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

fully edited. Content may change prior to final

publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 1

A Novel Dynamic Multiobjective Optimization

Algorithm with Hierarchical Response System
Han Li, Zidong Wang, Fellow, IEEE, Chengbo Lan, Peishu Wu, and Nianyin Zeng

Abstract—In this paper, a novel dynamic multi-objective opti- behaviors, the DMOAs are required to timely update the
mization algorithm (DMOA) is proposed based on a designed obtained Pareto solutions to ensure the convergence in each
hierarchical response system (HRS). Named as HRS-DMOA, environment. In order to track the time-varying Pareto front,
the proposed algorithm mainly aims at integrating merits from
the mainstream ideas of dynamic behavior handling (i.e., the it is of vital significance for the DMOAs to make effective
diversity-, memory-, and prediction-based methods) so as to make responses to the environmental changes, which is quite a
flexible responses to environmental changes. In particular, by two challenging issue.
pre-defined thresholds, the environmental changes are quantified To address above basic and important issue in handling
as three levels. In case of a slight environmental change, the DMOPs, the existing DMOAs can be generally divided into the
previous Pareto set-based refinement strategy is recommended,
while the diversity-based re-initialization method is applied in memory- [45], [59], prediction- [42], [48]–[50], [57], [70], and
case of a dramatic environmental change. For changes occur- diversity-based methods [12], [35], [36]. Additionally, some
ring in a medium level, the transfer-learning-based response is hybrid algorithms have also been proposed [8], [21], [31], [53],
adopted to make full use of the historical searching experiences. [69], and in particular, a novel trend of developing DMOAs
The proposed HRS-DMOA is comprehensively evaluated on a has emerged recently, which combines the memory mechanism
series of benchmark functions, and the results show an improved
comprehensive performance as compared with four popular and the prediction method, where the transfer learning (TL)
baseline DMOAs in terms of both convergence and diversity, technique has been adopted to make full use of history knowl-
which also outperforms other two state-of-the-art DMOAs in edge to accelerate convergence in a new environment, see [25],
10 out of 14 testing cases, exhibiting the competitiveness and [26], [33], [39], [55], [60] for some successful applications.
superiority of the algorithm. Finally, extensive ablation studies For a clear inspection of above mainstream algorithms, their
are carried out, and from the results, it is found that as compared
with randomly selecting the response methods, the proposed HRS main ideas are summarized in Table I (see Section II-B for
enables more reasonable and efficient responses in most cases. more discussions of their advantages and disadvantages).
In addition, the generalization ability of the proposed HRS as
a flexible plug-and-play module to handle dynamic behaviors is TABLE I: Overview of the mainstream DMOAs
proven as well.
Mainstream algorithms Main idea
Index terms— Dynamic multiobjective optimization algo- Enrich or maintain diversity of
rithm (DMOA); transfer learning (TL); hierarchical response Diversity-based
the population for sufficient search
system; evolutionary algorithm Recall the useful historical Pareto
Memory-based
solutions in the new environment
I. I NTRODUCTION Predict the varying Pareto set
Prediction-based
in advance to accelerate convergence
Dynamic multiobjective optimization problems (DMOPs) Apply several different strategies
composed of conflicting objective functions are inevitably Hybrid methods
to collaboratively handle changes
encountered in many real-world scenes [9], [14], [62], [66],
[67], where both the objectives and constraints may change It should be highlighted that all of the existing mainstream
with time [5], [12], and this has attracted wide research DMOAs have already been proven effective in many cases,
attention to the design of effective dynamic multiobjective which are popular and reliable in coping with the environ-
optimization algorithms (DMOAs) [6], [15], [20], [32], [34], mental changes. In [35], a novel coevolutionary multi-swarm
[44]. The population-based evolutionary algorithms have been particle swarm optimizer has been proposed to solve DMOPs,
proven to be effective under various optimization scenarios where once a change is detected, 20% of the swarm is gen-
in searching for the optimal solutions [2], [4], [7], [38], [40], erated randomly to enhance the population diversity, and the
[54], [56]. Particularly, owing to the wide existence of dynamic experimental results have shown that the proposed algorithm
This work was supported in part by the Natural Science Foundation of performs well in the rapidly changing environment. In [45], an
China under Grant 62073271, the National Science and Technology Major explicit memory has been adopted to store non-dominated so-
Project of China under Grant J2019-I-0013-0013, the Independent Innovation lutions, where a novel minimum distance search-based updat-
Foundation of AECC of China under Grant ZZCX-2018-017, and the Korea
Foundation for Advanced Studies. (Corresponding author: Nianyin Zeng) ing technique is used. When there is an environmental change,
H. Li, C. Lan, P. Wu and N. Zeng are with the Department of Instrumental the stored solutions are reused in later stages for population
and Electrical Engineering, Xiamen University, Fujian 361005, China. Email: initialization, and from the results, the proposed algorithm has
[email protected]
Z. Wang is with the Department of Computer Science, Brunel University shown its competitiveness in tracking the true Pareto front.
London, Uxbridge UB8 3PH, U.K. Email: [email protected] An individual-based self-learning prediction method has been
Copyright © 2023 Institute of Electrical and Electronics Engineers (IEEE). Personal use of this material is permitted. Permission from IEEE must be
obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes,
creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works by
sending a request to [email protected]. For more information, see https://www.ieee.org/publications/rights/rights-policies.html
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 2

proposed in [42], whose major innovation is the employment to some other algorithms.
of adjustable reference points, which can effectively alleviate A noticeable issue is that the most existing methods have
the situation of inaccurate prediction caused by the non- paid a great many of attention to designing effective response
uniform Pareto front. From the experiments, it is found that strategies, while the in-depth analysis is relatively limited on
the proposed algorithm can well balance the convergence the dynamic behaviors (e.g., the changing severity), whose
and diversity. Similarly, an adaptive reference vector-based precise characterization is conducive to tracking the varying
adjustment strategy has been introduced in [70] along with a Pareto front [18], [46]. Even in those emerging hybrid al-
linear prediction strategy, whose effectiveness is demonstrated gorithms that have applied different novel response methods,
on twelve functions with diverse dynamic characteristics. few attention has been paid to the quantification of the envi-
Aiming at recovering diversity in a short time, the authors ronmental changes, which motivates us to cover this research
in [21] have devised a subspace-based diversity maintenance gap. Additionally, in case of dynamic behaviors, it would be
strategy, which can identify gaps between population distri- tough to figure out the useful historical experiences in the
bution to maintain the diversity regardless of the environ- absence of further analysis about response-making; in cases
mental changes. Moreover, another layered prediction strategy of some slight changes, introducing the new individuals seems
has been proposed in [21], which benefits making prompt unnecessary and moreover, the prediction-based responses
responses and improving the accuracy of predicted evolution- might lead the evolution to an inappropriate direction when
ary direction. Similarly, a hybrid DMOA of prediction- and dramatic changes exist in the environment. In this regard, the
diversity-based strategy has been developed in [69], where it is DMOAs should be sensitive to the environmental changes to
regarded that too much reliance on the prediction method may enable both timely and appropriate responses. Accounting for
reduce the diversity. Therefore, the authors have employed the respective merits of the previously mentioned methods,
several mechanisms to generate different seed sets, which a promising way to endow the DMOAs with comprehensive
obtains satisfactory results on series of benchmark evalua- performances is to make efficient and rational integration of
tions, and the proposed algorithm is also promising to handle diverse response mechanisms.
the unpredictable DMOPs. Additionally, the memory- and Motivated by above discussions, in this study, we aim to
prediction-based strategy have been combined in [31], where design a novel DMOA capable of quantitatively measuring
the similarity among the environmental changes is compared. the environmental changes so as to take appropriate response
If the current detected change is dissimilar to any historical strategies accordingly. To be specific, in case of a slight en-
records, then individuals in the new environment are predicted vironmental change, the previous Pareto set-based refinement
based on the differential population center in the previous two strategy is recommended to save unnecessary computations
environments. Otherwise, memory mechanism is adopted to and track the almost unchanged Pareto front. On the contrary,
deal with the similar changes, and the experimental results in case of a dramatic change, an intuitive idea (of introducing
have shown the robustness of the proposed hybrid strategy. individuals in the new environment) is adopted due to the fact
Recently, more and more advanced DMOAs have been that searching for Pareto solutions in a totally different envi-
proposed. To cope with the irregular patterns in stochastic ronment can be directly regarded as solving a new problem. In
changes, a Mahalanobis distance-based approach has been addition to above two extreme cases, the TL-based response
developed in [22] to estimate the correlation between the strategy is applied to accelerate the convergence, whose main
current environment and the previous ones, which performs idea is to transfer useful history knowledge to provide a
significantly better than some latest algorithms when handling high-quality initial population in the new environment. It is
the stochastic DMOPs. In [61], a knowledge-guided Bayesian noticeable that different from those hybrid methods where
(KGB) classification method has been proposed to make several types of advanced strategies have been designed, the
robust prediction, which follows the same idea of TL-based main innovation of this study is to rationally adopt appropriate
DOMAs to sufficiently exploit the history information, and responses based on the quantification of the changes, which is
the effectiveness of the proposed KGB is demonstrated on supposed to encourage and promote more in-depth investiga-
different test suites. tions of the dynamic behaviors in DMOPs.
In [60], the authors have developed a clustering difference- In this paper, a novel dynamic multi-objective optimization
based TL method to solve DMOPs, where the k-means al- algorithm is proposed based on a hierarchical response system
gorithm is applied to divide the population into five clusters, (HRS), where the environmental changes are quantified and
whose centroid is then adopted to construct the target domain divided into three levels. The proposed algorithm (named
by the first-order difference. By increasing the similarity be- as the HRS-DMOA) is, in essence, a two-stage algorithm
tween source and target domains, the phenomenon of negative where 1) half of the non-dominated solutions in the previous
transfer is expected to be alleviated. For the same purpose, environment are firstly selected as the sensors to quantify the
a knowledge reconstruction method has been proposed in dynamic behaviors; and accordingly, based on two predefined
[17], where the fuzzy neural network is applied to extract thresholds, 2) three response strategies can be later adopted
domain knowledge from two successive Pareto sets, which to generate the initial population in the new environment to
is evaluated and screened via a pre-designed mechanism. accelerate the convergence. In particular, the diversity-based
Consequently, the suitable knowledge can be selected to guide strategy is adopted in case of severe changes, and when the
the evolution, and the proposed algorithm has presented both changes are in a medium level, the novel frontier of TL-based
better convergence and diversity performance in comparison response method is considered, which combines the popular
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 3

