IEEE TRANSACTIONS ON CYBERNETICS, VOL. 51, NO.

7, JULY 2021 3417

A Fast Dynamic Evolutionary Multiobjective

Algorithm via Manifold Transfer Learning
Min Jiang , Senior Member, IEEE, Zhenzhong Wang, Liming Qiu, Shihui Guo , Member, IEEE,
Xing Gao , Member, IEEE, and Kay Chen Tan , Fellow, IEEE

Abstract—Many real-world optimization problems involve I. I NTRODUCTION

multiple objectives, constraints, and parameters that may
YNAMIC multiobjective optimization problems
change over time. These problems are often called dynamic
multiobjective optimization problems (DMOPs). The difficulty
in solving DMOPs is the need to track the changing Pareto-
D (DMOPs) have received increasing attention in recent
years because many real-world problems can be formulated
optimal front efficiently and accurately. It is known that transfer as DMOPs [1], [2]. An example is a dynamic portfolio
learning (TL)-based methods have the advantage of reusing expe- optimization problem [3], which maximizes financial profit
riences obtained from past computational processes to improve
while minimizing risk. A variety of factors needs to be
the quality of current solutions. However, existing TL-based
methods are generally computationally intensive and thus time considered, including profit expectations, share/option prices,
consuming. This article proposes a new memory-driven mani- and risk obligations. These factors change over time, and
fold TL-based evolutionary algorithm for dynamic multiobjective it is a common demand to solve this optimization at high
optimization (MMTL-DMOEA). The method combines the mech- frequency. Therefore, accelerating the process of solving
anism of memory to preserve the best individuals from the past DMOPs is one of the research focuses on computational
with the feature of manifold TL to predict the optimal indi-
viduals at the new instance during the evolution. The elites of intelligence.
these individuals obtained from both past experience and future Among the recent progress of dynamic multiobjective
prediction will then constitute as the initial population in the optimization algorithms (DMOAs), the learning-based meth-
optimization process. This strategy significantly improves the ods are particularly promising. This class of approaches allows
quality of solutions at the initial stage and reduces the compu- the heuristic algorithms to improve their performance with
tational cost required in existing methods. Different benchmark
problems are used to validate the proposed algorithm and the
machine-learning techniques. Generally, a prediction model
simulation results are compared with state-of-the-art dynamic can be created via machine-learning techniques, and assist the
multiobjective optimization algorithms (DMOAs). The results standard DMOAs by reducing the computational cost while
show that our approach is capable of improving the computa- maintaining a comparable performance even if the environ-
tional speed by two orders of magnitude while achieving a better ment changes over time. For example, researchers proposed
quality of solutions than existing methods.
a memory-based evolutionary algorithm (EA) by introducing
Index Terms—Dynamic multiobjective, manifold learning, two prediction models: the first one used a regression model
transfer learning (TL). to predict the occurrence timing of environmental changes;
and the second model was based on Markov chains and used
to forecast changes [4]. Another work suggested integrating
motion information into an EA so that the algorithm can track
a time-changing optimum [5].
However, it is still an open challenge to solve DMOPs with
Manuscript received November 1, 2019; revised February 8, 2020 and April a demanding goal of minimal time cost (e.g., optimizing the
5, 2020; accepted April 8, 2020. Date of publication May 20, 2020; date of investment portfolio in high-frequency trading). One of the
current version June 23, 2021. This work was supported in part by the National
Natural Science Foundation of China under Grant 61673328 and Grant reasons is that the solutions of a given DMOP at different
61876162, in part by the Shenzhen Scientific Research and Development times are nonindependent identically distributed (Non-IID).
Funding Program under Grant JCYJ20180307123637294, and in part by the This indicates that although there may exist a latent relation-
Research Grants Council of the Hong Kong under Grant CityU11202418 and
Grant CityU11209219. This article was recommended by Associate Editor ship between the solution distributions at different times, they
Y. S. Ong. (Corresponding authors: Liming Qiu; Kay Chen Tan.) are not identical.
Min Jiang, Zhenzhong Wang, Shihui Guo, and Xing Gao are with the An algorithmic framework, called Tr-DMOEA [6],
School of Informatics, Xiamen University, Xiamen 361005, China (e-mail:
[email protected]). addresses this challenge by integrating transfer learn-
Liming Qiu is with the LenzTech Company, Beijing 100026, China (e-mail: ing (TL) [7] and population-based EAs. This approach speeds
[email protected]). up the evolutionary process by reusing past experiences with
Kay Chen Tan is with the Department of Computer Science, City University
of Hong Kong, Hong Kong, and also with the City University of Hong Kong TL and effectively improving the quality of the initial popu-
Shenzhen Research Institute, Shenzhen, China (e-mail: [email protected]). lation. This framework can benefit general population-based
This article has supplementary material provided by the multiobjective algorithms without requiring extensive modi-
authors and color versions of one or more figures available at
https://doi.org/10.1109/TCYB.2020.2989465. fications. The experimental results confirm the effectiveness
Digital Object Identifier 10.1109/TCYB.2020.2989465 of the proposed design for DMOPs. However, this solution
2168-2267 
c 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on January 24,2024 at 06:18:07 UTC from IEEE Xplore. Restrictions apply.
3418 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 51, NO. 7, JULY 2021

still suffers from two significant defects that prevent its Definition 2 [Dynamic Pareto-Optimal Set (DPOS)]: Both
application in real-world tasks. First, the TL technique [7] x and x∗ are decision vectors, and if a decision vector x∗ is
solves the Non-IID problem by searching for a latent space, said to be nondominated at time t if and only if there is no
and the determination of the optimal latent space involves other decision vector x such that x t x∗ at time t. The DPOS
choosing the optimal values of numerous hyperparameters. is the set of all Pareto-optimal solutions at time t, that is
 
This process of parameter tuning usually requires trial and DPOS = x∗ | ∃x, x t x∗ .
error and often takes a significant amount of time. Second,
even if the optimal latent space is found, Tr-DMOEA still Definition 3 [Dynamic Pareto-Optimal Front (DPOF)]: At
needs to invoke the optimization algorithm to produce the time t, the DPOF is the corresponding objective vectors of the
initial population, which also consumes excessive computing DPOS
   
