment problem (LAP), many-objective evolutionary searching for a set of uniformly distributed solutions which
optimization algorithm, nadir point. are closely approximating the PF. Because the MaOPs widely
exist in diverse real-world applications, such as policy manage-
ment in land exploitation with 14-objective [2] and calibration
of automotive engine with ten-objective [3], to name a few,
various algorithms for solving MaOPs have been developed.
Among these algorithms, the evolutionary paradigms are
Manuscript received March 22, 2017; revised July 19, 2017, September considerably preferable due to their population-based meta-
18, 2017, and November 11, 2017; accepted December 28, 2017. Date of
publication January 8, 2018; date of current version March 29, 2019. This heuristic characteristics obtaining a set of quality solutions in
work was supported in part by the China Scholarship Council under Grant a single run.
201506240048, in part by the Miaozi Project in Science and Technology During the past decades, various multiobjective evolution-
Innovation Program of Sichuan Province under Grant 16-YCG061, in part
by the National Natural Science Foundation Fund of China for Distinguished ary algorithms (MOEAs), such as elitist nondominated sorting
Young Scholar under Grant 61625204, and in part by the National Natural genetic algorithm (NSGA-II) [4], advanced version of strength
Science Foundation of China under Grant 61432012 and Grant U1435213. Pareto evolutionary algorithm [5], among others, have been
(Corresponding author: Gary G. Yen.)
Y. Sun is with the College of Computer Science, Sichuan University, proposed to effectively solve multiobjective optimization prob-
Chengdu 610065, China, and also with the School of Engineering and lems (MOPs). Unfortunately, these MOEAs do not scale well
Computer Science, Victoria University of Wellington, Wellington 6140, with the increasing number of objectives, mainly due to the loss
New Zealand (e-mail: yanan.sun@ecs.vuw.ac.nz).
G. G. Yen is with the School of Electrical and Computer Engineering, of selection pressure. To be specific, the number of nondomi-
Oklahoma State University, Stillwater, OK 74078 USA (e-mail: nated solutions in MaOPs accounts for a large proportion of the
gyen@okstate.edu). current population because of the dominance resistance phe-
Z. Yi is with the College of Computer Science, Sichuan University, Chengdu
610065, China (e-mail: zhangyi@scu.edu.cn). nomenon [6] caused by the curse of dimensionality [7], so that
This paper has supplementary downloadable multimedia material available the traditional elitism mechanism based on Pareto-domination
at http://ieeexplore.ieee.org provided by the authors. cannot effectively differentiate which solutions should survive
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. into the next generation. As a result, the density-based diversity
Digital Object Identifier 10.1109/TEVC.2018.2791283 promotion mechanism is considered the sole mechanism for
1089-778X  c 2018 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.
Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.
174 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 23, NO. 2, APRIL 2019

mating and environmental selections [8]. However, the solutions of MOEA in solving MaOPs. In [28], a modification of den-
with good diversity in MaOPs are generally not only distant sity estimation, termed as shift-based density estimation, was
from each other but also away from the PF. Consequently, the proposed to make the dominance comparison better suited
evolution with the solutions generated by the activated diver- for solving MaOPs. Furthermore, the favorable convergence
sity promotion is stagnant or even far away from the PF [9]. scheme was proposed in [29] to improve the selection pres-
To this end, various many-objective evolutionary algorithms sure in mating and environmental selections. Recently, a knee
(MaOEAs) specifically designed for addressing MaOPs have point-based algorithm (KnEA) [30] was presented as a sec-
been proposed in recent years. ondary selection scheme to enhance the selection pressure.
Generally, these MaOEAs can be divided into four different In summary, these algorithms introduced new comparison
categories. The first category covers the algorithms employing methods, designed effective selection mechanisms, or relaxed
reference prior to enhancing the diversity promotion which in the original comparison approach to improve the selection
turn improve the convergence. For example, the MaOEA using pressure in addressing MaOPs.
reference-point-based nondominated sorting approach (NSGA- The fourth category is known as the indicator-based methods.
III) [10] employs a set of reference vectors to assist the For instance, several MOEAs based on the hypervolume (HV)
algorithm to select solutions which are close to these reference were proposed in [31]–[33]; however, their major disadvantages
vectors. Yuan et al. [11] proposed the reference line-based were the costly overhead in calculating the HV values especially
algorithm which not only adopted the diversity improvement in solving MaOPs. To this end, Bader and Zitzler [34] proposed
mechanism like that in NSGA-III but also introduced con- the HypE method with the Monte Carlo simulation to estimate
vergence enhancement scheme by measuring the distance the HV value. Consequently, the computational cost was largely
between the origin to the solution projections on the corre- lowered compared to its predecessors whose HV values were
sponding reference line. In addition, a reference line-based calculated exactly. In [35], an p indicator-based algorithm
estimation of distribution algorithm was introduced in [12] for (p-EMOA) was proposed for solving bi-objective optimization
explicitly promoting the diversity of an MaOEA. Furthermore, problems, and then extended further for tri-objective prob-
an approach (RVEA) was presented in [13] to adaptively lems [36]. Furthermore, Villalobos and Coello [37] integrated
revise the reference vector positions based on the scales of the the p indicator with the differential evolution [38] to solve
objective functions to balance the diversity and convergence. MaOPs with up to ten objectives. Recently, an inverse genera-
The second category refers to the decomposition-based tional distance plus (IGD+ ) [39] indicator-based evolutionary
algorithms which decompose an MaOP into several single- algorithm (IGD+ -EMOA) was proposed in [40] for address-
objective optimization problems, such as the MOEA based on ing MaOPs with no more than eight objectives. Basically, the
decomposition (MOEA/D) [14] which was initially proposed IGD+ indicator is viewed as a variant of the inverse generational
for solving MOPs but scaled well for MaOPs. Specifically, distance (IGD) indicator.
MOEA/D transformed the original MOP/MaOP with m objec- Although the MaOEAs mentioned above have experimen-
tives into a group of single-objective optimization problems, tally demonstrated their promising performance, major issues
and each subproblem was solved in its neighboring region are easily to be identified in solving real-world applications.
which constrained by their corresponding reference vec- For example, it is difficult to choose the converting strategy
tors. Recently, diverse variants [15]–[20] of MOEA/D were of the MaOEAs from the second category, which motivates
proposed for improving the performance much further. multiple variants [15], [17], [19], [20], [41], [42] to be devel-
The third category is known as the convergence oped further. In addition, the MaOEAs from the first and
enhancement-based approaches. More specifically, the tradi- third categories only highlighted one of the characters in
tional Pareto dominance comparison methods widely utilized their designs (i.e., only the diversity promotion is explicitly
in MOEAs are not effective in discriminating populations with concerned in the MaOEAs from the first category, and the
good proximity in MaOPs. A natural way is to modify this convergence from the third category). However, both diversity
comparison principle to promote the selection mechanism. For and convergence are concurrently desired by the MaOEAs. In
example, the -dominance method [21] employed a relaxed this regard, some performance indicators, which are capable of
factor  to compare the dominance relation between solutions. simultaneously measuring the diversity and convergence, such
di Pierro et al. [22] proposed the preference order rank- as the HV and IGD indicators, are preferred to be employed
ing approach to replace the traditional nondominated sorting. for designing MaOEAs. However, the major issue of HV is
Furthermore, the fuzzy dominance methods [23], [24] studied its high computational complexity. Although the Monte Carlo
the fuzzification of the Pareto-dominance relation to design the simulation has been employed to mitigate this adverse impact,
ranking scheme to select promising solutions; the L-optimality the calculation is still impracticable when the number of objec-
paradigm was proposed in [25] to pick up solutions whose tives is more than 10 [34], while the calculation of IGD is
objectives were with the same importance by considering their scalable without these deficiencies.
objective value improvements. In addition, Yang et al. [26] In this paper, an IGD indicator-based MaOEA
proposed the grid-based approach to select the solutions that (MaOEA/IGD) has been proposed for effectively addressing
have the higher priority of dominance, and control the pro- MaOPs, and the contributions are outlined as follows.
portion of Pareto-optimal solutions by adjusting the grid size. 1) A decomposition-based nadir point estimation method
Meanwhile, López et al. [27] alternated the achievement func- (DNPE) has been presented to estimate the nadir
tion and the -indicator method to improve the performance points to facilitate the calculation of IGD indicator.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.
SUN et al.: IGD INDICATOR-BASED EVOLUTIONARY ALGORITHM FOR MaOPs 175