ideas of memory and prediction in solving DMOPs. It should A. Formulation of DMOP

be pointed out that the flexible response strategies can make Without loss of generality, a minimized DMOP is defined
the proposed HRS-DMOA adaptive to different situations, as:
which enables the responses to have little blindness so as to (
efficiently handle the dynamic behaviors. min F (x, t) = {f1 (x, t), f2 (x, t), ..., fm (x, t)}
(1)
The major contributions of this paper are listed as follows. s.t. g(x, t) ≤ 0, h(x, t) = 0
1. The dynamic behaviors in the DMOPs are characterized where x ∈ Rn is the decision vector, F (·) is a set of
both qualitatively and quantitatively in the proposed HRS- m objective functions, and g(·) and h(·) are inequality and
DMOA. equality constraints, respectively. t is the time variable that
2. Advantages complementation of the mainstream response can occur in both objectives and constraints. Based on the
methods in existing DMOAs is realized in the developed above formulation, some definitions are provided as follows.
hierarchical response system. Definition 1: Dynamic Pareto domination.
3. The established flexible response modes can endow the At time t, the decision vector x1 dominates x2 (denoted as
proposed algorithm with adaptivity (to diverse situations) and x1 ≺t x2 ) only in conditions of:
strong generalization ability. (
4. The proposed HRS is proven a reliable and effective plug- ∀ i ∈ {1, 2, ..., m}, fi (x1 , t) ≤ fi (x2 , t)
(2)
and-play module to solve DMOPs, which can be integrated ∃ j ∈ {1, 2, ..., m}, fj (x1 , t) < fj (x2 , t)
with different static optimizer.
Definition 2: Dynamic Pareto set (PS).
The remainder of this paper is organized as follows. Some
Pareto set at time t (denoted as P St ) is composed of current
preliminaries of this work are provided in Section II. The pro-
Pareto solutions x that cannot be dominated by any other
posed HRS-DMOA is elaborated in Section III, experimental
decision vector x′ ∈ Rn , which is defined as
results and discussions are presented in Section IV. Finally,
conclusions are drawn in Section V. P St = {x∗ |¬∃ x ∈ Rn , x ≺t x∗ } (3)
Definition 3: Dynamic Pareto front (PF).
II. P RELIMINARIES Pareto front at time t (denoted as P Ft ) can be obtained via
mapping P St to the objective space:
In this section, some preliminaries related to this study are
P Ft = {F (x, t)|x ∈ P St } (4)
provided. To make this paper easy to follow, annotations of
the frequently used symbols are presented in Table II.
B. Responses to Dynamic Behaviors
TABLE II: Major symbols and corresponding annotations
The existence of the dynamic behaviors makes it difficult to
Symbols Annotations handle DMOPs, and thus has aroused great research interests
x, z Decision variable on how to make effective responses to the environmental
F Objective function set changes. When the population enters a new environment,
fi The i-th objective diversity enhancement is a natural response manner. In [12],
population initialization in the new environment has been real-
g (·) Inequality constraints
ized by random generation and Gaussian mutation. As a result,
h (·) Equality constraints
partial previous individuals in the population are replaced by
m The number of objective functions
the newly generated ones so that the diversity is increased.
n The dimension of decision space Another intuitive response strategy is to directly adopt the
t Time variable obtained Pareto solutions in the previous environment (as the
nt Change severity initial population) to search for the updated Pareto solution
τt Change frequency in the new environment such that population state is main-
τ The maximum generation tained and no extra computations are required for subsequent
P St Pareto set at time t evolvement. It is worth mentioning that the above methods
P Ft Pareto front at time t are somehow blind due to their common procedure of simply
CD Change degree re-starting the search without analysis on the environmental
LT Lower-threshold of CD changes, which may lead to poor convergence. Though, in
HT Higher-threshold of CD some situations with dramatic environmental changes, the
Ds Source domain
large diversity in the population is able to facilitate a thorough
search in the new environment.
Dt Target domain
The memory-based methods aim to store and recall the his-
Q (·) Quality factor
torical searching experiences to form the initial population in
Pini Initialized population
the new environment [59], and thus are feasible to solve those
N Population size periodically changing DMOPs. In dynamic situations without
the periodic changes, it is hard to realize rapid but efficient
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 4

responses with the stored solutions in the memory, which mo- A. Environmental Change Quantification
tivates the development of the prediction-based methods due As previously mentioned, one of the most important issues
to correlations between consecutive environments may benefit in handling DMOPs is to make effective responses to the
further exploration on the stored solutions. For example, in dynamic behaviors therein, of which the premise is to detect
[72] and [68], the knee- and center-points of the population the environmental changes. The popular change detection
have been employed to predict the trace of PS, respectively. manner contains the population- and sensor-based methods [1],
In the predictive strategy proposed in [28], different types [44], and in this study, we hold the belief that it is important to
of special points have been sufficiently utilized with three not only detect but also quantify the environmental changes,
mechanisms. An ensemble learning-based prediction strategy which is conducive to adopting appropriate response strategies.
has been proposed in [52], which aims at overcoming the In this regard, half of the non-dominated solutions in the
shortcomings of inaccurate and unstable prediction. Similarly, previous environment are selected to form the sensor set S,
several sub-prediction models have been integrated in [16], which is used to estimate the change degree (denoted by CD)
and the proposed ensemble method is applied to estimate of objective function as [46]:
the fitness of individuals in the new environment, which has
X fi,j (t) − fi,j (t − 1)
presented better robustness than the single prediction model. CDi = , i = 1, 2, ..., m (5)
Above methods have concentrated on developing various novel fi,j (t − 1) + µ
j∈S
prediction strategies, which can accelerate the convergence in
the new environment and accordingly save some searching where m is the number of objective functions, fi,j (t) denotes
efforts, while as pointed out in some work, the prediction the i-th fitness value of sensor j in environment t, and
model may also suffer from the inaccurate tracking of PS and µ = 0.001 is a smoothing value that avoids the denominator
the extra computational costs [16], [52], especially when the equaling to zero. Then, the overall change degree of the
change is either too severe or slight. environment is defined as:
Considering that those historical experiences may still have CD = λ max {CDk } (6)
1≤k≤m
some reference values to search in the new environment,
the TL technique has paved a feasible path in solving the where λ is an amplification factor, which is set to m − 1 as
DMOPs by combining the memory- and prediction-based the same in [46].
strategy, whose idea is to transfer the useful knowledge to
assist learning in similar tasks. Under the assumption that there
B. Hierarchical Response System
may be inherent associations among individuals in successive
environments, developing TL-based DMOAs is an emerging In the proposed HRS-DMOA, the essence of response is to
and promising direction [51], where an important concern is re-initialize the population to start searching in a new environ-
to alleviate the negative transfer phenomenon, which may lead ment, where the major concern is how to improve reliability
the evolution towards wrong directions. of the initialized population. Hence, a hierarchical response
Based on above discussions, it is clear that each kind mode is taken to realize tailored and appropriate reactions
of the mainstream strategy in solving DMOPs has the own to environmental changes in different extents. To be specific,
advantages and disadvantages without a certain one always based on the change degree CD of environment in Eq. (6), two
being the best choice. More importantly, it seems feasible predefined thresholds LT and HT are used to divide CD into
and promising to realize the complementary performance by three levels, and accordingly, the response modes of refinement
quantifying the environmental changes. As previously men- (if CD < LT ), TL (if LT ≤ CD ≤ HT ), and re-initialization
tioned, if the environment is dramatically changed, it might be (if CD > HT ) are adopted.
better to re-initialize the population to restart searching than When CD is fewer than the lower-threshold LT , it is
to figure out the useful history solutions to accomplish the deemed that the change is negligible, that is, the new envi-
initialization. On the contrary, in case of a very slight change, ronment (denoted as t) is similar to the previous one (denoted
it might be unnecessary to apply any advanced prediction as t − 1). Therefore, P St−1 is encouraged to keep improving
models. Consequently, in this paper, we aim to realize an the convergence and searching for the Pareto solutions, where
efficient and rational integration of different response strategies mutation operators are adopted to further supplement the
based on the quantification of environmental changes, which diversity. If CD is larger than the higher-threshold HT , then
provides a novel idea to develop a competitive DMOA with the change is regarded so dramatic that most of the previous
comprehensive performance. searching experience is no longer useful. As a result, such
situation is treated as a new optimization problem and the
evolution is re-started, where only a few solutions in P St−1
are reserved in the re-initialized population. In particular, if
III. M ETHODOLOGY CD falls between LT and HT , then the TL-based response is
applied, where knowledge acquired in previous environments
In this section, the proposed HRS-DMOA is elaborated, is used to train a prediction model, and more details are
whose main idea is to grade the environmental changes via presented in the next subsection.
some quantification procedure so that hierarchical responses It is noticeable that, in essence, LT and HT determine the
can be adopted accordingly. probability of selecting above three response strategies, and
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 5