resources. DPOF = F x∗ , t | x∗ ∈ DPOS .
The motivation of this article is to accelerate DMOPs-
The purpose of a DMOA is to produce a set of solutions as
solving without deteriorating the solution quality. We propose
close as possible to the changing POF while maximizing the
a method for solving DMOPs by combining a memory mech-
diversity of the solution.
anism with manifold TL. To the best of our knowledge, this
is the first attempt to explore an integrative approach of
B. Manifold-Based Transfer Learning
both the techniques. These two approaches complement each
other and this integration provides two benefits. First, the 1) Manifold and Geodesic: A manifold is a topological
proposed method compresses high-dimensional data in a low- space that is locally homeomorphic to a Euclidean space. In
dimensional space with a special TL technique and it requires other words, at any point of a d-dimensional manifold, it is
a few hyperparameters. Second, computing resources are spent locally homeomorphic to the Euclidean space Rd . For exam-
on knowledge transfer of elite individuals via the memory ple, the surface of the Earth as a 2-D manifold, and any part
mechanism, which greatly improves the speed of finding the of it can be considered as a 2-D plane. A geodesic is a smooth
solutions in evolutionary optimization. curve on the manifold, that is locally the shortest curve con-
The remainder of this article is organized as follows. necting two points. A geodesic can be considered as a “bent”
Section II introduces the fundamental concepts of dynamic straight line on a manifold. Calculating the distance between
optimization problems and discusses the existing works. two points on the manifold is to calculate the geodesic dis-
Section III presents background on manifold TL and explains tance. For more details about manifold learning and geodesic,
the details of our method memory-driven manifold TL- please refer to [8].
based EA for dynamic multiobjective optimization (MMTL- 2) Sample Geodesic Flow: The manifold-based TL algo-
DMOEA). In Section IV, we present the experimental results rithm used in this article is sample geodesic flow (SGF) [9].
of the proposed algorithm on different test functions and The SGF algorithm is inspired by incremental learning: a
compare with state-of-the-art algorithms. Section V draws a learning system can continuously learn from the samples and
summary of this article and outlines future research directions. keep most knowledge of them. In order to apply the knowl-
edge learned from the source domain to the target domain,
the SGF algorithm constructs a geodesic path between the
II. P RELIMINARIES AND R ELATED W ORKS two domains and then transports the knowledge through this
A. Dynamic Multiobjective Optimization Problems path to the target domain. In other words, the source and target
The essential characteristic of DMOPs is that the objectives domains are mapped to the starting and ending points, and the
change over time or under different environments. Without algorithm constructs a geodesic flow between the two points
loss of generality, the general form of DMOPs is defined as in the manifold. After that, it will select k intermediate points
on the geodesic flow, and finally, transform the data of the
min F(x, t) = <f1 (x, t), f2 (x, t), . . . , fm (x, t)> source and target domains to these intermediate points.
s.t. x ∈ X (1) Fig. 1 presents an illustrative diagram of the SGF method.
First, the algorithm maps the source and target domains to
where x = x1 , x2 , . . . , xn  is the n-dimensional decision the two points S0 and S1 on a Grassmann manifold G(d, n),
variable, and X ∈ Rn is the decision space. t is the time and the algorithm will construct a geodesic flow from S0 to
(environment) variable, f1 , f2 , . . . , fm are objective functions, S1 (shown as a dotted line in the figure). Next, it will select
and m is the number of objectives. k intermediate points, the red dot S0.2 , S0.4 , S0.6 , S0.8 on the
A set of solutions is defined as a dynamic Pareto-optimal set geodesic flow, and transform the training and testing data to
(POS) if this solution set exists in a DMOP at a given time t, these intermediate subspaces. The data onto these subspaces
and the decision vector in the set cannot be further optimized will be analyzed to perform classification.
in any dimension.
Definition 1 (Dynamic Decision Vector Domination): At C. Related Works
time t, a decision vector x1 Pareto dominates another vector x2 ,
denoted by x1 t x2 , if and only if Existing DMOAs can be divided into the following
 categories: maintaining diversity methods, multipopulation-
∀i = 1, . . . , m, fi (x1 , t) ≤ fi (x2 , t) based methods, memory-based methods, and prediction-based
(2)
∃i = 1, . . . , m, fi (x1 , t) < fi (x2 , t). methods.

Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on January 24,2024 at 06:18:07 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: FAST DYNAMIC EVOLUTIONARY MULTIOBJECTIVE ALGORITHM VIA MANIFOLD TL 3419

The memory-based methods use additional storage to

implicitly or explicitly store useful information from previous
generations to improve the algorithm performance. Wang and
Li [18] proposed a multistrategy dynamic multiobjective EA
(MS-MOEA). It uses a progeny generation mechanism based
on adaptive genetics and differential operators to accelerate
convergence. Chen et al. [16] implemented a dynamic two-
archive EA that maintains two co-evolving populations simul-
taneously. Two complementary populations concern about
convergence and diversity, respectively.
The prediction-based DMOAs have received much atten-
Fig. 1. Illustrative diagram of the SGF method. A, B, and C are the labeled tion in the field of DMOPs, and this class of methods predicts
samples from the source domain. D and E are the samples from the target
domain that need to be classified.
the state of the changing environment and then makes a deci-
sion such that the algorithms can accommodate the changes
in advance. The prediction typically introduces machine-
Maintaining population diversity is an effective approach learning techniques to utilize existing information in the
for minimizing the likeliness of the algorithm to be trapped memory.
in local optima. This strategy assumes that a population with Rong et al. [19] presented a multidirectional prediction
higher diversity should have better adaptability and aims to strategy to predict the moving location of the POS accu-
maintain population diversity. A representative work in this rately. A classification strategy clusters the population into
category is the dynamic NSGA-II (D-NSGA-II) proposed by a number of representative groups and adapts the number
Deb et al. [10]. This algorithm has two versions, and the first of clusters according to the intensity of the environmental
version is DNSGA-II-A, which increases diversity by replac- change. Koo et al. [20] proposed a prediction strategy called
ing some individuals in the population, ζ % of the population, the dynamic gradient prediction strategy. It defines a set of
with randomly generated individuals. The second version, gradient prediction vectors that can be used to correlate the
DNSGA-II-B, ensures diversity by substituting mutated indi- current solution with the previously obtained optimal solution.
viduals for ζ % of the population. The actual performance of Rong et al. [21] presented a multimodel prediction approach
these two versions depends on the choice of ζ . Considering the (MMP), and the method detects the type of change and then
disconnectedness of the POF of DMOPs with constraints may selects an appropriate prediction model to generate an initial
cause the population to sink into local regions, decreasing the population.
population diversity. Chen et al. [11] designed a selection oper- Zhou et al. [22] proposed a method called a population
ator which can obtain nondominated solutions with diversity prediction strategy (PPS), for predicting the entire population.
when the environment changes. The mating selection strat- When a change in the environment is detected, the next center
egy and population selection operator can handle infeasible point is predicted by an autoregressive model using a series of
solutions adaptively. center points throughout the search process, and at the same
Recently, Ruan et al. [12] also proposed an algorithm based time, the previous manifold is used to estimate the next mani-
on maintaining population diversity. The method uses the posi- fold. The main problem of this method is the lack of sufficient
tion and direction of the center point of the optimal solution historical information at the beginning stage, which may lead
at the previous moment to predict future solutions. Jiang and to poor convergence.
Yang [13] proposed a strategy based on steady state and pop- Muruganantham et al. [23] proposed a predictive dynamic
ulation diversity, called SGEA. This algorithm detects the multiobjective EA (MOEA/D-KF) based on the Kalman fil-
incidence of environmental changes and uses a steady-state ter. The Kalman filter guides the search for a new POS
method to adapt to the new environment quickly. and generates a new initial population. The optimal solu-
The multipopulation strategy is an efficient solution for tions can be found via a decomposition-based differential
DMOPs, especially in the case of multiple peaks and com- evolution algorithm (MOEA/D-DE) [24]. Zou et al. [25]
peting peaks. Goh and Tan [14] proposed a coevolutionary proposed a prediction strategy based on center points and
multiobjective algorithm (dCOEA) based on competition and knee points (CKPSs) to predict the initial population of the
collaboration, which breaks the problem into several sub- next moment. Xu et al. [26] introduced a method in which
problems. Yazdani et al. [15] proposed a cooperative coevo- all decision variables are partitioned into two subcomponents
lutionary multipopulation framework for solving large-scale according to their interrelation with the environment. The two
dynamic optimization problems. The framework breaks a subcomponents are optimized with two prediction methods.
large-scale dynamic optimization problem into a set of lower Woldesenbet and Yen [27] distinguished decision variables by
dimensional components and controls the budget assignment their average sensitivities to the change in the objective space
to components [16] for tracking multiple moving optima. and relocated individuals.
Gong et al. [17] proposed a framework in which all the deci- The prediction-based algorithms show their advantages in
sion variables are divided into two subpopulation according maintaining the population quality, however, it may not pro-
to the interval similarity between each decision variable and duce satisfactory results in problems with Non-IID. An inac-
interval parameters. curate prediction model will likely lead the search process in

Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on January 24,2024 at 06:18:07 UTC from IEEE Xplore. Restrictions apply.
3420 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 51, NO. 7, JULY 2021

the wrong direction, which means actual results will be worse Process 1 MMTL-DMOEA
than a method that does not use the prediction technique. Input: The dynamic optimization function: F(x, t); the size of
TL is a powerful weapon that solves the dilemma caused by the population: N; an external storage P and its capacity
the Non-IID hypothesis, and recent research has also proved C; the number of clusters: L.
that transferring acquired knowledge in the optimization pro- Output: The Solutionst at time t.
cess is promising [28], especially for solving dynamic [6] 1: Initialization;
or multiple optimization tasks simultaneously [29]–[31]. 2: Solutions0 = SMOA(POP0 , F(x, 0), N);
Jiang et al. [6] proposed a framework called Tr-DMOEA for 3: Store Solutions0 into P;
solving DMOPs. This framework integrates between TL and 4: while Change detected do
classical evolutionary multiobjective optimization algorithms. 5: t = t + 1;
TL effectively generates a high-quality initial population pool 6: LastBestSol = FindBestSol(P, F(x, t), N);
via reusing past experience to speed up the evolutionary 7: TransSol = Transfer(LastBestSol, F(x, t), N, L);
process. 8: POPt = LastBestSol ∪ TransSol;
Bali et al. [31] proposed an online transfer parameter 9: (Solutionst ) = SMOA(POPt , F(x, t), N);
estimation approach for the multitasking EA. The proposed 10: if size(P ∪ Solutionst ) ≤ C then
algorithm evaluates the similarities between different task 11: Store Solutionst into P;
instances, therefore, the extent of knowledge transfer can be 12: else
adapted based on the optimal mixture of probabilistic mod- 13: Replace the earliest stored individuals with
els. Da et al. [32] demonstrated a framework for black-box Solutionst ;
transfer optimization. The framework measures the similari- 14: end if
ties between the source instance and the target instance and 15: end while
reveals latent synergies during optimizing. When faced with 16: return Solutionst ;
multiple sources, the positive and negative transfer is automat-
ically identified so as to curb the negative transfer, thereby the
excellent performance is kept. of this method is to accelerate the process of solution search-
Inspired by the recent autoencoding evolutionary ing while ensuring the solution quality. The main contribution
search [29], Zhou et al. [33] proposed an improved of MMTL-DMOEA is the integration of manifold TL with
adaptive indicator-based EA. The proposed algorithm is the elite memory mechanism, to produce an initial population
capable of transferring the past search information by for the next generation. This initial population can help any
adapting some obtained solutions with high hypervolume population-based multiobjective optimization algorithms find
value. Min et al. [34] designed a transfer stacking of the the POS of the optimization problem faster.
Gaussian process surrogate models by reusing the acquired Fig. 2 shows the algorithm flow of MMTL-DMOEA. This
knowledge. The adaptive knowledge gained from expensive algorithm first stores the optimal solutions obtained at differ-
problem-solving information is introduced to improve the ent times in external memory. Once the environment changes
performance of the optimization process of the target task. are detected, the algorithm selects a certain number of indi-
Liu et al. [35] proposed a neural-network-based information viduals from external storage according to predefined criteria.
transfer method to reuse past solutions. When the environment In the third step, the algorithm predicts individuals through
changes, the proposed algorithm collects the solutions from the principal component analysis (PCA) and the manifold TL
both the previous environment and the new environment. algorithm. After generating the predicted individuals, the elite
Then, the neural-network transfers acquired solutions for individuals selected from the memory are merged with them
guiding the search in the new environment. to constitute an initial population.
Introducing TL into the field of dynamic optimization The detailed steps of MMTL-DMOEA are shown in
shows its potential. However, at the same time, some new Process 1. It is worth pointing out that any population-based
problems arise. For example, Tr-DMOEA often leads to static multiobjective algorithm (SMOA) can be used as a rou-
diminishing diversity in solutions. Researchers proposed a tine for processing static multiobjective optimization problems
dynamic multiobjective estimation of distribution algorithm, in our proposed algorithm.
called DANE-EDA [36]. This design combines the Monte Remark 1: The parameter C refers to the maximum number
Carlo method with the TL technique, and this combination of individuals that can be stored in the external memory.
maintains a tradeoff of exploration–exploitation from temporal Remark 2: When the external memory overflows, the algo-
and spatial perspectives. Although researchers made progress rithm replaces the earliest stored individuals with the newly
in the DMOP domain, existing methods are generally time generated Solutionst .
consuming and not suitable for some applications in the real Our proposed algorithm follows the elite reservation mech-
world. anism. In other words, we will preserve the optimal solutions
obtained at previous moments. Although solutions at different
times are varying, there should be some inherent relationship
III. MMTL-DMOEA A LGORITHM between these solutions. Therefore, it is possible to obtain a
This article proposes the method of memory-driven mani- better initial population by reserving the best individuals and
fold TL (MMTL-DMOEA) to address DMOPs. The purpose making a reasonable transformation of these individuals.

Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on January 24,2024 at 06:18:07 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: FAST DYNAMIC EVOLUTIONARY MULTIOBJECTIVE ALGORITHM VIA MANIFOLD TL 3421

Fig. 2. Main idea of MMTL-DMOEA. Step 1: MMTL-DMOEA will preserve the nondominated solutions obtained from time 0 to time t. Step 2: Some
individuals are selected from the stored elite set according to the predefined mechanism. Step 3: The selected individuals are mapped to the predicted individuals
by using the manifold-based TL method. Step 4: MMTL-DMOEA generates an initial population by merging predicted and selected individuals.