In DNPE, the estimation focuses only on the extreme A. Nadir Point Estimation Methods
point areas and transforms the computation of an m- According to [45] and [46], the approaches for estimating
objective optimization problem into m single-objective the nadir points can be divided into three categories including
optimization problems. Therefore, less computational the surface-to-nadir, edge-to-nadir, and extreme-point-to-nadir
cost is required compared to its peer competitors (exper- schemes. In the surface-to-nadir scheme, the nadir points
iments are demonstrated in Section IV-E). are constructed from the current Pareto-optimal solutions,
2) A comparison scheme for the nondominated sorting has and updated as the corresponding algorithms evolve toward
been designed for improving the convergence of the the PF. MOEAs in [4], [18], and [47]–[49] and MaOEAs
proposed algorithm. In this scheme, the dominance rela- in [10], [13], and [15] belong to this category. However, these
tions of solutions are not obtained by the comparisons MOEAs are shown to perform poorly in MaOPs due to the
among all the solutions but the solutions to the reference curse of dimensionality [50]. In addition, the MaOEAs related
points. Therefore, the computational complexity of the methods are not suitable for the proposed algorithm because
presented comparison scheme is significantly lessened the MaOPs have been solved prior to the nadir point esti-
compared to the traditional comparison means because mation, while the nadir points in this paper are targeted for
the number of reference points is generally much less addressing MaOPs.
than that of the whole population. The edge-to-nadir scheme covers the Szczepański and
3) Three types of proximity distance assignment mecha- Wierzbicki’s approach [51], extremized crowded NSGA-II
nisms are proposed for the solutions according to their (EC-NSGA-II) [43], and the recently proposed empha-
PF rank values, which make the solutions with good sized critical region (ECR) approach [45]. Specifically,
convergence in the same PF to have higher chances Szczepański and Wierzbicki’s approach decomposed an m-
to be selected. Furthermore, these assignment mecha- objective problem into Cm 2 subproblems to estimate the nadir
nisms collectively assure the proposed IGD indicator to 2
point from the Cm edges, in which the major issues were the
be Pareto compliance. poor quality in nadir point found and the impractical compu-
4) Based on the proposed dominance comparison scheme tation complexity beyond three objectives [45]. EC-NSGA-II
and the proximity distance assignments, the selection modified the crowding distance of NSGA-II by assigning
mechanism which is employed for the mating selec- large rank values to the solutions which had the minimum
tion and the environmental selection is proposed to or maximum objective values. The ECR emphasized the solu-
concurrently facilitate the convergence and the diversity. tions lying at the edges of the PF (i.e., the critical regions)
The reminder of this paper is organized as follows. First, with the adopted MOEAs. Although EC-NSGA-II and ECR
related works are reviewed, and the motivation of the proposed have been reported to be capable of estimating the nadir
DNPE is presented in Section II. Then the details of the points in MaOPs, they required a significantly large number
proposed algorithm are documented in Section III. To evalu- of functional evaluations [45].
ate the performance of the proposed algorithm in addressing The extreme-point-to-nadir approaches refer to employ a
MaOPs, a series of experiments over scalable benchmark test direct means to estimate the extreme points based on which the
suits are performed against state-of-the-art MaOEAs, and their nadir points are derived, such as WC-NSGA-II [43] in which
results are measured by commonly chosen performance metrics the worst crowded solutions (extreme points) were preferred
and then analyzed in Section IV. In addition, the performance by ranking their crowding distances with large values. In
of the proposed MaOEA/IGD is also demonstrated by solving WC-NSGA-II, it was hopeful that the extreme points were
a real-world application, and the performance of the proposed obtained when the evolution terminated. However, emphasiz-
DNPE in nadir point estimation is investigated against its peer ing the extreme points easily led to the WC-NSGA-II losing
competitors. Finally, the proposed algorithm is concluded and the diversity which inadvertently affected the convergence in
the future works are illustrated in Section V. turn. In addition, Singh et al. [44] proposed PCSEA to look
for the nadir points with the corner-sort ranking method for
the MaOPs whose objective values were required to be with
II. R ELATED W ORKS AND M OTIVATION the identical scales.
Literatures related to the nadir point estimation and IGD In addition, there are also various methods not falling into
indicator-based EAs are thoroughly reviewed in this sec- the above categories. For example, Benayoun et al. [52] esti-
tion. Specifically, the worst crowded NSGA-II (WC-NSGA- mated the nadir points with the pay-table in which the jth row
II) [43] and the Pareto corner search evolutionary algorithm denoted the objective values of the solution which had the
(PCSEA) [44] would be reviewed and criticized in detail, minima on its jth objective. In addition, other related works
because the insightful observations of the deficiencies of were suggested in [53]–[55] for the problems assuming a lin-
these two approaches naturally lead to the motivation of the ear relationship between the objectives and variables. On the
proposed DNPE design. In addition, the IGD+ -EMOA is contrary, most of the real-world applications are nonlinear in
reviewed as well to highlight the utilization of the IGD+ indi- nature.
cator and the reference points sampling for the calculation of Because the nadir point estimation is a critical part of the
IGD in the proposed MaOEA/IGD. Please note that all the proposed algorithm for solving MaOPs, an approach with
discussions in this section are with the context of the problem a high computational complexity is certainly not preferable.
formulation in (1). Furthermore, the nadir points are employed for constructing

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.
176 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 23, NO. 2, APRIL 2019

Fig. 1. Example with bi-objective optimization problem to illustrate the ideal

point, extreme point, worst point, nadir point, and the PF. Fig. 2. Bi-objective example to illustrate the deficiency of PCSEA in
addressing the problem with different objective value scales.

the Utopian PF, while the reference point of IGD would come
from the PF. As a consequence, nadir points with a high is that over-emphasis on these solutions with maximal fit-
accuracy are not necessary a guarantee. Considering the bal- ness leads to the lack of diversity, which in turn affects the
ance between the computational complexity and the estimation convergence of the generated extreme points, i.e., the gen-
accuracy, we have proposed in this paper a DNPE method erated extreme points in WC-NSGA-II are not necessarily
by transforming an m-objective MaOP into m single-objective Pareto-optimal.
optimization problems to search for the respective m extreme PCSEA employed the corner-sorting to focus on the extreme
points. Thereafter the nadir point is derived. points during the evolution. Specifically, there were 2m
Specifically, because the proposed nadir point estimation ascended lists during the executions of PCSEA. The first
method (i.e., the DNPE) is based on the extreme-point-to-nadir m lists were about the m objectives of the solutions, while
scheme, the WC-NSGA-II and the PCSEA which are with the other m lists were about the excluded square L2 norm
the similar scheme are detailed further. For convenience of with each objective. Furthermore, the jth objective of the
reviewing the related nadir point estimation methods, multiple problem mto be optimized was with the excluded square L2
fundamental concepts of the MaOPs are given first. Then the norm f
i=1,i=j i (x) 2 . From these 2m lists, solutions with

WC-NSGA-II and PCSEA are discussed. smaller rank values which were equal to their positions in
Definition 1: Generally, there are m extreme points denoted these lists were selected until there was no available slot.
as yext 1 , . . . , ym in an m-objective optimization problem,
ext Experimental results have shown that PCSEA performs well
yext
i = f(x ext ), and xext = arg max f (x), where x ∈ PS and
i i x i in MaOPs due to the utilization of corner-sorting other than
i ∈ {1, . . . , m}. the nondominated sorting which easily leads to the loss of
Definition 2: The nadir point is defined as znad = selection pressure. However, the corner-sorting can be viewed
1 , . . . , zm }, where zi
{znad nad nad = f (xext ).
i i as to minimize the square L2 norm of all objectives, which
Definition 3: The worst point is defined as zw = deteriorates the performance of PCSEA in solving the prob-
{zw 1 , . . . , zm }, where zi = max fi (x) and x ∈ .
w w lems with different objective value scales and nonconcave PF
Definition 4: The ideal point is defined as z∗ = shapes. For example in Fig. 2 which illustrates an example of
{z1 , . . . , z∗m }, where z∗i = min fi (x) and x ∈ .
∗ bi-objective optimization problem, the arc AB denotes the PF,
Furthermore, the ideal point, extreme point, worst point, points A and B are with different values. It is clearly observed
nadir point, and the PF are plotted with a bi-objective that if the minimization of L2 norm regarding f1 (x) and f2 (x)
optimization problem in Fig. 1 for intuitively understanding are emphasized, only the extreme point A would be obtained
their significance. With these fundamental definitions, a cou- while the other one (point B) would be missed. This defi-
ple of nadir point estimation algorithms, WC-NSGA-II and ciency can also be seen in the problems with nonconcave PF
PCSEA, which are in the extreme-point-to-point scheme are shapes.
discussed as follows. Briefly, major concerns in these two nadir point esti-
WC-NSGA-II was designed based on NSGA-II by mod- mation algorithms are summarized as: 1) over-emphasizing
ifying its crowding distance assignment. According to the extreme points leads to the loss of diversity which in turn
definition of nadir point in Definition 2, WC-NSGA-II nat- deteriorates the convergence of the found nadir points and
urally emphasized the solutions with maximal objectives 2) simultaneously minimizing the objectives does not scale to
front-wise. Specifically, solutions on a particular nondom- problems with different objective value scales and nonconcave
inated front were sorted with an increasing order based PF shapes. To this end, a natural approach is recommended
on their fitness, and rank values equal to their positions by: 1) decomposing the problem to be solved into several
in the ordered list were assigned. Then the solutions with single-objective optimization problems in which the diversity
larger rank values were preferred in each generation during is not required and 2) assigning different weights to the objec-
the evolution. By this emphasis mechanism, it was hopeful tives. In the proposed DNPE, the m respective extreme points
that nadir point was obtained when the evolution of WC- are estimated by decomposing the m-objective MaOP into m
NSGA-II was terminated. However, one major deficiency single-objective problems associated with different weights.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.
SUN et al.: IGD INDICATOR-BASED EVOLUTIONARY ALGORITHM FOR MaOPs 177

y = {y1 , . . . , ym } is from the nondominated solutions of the

current population. However, this approximate approach for
generating the PF performs badly in MaOPs where multiple
local Pareto-optimal exist, which leads IGD+ -EMOA only to
solve MaOPs with no more than 8 objectives [40].
In summary, we first introduce details of the proposed
DNPE which is motivated by the insightful observations in
deficiencies of WC-NSGA-II and PCSEA. With the help of
the estimated nadir point, the Utopian PF is constructed and
the reference points are sampled for the calculation of the
proposed MaOEA/IGD. Compared to the approximation of the
Fig. 3. Example of bi-objective optimization problem, A, B are the extreme
PF in IGD+ -EMOA, our proposed MaOEA/IGD is capable
points, and C, D are the worse points. The shade region denotes the feasible of solving problems with many more objectives. Considering
objective space. the superiority of IGD+, its design principle is employed in
the proposed MaOEA/IGD when the generated solutions are
nondominated to the sampled reference points.
Specifically, the ith extreme point estimation is with the form
formulated as
III. P ROPOSED A LGORITHM