in this paper, (LT, HT ) is set to (0.03, 0.78). In addition, Algorithm 2 Construction of the target domain
the framework of the proposed hierarchical response system Require:
is summarized in Algorithm. 1. Two populations P1 = {x1i }N 2 N
i=1 and P2 = {xi }i=1 in the
new environment, mutation probability pm
Algorithm 1 Hierarchical response system Ensure:
Require: Target domain Dt
Predefined lower-threshold LT and higher-threshold HT 1: Initialize Dt = ∅
Ensure: 2: For n from 1 to N
Response strategy 3: Crossover x1n and x2n to generate x̃1n and x̃2n
1: Calculate change degree (CD) according to 4: Define competition group Gc = {x1n , x2n , x̃1n , x̃2n }
Eqs. (5) and (6) 5: Calculate Q(x), x ∈ Gc according to Eq. (7)
2: If CD < LT 6: Dt ← arg max Q(x)
x∈Gc
3: Refinement 7: EndFor
4: else if CD > HT 8: For each individual x̃ ∈ Dt
5: Re-initialization 9: Generate a random number rand
6: else 10: If rand < pm
7: TL 11: x̃′ = mutation(x̃)
8: EndIf 12: If Q(x̃′ ) ≥ Q(x̃)
13: Substitute x̃ with x̃′
14: EndIf
C. TL-based Population Initialization 15: EndIf
16: EndFor
The essence of transfer learning is to apply the learned
17: Return Dt
knowledge from the source domain Ds to assist solving related
but not the same tasks in the target domain Dt . In the pro-
posed HRS-DMOA, the main idea of the TL-based response
the implementation details of above TL-based initialization
is to make full use of the historical searching experiences
strategy are presented.
to initialize population in the new environment, which is
required to have high quality to accelerate the convergence
[26]. Consequently, PS in the previous environment with some D. Overall Framework of HRS-DMOA
mutations is selected as the source domain Ds . In addition, to The overall flowchart of the proposed HRS-DMOA is shown
guarantee knowledge is transferred towards a correct direction, in Fig. 1, where the main contributions (i.e., the proposed
another group of transfer reference points (TRPs) is screened HRS) are illustrated in the red dashed box. In addition, the
in the new environment to form the target domain Dt . To support vector machine (SVM) [10] is adopted as the base
be specific, a local search strategy is applied to assign each learner in the TL-based response strategy, and the MOEA/D
individual with a quality factor Q(·) as [3]: [64] is employed as the static optimizer, which is a rep-
  resentative multi-objective problem solver that decomposes
Q(x) = min max fj (z, t)−fj (x, t) , j ∈ {1, 2, · · · , m} the problem into several scalar sub-problems. Since that the
z∈P \{x}
(7) MOEA/D has been successfully applied in various situations,
where P stands for the population, and the individuals with it is deemed that MOEA/D is competent in searching for the
larger Q(·) value are deemed to have better quality. Then, Pareto solutions in each individual environment when solving
based on Eq. (7), tournament is performed between two a DMOP.
populations in the new environment to obtain a set of TRPs to Explanations of the proposed hierarchical response modes
form the Dt . The details are displayed in Algorithm. 2, where are further summarized as follows.
simulated binary crossover and polynomial mutation operators (1) Refinement. At time t, polynomial mutation is performed
are applied to further enhance the diversity. on P St−1 to enhance the diversity, and Pini is obtained
Let T = {X, Y } denote the training set, where X = by selection from the augmented P St−1 .
Ds ∪ Dt and Y = {0, 1} is the ground-truth label determined (2) TL. At time t, the TL-based initialization is applied to
by the domination relationship at the new time t (Y = 1 generate Pini , where the augmented P St−1 is adopted
for non-dominated individuals). In the proposed HRS-DMOA, as the source domain, and a group of TRPs is screed to
the transfer learning process is realized via the sample-based form the target domain (see Section III-C for details).
TrAdaboost technique [11], where several base learners are (3) Re-initialization. At time t, only a few of the individuals
trained by approximating the mapping from X to Y , and in P St−1 are reserved in Pini , whereas the rest of Pini
the weight of training samples is updated according to the is directly generated by random initialization.
weighted errors on Dt of the trained classifier. Then, the It is noticeable that the TL technique is only adopted
ensemble learning method is adopted to form a strong classifier when the environmental changes are regarded in a medium
H(·) which is later used to generate an initial population level. The merits of this setting contain mainly two aspects.
with high quality in the new environment. In Algorithm. 3, On one hand, if the environment slightly changes, it seems
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 6

Fig. 1: Flowchart of HRS-DMOA, where the proposed HRS (in the red dashed box) is responsible for generating the initial
population (Pini ) to the static optimizer, which accomplishes the searching in each environment and can be any
multi-objective problem solver (the MOEA/D is employed in this study). By comparing the calculated change degree (CD)
with the two thresholds, the responses based on TL, refinement, and re-initialization are adopted when LT ≤ CD ≤ HT ,
CD < LT , and CD > HT , respectively. Additionally, the “TRP Acquisition” refers to the construction of target domain Dt .

unnecessary to introduce the prediction model, which will no 5) In the TL-based initialization, SVM is employed as the
doubts burden the computation. On the other hand, dramatic base learner, whose computational complexity is O(N 2 mn)
environmental changes may lead to few correlations between [10].
the successive environments, and in this case, the negative 6) In the re-initialization response mode, the computational
transfer phenomenon easily occurs. As a result, the advantages complexity is O(N mn).
of different strategies are effectively integrated in the proposed 7) In the refinement response mode, only polynomial mu-
hierarchical response system, which is, in fact, a rational tation operation and a selection process are involved. Hence,
combination of them. In Algorithm. 4, a brief pseudocode of the computational complexity is O(N n) + O(N ).
the proposed HRS-DMOA is presented for a clear view. In general, N is far larger than m and n. Consequently,
Remark 1: In the first environment t = 1, the initial the computational complexity of the proposed HRS-DMOA is
population Pini is directly generated by the static optimizer O(N 2 mn).
Opt(·). In other cases, Pini is provided to Opt(·) by the
adopted response strategy. IV. E XPERIMENTAL R ESULTS AND D ISCUSSIONS
In this section, benchmark evaluations are performed to
validate the effectiveness of the proposed HRS-DMOA in
E. Complexity Analysis solving DMOPs, and sufficient ablation studies are carried out
Given that N is the population size, m is the number to verify the superiority of the developed HRS.
of objectives and n is the dimension of decision vector.
The following cases are mainly considered to estimate the A. Experimental Environment
computational complexity of the proposed HRS-DMOA. The proposed HRS-DMOA is comprehensively evaluated
1) MOEA/D is used to search for Pareto solutions in each on 14 DMOPs in the CEC2018 test suites [27], where the
environment, whose computational complexity is O(N mT ) provided benchmark DMOPs can well characterize the prop-
[64], where T is the size of applied neighborhood. erties of dynamic problems in various real-world optimization
2) The computational complexity is O(n) for both polyno- scenarios, such as time-varying PS and disconnected PF, etc.
mial mutation and simulated binary crossover. The 14 benchmark functions are named as DF1 to DF14,
3) The computational complexity is O(N ) when calculating and one can find more information of the applied test suites
the fitness value of the sensors in a new environment. in [27]. For the evaluation metrics, the inverted generational
4) The computational complexity is O(N m) when calculat- distance (IGD) and the hypervolume (HV) are adopted. To
ing the quality factor Q(·) to select transfer reference points be specific, IGD reflects the convergence of algorithms by
to form Dt . measuring distance between the obtained PF and the real one.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 7

Algorithm 3 Transfer learning-based initialization t is calculated as:

Require:
P
x∈P˜Ft dist(x, P Ft )
The number of base learner Nb , source domain Ds , target IGDt = (8)
|P˜Ft |
domain Dt , and population size N
Ensure: where dist(·) is the Euclidean distance and P˜Ft stands for
Initialized population Pini in the new environment the true PF at time t. HV refers to the volume of the
1: Calculate ns = |Ds | and nt = |Dt | hypercube surrounded by individuals in the obtained PF and
2: Obtain training set T = {X, Y } where X = Ds ∪ Dt and corresponding reference point in the objective space. Let ν(·)
corresponding ground-truth label Y = {yi }ni=1
s +nt
denote the volume of the mentioned hypercube, then HV at
3: Initialize weight for each training sample by time t is given as:
( [
1/ns , x ∈ Ds HVt = ν(x) (9)
ω(x) = x∈P Ft
1/nt , x ∈ Dt
In addition, another commonly used performance indicator
4: For j from 1 to Nb P MS (short for the maximum spread) is adopted to compre-
5: Weight normalization by ω(x) = ω(x)/ z ω(z)
hensively evaluate the diversity of the five DMOAs, which
6: Train a base learner Lj with (T ; ω)
measures the scope that the true PF is covered by the obtained
7: Calculate error εj of P
trained Lj on target domain Dt
one. The larger MS, the better diversity, whose definition is
according to εj = ω(x)|Lj (x) − y|
(x,y)∈Dt given as:
εj 1
8: Define βj = 1−εj and β = √ v
m h
1+ ln(n2s )/Nb
min{Fkmax , fkmax } − max{Fkmin , fkmin } i2
u
u1 X
9: Update weight for each training sample based on M St = t
m Fkmax − Fkmin
k=1
ω(x) · β |Lj (x)−y| , (x, y) ∈ Ds
(
ω(x) = (10)
|L (x)−y|
ω(x) /βj j , (x, y) ∈ Dt where Fkmax and Fkmin , fkmax and fkmin represent the max-
imum and minimum values of the k-th objective in the true
10: EndFor
and obtained PF at time t, respectively.
11: Integrate base learners to form a strong classifier as
PN In the average level, the above three indicators are reported
H(x) = sgn[ j=1 b
−ln(βj )Lj (x)]
and calculated as:
12: Generate a selection pool Ps in the new environment
τ
(|Ps | >> N ) 1X
M IGD = IGDt
13: Screen high-quality solutions from Ps based on P˜s = τ t=1
{z| H(z) = 1, z ∈ Ps } until |P˜s | = N τ
14: Return Pini = P˜s 1X
M HV = HVt (11)
τ t=1
τ
1X
MMS = M St
Algorithm 4 Framework of HRS-DMOA τ t=1
Require: where τ is the maximum generation.
Static optimizer Opt(·), thresholds LT and HT , optimiza- Moreover, to further verify the competitiveness of the pro-
tion problem F , and the time variables t = {1, 2, ..., ⌊ ττt ⌋} posed HRS-DMOA, other four popular DMOAs are employed
Ensure: as baseline models for comparison, including the second ver-
Pareto sets {P St } of all environments sion of the dynamic non-dominated sorting genetic algorithm
1: Initialization procedure
(NSGA)-II [12], the change-responsive NSGA-II [46], the TL-
2: While t ≤ ττ
t based DMOA in [25], and the knee-point-based imbalanced
3: If t = 1 TL-DMOA in [26]. For convenience, above four algorithms
4: P St ← Opt(F , t) are denoted as DNSGA-II-B, CR-DMOEA, Tr-DMOEA, and
5: else KT-DMOEA, respectively. Notice that the response strategy
6: Determine the response mode via Algorithm. 1 based on diversity is applied in the former two methods, and
7: Generate the Pini by the adopted strategy the TL-based response strategy (which can be regarded as the
8: P St ← Opt(F , t, Pini ) combination of the memory- and prediction-based ones) is
9: EndIf adopted with different implementation manners in the latter
10: t←t+1 two algorithms. Hence, above four models are employed as
11: EndWhile
⌊ ττ ⌋ baselines for comparison to validate whether the proposed
12: Return {P St }t=1 t
HRS-DMOA can realize better comprehensive performance
via the rational integration of different mainstream strategies.
In addition, the dynamic behavior of DMOPs is depicted
as t = n1t · ⌊ ττt ⌋, where the number of the environments is
Considering the time-varying property of DMOPs, IGD at time set to 50 (τ = 50 × τt ), τt and nt stand for the frequency
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 8

and severity of change, respectively. On each benchmark, three worse than the KT-DMOEA, which shows the competitiveness
groups of evaluation are performed based on different dynamic of our method.
parameters, including (nt = 5, τt = 10), (nt = 10, τt = 5) 2) Diversity: According to Table IV, HRS-DMOA achieves
and (nt = 10, τt = 10). For a fair comparison, all algorithms half of the best results in terms of MHV, and the sub-optimal
are evaluated with the same parametric conditions on each algorithm Tr-DMOEA wins in 14 cases, which implies the
benchmark, where the dimension of test problems is set to advantage of the proposed algorithm regarding to the compre-
10, the population size for bi- and tri-objective optimization hensive performance. Furthermore, as displayed in Table V,
problem is set to 100 and 150, respectively. 10 epochs of it is found that the proposed HRS-DMOA obtains 22 best
search in each individual environment is fixed for all DMOAs, MMS out of the 42 results, and in totally 106 cases, our
that is, the searching times for Pareto solutions by the applied method performs significantly better than other comparison
static optimizer in all algorithms are the same and, by doing so, models. Combining the results reported in both Table IV and
the effectiveness of the responses to the environmental changes Table V, it can be concluded that the proposed algorithm yields
can be reflected to some extents. Moreover, considering that considerable diversity as well. While on the problem DF6, the
while applying the TL technique to make responses, it is MHV of our method is not so satisfactory as compared with
inevitable to consume some evaluations in advance at the other algorithms, which may due to that the time-varying PF
new environment to establish the target domain, for those of DF6 has the geometric property of long-tails and knee-
algorithms that do not involve extra evaluations, six more regions. By observing the response manner of our method
epochs of search for the static optimizer are added to promote when solving DF6, it is found that in almost 65.3% of the total
the fairness of comparison. 50 environment, the initial population is randomly generated
All algorithms have run 20 times individually on each within the re-initialization mode. Simultaneously, it can be
benchmark to alleviate the influence of randomness, and inferred that the negative transfer phenomenon may occur in
results in average level are reported along with the Wilcoxon the left 34.7% cases where the TL-based response is adopted,
rank sum test at the significance level of 0.05 [13]. Comparison as in the unstable environment it is hard to accurately identify
results in terms of MIGD, MHV, and MMS are presented in the useful knowledge. Consequently, the random initialization
Table III, Table IV, and Table V, respectively, where “+/-” is performed in most cases, and recalling the historical experi-
indicates that the proposed HRS-DMOA is significantly bet- ences makes few contributions, which collaboratively lead to
ter/worse than the corresponding algorithm, and “=” denotes a poor performance.
there is no significant difference between the two algorithms It is also noticeable that the PF of problem DF2 keeps
in a statistical sense. unchanged in a period of time (while the PS changes), which
requires the algorithms to not only monitor environmental
B. Benchmark Evaluations Results changes, but also maintain the population diversity to obtain
1) Convergence: As reported in Table III, HRS-DMOA considerable result. From the reported results, with different
achieves 26 out of 42 best results in terms of MIGD, which dynamic parameter settings, the proposed HRS-DMOA ob-
performs significantly better than KT-DMOEA, Tr-DMOEA, tains almost all the best MHV and MMS on DF2, exhibiting
CR-DMOEA and DNSGA-II-B in 32, 35, 29 and 39 cases, the superiority of HRS in handling dynamic behaviors with
respectively, reflecting the reliability of the proposed method multiple concerns.
in adapting to the varying environments in DMOPs. According To sum up, the proposed HRS-DMOA ranks first in most
to the Wilcoxon rank sum test, there are 135 cases where HRS- testing cases, which is competent in handling complex dy-
DMOA performs significantly better than other algorithms, namic behaviors in DMOPs with satisfactory comprehensive
which shows that the proposed method is a competitive DMOA performance. Moreover, according to the above results, it is
with outstanding convergence. Notice that on the problems found that the DNSGA-II-B has achieved few best results in
DF7, the DNSGA-II-B performs slightly better than the pro- terms of MIGD although a competitive diversity is presented,
posed HRS-DMOA in two situations, which may mainly which indicates that the diversity-based response strategy does
due to the time-varying PF ranges of DF7 are in dissimilar have some limitations in practice. Meanwhile, in various dy-
scales, and simultaneously the corresponding PS center is namic situations, it is also hard to guarantee the performances
fixed, thereby making it extremely hard to realize efficient of DMOAs with only TL-based strategy. Hence, it can be
knowledge transfer towards a proper direction. concluded that the designed HRS is effective and practical,
In addition, for the tri-objective problem DF11, not only the which successfully integrates the merits of different response
centroid of its PF oscillates by expanding and shrinking with strategies.
multiple scales, but also the density of solutions changes with
time, which brings great challenges to the algorithms. On one C. Comparisons with Other Hybrid DMOA
hand, the population is required to rapidly converge with a In this subsection, two additional hybrid methods SVM-
considerable diversity so that the obtained PF can well cover DMOA [24] and PPS-RM-MEDA [71] are employed for
the true one. On the other hand, the population should also comparisons. In Table VI, the results on DF series testing
realize timely escape from the local optimum. Such a high problems are presented, where the number in parenthesis
requirement on maintaining diversity in the objective-space denotes the rank (from 1st to 3rd) of corresponding algorithm,
makes DF11 quite hard to handle, whereas the proposed HRS- and the data of above two comparison models are cited from
DMOA still ranks second on DF11 and performs only slightly [37] and [58], respectively.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 9