Lines 10–14 in Process 1 implement our elite reserving Process 2 FindBestSol

strategy. When the external memory overflows, the algorithm Input: The dynamic optimization function: F(x, t);
replaces the earliest stored individuals with the newly gener- Number of individuals: N;
ated ones. This straightforward strategy builds on the intuition External storage: P.
that the optimal solution for the next moment may be more Output: Excellent individuals LastBestSol.
relevant to the recently obtained individuals. 1: Uniformly sample ns solutions XT from the decision space;
Selecting the most suitable individual from external memory 2: Call a SVR to construct the estimator E with
is critical in MMTL-DMOEA. The results of the selection {XT , F(XT , t)};
algorithm (Process 2) provide samples of the source domain 3: Estimate objectives of P: Y = E(P);
for TL, and part of selected solutions directly constitute the 4: Find non-dominated solutions LastBestSol in Y;
final population. When changes are detected, to avoid extra 5: while N/2 < size(LastBestSol) do
function evaluations of individuals in the external storage, an 6: Delete individual in LastBestSol;
estimator for estimating the objective values of individuals 7: end while
from the external memory is constructed by support vector 8: while N/2 > size(LastBestSol) do
regression (SVR) with only a small number of individuals sam- 9: Add Gaussian noise with individuals in LastBestSol to
pled from the decision space in the new environment. Then, LastBestSol;
according to the estimated objective values, all individuals in 10: end while
the external storage are ranked by nondominated sorting, and 11: return LastBestSol;
the nondominated solutions are chosen. In order to ensure the
diversity of solutions, the selection algorithm will eliminate
individuals with a higher degree of congestion. When the num- when 0 < k < 1
ber of individuals in the nondominated solution set is too large,
φ(k) = PS U1 (k) − RS U2 (k) (3)
the algorithm will randomly select some solutions to remove
from the set; otherwise, some random noise will be generated where U1 and U2 are a pair of orthogonal matrices obtained
to supplement the solution set. by SVD decomposition. (k) and (k) are d × d dimen-
MMTL-DMOEA also introduces TL to predict the initial sional diagonal matrices with diagonal elements cos(kθi ) and
population of the next moment, as shown by Transfer in line 7 sin(kθi ), i = 1, 2, . . . , d, and θi ∈ [0, π/2] are the principal
of Process 1. The Transfer algorithm is an improved version angle of PS , PT .
of the TL method SGF. After obtaining the geodesic model, the point where the
In SGF, the difficulty lies in constructing a geodesic model. original data are mapped to the geodesic can be calculated
First, the PCA method is used to produce the feature vector through
corresponding to first d eigenvalues. Descending eigenvalues
xk = xT φ(k), k ∈ (0, 1) (4)
sort the covariance matrix as a set of subspaces, which are
denoted as PS and PT , respectively. PS ∈ Rn×d , PT ∈ Rn×d . where x ∈ Rnis the original data (training data or testing
RS ∈ Rn×(nd) is the orthogonal complement of PS . data), and xk is the feature data on the middle subspace where
The geodesic from the source domain to the target domain x corresponds to the position p on the geodesic flow (e.g., A0.6
can be represented by φ(k) with φ(0) = PS and φ(1) = PT in Fig. 1). The number of subspaces p is predefined.

Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on January 24,2024 at 06:18:07 UTC from IEEE Xplore. Restrictions apply.
3422 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 51, NO. 7, JULY 2021

Process 3 Transfer constructing the geodesic flow spends O(d2 ) and O(1), respec-
Input: The elite individuals: LastBestSol; tively. The computational cost of finding solutions by using
The dynamic optimization function: F(x, t); the interior-point algorithm consumes O(m3 n) [6]. Because
Number of individuals: N; the number of clusters is a constant set by the user, therefore,
Clustering number: L. the computational complexity of Transfer is O(m3 n). The total
Output: Predicted individuals: TransSol. computational complexity of the proposed prediction is thus
1: Initialization: dimension d = m − 1; TransSol = ∅; O(m3 n).
2: Clustering LastBestSol: LastBestSol1 , ..., LastBestSolL by
LPCA; IV. E XPERIMENTS
3: for j = 1 to L do
In order to verify the effectiveness of MMTL-DMOEA, we
4: Use PCA for LastBestSolj to get PS ;
carried out three sets of experiments. The first experiment
5: Generate set T containing N individuals of F(x, t);
compares the proposed MMTL-DMOEA with the state-of-the-
6: Use PCA for T to get PT ∈ Rd×N ;
art DMOAs. The second experiment verifies the effectiveness
7: Construct the geodesic flow φ(k) = PS U1 (k) −
of integrating the components of memory and manifold TL.
RS U2 (k), k ∈ (0, 1);
The third experiment compares with Tr-DMOEA [6] and
8: for x ∈ LastBestSolj do
analyzes performance improvement.
9: Project x to φ(·) and get x;
10: x = arg minx xT φ(·) − x;
11: TransSol = x ∪ TransSol; A. Comparison With Other DMOPs
12: end for The test functions used are FDA series [38], dMOP
13: end for series [14], and DF series [39]. The time parameter in the
14: return TransSol; test functions is t = (1/nt )τ/τt , where nt , τt , and τ are the
severity of the time change, the frequency of change, and the
maximum generation of the problems, respectively.
The algorithms used for comparison in this experiment are
The detailed steps for Transfer are shown in Process 3.
as follows: MOEA/D-SVR [40], Tr-DMOEA [6], MOEA/D-
We believe that the estimated nondominated solutions
KF [23], PPS [22], and the baseline algorithm MOEA/D [24],
LastBestSol can be viewed as m − 1-dimensional seg-
which are modified as RI-MOEA/D to adapt to dynamic
mented manifolds [37], where m is the number of objec-
change, that is, 10% of the population is randomly reinitial-
tives. Hence, first, the LastBestSol is clustered by LocalPCA
ized when the environment changes. For a fair comparison,
(LPCA) [37] into m − 1-dimensional L segmented manifolds:
the baseline algorithm in all the compared algorithms are
LastBestSol1 , . . . , LastBestSolL in line 2. For each cluster,
replaced by MOEA/D [24], and these compared algorithms
the geodesic flow φ(·) are constructed in lines 4–7. Then,
are denoted as KT-MOEA/D, SVR-MOEA/D, Tr-MOEA/D,
x ∈ LastBestSolj is mapped into the geodesic flow in line 9,
KF-MOEA/D, and PPS-MOEA/D, respectively, and most of
and we can obtain the mapped data x. In the dynamic environ-
the parameters of these algorithms are set according to their
ment, the manifolds LastBestSolj in the source domain may
original references.
be similar to those in the target domain, so the mapped tar-
There are three metrics used in this group of comparison
get domain data xT φ(·) is similar to x. Therefore, in line 10,
experiments, and three metrics are listed as follows.
we find a solution x , such that the mapped data xT φ(·) on
Inverted generational distance (IGD) measures the conver-
the geodesic flow is closet to x. This is a single-objective
gence and diversity of the solutions. A smaller value of IGD
optimization problem, and any single-objective optimization
indicates a better convergence of the solution obtained by
algorithm can be applied to solve the problem. In this article,
the algorithm and the higher the diversity. This metric is
we use the interior-point method to solve the problem.
computed as
1 
A. Computational Complexity IGD = min p − p∗2 (5)
nPt t
p∗ ∈P∗
p∈P
In the FindBestSol, constructing the SVR estimator with ns
samples needs O(n2s d), where d is the dimension of decision where represents the true POF of the problem, P∗ represents
Pt
variables. Using SVR to estimate the objective values con- the obtained POF by testing algorithms, and nPt represents the
sumes O(N 2 d), where N is the population size. According to number of solutions in the true POF.
the estimated objective values, the cost of finding nondomi- The MIGD metric is a variant of IGD and is defined as the
nated solutions by the fast nondominated sorting is O(N 2 m), mean IGD values in time steps
where m is the number of objectives. Because N  ns   1   
and d  m, therefore, the computational complexity of MIGD POF∗t , POFt = IGD POF∗t , POFt (6)
|T|
FindBestSol is O(N 2 d). t∈T
The computational cost of Transfer involves clustering, con- where T is a set of discrete-time points in a run and |T| is the
structing the geodesic flow, and finding solutions by using cardinality of T.
the interior-point algorithm. Clustering LastBestSol by LPCA Schott’s spacing metric (SP) measures the solution unifor-
calls for O(d2 ). For each cluster, mapping data via PCA and mity. A smaller SP value reflects a more uniform distribution

Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on January 24,2024 at 06:18:07 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: FAST DYNAMIC EVOLUTIONARY MULTIOBJECTIVE ALGORITHM VIA MANIFOLD TL 3423

TABLE I
M EAN AND S TANDARD D EVIATION VALUES OF MIGD M ETRIC O BTAINED BY C OMPARED A LGORITHMS FOR D IFFERENT DYNAMIC T EST S ETTINGS

of the solution obtained by the algorithm. This metric is these tables are the best metrics obtained by the compared
calculated as algorithm.
 The experimental results of Table I show that the MMTL-
nPOF∗
1  2 MOEA/D algorithm has achieved the best results in 20 out of
SP = Di − D (7)
nPOF − 1
∗ 24 test cases based on the MIGD values, indicating that the
i=1
optimal solution obtained by MMTL-MOEA/D is better than
where Di represents the Euclidean distance between the ith the existing algorithm in terms of convergence. Fig. 3 plots the
individual in POF∗ and its closest solution, and D represents IGD values obtained by different algorithms after each change.
the mean of Di . This metric is similarly modified to the IGD It shows that the curves obtained by the proposed method are
to act as a performance metric for evaluating DMOAs. at the bottom in most cases, and the curves of the proposed
Maximum spread (MS) measures the coverage extent of method are smoother, which means that the method not only
obtained solutions in the true POF. A larger MS value indi- performs better but also is more stable. Table II presents the
cates a higher coverage by the obtained solution in the real average SP metric values on the FDA and dMOP series of
POF. This metric is calculated as test functions. The results show that MMTL-MOEA/D is far
   ⎤2 better than the other five algorithms in 15 out of 24 test cases.
⎡
∗ ∗
1 ⎣ min Pk , Pk − max Pk , Pk ⎦
M t t This good performance shows that the distribution of solutions
MS = obtained by MMTL-MOEA/D is relatively uniform. Table III
M
k=1 Ptk − Ptk shows the average MS of the solutions obtained for the six
algorithms. The results show that the MMTL-MOEA/D has a
where Ptk and Ptk represent the maximum and minimum of the higher degree of coverage of the true POF than the other five
kth objective in true POF, respectively, and P∗k and P∗k rep- algorithms on most of the test cases.
resent the maximum and minimum of the kth objective, in To see whether the proposed prediction model could be
the obtained POF, respectively. This metric is similarly modi- integrated into other MOEAs, we also replace the baseline
fied to the IGD to act as a performance metric for evaluating SMOA by RM-MEDA [37] and these compared algorithms
DMOAs. are denoted as MMTL-RM-MEDA, SVR-RM-MEDA, Tr-RM-
The experiment parameters are set as follows. The size of MEDA, KF-RM-MEDA, and PPS-RM-MEDA, respectively.
population N is 100; and the external storage size C is set to The experimental results are presented in Tables I and II in the
10 × N. Most of the parameters of the SVR in FindBsetSol supplementary material. From these tables, we can obverse that
are set by default [41]; the number of sampling individuals ns our proposed model is still competitive in terms of MIGD, SP,
in FindBsetSol is 30; the number of manifold segments L in and MS metric values. We also plot the IGD values obtained
Transfer is 4; and the number of intermediate subspaces p in by different algorithms in Fig. 1 in the supplementary material.
Transfer is 5. We conducted experiments on the DF series test functions
The experimental results are shown in Tables I–III. The data under the configuration of nt = 10 and τt = 10. All DF
in the table are the MIGD, SP, and MS values of compared functions have a changing POS, which belong to type I (POS
algorithms under different configurations. The bold data in changes, but POF does not change) or type II (POS and POF

Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on January 24,2024 at 06:18:07 UTC from IEEE Xplore. Restrictions apply.
3424 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 51, NO. 7, JULY 2021

TABLE II
M EAN AND S TANDARD D EVIATION VALUES OF SP M ETRIC O BTAINED BY C OMPARED A LGORITHMS FOR D IFFERENT DYNAMIC T EST S ETTINGS

Fig. 3. IGD values of DMOAs at the configuration (τt = 10 and nt = 10).

change as well) DMOPs and meet the Non-IID problems. B. Ablation Study
The statistical results of MIGD, MS, and SP are shown in The main contribution of MMTL-DMOEA is to use the
Table IV in the supplementary material. It can be observed optimal solution information obtained from the previous
that in 20 out of 42 cases, MMTL-MOEA/D performs the best moment to generate the initial population at the next moment.
among all the algorithms, eight for RI-MOEA/D, six for KF- Specifically, our method produces the initial population by
MOEA/D, four for SVR-MOEA/D, three for Tr-MOEA/D, and two different mechanisms: 1) memory-based one and 2) TL-
only one for PPS-MOEA/D. The comparative studies on DF based one. In order to verify the effectiveness of these two
test functions can also verify the effectiveness of our proposed mechanisms, we designed an ablation experiment.
MMTL-DMOEA in solving Non-IID problems. In this experiment, we propose two modified algorithms
From the experimental results, it is obvious that MMTL- by using only one out of the two mechanisms mentioned
MOEA/D performs not very well on the FDA5 problem. above. For the cases in which the memory or the TL mecha-
FDA5 is triobjective problem and the POS and the POF nism alone generates all individuals of the initial population,
change as well, so FDA5 has more complicated POF and we respectively refer the case as MMTL-MOEA/DM or
POS. Hence, clustering solutions FDA5 into low-dimensional MMTL-MOEA/DT . During the operation of the algo-
segmented manifolds meets challenges and leads to a poor rithm, if the number of the predicted initial individuals
performance of Transfer. less than the population size, the MMTL-MOEA/DM and

Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on January 24,2024 at 06:18:07 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: FAST DYNAMIC EVOLUTIONARY MULTIOBJECTIVE ALGORITHM VIA MANIFOLD TL 3425

TABLE III
M EAN AND S TANDARD D EVIATION VALUES OF MS M ETRIC O BTAINED BY C OMPARED A LGORITHMS FOR D IFFERENT DYNAMIC T EST S ETTINGS

TABLE IV
M EAN AND S TANDARD D EVIATION VALUES OF MIGD, SP, AND MS M ETRICS O BTAINED BY
MMTL-MOEA/D, MMTL-MOEA/DM , MMTL-MOEA/DT , AND RI-MOEA/D

MMTL-MOEA/DT both add random individuals to replenish we can find that when the initial population of the next
the population so as to ensure sufficient diversity of the initial moment is generated using the TL or the memory-based
population. mechanism, the metric of MIGD is significantly improved,
We compare these two algorithms with MMTL-MOEA/D compared with the RI-MOEA/D algorithm. It shows that when
and the baseline algorithm RI-MOEA/D. The settings of the problem environment changes, the two mechanisms can
experimental parameters are consistent with the above exper- adapt the evolution algorithm to the new environment faster
iments, and the FDA series and dMOP series test functions and eventually improve the convergence and diversity of the
are used. solution.
Table IV lists the average MIGD, MS, and SP metric val- At the same time, the comparison among the three
ues for the experiment. Comparing the three metrics of the columns of MMTL-MOEA/D, MMTL-MOEA/DM , and
MOEA/D, the MMTL-MOEA/DM , and the MMTL-MOEA/DT , MMTL-MOEA/DT show that the combination of the two

Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on January 24,2024 at 06:18:07 UTC from IEEE Xplore. Restrictions apply.
3426 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 51, NO. 7, JULY 2021

TABLE V
RUNNING T IME OF T WO A LGORITHMS ON D IFFERENT B ENCHMARK F UNCTIONS (U NIT: S ECONDS )

TABLE VI
RUNNING T IME OF DMOA S ON D IFFERENT F UNCTIONS (U NIT: S ECONDS )

mechanisms is superior to the use of only one mechanism. Tables I–IV in the supplementary material record the MIGD,
One thing we need to point out is that in this experi- SP, and MS values of the Tr-MOEA/D and MMTL-MOEA/D
ment, the MMTL-MOEA/DM algorithm is superior to the on FDA, dMOP, and DF problems. The above experimen-
MMTL-MOEA/DT algorithm. tal results show that the manifold transfer-based population
We think one of the reasons for this difference is that the prediction is effective in solving DMOPs. The reasons for the
MMTL-MOEA/DT is more sensitive to parameters, and it is superior performance of the proposed MMTL-MOEA/D are
possible to improve the performance of the MMTL-MOEA/DT two-fold. First, it compresses high-dimensional variables in a
by setting reasonable parameters, but how to find the optimal low-dimensional space, so MMTL-MOEA/D is more effective;
parameters is a tough problem. second, the proposed design can transfer information gradually
over geodesic flow so that it can improve its adaptability to
solution space under the new environment. Besides, from the
C. Compared With Tr-DMOEA previous ablation study, the memory-based mechanism and the
The MMTL-DMOEA stems from the framework of the Tr- TL-based mechanism can both help improve performance, so
DMOEA, therefore we want to prove that MMTL-DMOEA the combination of these two mechanisms can achieve better
outperforms the original algorithm in terms of the evolution performance in solving DMOPs.
speed and convergence. Both MMTL-MOEA/D and Tr-MOEA/D are prediction
In order to compare evolution speed, we recorded the run- methods based on TL. Compared with Tr-MOEA/D, one of the
ning time of the proposed Tr-MOEA/D and MMTL-MOEA/D differences is that Tr-MOEA/D requires a significant amount
under the same hardware configuration.1 Table V shows the of computation to obtain latent space to find optimal solutions,
mean running time of each environment of Tr-MOEA/D and while our MMTL-MOEA/D directly predicts the optimal solu-
MMTL-MOEA/D in solving the FDA series and dMOP series tions in the decision space. Therefore, MMTL-MOEA/D takes
test problems. The experimental results show that the proposed less time but achieves better performance.
method has greatly improved the speed of operation, and the
acceleration rate is hundreds of times.
Two reasons are accountable for this performance improve- D. Running Time Cost
ment. First, the TL method in the proposed method spent The good performance of MMTL-MOEA/D comes with
much less time than the Tr-MOEA/D algorithm. Second, a small price. We compare the mean running time cost of
MMTL-MOEA/D directly predicts the optimal solutions in each environment of different algorithms and present the
the decision space, while the Tr-MOEA/D method uses the results in Table VI. As shown in Table VI, KF-MOEA/D
POF to make the prediction. After locating the individual in and PPS-MOEA/D consume less time than the other com-
the objectives space, Tr-MOEA/D needs extra time to find the pared algorithms. MMTL-MOEA/D is faster than the other
corresponding individuals in the decision space. To ensure the compared algorithms on FDA5 and dMOP1 test functions. In
fairness of the comparison, we use the same settings mentioned other functions, MMTL-MOEA/D is only 1–2 s slower than
above to conduct the experiment. PPS-MOEA/D or KF-MOEA/D. But considering the good
performance of MMTL-MOEA/D in MIGD, SP and MS met-
1 The implementation environment is as follows: 1.6-GHz Intel Core i5, rics, MMTL-MOEA/D is still very competitive as compared
8-GB 1600-MHz DDR3. to other algorithms.

Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on January 24,2024 at 06:18:07 UTC from IEEE Xplore. Restrictions apply.
JIANG et al.: FAST DYNAMIC EVOLUTIONARY MULTIOBJECTIVE ALGORITHM VIA MANIFOLD TL 3427

V. C ONCLUSION [16] R. Chen, K. Li, and X. Yao, “Dynamic multiobjectives optimization with
a changing number of objectives,” IEEE Trans. Evol. Comput., vol. 22,
TL-based methods have shown their promising applica- no. 1, pp. 157–171, Feb. 2018.
tions in solving DMOPs. However, existing methods are faced [17] D. Gong, B. Xu, Y. Zhang, Y. Guo, and S. Yang, “A similarity-
with various challenges particularly on the speed of conver- based cooperative co-evolutionary algorithm for dynamic interval multi-
objective optimization problems,” IEEE Trans. Evol. Comput., vol. 24,
gence. To overcome the issue, a memory-driven manifold no. 1, pp. 142–156, Feb. 2020.
TL algorithm, called MMTL-DMOEA, has been proposed. [18] Y. Wang and B. Li, “Multi-strategy ensemble evolutionary algorithm for
The MMTL-DMOEA combines the manifold-based TL algo- dynamic multi-objective optimization,” Memetic Comput., vol. 2, no. 1,
pp. 3–24, 2010.
rithm with a memory-based prediction method to generate [19] M. Rong, D. Gong, Y. Zhang, Y. Jin, and W. Pedrycz, “Multidirectional
the initial population during the evolution. We conducted sys- prediction approach for dynamic multiobjective optimization problems,”
tematic experiments and showed that the solutions obtained IEEE Trans. Cybern., vol. 49, no. 9, pp. 3362–3374, Sep. 2019.
[20] W. T. Koo, C. K. Goh, and K. C. Tan, “A predictive gradient strategy for
by the proposed MMTL-DMOEA are better than state-of-the- multiobjective evolutionary algorithms in a fast changing environment,”
art algorithms in terms of convergence, diversity, uniformity, Memetic Comput., vol. 2, no. 2, pp. 87–110, 2010.
and convergence speed in particular. The integration of both [21] M. Rong, D. Gong, W. Pedrycz, and L. Wang, “A multi-model prediction
method for dynamic multi-objective evolutionary optimization,” IEEE
memory and manifold TL can retain the effective strategy of Trans. Evol. Comput., vol. 24, no. 2, pp. 290–304, Apr. 2020.
elitism in traditional EA, while at the same time making full [22] A. Zhou, Y. Jin, and Q. Zhang, “A population prediction strategy for evo-
use of the advantages in TL. lutionary dynamic multiobjective optimization,” IEEE Trans. Cybern.,
vol. 44, no. 1, pp. 40–53, Jan. 2014.
[23] A. Muruganantham, K. C. Tan, and P. Vadakkepat, “Evolutionary
dynamic multiobjective optimization via Kalman filter prediction,” IEEE
R EFERENCES Trans. Cybern., vol. 46, no. 12, pp. 2862–2873, Dec. 2016.
[24] Q. Zhang and H. Li, “MOEA/D: A multiobjective evolutionary algorithm
[1] R. Azzouz, S. Bechikh, and L. B. Said, “Dynamic multi- based on decomposition,” IEEE Trans. Evol. Comput., vol. 11, no. 6,
objective optimization using evolutionary algorithms: A survey,” in pp. 712–731, Dec. 2007.
Recent Advances in Evolutionary Multi-Objective Optimization. Cham, [25] J. Zou, Q. Li, S. Yang, H. Bai, and J. Zheng, “A prediction strat-
Switzerland: Springer, 2017, pp. 31–70. egy based on center points and knee points for evolutionary dynamic
[2] Z. Yang, Y. Jin, and K. Hao, “A bio-inspired self-learning coevolutionary multi-objective optimization,” Appl. Soft Comput., vol. 61, pp. 806–818,
dynamic multiobjective optimization algorithm for Internet of Things Dec. 2018.
services,” IEEE Trans. Evol. Comput., vol. 23, no. 4, pp. 675–688, [26] B. Xu, Y. Zhang, D. Gong, Y. Guo, and M. Rong, “Environment
Aug. 2019. sensitivity-based cooperative co-evolutionary algorithms for dynamic
[3] J. Xu, P. B. Luh, F. B. White, E. Ni, and K. Kasiviswanathan, “Power multi-objective optimization,” IEEE/ACM Trans. Comput. Biol.
portfolio optimization in deregulated electricity markets with risk Bioinformat., vol. 15, no. 6, pp. 1877–1890, Nov. 2018.
management,” IEEE Trans. Power Syst., vol. 21, no. 4, pp. 1653–1662, [27] Y. Woldesenbet and G. Yen, “Dynamic evolutionary algorithm with vari-
Nov. 2006. able relocation,” IEEE Trans. Evol. Comput., vol. 13, no. 3, pp. 500–513,
[4] A. Simões and E. Costa, “Prediction in evolutionary algorithms for Jun. 2009.
dynamic environments,” Soft Comput, vol. 18, no. 8, pp. 1471–1497, [28] A. Gupta, Y. Ong, and L. Feng, “Insights on transfer optimization:
Oct. 2013. Because experience is the best teacher,” IEEE Trans. Emerg. Topics
[5] C. Rossi, M. Abderrahim, and J. C. Díaz, “Tracking moving optima Comput. Intell., vol. 2, no. 1, pp. 51–64, Feb. 2018.
using Kalman-based predictions,” Evol. Comput., vol. 16, no. 1, [29] L. Feng et al., “Evolutionary multitasking via explicit autoencoding,”
pp. 1–30, Mar. 2008. IEEE Trans. Cybern., vol. 49, no. 9, pp. 3457–3470, Sep. 2019.
[6] M. Jiang, Z. Huang, L. Qiu, W. Huang, and G. G. Yen, “Transfer [30] L. Zhou et al., “Toward adaptive knowledge transfer in multifactorial
learning-based dynamic multiobjective optimization algorithms,” IEEE evolutionary computation,” IEEE Trans. Cybern., early access, Mar. 6,
Trans. Evol. Comput., vol. 22, no. 4, pp. 501–514, Aug. 2018. 2020, doi: 10.1109/TCYB.2020.2974100.
[7] S. J. Pan, I. W. Tsang, J. T. Kwok, and Q. Yang, “Domain adaptation [31] K. K. Bali, Y. Ong, A. Gupta, and P. S. Tan, “Multifactorial evolutionary
via transfer component analysis,” IEEE Trans. Neural Netw., vol. 22, algorithm with online transfer parameter estimation: MFEA-II,” IEEE
no. 2, pp. 199–210, Feb. 2011. Trans. Evol. Comput., vol. 24, no. 1, pp. 69–83, Feb. 2020.
[32] B. Da, A. Gupta, and Y. Ong, “Curbing negative influences online
[8] T. Lin and H. Zha, “Riemannian manifold learning,” IEEE Trans. Pattern
for seamless transfer evolutionary optimization,” IEEE Trans. Cybern.,
Anal. Mach. Intell., vol. 30, no. 5, pp. 796–809, May 2008.
vol. 49, no. 12, pp. 4365–4378, Dec. 2019.
[9] R. Gopalan, R. Li, and R. Chellappa, “Domain adaptation for object [33] W. Zhou et al., “A preliminary study of adaptive indicator based
recognition: An unsupervised approach,” in Proc. IEEE Int. Conf. evolutionary algorithm for dynamic multiobjective optimization via
Comput. Vis. (ICCV), 2011, pp. 999–1006. autoencoding,” in Proc. IEEE Congr. Evol. Comput. (CEC), 2018,
[10] K. Deb, U. N. B. Rao, and S. Karthik, “Dynamic multi-objective pp. 1–7.
optimization and decision-making using modified NSGA-II: A case [34] A. T. W. Min, Y. Ong, A. Gupta, and C. Goh, “Multiproblem surrogates:
study on hydro-thermal power scheduling,” in Proc. Int. Conf. Evol. Transfer evolutionary multiobjective optimization of computationally
Multi Crit. Optim., 2007, pp. 803–817. expensive problems,” IEEE Trans. Evol. Comput., vol. 23, no. 1,
[11] Q. Chen, J. Ding, S. Yang, and T. Chai, “A novel evolution- pp. 15–28, Feb. 2019.
ary algorithm for dynamic constrained multiobjective optimization [35] X.-F. Liu et al., “Neural network-based information transfer for dynamic
problems,” IEEE Trans. Evol. Comput., early access, Dec. 6, 2019, optimization,” IEEE Trans. Neural Netw. Learn. Syst., early access,
doi: 10.1109/TEVC.2019.2958075. Jul. 19, 2019, doi: 10.1109/TNNLS.2019.2920887.
[12] G. Ruan, G. Yu, J. Zheng, J. Zou, and S. Yang, “The effect of diver- [36] M. Jiang, L. Qiu, Z. Huang, and G. G. Yen, “Dynamic multi-objective
sity maintenance on prediction in dynamic multi-objective optimization,” estimation of distribution algorithm based on domain adaptation and
Appl. Soft Comput., vol. 58, pp. 631–647, Sep. 2017. nonparametric estimation,” Inf. Sci., vol. 435, pp. 203–223, Apr. 2018.
[13] S. Jiang and S. Yang, “A steady-state and generational evolutionary [37] Q. Zhang, A. Zhou, and Y. Jin, “RM-MEDA: A regularity model-based
algorithm for dynamic multiobjective optimization,” IEEE Trans. Evol. multiobjective estimation of distribution algorithm,” IEEE Trans. Evol.
Comput., vol. 21, no. 1, pp. 65–82, Feb. 2017. Comput., vol. 12, no. 1, pp. 41–63, Feb. 2008.
[14] C.-K. Goh and K. C. Tan, “A competitive-cooperative coevolutionary [38] M. Farina, K. Deb, and P. Amato, “Dynamic multiobjective optimization
paradigm for dynamic multiobjective optimization,” IEEE Trans. Evol. problems: Test cases, approximations, and applications,” IEEE Trans.
Comput., vol. 13, no. 1, pp. 103–127, Feb. 2009. Evol. Comput., vol. 8, no. 5, pp. 425–442, Oct. 2004.
[15] D. Yazdani, M. N. Omidvar, J. Branke, T. T. Nguyen, and X. Yao, [39] S. Jiang, S. Yang, X. Yao, K. Tan, M. Kaiser, and N. Krasnogor,
“Scaling up dynamic optimization problems: A divide-and-conquer “Benchmark problems for cec2018 competition on dynamic
approach,” IEEE Trans. Evol. Comput., vol. 24, no. 1, pp. 1–15, multiobjective optimisation,” in Proc. CEC2018 Competition, 2018,
Feb. 2020. pp. 1–8.

Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on January 24,2024 at 06:18:07 UTC from IEEE Xplore. Restrictions apply.
3428 IEEE TRANSACTIONS ON CYBERNETICS, VOL. 51, NO. 7, JULY 2021

[40] L. Cao, L. Xu, E. D. Goodman, C. Bao, and S. Zhu, “Evolutionary Shihui Guo (Member, IEEE) received the B.S.
dynamic multiobjective optimization assisted by a support vector regres- degree in electrical engineering from Peking
sion predictor,” IEEE Trans. Evol. Comput., vol. 24, no. 2, pp. 305–319, University, Beijing, China, in 2010, and the Ph.D.
Apr. 2020. degree in computer animation from the National
[41] C.-C. Chang and C.-J. Lin, “LIBSVM: A library for support vector Centre for Computer Animation, Bournemouth
machines,” ACM Trans. Intell. Syst. Technol., vol. 2, no. 3, pp. 1–27, University, Poole, U.K., in 2015.
May 2011. He is an Associate Professor with the School of
Informatics, Xiamen University, Xiamen, China. His
research interests include computational intelligence
and human–computer interaction.

Min Jiang (Senior Member, IEEE) received the

bachelor’s and Ph.D. degrees in computer science
from Wuhan University, Wuhan, China, in 2001
and 2007, respectively.
He was a Postdoctoral Fellow with the
Department of Mathematics, Xiamen University,
Xiamen, China, where he is currently a Professor
with the Department of Artificial Intelligence.
His main research interests are machine learning,
computational intelligence, and robotics. He
has a special interest in dynamic multiobjective
optimization, transfer learning, and the software development and in the Xing Gao (Member, IEEE) received the Ph.D.
basic theories of robotics. degree in computer science from the Harbin Institute
Prof. Jiang received the Outstanding Reviewer Award from the IEEE of Technology, Harbin, China, in 2009.
T RANSACTIONS ON C YBERNETICS in 2016. He is the Chair of the IEEE He is currently an Associate Professor with the
CIS Xiamen Chapter. He is currently serving as an Associate Editor for the School of Informatics, Xiamen University, Xiamen,
IEEE T RANSACTIONS ON N EURAL N ETWORKS AND L EARNING S YSTEMS China. His research interests include computing
and the IEEE T RANSACTIONS ON C OGNITIVE AND D EVELOPMENTAL intelligence and computer graphics.
S YSTEMS.

Zhenzhong Wang received the bachelor’s degree in

computer science and technology from Northeastern
University, Shenyang, China, in 2017. He is cur-
rently pursuing the master’s degree with the School
of Informatics, Xiamen University, Xiamen, China.
His research interests include computational intel-
ligence and machine learning.
Kay Chen Tan (Fellow, IEEE) received the B.Eng.
degree (First Class Hons.) in electronics and electri-
cal engineering and the Ph.D. degree in evolutionary
computation and control systems from the University
Liming Qiu received the bachelor’s degree from of Glasgow, Glasgow, U.K., in 1994 and 1997,
Fujian Agriculture and Forestry University, Fuzhou, respectively.
China, in 2015, and the master’s degree from the He is a Full Professor with the Department of
School of Informatics, Xiamen University, Xiamen, Computer Science, City University of Hong Kong,
China, in 2018. Hong Kong. He has published over 200 refereed arti-
He is currently an Algorithm Engineer with cles and six books.
LenzTech, Beijing, China. His research interests Prof. Tan is the Editor-in-Chief of the IEEE
include computational intelligence and machine T RANSACTIONS ON E VOLUTIONARY C OMPUTATION, was the Editor-in-
learning. Chief of the IEEE Computational Intelligence Magazine from 2010 to 2013,
and currently serves as the Editorial Board Member of over ten journals. He
was an Elected Member of IEEE CIS AdCom from 2017 to 2019.

Authorized licensed use limited to: BEIHANG UNIVERSITY. Downloaded on January 24,2024 at 06:18:07 UTC from IEEE Xplore. Restrictions apply.