m
 2
min | fi (x)| + λ fj (x) (2) In this section, the proposed IGD indicator-based evo-
j=1,j=i lutionary algorithm for addressing MaOPs (in short for
MaOEA/IGD) is presented. To be specific, the framework
where λ is a factor with the value greater than 1 to highlight
of the proposed algorithm is outlined first. Then the details
the priority of solving its associated term. In order to better
of each step in the framework are documented. Next, the
justify our motivation, an example with bi-objective is plot-
computational complexity of the proposed algorithm is ana-
ted in Fig. 3 in which A, B are the extreme points, C, D are
lyzed. Finally, the mechanisms of promoting the diversity
the worse points, and the shaded region denotes the feasible
and the convergence in the proposed algorithm are discussed.
objective space. To obtain the extreme point on the f1 objec-
Noted here that, the proposed algorithm is described within
tive, it is required to minimize |f1 (x)| + λ(f2 (x))2 according
the context formulated by (1).
to (2). Because λ is greater than 1, the term (f2 (x))2 is opti-
mized with a higher priority. Consequently, solutions locating
in line BC are obtained, based on which |f1 (x)| is minimized A. Framework of the Proposed Algorithm
much further, then the extreme point B is obtained. Because the proposed algorithm is based on the IGD indi-
cator, a set of uniformly distributed points which is generated
B. IGD+ -EMOA from the PF is required. However, exact points are difficult to
Prior to the introduction of IGD+ -EMOA, it is neces- be obtained due to the unknown analytical form of the PF for
sary to compare the differences between the IGD and IGD+ a given real-world application. In the proposed algorithm, a
indicators. For this purpose, we first list their respective math- set of reference points, denoted by p∗ , which are evenly dis-
ematical formulations. Then the superiority of IGD+ indicator tributed in the Utopian PF is generated first (Section III-B).
is highlighted. Finally, the IGD+ -EMOA is discussed much Then the population with the predefined size is randomly ini-
further. tialized in the feasible space, and their fitness are evaluated.
Basically, the IGD indicator is with the form formulated by Next, the population begins to evolve in pursuing the optima
 until the stopping conditions are satisfied. When the evolu-
p∈p∗ dist(p, PF) tion terminates, a number of promising solutions which are
IGD = (3)
|p∗ | hopeful to uniformly distributed in the PF with good prox-
where p∗ denotes a set of reference points in the calculation imity is obtained. In order to remedy the shortage caused by
of IGD, PF denotes the nondominated solutions generated by using the p∗ as the Pareto-optimal solutions with promising
the algorithm, dist(p, PF) denotes the nearest distance from p diversity for IGD indicator, the rank of each solution as well
to solutions in PF, and the distance as its proximity distances to all the points in p∗ are assigned
 from p to the solution y (Section III-C) first during each generation of the evolution.
m
in PF is calculated by d(p, y) = j=1 (pj − yj ) . It has been
2
Then new offspring are generated from their parents selected
pointed out in [39] that IGD cannot differentiate the quality of based on the comparisons over their corresponding rank val-
generated solutions when they are nondominated to the solu- ues and proximity distances (Section III-D). Next, the fitness,
tions in p∗ , and the IGD+ indicator
 is proposed by changing ranks, and the proximity distances of the generated offspring
m
j=1 max(yj − pj , 0) .
2
the calculation of d(p, y) to to the solutions in p∗ are calculated. Finally, a limit number of
+ +
IGD -EMOA employed the IGD indicator as its selection individuals is selected from the current population to survive
mechanism. In addition, the p∗ in IGD+ -EMOA is sam- into the next generation with the operation of environmen-
pled from the approximate PF. Specifically, it supposed that tal selection (Section III-E). In summary, the details of the
the PF is obtained by solving yr1 + · · · + yrm = 1, where framework are listed in Algorithm 1.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.
178 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 23, NO. 2, APRIL 2019

Algorithm 1: Framework of the Proposed Algorithm Algorithm 3: Estimate Extreme Points

1 p
∗ ← Uniformly generate reference points for IGD Input: Optimization problem f(x) = (f1 (x), · · · , fm (x)).
indicator; Output: Extreme points {yext 1 , · · · , ym }.
ext

2 P0 ← Randomly initialize the population; 1 ϒ ← ∅;

3 Fitness evaluation on P0 ; 2 for i ← 1 to m do
4 t ← 0; 3  ← ∅;
5 while stopping criteria are not satisfied do 4 for k ← 1 to m do
6 Assign the ranks and the proximity distances for 5 if k = i then
individuals in Pt ; 6  ←  ∪ fk (x);
7 Qt ← Generate offspring from Pt ; 7 end
8 Fitness evaluation on Qt ; 8 end
i = arg minx λ  2 + |fi (x)|;
xext 2
9 Assign the rank and the proximity distance for each 9
individual in Qt ; 10 yi ← fi (xi );
ext ext

10 Pt+1 ← Environmental selection from Qt ∪ Pt ; 11 ϒ ← ϒ ∪ yext i ;

11 t ← t + 1; 12 end
12 end 13 Return ϒ.
13 Return Pt .

Algorithm 2: Uniformly Generate p∗ for IGD Indicator

Input: Optimization problem f(x) = (f1 (x), · · · , fm (x)); operator. Finally, the extreme point yext
i is obtained by line 10
the size k of p∗ . in Algorithm 3, where λ is a factor with the value greater
Output: p∗ . than 1 to highlight the weight of the corresponding term in
1 Estimate the extreme points of f(x) with Algorithm 3;
the optimization. When all the extreme points have been esti-
∗ ∗
2 z* = {z1 , · · · , zm } ← Extract the ideal point;
mated, the nadir point and the ideal point are extracted from
= {z1 , · · · , znad the extreme points based on Definitions 2 and 4, respectively.
m } ← Extract the nadir point;
3 z
nad nad
∗
4 p ← Uniformly generate k points from the constrained
This is followed by generating a set of uniformly distributed
hyperplane; reference points denoted by p∗ from the (m − 1)-dimensional
5 for i ← 1 to k do
constrained hyperplane which is contoured by m lines with the
6 for j ← to m do unit intercepts in the positive part of the quadrant. Noted here
7 (p∗ )ij = (p∗ )ij × (znad − z∗j ) + z∗j that the Das and Dennis’s [56] method, which is widely used
j
by some state-of-the-art MaOEAs, such as MOEA/D [14] and
8 end
NSGA-III [10], is employed for the generation of p∗ (line 4
9 end
∗ of Algorithm 2). Ultimately, all the points in p∗ are trans-
10 Return p .
formed into the Utopian PF, which are detailed in lines 5–9
of Algorithm 2.

B. Uniformly Generating Reference Points

In order to obtain the p∗ , the extreme points of the problem
f(x) to be optimized are calculated first. Then the ideal point
and the nadir point are extracted. Next, a set of solutions is C. Assigning Ranks and Proximity Distances
uniformly sampled from the constrained (m − 1)-dimensional When p∗ has been generated, the population is randomly
hyperplane. Finally, these solutions are transformed into the initialized in the feasible search space first, and then the fit-
Utopian PF for the usage of the IGD indicator based on the ness of individuals are evaluated. Next, the rank value and the
ideal point and the nadir point. Furthermore, these steps are proximity distances of each individual are assigned. It is noted
listed in Algorithm 2, and all the details of obtaining the p∗ here that, the rank values are used to distinguish the proximity
are illustrated as follows. of the solutions to the Utopian PF from the view of reference
To estimate the extreme points, the motivation mentioned in points, and the proximity distances are utilized to indicate
Section II-A is implemented, and the details are presented in which individuals are with better convergence and diversity
Algorithm 3. Specifically, the m extreme points are estimated in the subpopulation in which the solutions are with the same
individually based on the m objectives of the optimization rank values. More details are discussed in Section III-G.
problem. Furthermore, to estimate the ith extreme point yext i , Particularly, three rank values, denoted by r1 , r2 , and
the square L2 norm of {fk (x)|k = 1, . . . , i − 1, i + 1, . . . , m} r3 , exist in the proposed algorithm for all the individuals.
is calculated first. Then the absolute value of fi (x) is cal- Specifically, the way to rank individual s is based on the
culated. Mathematically,
 these two steps are formulated by definitions given as follows.
fl2 (x) = m ||f
k=1,k=i k (x)|| 2 and f (x) = |f (x)|, respectively,
2 l1 i Definition 5: Individual s is ranked as r1 , if it dominates at
where · 2 is the L2 norm operator, and |·| is the absolute value least one solution in p∗ .