TABLE III: Performance of five DMOAs on MIGD

Algorithms
Problems nt , τt
DNSGA-II-B [12] Tr-DMOEA [25] CR-DMOEA [46] KT-DMOEA [26] HRS-DMOA
5, 10 0.4305±1.18e-01(+) 0.2919±1.25e-01(+) 0.0849±1.22e-02(+) 0.1427±2.17e-02(+) 0.0740±1.62e-02
DF1 10, 5 2.3232±7.40e-01(+) 0.2727±6.34e-02(+) 0.1102±5.41e-02(=) 0.1326±2.27e-02(+) 0.0859±1.94e-02
10, 10 2.0923±6.29e-01(+) 0.3594±6.94e-02(+) 0.1038±2.54e-02(=) 0.1261±1.94e-02(+) 0.0884±1.80e-02
5, 10 0.2667±7.12e-02(+) 0.5891±1.00e-01(+) 0.0242±4.11e-03(-) 0.1093±1.29e-02(+) 0.0505±1.68e-02
DF2 10, 5 1.2338±5.04e-01(+) 0.5183±7.52e-02(+) 0.0347±1.87e-02(=) 0.1196±1.04e-02(+) 0.0419±1.31e-02
10, 10 1.3458±5.50e-01(+) 0.5227±7.06e-02(+) 0.0402±1.66e-02(=) 0.1095±1.09e-02(+) 0.0482±1.37e-02
5, 10 0.8344±1.86e-01(+) 0.7486±3.67e-01(=) 0.6393±3.50e-01(=) 0.8366±8.29e-02(+) 0.5182±1.33e-01
DF3 10, 5 2.4443±8.07e-01(+) 0.6780±1.75e-01(=) 0.3807±7.25e-02(-) 0.7350±1.24e-01(+) 0.5680±1.44e-01
10, 10 2.7178±7.92e-01(+) 0.8917±4.27e-01(=) 0.3943±1.18e-01(-) 0.7128±9.16e-02(=) 0.5916±1.57e-01
5, 10 1.6360±2.44e-01(+) 3.7776±4.93e-01(+) 1.2742±1.41e-01(+) 1.6231±1.43e-01(+) 0.3620±7.49e-02
DF4 10, 5 1.8601±3.52e-01(+) 5.1859±2.01e-01(+) 1.3296±1.47e-01(+) 1.7109±9.97e-02(+) 0.4296±6.97e-02
10, 10 1.7133±3.27e-01(+) 5.4790±2.25e-01(+) 1.3684±1.81e-01(+) 1.7089±1.14e-01(+) 0.4310±7.83e-02
5, 10 0.3309±5.93e-02(+) 0.1653±2.79e-02(+) 0.0698±5.30e-02(+) 0.4060±4.98e-02(+) 0.0361±7.22e-03
DF5 10, 5 1.5103±5.20e-01(+) 0.1605±3.31e-02(+) 0.0594±2.09e-02(+) 0.3623±6.79e-02(+) 0.0367±9.04e-03
10, 10 1.4960±4.17e-01(+) 0.1916±3.56e-02(+) 0.0655±4.32e-02(+) 0.3501±6.63e-02(+) 0.0366±5.65e-03
5, 10 6.2760±1.45e+00(+) 2.2846±6.85e-01(+) 2.7808±2.50e+00(=) 3.2694±3.06e-01(+) 1.5302±6.95e-01
DF6 10, 5 1.8258±6.45e-01(=) 3.9235±1.05e+00(+) 5.9046±4.39e+00(+) 3.7916±4.24e-01(+) 1.7784±6.44e-01
10, 10 1.8768±1.04e+00(+) 1.5666±4.28e-01(+) 6.8773±3.12e+00(+) 3.9172±3.90e-01(+) 1.2066±6.06e-01
5, 10 2.8578±5.82e-01(+) 3.5497±9.14e-01(+) 2.3553±7.25e-01(+) 2.9380±7.64e-01(+) 0.9094±2.12e-01
DF7 10, 5 1.1209±1.35e-01(-) 3.6926±9.29e-01(+) 9.1002±7.34e+00(+) 4.4453±7.07e-01(+) 1.2286±1.74e-01
10, 10 1.1154±1.53e-01(=) 2.1510±7.00e-01(+) 5.7568±2.25e+00(+) 2.9671±4.29e-01(+) 1.1298±1.53e-01
5, 10 0.3023±5.64e-02(+) 0.6467±1.64e-02(+) 0.9582±1.32e-01(+) 1.0937±1.53e-02(+) 0.1373±6.89e-02
DF8 10, 5 0.2783±5.55e-02(+) 0.5996±2.09e-02(+) 1.0822±1.38e-01(+) 1.0736±2.32e-02(+) 0.1232±3.16e-02
10, 10 0.2749±5.93e-02(+) 0.6068±2.63e-02(+) 1.0353±1.20e-01(+) 1.0904±1.78e-02(+) 0.1108±3.15e-02
5, 10 1.1643±2.73e-01(+) 2.1294±2.93e-01(+) 0.2699±1.11e-01(=) 0.6778±7.75e-02(+) 0.2189±3.09e-02
DF9 10, 5 1.1467±2.80e-01(+) 3.1613±5.44e-01(+) 0.2797±1.12e-01(+) 0.6427±1.09e-01(+) 0.1993±3.56e-02
10, 10 1.0891±2.99e-01(+) 3.1626±4.40e-01(+) 0.2953±1.40e-01(+) 0.6567±8.32e-02(+) 0.1965±2.64e-02
5, 10 0.8708±1.70e-01(+) 0.1923±3.90e-02(-) 0.4345±8.85e-02(+) 0.3086±1.91e-02(+) 0.2661±7.70e-02
DF10 10, 5 1.2816±3.47e-01(+) 0.1280±1.72e-02(-) 0.3933±7.34e-02(+) 0.2836±2.15e-02(-) 0.3283±2.28e-02
10, 10 1.2348±3.28e-01(+) 0.1158±1.45e-02(-) 0.3504±6.57e-02(=) 0.2946±1.33e-02(-) 0.3230±2.76e-02
5, 10 0.7717±1.56e-01(+) 0.3433±3.16e-02(+) 0.3868±5.11e-03(+) 0.1636±5.54e-03(+) 0.1492±3.68e-03
DF11 10, 5 0.8730±1.55e-01(+) 0.4038±4.48e-02(+) 0.4847±8.07e-03(+) 0.1634±8.18e-03(-) 0.3677±2.78e-03
10, 10 0.9039±9.84e-02(+) 0.3728±4.60e-02(=) 0.4776±1.52e-02(+) 0.1643±8.74e-03(-) 0.3666±2.20e-03
5, 10 0.8208±6.17e-02(+) 2.5764±1.73e-01(+) 0.3076±1.14e-02(-) 0.6093±5.28e-02(=) 0.5108±1.78e-01
DF12 10, 5 0.8656±7.23e-02(+) 1.4244±9.61e-02(+) 0.3051±3.42e-03(=) 0.6321±5.39e-02(=) 0.4643±1.81e-01
10, 10 0.9036±7.61e-02(+) 1.4266±1.16e-01(+) 0.3074±3.26e-03(-) 0.6358±7.12e-02(=) 0.6236±1.69e-01
5, 10 0.5057±1.04e-01(+) 0.3911±3.42e-02(+) 0.2554±1.82e-02(+) 0.4067±4.16e-02(+) 0.2409±7.86e-03
DF13 10, 5 1.6747±4.90e-01(+) 0.3841±2.29e-02(+) 0.3052±1.93e-02(+) 0.3789±3.80e-02(+) 0.2556±1.20e-02
10, 10 1.6450±6.22e-01(+) 0.4262±4.73e-02(+) 0.3031±9.74e-03(+) 0.3741±3.30e-02(+) 0.2527±1.26e-02
5, 10 0.4126±1.07e-01(+) 0.2410±5.64e-02(+) 0.1246±2.11e-02(+) 0.1310±1.44e-02(+) 0.0972±4.35e-03
DF14 10, 5 3.0028±8.52e-01(+) 0.2174±5.35e-02(+) 0.1570±2.06e-02(+) 0.1253±1.16e-02(=) 0.1216±3.71e-03
10, 10 3.0825±1.14e+00(+) 0.1938±3.51e-02(+) 0.1614±2.42e-02(+) 0.1210±1.32e-02(=) 0.1231±3.97e-03
+/-/= \ 39 / 1 / 2 35 / 3 / 4 29 / 4 / 9 32 / 4 / 6 \