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.
SUN et al.: IGD INDICATOR-BASED EVOLUTIONARY ALGORITHM FOR MaOPs 179

Definition 6: Individual s is ranked as r2 , if it is nondomi-

nated to all the solutions in p∗ .
Definition 7: Individual s is ranked as r3 , if it is dominated
by all the solutions in p∗ , or dominated by a part of solutions
in p∗ but nondominated to the remaining solutions.
With Definitions 5–7, it is concluded that Pareto-optimal
solutions are all with rank values r1 , r2 , and r3 , if the PF is con-
vex, a hyperplane, and concave, respectively.1 To be specific, if
the PF of a minimization problem is a hyperplane, the Utopian
PF is obviously equivalent to the PF. Consequently, the Pareto-
optimal solutions lying at the PF are all nondominated to
the reference points which are sampled from the Utopian PF.
Based on Definition 6, the Pareto-optimal solutions are ranked
with r2 . This is also held true for the Pareto-optimal solutions
ranked with r1 for convex PF and Pareto-optimal solutions Fig. 4. Bi-objective optimization problem is illustrated to show the motivation
ranked with r3 for concave PF. of the proximity distance assignment for the individuals with rank value r2 .
Based on the conclusion mentioned above, the proximity
distances of each individual with different ranks in the popu- Algorithm 4: Assign Proximity Distance
lation are calculated. For convenience, it is assumed that there Input: Current population Pt with size q; reference
are k solutions in p∗ , and q individuals in the current popula- points p∗ with size k.
tion. Consequently, there will be q×k proximity distances. Let Output: Proximity distances matrix d.
dji denote the proximity distance of f(xi ) = (f1 (xi ), . . . , fm (xi )) 1 for i ← 1 to q do
(xi refers to the ith individual) to the jth point in (p∗ )j = 2 xi ← Pit ;
((p∗ )1 , . . . , (p∗ )m ), where i = {1, . . . , q} and j = {1, . . . , k}.
j j
3 r ← Calculate the rank of xi ;
Due to one individual with multiple proximity distances, the 4 for j ← 1 to k do
corresponding minimal proximity distance would be employed 5 if r = r1 then

when two individuals are compared upon their proximity dis- m ∗ j 2
6 dji ← − l=1 (fl (x ) − (p )l ) ;
i
tances. Because the proximity distance assignment is used to
differentiate the convergence of individuals with the same rank 7 else if r =r2 then
m ∗ j
values by comparing their associate proximity distances when 8 dji ← l=1 max(fl (x ) − (p )l , 0) ;
i 2

a prior knowledge of the PF is unknown in advance, the rank ri else

9

of xi is confirmed first in order to calculate the dji . Particularly, m ∗ j 2
10 dji ← l=1 (fl (x ) − (p )l ) ;
i
the proximity distance assignment is designed as follows. If
ri is equal to r3 , dji is set to the Euclidean distance between 11 end
end
f(xi ) and (p∗ )j ; if ri is equal to r1 , dji is set to the negative
12
13 end
value of the Euclidean distance between f(xi ) and (p∗ )j ; if ri
14 Return d.
is equal to r2 , dji is calculated by

m
  j 2
dji = max fl xi − (p∗ )l , 0 . (4) select the desirable individual x2 which has the most promis-
l=1 ing convergence to the PF. However, if the proximity distance
For an intuitive understanding, an example is illustrated in quantified by (4) is employed, it is clearly that the individ-
Fig. 4 to present the motivation of the proximity distance ual x2 is with the smallest minimal proximity distance (the
assignment for individuals who are ranked as r2 . In Fig. 4, minimal proximity distances of individuals x1 , x2 , and x3 are
the black solid circles refer to the reference points, the black 2, 0.5, and 2, respectively), which satisfies the motivation of
triangles marked by 1, 2, and 3 refer to the individuals x1 , the proximity distance assignment that a smaller value implies
x2 , and x3 which are with the rank r2 (these three individuals a better convergence. Noted here that, the proximity distance
are nondominated to the reference points). In this situation, assignment for the individuals with rank r2 is also employed
it is clearly shown that the individual x1 is with the smallest in the IGD+ indicator [39]. In summary, smaller proximity
minimal proximity distance if the Euclidean distance metric distance reveals the better proximity when the exact PF of
is employed (the minimal proximity distances of individuals the problem to be optimized is unknown. Furthermore, the
x1 , x2 , and x3 are 2.2361, 2.5495, and 2.8284, respectively). algorithm of the proximity distance assignment is presented
Consequently, both distance measurements of the individuals in Algorithm 4.
with ranks r1 and r3 in this situation cannot be utilized to
D. Generating Offspring
1 This is considered in the context of a minimization problem with con-
tinuous PF, and the extreme points are excluded from the Pareto-optimal The process of generating offspring in the proposed algo-
solutions. rithm is similar to that in genetic algorithms, in addition to

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.
180 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 23, NO. 2, APRIL 2019

Algorithm 5: Filling Up the Gene Pool from x1 and x2 is randomly selected being as one potential
Input: Current population Pt ; Gene pool size g. parent solution to be put into the gene pool (lines 17–19).
Output: Gene pool G. When the gene pool is full, two parent solutions are randomly
1 G ← ∅; selected from the gene pool for generating offspring, and then
2 while the size of G is less than g do these selected parent solutions are removed from the gene pool
3 {x1 , x2 } ← Randomly select two individuals from Pt ; until the gene pool is empty. Noted here that, the SBX [58]
4 rx1 ← Obtain the rank of x1 ; and the polynomial mutation [59] operators are employed for
5 rx2 ← Obtain the rank of x2 ; the corresponding crossover and mutation operations in the
6 d1 ← Obtain the proximity distances of x1 ; proposed algorithm. It has been reported that two solutions
selected in a large search space is not necessary to gener-
7 d2 ← Obtain the proximity distances of x2 ;
ate promising offspring [7], [60]. Generally, two ways can be
8 if rx1 < rx2 then
employed to solve this problem. One is the mating restriction
9 G ← G ∪ x1 ;
method to limit the offspring to be generated by the neighbor
10 else if rx1 > rx2 then
solutions [61]. The other one is to use SBX with a large dis-
11 G ← G ∪ x2 ;
tribution index [10]. In the proposed algorithm, the latter one
12 else
is utilized due to its simplicity.
13 if min(d1 ) < min(d2 ) then
14 G ← G ∪ x1 ;
E. Environmental Selection
15 else if min(d1 ) > min(d2 ) then
16 G ← G ∪ x2 ; When the offspring have been generated, the size of the
17 else current population is greater than that of the available slots.
18 x ← Randomly select one individual from As a consequence, the environmental selection takes effects to
{x1 , x2 }; select a set of representatives to survive to the next genera-
19 G ← G ∪ x; tion. In summary, the individuals are selected from the current
20 end population according to their assigned rank values and prox-
21 end imity distances. For convenience, it is assumed that there are
22 end N available slots, the selected individuals are to be stored in
23 Return G. Pt+1 , and the individuals with ranks r1 , r2 , as well as r3 are
grouped into Fr1 , Fr2 , and Fr3 nondominated fronts, respec-
tively.
i To be specific, the counter i is increased by one until
i−1
j=1 |Frj | > N where |·| is a countable operator. If j=1 |Frj |
is equal to N, the individuals in Fr1 , . . . , Fri−1 are copied into
the selection of individuals for filling up the gene pool from
Pt+1 and the environmental selection is terminated. Otherwise,
which the parent solutions are selected to generate offspring.
the individuals in Fr1 , . . . , Fri−1 are copied into Pt+1 first, then
In this section, the processes of generating offspring are elab- i−1
A = N − j=1 |Frj | individuals are selected from Fri . In sum-
orated in steps 1–5. Then, the details for filling up the gene
mary, the details of the environmental selection are presented
pool are presented in Algorithm 5.
in Algorithm 6. Furthermore, line 11 is confirmed by finding
Step 1: Select solutions from the current population to fill
A individuals who have the minimal total proximity distances
up the gene pool, until it is full.
to the A reference points r (line 10), which involves a linear
Step 2: Select two parent solutions from the gene pool and
assignment problem (LAP). In the proposed algorithm, the
remove them from the gene pool.
Hungarian method [62] is employed to solve this LAP.
Step 3: Employ the simulated binary crossover (SBX)
operator to generate offspring with the selected
parent solutions. F. Computational Complexity
Step 4: Employ the polynomial mutation operator to In this section, the computational complexity of the
mutate the generated offspring. proposed algorithm is analyzed. For convenience, it is assumed
Step 5: Repeat steps 2–4 until the gene pool is empty. that the problem to be optimized is with m objectives, n
The binary tournament selection [57] approach is employed decision variables, N desired solutions for decision-makers,
in Algorithm 5 to select individuals from the current pop- and the computational complexity is analyzed in the context
ulation to fill up the gene pool. In other words, the binary of Algorithm 1. To estimate each extreme point, the genetic
tournament selection [57] approach is employed to select algorithm is employed, and the SBX as well as polynomial
individuals from the current population. Specifically, two indi- mutation are used as the genetic operators. Furthermore, it is
viduals which are denoted by x1 and x2 are randomly selected assumed that the population size for estimating extreme points
from the current population first (line 3). Then, their ranks is set to be N, and the generation is set to be t1 . Consequently,
and the proximity distances are obtained (lines 4–7). Next, the total computation cost of uniformly generating reference
the individual with smaller rank value is selected to be copied points for IGD indicator (line 1) is O(t1 m2 N). Furthermore,
to the gene pool (lines 8–11). If x1 and x2 have the same rank lines 2 and 3 require O(nN) and O(mN) computations, respec-
values, the individual who has the smaller minimal proxim- tively. Because the number of the reference points is equal to
ity distance is selected (lines 13–16). Otherwise, an individual that of the desired solutions, the computational complexity of

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.
SUN et al.: IGD INDICATOR-BASED EVOLUTIONARY ALGORITHM FOR MaOPs 181

Algorithm 6: Environmental Selection When the number of solutions to be selected is larger than
Input: Fr1 , Fr2 , and Fr3 ; Available slots size N. the available slots, the representatives are chosen from a global
Output: Pt+1 . view in the proposed algorithm. For convenience of under-
1 Pt+1 ← ∅; standing, it is first assumed that a representatives need to be
2 i ← 1; selected from b solutions where b > a. Then the selection of
3 while |Pt+1 | + |Fri | < N do a representatives is simultaneously considered by the calcu-
4 Pt+1 ← Pt+1 ∪ Fri ; lation of IGD indicator, as oppose to choosing one by one.
5 i ← i + 1; Simultaneously selecting a representatives involves a LAP.
6 end By this linear assignment, each selected reference point can
7 if |Pt+1 | + |Fri | = N then have one distinct individual, which improves the diversity and
8 Pt+1 ← Pt+1 ∪ Fri ; the convergence simultaneously and this conclusion can also
9 else be found in literatures [37], [40], [63]. If the individuals are
i−1 selected by finding the individual who has the least distance to
10 r ← Uniformly select A = N − j=1 |Frj | reference
points from p∗ ; the reference points, the diversity is not necessarily guaranteed.
11 R ← Select A individuals from Fri ;
12 Pt+1 ← Pt+1 ∪ R; IV. E XPERIMENTS
13 end To evaluate the performance of the proposed algorithm
14 Return Pt+1 . in solving MaOPs, a series of experiments is performed.
Particularly, NSGA-III [10], MOEA/D [14], HypE [34],
RVEA [13], and KnEA [30] are selected as the state-of-the-art
peer competitors. Although the IGD+ -EMOA can be viewed
assigning ranks and proximity distances in line 6 are O(mN 2 ) as the peer algorithm based on IGD indicator, it is merely
and O(nN 2 ), respectively. Furthermore, generating offspring capable of solving MaOPs with no more than eight objec-
(line 7) needs O([N/2](n + n)) computations because the size tives. As a consequence, IGD+ -EMOA is excluded from the
of gene pool is set to be N. Since only the fitness, ranks list of peer competitors in our experiments.
and the proximity distances of the generated offspring need The remaining of this section is organized as follows. At
to be calculated, as a consequence, lines 8 and 9 consume first, the selected benchmark problems used in this experi-
O([N/2]m) and O([N/2]Nm) + O([N/2]Nn), respectively. In ment are introduced. Then, the chosen performance metric is
the environmental selection, the best case scenario in computa- given to measure the quality of the approximate Pareto-optimal
tional complexity is O(N), while the worst is O(N 3 ) given that solutions generated by the competing algorithms. Next, the
N individuals are linearly assignment to the reference points. parameter settings employed in all the compared algorithms
Furthermore, it is considered common that N is greater than are listed, and experimental results measured by the consid-
n, and N >> m in MaOPs. Therefore, lines 5–12 overall need ered performance metric are presented and analyzed. Finally,
O(tN 3 ) computations with the generation t. In summary, the the performance of the proposed algorithm in solving a real-
computational complexity of the proposed algorithm is O(tN 3 ) world MaOP is shown (in Section III of the supplemental
where t is the number of the generation and N is the number materials), and the performance on the proposed DNPE in
of solutions. estimating nadir point is empirically investigated.

G. Discussion A. Benchmark Test Problems

Loss of selection pressure is a major issue for traditional The widely used scalable test problems DTLZ1–DTLZ7
MOEAs in effectively solving MaOPs because of the tradi- from the DTLZ benchmark test suite [64] and WFG1–WFG9
tional domination comparisons between individuals giving a from the WFG benchmark test suite [65] are employed in
large proportion of nondominated solutions. In the proposed our experiments. Specifically, each objective function in one
algorithm, the dominance relation of all the individuals are given m-objective test problem of DTLZ has n = k + m − 1
compared to the reference points which are employed for the decision variables, and k is set to be 5 for DTLZ1, 10 for
calculation of IGD indicator. However, the exact reference DTLZ2–DTLZ6, and 20 for DTLZ7 problems. Moreover, each
points which are uniformly distributed in the PF are diffi- objective function of a given problem in WFG test suite has
cult to obtain. For this purpose, a set of points which are n = k + l decision variables, and k is set to be (m − 1) and l
evenly distributed in the Utopian PF are sampled. Furthermore, is set to be 20 based on the suggestion from [65].
in order to address this inefficiency given by these approxi-
mated reference points, three proximity distances are designed B. Performance Metric
according to their dominance relation to the approximated ref- The widely used HV [66] which simultaneously measures
erence points. This is in hope that less value of the proximity the convergence and diversity of the MaOEAs is selected
distance means that the corresponding individual is with a bet- as the performance metric in these experiments. Specifically,
ter proximity. Specifically, if the solutions with rank r2 are still the reference points for the calculation of HV are set to be
with the distance calculation of that with ranks r1 or r3 , the {1, . . . , 1} for DTLZ1, {2, . . . , 2} for DTLZ2-DTLZ6, and
convergence will be lost in the proposed algorithm [39]. {3, . . . , 2m + 1} for DTLZ7 as well as WFG1-WFG9 test

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.
182 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 23, NO. 2, APRIL 2019

TABLE I
C ONFIGURATIONS OF T WO L AYERS S ETTING in HypE and KnEA are set to be the same as that in others. In
the experiment, the reference points and reference vectors are
sampled with the two-layer method [10] and the configurations
are listed in Table I.
5) Genetic Operators: The SBX [58] and polynomial muta-
tion [69] are employed as the genetic operators. Moreover,
the probabilities of the crossover and mutation are set to be
problems. Please note that the solutions are discarded for the 1 and 1/n, respectively. The distribution indexes of mutation
calculation of HV when they are dominated by the predefined and crossover are set to be 20, in addition to NSGA-III whose
reference points. Because the computational cost increases sig- mutation distribution index is specifically set to be 30 based
nificantly as the number of objectives grows, Monte Carlo on the recommendation from its developers [10].
simulation [34]2 is applied for the calculation when m ≥ 10, In solving the proposed nadir point estimation method
otherwise the exact approach proposed in [67] is utilized.3 In for constructing the Utopian PF, evolutionary algorithm is
our experiments, all the HV values are normalized to [0, 1] employed. To be specific, the SBX and polynomial mutation,
by dividing the HV value of the origin with the corresponding both of whose distribution index are set to be 20, and proba-
reference point. Moreover, higher HV values indicate a better bilities for crossover and mutation are set to be 0.9 and 1/n,
performance of the corresponding MaOEA. respectively, are utilized as the genetic operators. In addition,
the population sizes are set to be the same to those in Table I,
C. Parameter Settings and the numbers of generations for all are set to be 1000.
Besides, the balance parameter λ in (2) is specified as 100.
In this section, the baseline parameter settings which are
adopted by all the compared MaOEAs are declared first. Then
the special parameter settings required by each MaOEA are D. Experimental Results and Analysis
provided. In this section, the results, which are generated by com-
1) Number of Objectives: Test problems with 8, 15, and peting algorithms over considered test problems with spe-
20 objectives are considered in the experiments because the cific objective numbers and then measured by the selected
proposed algorithm aims specifically at effectively solving performance metric, are presented and analyzed to high-
MaOPs. light the superiority of the proposed algorithm in addressing
2) Number of Function Evaluations and Stop Criterion: All MaOPs. Specifically, the mean values as well as the stan-
compared algorithms are individually executed 30 independent dard deviations of HV results over DTLZ1–DTLZ7 and
times. The maximum number of function evaluations for each WFG1–WFG9 test problems are listed in Tables II and III,
compared MaOEA in one independent run is set to be 2.3 × respectively. Furthermore, the numbers with bold face imply
106 , 4.3 × 106 , and 5.5 × 106 for 8-, 15-, and 20-objective, the best mean values over the corresponding test problem with
respectively, which is employed as the termination criterion. a given objective number (the second and third columns in
Noted that, the parameter settings here are set based on the Tables II and III) against all compared algorithms. Moreover,
convention that the maximum generations for the MaOEAs the symbols “+,” “−,” and “=” indicate whether the null
with more than ten objectives are generally in the order of hypothesis of the results, which are generated by the proposed
103 (the generation number set here is approximately to 1200). algorithm and corresponding compared peer competitor, is
Because of the proposed algorithm includes the phases of nadir accepted or rejected with the significance level 5% by
point estimation and the optimization for MaOPs, the function the considered rank-sum test. In addition, the last rows in
evaluations specified here will be shared by these two phases Tables II and III present the summarizations indicating how
for a fair comparison. many times the proposed algorithm performs better than,
3) Statistical Approach: Because of the heuristic character- worse than or equal to the chosen peer competitor, respec-
istic of the peer evolutionary algorithms, all the results, which tively. In order to conveniently investigate the experimental
are measured by the performance metric over 30 independent results of the well-designed proximity distance assignments in
runs for each competing algorithm, are statistically evaluated. the proposed MaOEA/IGD and the conclusion in Section III-C,
In this experiment, the Mann–Whitney–Wilcoxon rank-sum test problems are group into “convex,” “linear,” and “concave”
test [68] with a 5% significance level is employed for this based on the respective test problem features, and displayed in
purpose. the first columns of Tables II and III. Noted that, although the
4) Population Size: In principle, the population size can PFs of DTLZ7 and WFG1 are mixed, they are classified into
be arbitrarily assigned. However, the population size of the the linear category due to their PF shapes being more similar
proposed algorithm, NSGA-III, MOEA/D, and RVEA depends to linear.
on the number of the associated reference points or refer- From the results measured by HV on DTLZ1–DTLZ7 test
ence vectors. For a fair comparison, the sizes of population problems (Table II), it is clearly shown that MaOEA/IGD
2 The source code is available at: http://www.tik.ee.ethz.ch/sop/download/
achieves the best performance among its peer competitors
supplementary/hype/.
upon 8- and 20-objective DTLZ1 and DTLZ7, while per-
3 The source code is available at: http://www.wfg.csse.uwa.edu.au/ forms slightly worse upon 15-objective DTLZ1 by KnEA
hypervolume/. and DTLZ7 by RVEA. Furthermore, MaOEA/IGD also wins

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.
SUN et al.: IGD INDICATOR-BASED EVOLUTIONARY ALGORITHM FOR MaOPs 183

TABLE II
HV R ESULTS OF M AOEA/IGD AGAINST NSGA-III, MOEA/D, H YP E, RVEA, AND K N EA OVER DTLZ1–DTLZ7 W ITH 8-, 15-, AND 20-O BJECTIVE