As can be seen from Table VI, the proposed HRS-DMOA results, and according to our algorithm configuration, it is
ranks first and second in 13 and 14 cases, respectively. In difficult for the static optimizer to always obtain sufficient
particular, on the complex tri-objective benchmark problems high-quality Pareto solutions within only 10 epochs of search
DF13 and DF14, our method yields considerable convergence in each environment, which accounts for that our algorithm
performance, which owes to that the changing severity is may present performance declination as compared with the
taken into account when making response to the environmental optimal one in some situations.
changes, and it enables the generated initial population well In addition, the proposed HRS-DMOA is further compared
adapt to the new environment. It should be pointed out that with the MoE [43], which is a state-of-the-art DMOA that
the static optimizer can also make great contribution to the has employed multiple prediction mechanism. To make a fair
comparison, the basic experimental settings are made the same
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 10

TABLE IV: Performance of five DMOAs on MHV

Algorithms
Problems nt , τt
DNSGA-II-B [12] Tr-DMOEA [25] CR-DMOEA [46] KT-DMOEA [26] HRS-DMOA
5, 10 0.1838±5.31e-02(+) 0.2526±7.37e-02(+) 0.4500±1.54e-02(=) 0.3811±1.63e-02(+) 0.4587±3.15e-02
DF1 10, 5 0.0089±3.60e-02(+) 0.2493±3.75e-02(+) 0.4241±6.92e-02(=) 0.3909±1.69e-02(+) 0.4633±3.03e-02
10, 10 0.0108±3.39e-02(+) 0.2100±4.08e-02(+) 0.4236±3.82e-02(+) 0.3940±1.57e-02(+) 0.4697±4.19e-02
5, 10 0.2317±7.41e-02(+) 0.5531±1.11e-01(+) 0.6872±8.42e-03(=) 0.5862±1.04e-02(+) 0.6990±2.31e-02
DF2 10, 5 0.0516±1.14e-01(+) 0.6777±3.12e-02(-) 0.6743±2.50e-02(-) 0.5895±9.76e-03(+) 0.6595±2.02e-02
10, 10 0.0465±8.94e-02(+) 0.6439±8.38e-02(+) 0.6661±1.66e-02(+) 0.5976±1.28e-02(+) 0.6818±1.76e-02
5, 10 0.0305±4.37e-02(+) 0.1724±2.57e-02(=) 0.0878±7.50e-02(=) 0.1495±1.06e-02(-) 0.1196±8.54e-02
DF3 10, 5 0.0087±3.91e-02(+) 0.2009±1.54e-02(=) 0.1628±5.45e-02(-) 0.1655±1.40e-02(-) 0.1069±6.22e-02
10, 10 0.0050±2.33e-03(+) 0.2013±1.74e-02(=) 0.1609±7.74e-02(=) 0.1634±7.47e-03(=) 0.1304±5.94e-02
5, 10 0.1625±3.79e-02(+) 0.0209±2.64e-02(+) 0.6013±5.63e-02(-) 0.4960±3.08e-02(=) 0.5218±3.35e-02
DF4 10, 5 0.1489±6.04e-02(+) 0.0188±1.61e-02(+) 0.5185±5.31e-02(=) 0.4840±3.13e-02(+) 0.5193±2.72e-02
10, 10 0.1813±5.76e-02(+) 0.0027±2.89e-03(+) 0.5709±5.39e-02(-) 0.4705±3.54e-02(+) 0.5216±2.87e-02
5, 10 0.2538±4.22e-02(+) 0.4330±3.43e-02(+) 0.4971±5.73e-02(=) 0.2335±2.46e-02(+) 0.5293±1.22e-02
DF5 10, 5 0.0282±6.65e-02(+) 0.4303±3.15e-02(+) 0.4999±3.00e-02(+) 0.2719±2.37e-02(+) 0.5306±1.69e-02
10, 10 0.0166±4.68e-02(+) 0.4055±2.88e-02(+) 0.4932±5.37e-02(+) 0.2777±2.35e-02(+) 0.5284±1.08e-02
5, 10 0.0012±3.19e-03(+) 0.9480±7.03e-02(-) 0.2471±4.22e-02(-) 0.0271±1.19e-02(=) 0.0268±1.13e-02
DF6 10, 5 0.0635±9.28e-02(=) 0.8041±1.08e-01(-) 0.1745±4.44e-02(-) 0.0299±1.34e-02(=) 0.0260±9.25e-03
10, 10 0.0785±6.97e-02(=) 0.8315±1.03e-01(-) 0.2412±2.41e-02(-) 0.0289±1.22e-02(=) 0.0275±1.22e-02
5, 10 0.1285±3.17e-02(=) 0.9479±9.08e-02(-) 0.0124±1.25e-02(+) 0.2334±2.03e-02(-) 0.1269±1.62e-02
DF7 10, 5 0.1400±3.17e-02(=) 0.8986±7.02e-02(-) 0.0315±2.15e-02(+) 0.2692±3.50e-02(-) 0.1347±3.30e-02
10, 10 0.1398±2.99e-02(=) 0.8392±7.24e-02(-) 0.0379±1.78e-02(+) 0.2471±2.12e-02(-) 0.1382±2.96e-02
5, 10 0.6909±3.88e-02(+) 0.9042±1.10e-02(-) 0.9340±1.51e-02(+) 0.9115±6.78e-03(-) 0.9532±1.28e-02
DF8 10, 5 0.6484±5.18e-02(+) 0.8954±1.22e-02(-) 0.9439±1.84e-02(=) 0.9133±5.97e-03(-) 0.9442±1.69e-02
10, 10 0.6445±3.04e-02(+) 0.8988±7.67e-03(-) 0.9402±1.69e-02(-) 0.9123±7.45e-03(-) 0.9255±1.82e-02
5, 10 0.0555±3.79e-02(+) 0.5656±2.55e-02(-) 0.3232±9.77e-02(=) 0.1698±1.88e-02(+) 0.3161±3.47e-02
DF9 10, 5 0.0495±3.63e-02(+) 0.1104±3.65e-02(+) 0.3020±9.94e-02(=) 0.1958±2.34e-02(+) 0.3391±4.16e-02
10, 10 0.0634±4.50e-02(+) 0.1294±1.90e-02(+) 0.2753±1.20e-01(=) 0.1842±1.53e-02(+) 0.3443±3.07e-02
5, 10 0.0371±1.37e-01(+) 0.6399±2.59e-02(+) 0.9115±3.26e-02(=) 0.6142±2.17e-02(+) 0.8791±1.36e-01
DF10 10, 5 0.0530±1.75e-01(+) 0.8712±1.91e-02(+) 0.9066±1.73e-02(=) 0.6605±1.57e-02(+) 0.9150±1.83e-02
10, 10 0.0407±1.40e-01(+) 0.8841±1.14e-02(+) 0.9161±1.81e-02(+) 0.6581±1.26e-02(+) 0.9247±1.76e-02
5, 10 0.1093±2.03e-01(+) 0.1336±1.54e-02(+) 0.4875±8.77e-03(+) 0.2193±2.94e-03(+) 0.7670±1.62e-02
DF11 10, 5 0.0561±1.91e-01(+) 0.1402±2.17e-02(+) 0.6350±1.03e-02(+) 0.2224±3.19e-03(+) 0.7775±9.94e-03
10, 10 0.0548±1.90e-01(+) 0.1414±1.33e-02(+) 0.6307±2.01e-02(+) 0.2231±2.20e-03(+) 0.7728±1.65e-02
5, 10 0.9800±1.55e-02(-) 0.7480±4.12e-02(+) 0.8960±8.09e-03(-) 0.7784±1.38e-02(+) 0.8375±3.47e-02
DF12 10, 5 0.9486±3.99e-02(-) 0.6676±4.37e-02(+) 0.9089±3.54e-03(-) 0.7988±6.90e-03(=) 0.8371±5.48e-02
10, 10 0.9649±2.56e-02(-) 0.6630±6.25e-02(+) 0.9075±6.50e-03(-) 0.7957±8.83e-03(+) 0.8141±6.23e-02
5, 10 0.4643±1.16e-01(=) 0.5326±2.97e-02(-) 0.5135±2.02e-02(-) 0.4044±2.37e-02(+) 0.4549±1.63e-02
DF13 10, 5 0.0933±1.08e-01(+) 0.5011±1.81e-02(-) 0.3020±2.54e-02(+) 0.4223±2.17e-02(+) 0.4506±1.17e-02
10, 10 0.1045±1.08e-01(+) 0.4889±1.12e-02(-) 0.2953±1.99e-02(+) 0.4232±2.11e-02(+) 0.4544±6.67e-03
5, 10 0.0281±1.39e-02(+) 0.3516±6.45e-02(+) 0.4226±3.28e-02(+) 0.4063±1.91e-02(+) 0.4884±1.03e-02
DF14 10, 5 0.0027±1.22e-02(+) 0.3733±5.11e-02(+) 0.4116±1.59e-02(+) 0.4276±1.65e-02(+) 0.4801±9.70e-03
10, 10 0.0018±8.25e-03(+) 0.3960±3.58e-02(+) 0.4096±1.58e-02(+) 0.4326±1.38e-02(+) 0.4791±1.01e-02
+/-/= \ 33 / 3 / 6 25 / 14 / 3 17 / 12 / 13 28 / 8 / 6 \

as mentioned in [21], and the comparison results on JY series to realize rational and efficient integration of the mainstream
problems [23] are displayed in Table VII, where data of the strategies, thus only several basic response methods are applied
algorithm MoE are cited from [21]. in the proposed HRS-DMOA, which may limit the algorithm
performance as compared with the meticulously designed
According to Table VII, the proposed HRS-DMOA wins
advanced strategies in MoE.
the state-of-the-art MoE in 15 cases, and on problems of JY1
and JY5, the MoE yields better performance than our method. It is also noticeable that on the rest five JY problems,
Particularly, on these two problems, the results obtained by the proposed HRS-DMOA obtains the best results under all
the proposed HRS-DMOA merely reach the level of 10−1 , dynamic parameter settings, which is a promising and inspiring
whose major reason is that the main focus of this study is result as it implies that even the basic methods could realize
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 11