TABLE III
HV R ESULTS OF M AOEA/IGD AGAINST NSGA-III, MOEA/D, H YP E, RVEA, AND K N EA OVER WFG1–WFG9 W ITH 8-, 15-, AND 20-O BJECTIVE

the best scores on 8- and 15-objective DTLZ4 and DTLZ6, on WFG8 test problems. Although MaOEA/IGD does not
but is defeated by NSGA-III upon these two problems with show the best scores on WFG3 and WFG4, it obtains sim-
20-objective. Although NSGA-III and KnEA show better ilar statistical results compared to the respective winners
performance upon 8- and 20-objective DTLZ2 and DTLZ5, (i.e., KnEA and NSGA-III). For 15-objective test problems,
MaOEA/IGD is the winner upon 15-objective DTLZ2 and MaOEA/IGD shows a better performance on WFG5, WFG6,
DTLZ5. In addition, MaOEA/IGD achieves the best score WFG8, and WFG9 than competing algorithms, while worse
upon 20-objective DTLZ3. than RVEA on WFG2 and WFG3, NSGA-III on WFG2, and
The HV results from WFG1–WFG9 test problems gen- KnEA on WFG4. Although NSGA-III performs better than
erated by competing algorithms are listed in Table III. For MaOEA/IGD on WFG7, MaOEA/IGD performs better than
eight-objective WFG test problems, MaOEA/IGD shows a all other peer competitors. In addition, MaOEA/IGD wins
better performance on WFG1, WFG2, WFG7, and WFG9 over NSGA-III, MOEA/D, HypE, RVEA, and KnEA on 20-
than its peer competitors, and performs a little worse than objective WFG1–WFG6 and WFG9, but underperforms on
that of KnEA on WFG5, RVEA on WFG6, and NSGA-III WFG7 and WFG8 in which RVEA performs better.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.
184 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 23, NO. 2, APRIL 2019

Briefly, MaOEA/IGD wins nine times out of the 12 com-

parisons upon the test problems whose PF shapes are linear
(i.e., DTLZ1, DTLZ7, WFG1, and WFG3), which can be
interpreted that the sampled reference points from the Utopian
PF for the proposed algorithm are the Pareto-optimal solu-
tions due to the linear feature of the PF, and the proximity
distance assignment for the solutions with rank value r2 has
taken effects. Furthermore, MaOEA/IGD shows competitive
performance on WFG2 test problem whose feature of the
PF is convex. Because the sampled reference points on the
Utopian PF are all nondominated by the Pareto-optimal solu-
tions, the proximity distances for solutions with rank r1 in (a)
MaOEA/IGD take effects in this situation. In addition, it is
no strange that MaOEA/IGD obtains better results on most of
other test problems whose PF features are concave because
the reference points utilized to maintain the diversity and con-
vergence of the proposed algorithm dominate the solutions
uniformly generated from the PF. In summary, the proposed
algorithm shows considerable competitiveness against consid-
ered competing algorithms in addressing selected MaOPs with
the results measured by the HV performance metric.
Theoretically, the major shortcoming of HV indicator
against IGD is its much higher computational complexity.
However, noted that from Tables II and III, the proposed
algorithm, which is designed based on the IGD indicator, (b)
outperforms HypE, which is motivated by the HV indicator,
upon all test problems with the selected numbers of objec-
tives, although the numbers of function evaluations regarding
HypE is set to be a much large number. The deficiencies
of HypE in this regard are explained as follows. First, it
has been reported in [11], [15], [70], and [71] that the HV
result is largely affected by the nadir points of the problem
to be optimized. In HypE, the nadir points are determined as
the evolution continues. In this way, the obtained nadir point
would be inaccurate during the early evolution process (the
reasons have been discussed in reviewing the nadir point esti-
mation approaches in Section II), which leads to the worse
performance of HypE. Second, the HV results of HypE in solv- (c)
ing MaOPs are estimated by Monte Carlo simulation, while
the number of reference points in Monte Carlo simulation is Fig. 5. Numbers of function evaluations performed by WC-NSGA-II,
PCSEA, and DNPE on (a) DTLZ1, (b) DTLZ2, and (c) WFG2 with 8-, 10-,
critical to the successful performance [34]. In practice, that 15-, and 20-objective.
number is unknown and unavailable of such may lead to a
poor performance.

where zi denotes the ith element of the estimated nadir point

E. Investigation on Nadir Point Estimation derived from the extreme points generated by the compared
In this section, we will investigate the performance of the algorithm or 2) the maximum function evaluation numbers
proposed DNPE on estimating the nadir point. To be specific, 100 000 is met. The experimental results for DTLZ1, DTLZ2,
two peer competitors including WC-NSGA-II and PCSEA and WFG2 with 8-, 10-, 15-, and 20-objective are plotted in
which have been discussed in Section II are utilized to perform Fig. 5. Please note that the reason of choosing these three
comparisons on selected test problems. In these comparisons, test problems is that they cover the various shapes of PF
the numbers of function evaluations regarding each com- (i.e., DTLZ1, DTLZ2, and WFG2 are with linear, concave,
pared algorithm are counted until: 1) the metric E ≤ 0.01 and convex PF, respectively) and characteristics of objective
formulated by value scales (i.e., DTLZ1 and DTLZ2 are with the same objec-
 tive value scales while WFG2 is not). Specifically, the ideal

m
 nad 2  2 points of DTLZ1, DTLZ2, and WFG2 are {0, . . . , 0}, and the
E= zi − zi / znad
i − z∗i (5) nadir points are {0.5, . . . , 0.5}, {1, . . . , 1}, and {2, 4, . . . , 2m},
i=1 respectively. In addition, the population size is specified as

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.
SUN et al.: IGD INDICATOR-BASED EVOLUTIONARY ALGORITHM FOR MaOPs 185

200, the probabilities of SBX and polynomial mutation are DTLZ2, and WFG2) with challenging features in PF shapes
set to be 0.9 and 1/n, and both of distribution index are set to and objective value scales, the experimental results reveal the
be 20. Because the proposed DNPE is based on the decompo- satisfactory results obtained by the proposed nadir point esti-
sition to estimate the nadir point, E ≤ 0.01/m and maximum mation method. In near future, we will place our efforts mainly
function evaluation number with 100 000/m are set to be the on two essential aspects.
stopping criteria for estimating each extreme point. 1) Constructing more accurate PF with limited informa-
The results performed by compared nadir point estimation tion priori to obtaining the Pareto-optimal solutions to
methods on 8-, 10-, 15-, and 20-objective DTLZ1, DTLZ2, improve the development of indicator-based algorithms
and WFG2 are illustrated in Fig. 5(a)–(c), respectively. It is which require the uniformly distributed reference points.
clearly shown in Fig. 5(a) that these compared algorithms find 2) Extending the proposed algorithm to solve constrained
the satisfactory nadir points of the DTLZ1 which is with the MaOPs.
linear PF within the predefined maximum function evaluation
numbers, and the proposed DNPE takes the least numbers of
function evaluations over the four considered objective num- R EFERENCES
bers. Moreover, WC-NSGA-II cannot find the nadir point over [1] V. Khare, X. Yao, and K. Deb, “Performance scaling of multi-objective
DTLZ2 with concave PF and WFG2 with convex PF with 10-, evolutionary algorithms,” in Proc. Int. Conf. Evol. Multi-Criterion
Optim., 2003, pp. 376–390.
15-, and 20-objective, and PCSEA cannot find the nadir point
[2] O. Chikumbo, E. Goodman, and K. Deb, “Approximating a multi-
over WFG2 with different objective value scales, while the dimensional Pareto front for a land use management problem: A
proposed DNPE performs well on both test problems with all modified MOEA with an epigenetic silencing metaphor,” in Proc. IEEE
considered objective numbers. In addition, the proposed DNPE Congr. Evol. Comput. (CEC), 2012, pp. 1–9.
[3] R. J. Lygoe, M. Cary, and P. J. Fleming, “A real-world applica-
is scalable to the objective number in the estimating nadir tion of a many-objective optimisation complexity reduction process,”
points of the MaOPs, which can be seen from Fig. 5(a) and (b). in Evolutionary Multi-Criterion Optimization. Heidelberg, Germany:
In summary, the proposed DNPE shows quality performance Springer, 2013, pp. 641–655.
[4] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist
in estimating nadir point of MaOPs with different PF features multiobjective genetic algorithm: NSGA-II,” IEEE Trans. Evol. Comput.,
and objective scales. vol. 6, no. 2, pp. 182–197, Apr. 2002.
[5] E. Zitzler, M. Laumanns, and L. Thiele, “SPEA2: Improving the strength
Pareto evolutionary algorithm for multiobjective optimization,” in Proc.
Evol. Methods Design Optim. Control Appl. Ind. Problems (EUROGEN),
V. C ONCLUSION 2002, pp. 95–100.
In this paper, an IGD indicator-based evolutionary algorithm [6] C. M. Fonseca and P. J. Fleming, “Multiobjective optimization and
is proposed for solving MaOPs. In order to obtain a set of multiple constraint handling with evolutionary algorithms. I. A unified
formulation,” IEEE Trans. Syst., Man, Cybern. A, Syst., Humans, vol. 28,
uniformly distributed reference points for the calculation of no. 1, pp. 26–37, Jan. 1998.
the IGD indicator, a DNPE method is designed to construct [7] R. C. Purshouse and P. J. Fleming, “On the evolutionary optimization of
the Utopian PF in which the reference points can be easily many conflicting objectives,” IEEE Trans. Evol. Comput., vol. 11, no. 6,
pp. 770–784, Dec. 2007.
sampled. For solving the deficiency of the Utopian PF being [8] B. Li, J. Li, K. Tang, and X. Yao, “Many-objective evolutionary
as the PF in the phase of sampling the reference points, one algorithms: A survey,” ACM Comput. Surveys, vol. 48, no. 1, p. 13,
rank assignment mechanism is proposed to compare the dom- 2015.
[9] T. Wagner, N. Beume, and B. Naujoks, “Pareto-, aggregation-,
inance relation of the solutions to the reference points, based and indicator-based methods in many-objective optimization,” in
on which three types of proximity distance assignments are Evolutionary Multi-Criterion Optimization. Heidelberg, Germany:
designed to distinct the quality of the solutions with the same Springer, 2007, pp. 742–756.
[10] K. Deb and H. Jain, “An evolutionary many-objective optimization
front rank values. In addition, the linear assignment principle is algorithm using reference-point-based nondominated sorting approach,
utilized as the selection mechanism to choose representatives part I: Solving problems with box constraints,” IEEE Trans. Evol.
for concurrently facilitating the convergence and diversity of Comput., vol. 18, no. 4, pp. 577–601, Aug. 2014.
the proposed algorithm. In summary, based on the proposed [11] Y. Yuan, H. Xu, B. Wang, and X. Yao, “A new dominance relation-
based evolutionary algorithm for many-objective optimization,” IEEE
nadir estimation method, the proposed dominance comparison Trans. Evol. Comput., vol. 20, no. 1, pp. 16–37, Feb. 2016.
approach, rank value and proximity distance assignment, and [12] Y. Sun, G. G. Yen, and Z. Yi, “Reference line-based estimation of distri-
selection mechanism collectively improve the evolution of the bution algorithm for many-objective optimization,” Knowl.-Based Syst.,
vol. 132, pp. 129–143, Sep. 2017, doi: 10.1016/j.knosys.2017.06.021.
proposed algorithm toward the PF with promising diversity. [13] R. Cheng, Y. Jin, M. Olhofer, and B. Sendhoff, “A reference vector
In order to qualify the performance of the proposed algo- guided evolutionary algorithm for many-objective optimization,” IEEE
rithm, a series of well-designed experiments is performed over Trans. Evol. Comput., vol. 20, no. 5, pp. 773–791, Oct. 2016.
[14] Q. Zhang and H. Li, “MOEA/D: A multiobjective evolutionary algorithm
two widely used benchmark test suites with 8-, 15-, and 20- based on decomposition,” IEEE Trans. Evol. Comput., vol. 11, no. 6,
objective, their results measured by the selected performance pp. 712–731, Dec. 2007.
metric indicate that the proposed algorithm is with consid- [15] Y. Yuan, H. Xu, B. Wang, B. Zhang, and X. Yao, “Balancing conver-
gence and diversity in decomposition-based many-objective optimizers,”
erable competitiveness in solving MaOPs. In addition, we IEEE Trans. Evol. Comput., vol. 20, no. 2, pp. 180–198, Apr. 2016.
utilize the proposed algorithm to solve one real-world MaOP, [16] Z. Wang, Q. Zhang, M. Gong, and A. Zhou, “A replacement strategy
in which the satisfactory results demonstrate the superiority for balancing convergence and diversity in MOEA/D,” in Proc. IEEE
of the proposed algorithm. Moreover, experiments are per- Congr. Evol. Comput. (CEC), 2014, pp. 2132–2139.
[17] K. Li, Q. Zhang, S. Kwong, M. Li, and R. Wang, “Stable matching-based
formed by the proposed DNPE method against a couple of selection in evolutionary multiobjective optimization,” IEEE Trans. Evol.
competitors over three representative test problems (DTLZ1, Comput., vol. 18, no. 6, pp. 909–923, Dec. 2014.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.
186 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 23, NO. 2, APRIL 2019

[18] K. Li, S. Kwong, Q. Zhang, and K. Deb, “Interrelationship-based [42] Y. Sun, G. G. Yen, and Z. Yi, “Global view-based selection mechanism
selection for decomposition multiobjective optimization,” IEEE Trans. for many-objective evolutionary algorithms,” in Proc. IEEE Congr. Evol.
Cybern., vol. 45, no. 10, pp. 2076–2088, Oct. 2015. Comput. (CEC), 2017, pp. 427–434.
[19] M. Asafuddoula, T. Ray, and R. Sarker, “A decomposition-based evolu- [43] K. Deb, S. Chaudhuri, and K. Miettinen, “Towards estimating nadir
tionary algorithm for many objective optimization,” IEEE Trans. Evol. objective vector using evolutionary approaches,” in Proc. 8th ACM Annu.
Comput., vol. 19, no. 3, pp. 445–460, Jun. 2015. Conf. Genet. Evol. Comput., 2006, pp. 643–650.
[20] S. B. Gee, K. C. Tan, V. A. Shim, and N. R. Pal, “Online diversity [44] H. K. Singh, A. Isaacs, and T. Ray, “A Pareto corner search evolutionary
assessment in evolutionary multiobjective optimization: A geometrical algorithm and dimensionality reduction in many-objective optimization
perspective,” IEEE Trans. Evol. Comput., vol. 19, no. 4, pp. 542–559, problems,” IEEE Trans. Evol. Comput., vol. 15, no. 4, pp. 539–556,
Aug. 2015. Aug. 2011.
[21] M. Laumanns, L. Thiele, K. Deb, and E. Zitzler, “Combining conver- [45] H. Wang, S. He, and X. Yao, “Nadir point estimation for many-
gence and diversity in evolutionary multiobjective optimization,” Evol. objective optimization problems based on emphasized critical regions,”
Comput., vol. 10, no. 3, pp. 263–282, 2002. Soft Comput., vol. 21, no. 9, pp. 2283–2295, 2017.
[22] F. di Pierro, S.-T. Khu, and D. A. Savic, “An investigation on prefer- [46] K. Deb and K. Miettinen, “A review of nadir point estimation procedures
ence order ranking scheme for multiobjective evolutionary optimization,” using evolutionary approaches: A tale of dimensionality reduction,”
IEEE Trans. Evol. Comput., vol. 11, no. 1, pp. 17–45, Feb. 2007. Kanpur Genet. Algorithms Lab., Indian Inst. Technol., Kanpur, India,
[23] G. Wang and H. Jiang, “Fuzzy-dominance and its application in evolu- KanGAL Rep. 2008004, 2008.
tionary many objective optimization,” in Proc. IEEE Int. Conf. Comput. [47] L. Ke, Q. Zhang, and R. Battiti, “MOEA/D-ACO: A multiobjective evo-
Intell. Security Workshops. (CISW), 2007, pp. 195–198. lutionary algorithm using decomposition and antcolony,” IEEE Trans.
[24] Z. He, G. G. Yen, and J. Zhang, “Fuzzy-based Pareto optimality for Cybern., vol. 43, no. 6, pp. 1845–1859, Dec. 2013.
many-objective evolutionary algorithms,” IEEE Trans. Evol. Comput., [48] X. Ma et al., “A multiobjective evolutionary algorithm based on decision
vol. 18, no. 2, pp. 269–285, Apr. 2014. variable analyses for multiobjective optimization problems with large-
[25] X. Zou, Y. Chen, M. Liu, and L. Kang, “A new evolutionary algorithm scale variables,” IEEE Trans. Evol. Comput., vol. 20, no. 2, pp. 275–298,
for solving many-objective optimization problems,” IEEE Trans. Syst., Apr. 2016.
Man, Cybern. B, Cybern., vol. 38, no. 5, pp. 1402–1412, Oct. 2008. [49] N. Chen et al., “An evolutionary algorithm with double-level archives
[26] S. Yang, M. Li, X. Liu, and J. Zheng, “A grid-based evolutionary algo- for multiobjective optimization,” IEEE Trans. Cybern., vol. 45, no. 9,
rithm for many-objective optimization,” IEEE Trans. Evol. Comput., pp. 1851–1863, Sep. 2015.
vol. 17, no. 5, pp. 721–736, Oct. 2013. [50] K. Praditwong and X. Yao, “How well do multi-objective evolution-
[27] A. López, C. A. C. Coello, A. Oyama, and K. Fujii, “An alternative ary algorithms scale to large problems,” in Proc. IEEE Congr. Evol.
preference relation to deal with many-objective optimization problems,” Comput., 2007, pp. 3959–3966.
in Proc. Int. Conf. Evol. Multi-Criterion Optim., 2013, pp. 291–306. [51] M. Szczepański and A. P. Wierzbicki, “Application of multiple cri-
[28] M. Li, S. Yang, and X. Liu, “Shift-based density estimation for Pareto- teria evolutionary algorithms to vector optimisation, decision support
based algorithms in many-objective optimization,” IEEE Trans. Evol. and reference point approaches,” J. Telecommun. Inf. Technol., vol. 3,
Comput., vol. 18, no. 3, pp. 348–365, Jun. 2014. pp. 16–33, Mar. 2003.
[29] J. Cheng, G. G. Yen, and G. Zhang, “A many-objective evolutionary [52] R. Benayoun, J. De Montgolfier, J. Tergny, and O. Laritchev, “Linear
algorithm with enhanced mating and environmental selections,” IEEE programming with multiple objective functions: Step method (stem),”
Trans. Evol. Comput., vol. 19, no. 4, pp. 592–605, Aug. 2015. Math. Program., vol. 1, no. 1, pp. 366–375, 1971.
[30] X. Zhang, Y. Tian, and Y. Jin, “A knee point-driven evolutionary algo- [53] M. I. Dessouky, M. Ghiassi, and W. J. Davis, “Estimates of the minimum
rithm for many-objective optimization,” IEEE Trans. Evol. Comput., nondominated criterion values in multiple-criteria decision-making,”
vol. 19, no. 6, pp. 761–776, Dec. 2015. Eng. Costs Prod. Econ., vol. 10, no. 1, pp. 95–104, 1986.
[31] M. Emmerich, N. Beume, and B. Naujoks, “An EMO algorithm using [54] H. Isermann and R. E. Steuer, “Computational experience concerning
the hypervolume measure as selection criterion,” in Proc. Int. Conf. Evol. payoff tables and minimum criterion values over the efficient set,” Eur.
Multi-Criterion Optim., 2005, pp. 62–76. J. Oper. Res., vol. 33, no. 1, pp. 91–97, 1988.
[32] C. Igel, N. Hansen, and S. Roth, “Covariance matrix adaptation for [55] P. Korhonen, S. Salo, and R. E. Steuer, “A heuristic for estimating nadir
multi-objective optimization,” Evol. Comput., vol. 15, no. 1, pp. 1–28, criterion values in multiple objective linear programming,” Oper. Res.,
2007. vol. 45, no. 5, pp. 751–757, 1997.
[33] D. Brockhoff and E. Zitzler, “Improving hypervolume-based [56] I. Das and J. E. Dennis, “Normal-boundary intersection: A new method
multiobjective evolutionary algorithms by using objective reduction for generating the Pareto surface in nonlinear multicriteria optimization
methods,” in Proc. IEEE Congr. Evol. Comput., 2007, pp. 2086–2093. problems,” SIAM J. Optim., vol. 8, no. 3, pp. 631–657, 1998.
[34] J. Bader and E. Zitzler, “HypE: An algorithm for fast hypervolume-based [57] B. L. Miller and D. E. Goldberg, “Genetic algorithms, tournament selec-
many-objective optimization,” Evol. Comput., vol. 19, no. 1, pp. 45–76, tion, and the effects of noise,” Complex Syst., vol. 9, no. 3, pp. 193–212,
2011. 1995.
[35] K. Gerstl, G. Rudolph, O. Schütze, and H. Trautmann, “Finding [58] K. Deb and R. B. Agrawal, “Simulated binary crossover for continuous
evenly spaced fronts for multiobjective control via averaging Hausdorff- search space,” Complex Syst., vol. 9, no. 3, pp. 1–15, 1994.
measure,” in Proc. 8th Int. Conf. Elect. Eng. Comput. Sci. Autom. Control [59] K. Deb, Multi-Objective Optimization Using Evolutionary Algorithms,
(CCE), 2011, pp. 1–6. vol. 16. Chichester, U.K.: Wiley, 2001.
[36] H. Trautmann, G. Rudolph, C. Dominguez-Medina, and O. Schütze, [60] S. F. Adra and P. J. Fleming, “Diversity management in evolutionary
“Finding evenly spaced Pareto fronts for three-objective optimization many-objective optimization,” IEEE Trans. Evol. Comput., vol. 15, no. 2,
problems,” in EVOLVE—A Bridge Between Probability, Set Oriented pp. 183–195, Apr. 2011.
Numerics, and Evolutionary Computation II. Heidelberg, Germany: [61] K. Deb and D. E. Goldberg, “An investigation of niche and species for-
Springer, 2013, pp. 89–105. mation in genetic function optimization,” in Proc. 3rd Int. Conf. Genet.
[37] C. A. R. Villalobos and C. A. C. Coello, “A new multi-objective evo- Algorithms, 1989, pp. 42–50.
lutionary algorithm based on a performance assessment indicator,” in [62] R. Jonker and T. Volgenant, “Improving the Hungarian assignment
Proc. 14th ACM Annu. Conf. Genet. Evol. Comput., 2012, pp. 505–512. algorithm,” Oper. Res. Lett., vol. 5, no. 4, pp. 171–175, 1986.
[38] R. Storn and K. Price, Differential Evolution-A Simple and Efficient [63] J. A. M. Berenguer and C. A. C. Coello, “Evolutionary many-objective
Adaptive Scheme for Global Optimization Over Continuous Spaces, optimization based on Kuhn–Munkres’ algorithm,” in Proc. Int. Conf.
vol. 3. Berkeley, CA, USA: ICSI, 1995. Evol. Multi-Criterion Optim., 2015, pp. 3–17.
[39] H. Ishibuchi, H. Masuda, Y. Tanigaki, and Y. Nojima, “Modified distance [64] K. Deb, L. Thiele, M. Laumanns, and E. Zitzler, Scalable Test Problems
calculation in generational distance and inverted generational distance,” for Evolutionary Multiobjective Optimization. London, U.K.: Springer,
in Proc. Int. Conf. Evol. Multi-Criterion Optim., 2015, pp. 110–125. 2005.
[40] E. M. Lopez and C. A. C. Coello, “IGD+-EMOA: A multi-objective [65] S. Huband, P. Hingston, L. Barone, and L. While, “A review of
evolutionary algorithm based on IGD+,” in Proc. IEEE Congr. Evol. multiobjective test problems and a scalable test problem toolkit,” IEEE
Comput. (CEC), 2016, pp. 999–1006. Trans. Evol. Comput., vol. 10, no. 5, pp. 477–506, Oct. 2006.
[41] Y. Sun, G. G. Yen, H. Mao, and Z. Yi, “Manifold dimension reduction [66] E. Zitzler and L. Thiele, “Multiobjective evolutionary algorithms: A
based clustering for multi-objective evolutionary algorithm,” in Proc. comparative case study and the strength Pareto approach,” IEEE Trans.
IEEE Congr. Evol. Comput. (CEC), 2016, pp. 3785–3792. Evol. Comput., vol. 3, no. 4, pp. 257–271, Nov. 1999.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.
SUN et al.: IGD INDICATOR-BASED EVOLUTIONARY ALGORITHM FOR MaOPs 187