TABLE V: Performance of five DMOAs on MMS

Algorithms
Problems nt , τt
DNSGA-II-B [12] Tr-DMOEA [25] CR-DMOEA [46] KT-DMOEA [26] HRS-DMOA
5, 10 0.7769±5.89e-02(+) 0.7443±1.39e-01(+) 0.8912±8.68e-02(=) 0.7694±1.82e-02(+) 0.9084±4.65e-02
DF1 10, 5 0.9113±6.31e-02(=) 0.9064±2.50e-02(=) 0.8752±1.00e-01(+) 0.7654±2.86e-02(+) 0.9137±3.56e-02
10, 10 0.8646±1.08e-01(+) 0.8606±4.52e-02(+) 0.8781±8.70e-02(+) 0.7701±4.12e-02(+) 0.9040±2.95e-02
5, 10 0.8513±4.88e-02(+) 0.5661±1.10e-01(+) 0.9174±4.79e-02(=) 0.7866±2.54e-02(+) 0.9216±3.17e-02
DF2 10, 5 0.8047±5.25e-02(+) 0.5548±8.51e-02(+) 0.9271±6.30e-02(=) 0.7382±2.55e-02(+) 0.9373±2.77e-02
10, 10 0.8112±7.64e-02(+) 0.5510±1.25e-01(+) 0.8940±7.61e-02(=) 0.7531±2.40e-02(+) 0.9297±4.15e-02
5, 10 0.3554±9.02e-02(+) 0.5532±3.83e-02(-) 0.3715±1.51e-01(+) 0.5541±3.14e-02(-) 0.4140±1.09e-01
DF3 10, 5 0.8493±1.52e-01(-) 0.6196±7.01e-02(-) 0.4165±1.06e-01(=) 0.5297±7.05e-02(-) 0.4244±1.04e-01
10, 10 0.8780±1.76e-01(-) 0.6380±4.57e-02(-) 0.4259±1.28e-01(=) 0.4953±4.04e-02(=) 0.4879±8.93e-02
5, 10 0.4351±5.95e-02(+) 0.5384±1.22e-01(+) 0.7662±8.33e-02(+) 0.4607±2.44e-02(+) 0.9429±3.37e-02
DF4 10, 5 0.3068±8.30e-02(+) 0.5556±9.31e-02(+) 0.7601±1.01e-01(+) 0.4760±3.10e-02(+) 0.9235±5.10e-02
10, 10 0.2836±8.99e-02(+) 0.3909±6.13e-02(+) 0.7400±1.00e-01(+) 0.4613±2.41e-02(+) 0.9060±4.01e-02
5, 10 0.9698±4.51e-04(=) 0.9637±3.60e-04(+) 0.9292±1.31e-01(=) 0.8683±1.95e-02(+) 0.9808±1.22e-02
DF5 10, 5 0.9686±2.53e-03(+) 0.9719±1.62e-04(=) 0.9551±6.32e-02(=) 0.8760±1.80e-02(+) 0.9754±1.23e-02
10, 10 0.9491±1.70e-03(+) 0.9779±3.22e-04(=) 0.9661±4.31e-02(=) 0.8626±2.51e-02(+) 0.9788±1.26e-02
5, 10 0.9988±2.65e-03(-) 0.9940±4.71e-03(=) 0.8729±1.88e-01(+) 0.7458±3.74e-02(+) 0.9552±6.03e-02
DF6 10, 5 0.9997±1.34e-03(=) 0.9948±3.50e-03(-) 0.9153±2.39e-01(+) 0.7610±4.17e-02(+) 0.9913±2.73e-02
10, 10 0.9999±5.24e-04(-) 0.9927±3.42e-03(=) 0.9989±2.31e-03(-) 0.7573±3.31e-02(+) 0.9576±6.68e-02
5, 10 0.8251±3.56e-02(+) 0.8178±1.27e-01(=) 0.9127±5.43e-02(=) 0.7114±4.08e-02(+) 0.8878±3.69e-02
DF7 10, 5 0.7918±2.53e-02(+) 0.6464±1.32e-01(+) 0.8258±6.04e-02(+) 0.6057±4.41e-02(+) 0.8778±3.10e-02
10, 10 0.7999±3.07e-02(+) 0.6598±8.29e-02(+) 0.8061±5.45e-02(+) 0.6271±5.98e-02(+) 0.8727±2.46e-02
5, 10 0.9321±2.87e-02(-) 0.6392±2.43e-02(+) 0.3383±1.04e-01(+) 0.2264±1.37e-02(+) 0.8508±3.56e-02
DF8 10, 5 0.9674±2.15e-02(-) 0.6378±1.65e-02(+) 0.3006±1.46e-01(+) 0.2200±1.05e-02(+) 0.8671±1.43e-02
10, 10 0.9683±2.06e-02(-) 0.6421±2.04e-02(+) 0.3689±1.49e-01(+) 0.2238±1.22e-02(+) 0.8739±2.29e-02
5, 10 0.9946±1.28e-02(-) 0.9530±4.65e-02(=) 0.7751±1.87e-01(+) 0.8053±3.60e-02(+) 0.9261±3.81e-02
DF9 10, 5 0.9983±2.14e-03(-) 0.9759±8.82e-03(=) 0.8130±1.73e-01(+) 0.7497±3.13e-02(+) 0.9414±3.94e-02
10, 10 0.9992±1.60e-03(-) 0.9707±1.53e-02(-) 0.9023±1.29e-01(=) 0.7809±5.64e-02(+) 0.9170±4.44e-02
5, 10 0.8952±1.15e-02(+) 0.9087±3.85e-03(-) 0.6635±1.44e-01(+) 0.8159±3.05e-02(+) 0.9102±2.08e-02
DF10 10, 5 0.8826±3.44e-02(+) 0.8935±5.35e-05(+) 0.7003±1.98e-01(+) 0.8094±2.06e-02(+) 0.9191±2.88e-02
10, 10 0.8850±3.65e-02(+) 0.8772±8.68e-09(+) 0.8149±1.20e-01(+) 0.8027±2.25e-02(+) 0.9153±2.44e-02
5, 10 0.9725±2.86e-02(+) 0.9882±1.28e-02(=) 0.8530±1.44e-02(+) 0.9606±5.77e-03(+) 0.9986±1.39e-03
DF11 10, 5 0.7343±1.99e-02(+) 0.9883±5.09e-03(-) 0.7523±6.39e-03(+) 0.9592±6.01e-03(-) 0.9523±4.30e-03
10, 10 0.7247±4.19e-02(+) 0.9920±4.66e-03(-) 0.7639±2.22e-02(+) 0.9612±9.65e-03(-) 0.9513±2.03e-03
5, 10 0.5946±1.16e-01(+) 0.7350±8.84e-02(+) 0.5985±5.25e-02(+) 0.6382±3.05e-02(+) 0.7943±6.41e-02
DF12 10, 5 0.7137±9.75e-02(+) 0.6971±4.75e-02(+) 0.5514±7.51e-03(+) 0.6026±3.71e-02(+) 0.7986±7.50e-02
10, 10 0.7094±5.77e-02(=) 0.6731±4.06e-02(+) 0.5543±7.65e-03(+) 0.6036±4.31e-02(+) 0.7499±9.66e-02
5, 10 0.9852±2.21e-02(-) 0.9962±1.61e-03(-) 0.9265±5.56e-02(+) 0.8499±2.20e-02(+) 0.9526±2.01e-02
DF13 10, 5 0.9982±5.58e-03(-) 0.9968±1.07e-03(-) 0.9470±4.43e-02(=) 0.8431±1.68e-02(+) 0.9504±2.83e-02
10, 10 0.9997±1.14e-03(-) 0.9966±4.10e-03(-) 0.9184±5.90e-02(+) 0.8378±1.61e-02(+) 0.9563±1.89e-02
5, 10 0.8935±1.65e-02(+) 0.9245±1.06e-02(+) 0.7087±8.65e-02(+) 0.9215±1.40e-02(=) 0.9326±1.13e-02
DF14 10, 5 0.9930±3.10e-02(-) 0.9415±1.63e-02(-) 0.6398±7.91e-02(+) 0.9139±1.12e-02(-) 0.8729±1.12e-02
10, 10 0.9926±3.28e-02(-) 0.9582±4.06e-02(-) 0.6268±5.33e-02(+) 0.9126±1.44e-02(-) 0.8723±1.23e-02
+/-/= \ 23 / 15 / 4 20 / 13 / 9 29 / 1 / 12 34 / 6 / 2 \

considerable overall performance via rational integration. In strategies.