[67] L. While, L. Bradstreet, and L. Barone, “A fast way of calculating exact Zhang Yi (F’16) received the Ph.D. degree in
hypervolumes,” IEEE Trans. Evol. Comput., vol. 16, no. 1, pp. 86–95, mathematics from the Institute of Mathematics,
Feb. 2012. Chinese Academy of Science, Beijing, China, in
[68] R. G. D. Steel, J. H. Torrie, and D. A. Dickey, Principles and 1994.
Procedures of Statistics A Biometrical Approach. New York, NY, USA: He is currently a Professor with the Machine
McGraw-Hill, 1997. Intelligence Laboratory, College of Computer
[69] K. Deb and M. Goyal, “A combined genetic adaptive search (GeneAS) Science, Sichuan University, Chengdu, China. He
for engineering design,” Comput. Sci. Informat., vol. 26, no. 4, has co-authored three books entitled Convergence
pp. 30–45, 1996. Analysis of Recurrent Neural Networks (Kluwer
[70] A. Auger, J. Bader, D. Brockhoff, and E. Zitzler, “Theory of the Academic Publishers, 2004), Neural Networks:
hypervolume indicator: Optimal μ-distributions and the choice of the Computational Models and Applications (Springer,
reference point,” in Proc. 10th ACM SIGEVO Workshop Found. Genet. 2007), and Subspace Learning of Neural Networks (CRC Press, 2010). His
Algorithms, 2009, pp. 87–102. current research interests include neural networks and big data.
[71] H. Ishibuchi, Y. Hitotsuyanagi, N. Tsukamoto, and Y. Nojima, Dr. Yi was an Associate Editor of the IEEE T RANSACTIONS ON N EURAL
“Many-objective test problems to visually examine the behavior of N ETWORKS AND L EARNING S YSTEMS from 2009 to 2012. He has been an
multiobjective evolution in a decision space,” in Parallel Problem Associate Editor of the IEEE T RANSACTIONS ON C YBERNETICS since 2014.
Solving From Nature, PPSN XI. Heidelberg, Germany: Springer, 2010,
pp. 91–100.

Yanan Sun (S’15–M’18) received the Ph.D. degree

in engineering from Sichuan University, Chengdu,
China, in 2017.
From 2015 to 2017, he was a joint Ph.D. stu-
dent financed by the China Scholarship Council in
the School of Electrical and Computer Engineering,
Oklahoma State University, Stillwater, OK, USA. He
is currently a Post-Doctoral Research Fellow with
the School of Engineering and Computer Science,
Victoria University of Wellington, Wellington,
New Zealand. His current research interests include
many-objective optimization and deep learning.

Gary G. Yen (S’87–M’88–SM’97–F’09) received

the Ph.D. degree in electrical and computer engineer-
ing from the University of Notre Dame, Notre Dame,
IN, USA, in 1992.
He is currently a Regents Professor with the
School of Electrical and Computer Engineering,
Oklahoma State University (OSU), Stillwater, OK,
USA. Before joining OSU in 1997, he was with the
Structure Control Division, U.S. Air Force Research
Laboratory, Albuquerque, NM, USA. His current
research interests include intelligent control, com-
putational intelligence, conditional health monitoring, signal processing, and
their industrial/defense applications.
Dr. Yen was a recipient of the Andrew P Sage Best Transactions
Paper Award from the IEEE Systems, Man and Cybernetics Society in
2011, and the Meritorious Service Award from the IEEE Computational
Intelligence Society in 2014. He was an Associate Editor of the IEEE Control
Systems Magazine, the IEEE T RANSACTIONS ON C ONTROL S YSTEMS
T ECHNOLOGY, Automatica, Mechantronics, the IEEE T RANSACTIONS ON
S YSTEMS , M AN , AND C YBERNETICS —PART A: S YSTEMS AND H UMANS,
the IEEE T RANSACTIONS ON S YSTEMS , M AN , AND C YBERNETICS —
PART B: C YBERNETICS, and the IEEE T RANSACTIONS ON N EURAL
N ETWORKS. He is currently serving as an Associate Editor for the
IEEE T RANSACTIONS ON E VOLUTIONARY C OMPUTATION and the IEEE
T RANSACTIONS ON C YBERNETICS. He served as the General Chair for the
2003 IEEE International Symposium on Intelligent Control held in Houston,
TX, USA, and the 2006 IEEE World Congress on Computational Intelligence
held in Vancouver, BC, Canada. He served as the Vice President for the
Technical Activities from 2005 to 2006 and then the President from 2010 to
2011 of the IEEE Computational intelligence Society. He was the Founding
Editor-in-Chief of the IEEE Computational Intelligence Magazine from 2006
to 2009.

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO PARANA. Downloaded on January 08,2024 at 17:43:54 UTC from IEEE Xplore. Restrictions apply.