addition, it can be inferred that to sufficiently exploit the
merits of those meticulously designed advanced methods, an
in-depth analysis on the specific situation is necessary and D. Generalization Ability of HRS
helpful, which indicates the reliability and superiority of the
As a comprehensive approach to deal with the dynamic
core idea in the proposed quantification-based hierarchical
behaviors in DMOPs, the proposed HRS is proven com-
scheme. Simultaneously, above inference motivates us to pro-
petitive in above presented experimental results. To further
vide more thorough insights on the dynamic behaviors, and
validate the effectiveness and generalization ability of HRS,
to seek potential integration of some state-of-the-art response
in this subsection, another well-known algorithm NSGA-II is
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 12

TABLE VI: Comparisons with other hybrid algorithms on two state-of-the-art methods, namely KGB [61] and TCD [60],
DF series problems in terms of MIGD respectively, which are integrated to the NSGA-II algorithm
Algorithms as well to form the DMOAs.
Problems nt , τt
SVM-DMOA [24] PPS-RM-MEDA [71] HRS-DMOA According to the results reported in Table VIII, it is found
10, 10 0.4310±4.85e-02(3) 0.0365±7.34e-03(1) 0.0884±1.80e-02(2)
DF1 that our HRS-NSGA-II yields 10 best MIGD out of the 14
10, 5 1.3278±8.55e-02(3) 0.0949±2.11e-02(2) 0.0859±1.94e-02(1)
10, 10 0.2962±2.26e-02(3) 0.0408±8.58e-03(1) 0.0482±1.37e-02(2) benchmarks, which is also an inspiring result. On one hand,
DF2
10, 5 0.8960±6.07e-02(3) 0.0917±7.86e-03(2) 0.0419±1.31e-02(1) it exhibits the superiority of the proposed HRS in comparison
DF3
10, 10 0.5159±5.47e-02(2) 0.2012±6.02e-03(1) 0.5916±1.57e-01(3) to other advanced response approaches to the environmental
10, 5 1.2482±2.07e-01(3) 0.2809±3.57e-02(1) 0.5680±1.44e-01(2)
changes; on the other hand, the proposed HRS is demonstrated
10, 10 0.1219±3.83e-03(1) 1.0118±1.78e-02(3) 0.4310±7.83e-02(2)
DF4
10, 5 0.1990±4.03e-02(1) 1.0861±1.48e-02(3) 0.4296±6.97e-02(2) a flexible plug-and-play module with considerable general-
DF5
10, 10 0.1482±4.32e-02(2) 1.1942±1.30e-02(3) 0.0366±5.65e-03(1) ization ability, which can be integrated with any other static
10, 5 1.2815±2.70e-01(2) 1.3787±6.25e-02(3) 0.0367±9.04e-03(1) optimizer to effectively solve the DMOPs.
10, 10 3.9768±5.77e-01(3) 3.0892±3.07e-01(2) 1.2066±6.06e-01(1)
DF6
10, 5 9.5840±1.09e+00(3) 6.5176±5.38e-01(2) 1.7784±6.44e-01(1)
10, 10 0.5220±5.70e-02(1) 2.9774±2.83e-01(3) 1.1298±1.53e-01(2) E. Sensitivity Analysis of Thresholds LT and HT
DF7
10, 5 0.6959±1.22e-01(1) 5.7018±5.14e-01(3) 1.2286±1.74e-01(2) In the proposed HRS-DMOA, the thresholds LT and HT
10, 10 0.0752±1.12e-02(1) 0.8673±7.93e-03(3) 0.1108±3.15e-02(2)
DF8
10, 5 0.1940±1.42e-02(2) 0.7860±2.10e-02(3) 0.1232±3.16e-02(1)
play important roles in grading the change degree to adopt dif-
10, 10 0.4151±6.63e-02(2) 1.5801±1.13e-02(3) 0.1965±2.64e-02(1) ferent response strategies. Hence, in this subsection, influences
DF9
10, 5 1.0983±1.63e-01(2) 1.6691±3.88e-02(3) 0.1993±3.56e-02(1) of the two thresholds are investigated.
10, 10 0.6037±3.84e-02(3) 0.1389±1.84e-03(1) 0.3230±2.76e-02(2)
DF10
10, 5 0.6188±3.34e-02(3) 0.1874±2.68e-03(1) 0.3283±2.28e-02(2)
10, 10 0.6717±2.15e-03(3) 0.1565±1.09e-02(1) 0.3666±2.20e-03(2)
DF11
10, 5 0.6933±3.06e-03(3) 0.1783±7.75e-03(1) 0.3677±2.78e-03(2)
10, 10 0.3829±3.86e-03(1) 1.1756±5.61e-03(3) 0.6236±1.69e-01(2)
DF12
10, 5 0.4262±6.10e-03(1) 1.1760±5.61e-03(3) 0.4643±1.81e-01(2)
10, 10 0.5733±5.07e-02(2) 1.3815±2.58e-02(3) 0.2527±1.26e-02(1)
DF13
10, 5 1.4883±2.57e-01(2) 1.6325±3.88e-02(3) 0.2556±1.20e-02(1)
10, 10 0.2556±5.73e-02(2) 0.8579±7.83e-03(3) 0.1231±3.97e-03(1)
DF14
10, 5 0.8887±2.44e-01(2) 1.0334±4.22e-02(3) 0.1216±3.71e-03(1)

TABLE VII: Comparisons with the state-of-the-art MoE

algorithm on JY series problems in terms of MIGD
Algorithms
Problems nt , τt
MoE [43] HRS-DMOA
10, 10 2.43e-2±4.24e-3 1.16e-1±2.05e-2
JY1 10, 20 9.36e-3±6.17e-4 1.10e-1±1.99e-2
10, 30 7.01e-3±2.43e-4 1.01e-1±1.80e-2
10, 10 1.68e-1±1.63e-3 3.11e-2±1.39e-3
JY2 10, 20 1.64e-1±1.46e-4 3.28e-2±1.51e-3
10, 30 1.63e-1±7.22e-5 3.20e-2±1.87e-3
10, 10 3.16e-1±6.93e-3 1.51e-2±2.91e-2
JY3 10, 20 3.12e-1±1.88e-3 1.56e-2±2.97e-2
10, 30 3.13e-1±2.77e-3 1.44e-2±2.46e-2
10, 10 1.51e-1±1.78e-3 9.09e-2±2.60e-2
JY4 10, 20 1.36e-1±2.86e-4 7.43e-2±1.76e-2
10, 30 1.35e-1±8.84e-5 8.47e-2±2.01e-2
10, 10 9.89e-3±4.54e-4 1.51e-1±4.95e-2
JY5 10, 20 7.59e-3±1.99e-4 1.73e-1±6.03e-2
10, 30 7.12e-3±1.70e-4 1.65e-1±3.97e-2
10, 10 1.64e+0±1.35e-1 5.34e-1±6.53e-1
JY6 10, 20 9.45e-1±1.01e-1 3.96e-1±6.13e-1
10, 30 5.86e-1±4.90e-2 2.90e-1±1.40e-1
10, 10 2.69e+0±5.09e-1 1.20e+0±1.88e-1 Fig. 2: Change degree of 14 DF problems.
JY7 10, 20 2.27e+0±8.53e-1 1.31e+0±2.65e-1
10, 30 2.20e+0±6.58e-1 1.30e+0±1.96e-1 As previously mentioned that, in essence, LT and HT
determine the proportion of selecting three response strategies,
which also depends on the change degree CD of the environ-
employed as the static optimizer, and the integrated HRS- ment. In Fig. 2, the changing curves of CD are illustrated in
NSGA-II is evaluated on the 14 benchmark problems with the three given dynamic conditions of all test problems. Based
(nt , τt , τ ) = (10, 10, 200). Moreover, the evaluation results are on the CD values shown in Fig. 2, three proportions of adopt-
compared with the baseline algorithm D-NSGA-II and other ing refinement, TL and re-initialization response modes are
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 13

TABLE VIII: Comparisons against other DMOAs with NSGA-II serving as the static optimizer in terms of MIGD
Algorithms
Problems
KGB-NSGA-II TCD-NSGA-II DNSGA-II-B HRS-NSGA-II
DF1 0.3113±1.39e-03 0.1887±4.93e-02 0.7182±1.54e-01 0.2257±4.71e-02
DF2 0.2198±1.64e-03 0.1146±1.74e-02 0.4638±1.12e-01 0.1380±1.91e-02
DF3 0.9593±1.19e-02 0.5371±6.67e-02 1.0556±2.24e-01 0.4983±2.02e-01
DF4 3.1503±2.84e-01 0.1487±4.20e-02 2.0526±2.28e-01 0.1812±1.76e-02
DF5 0.4004±3.05e-03 0.3427±1.46e-01 0.6813±6.67e-02 0.0232±4.40e-03
DF6 19.8500±4.97e+00 6.6381±6.96e-01 6.0010±1.07e+00 3.3446±9.41e-01
DF7 0.4575±5.85e-03 8.0032±3.77e+00 1.1040±1.23e-01 0.5021±5.62e-02
DF8 0.2130±2.51e-04 0.3057±2.54e-02 0.4431±1.08e-01 0.0441±8.75e-03
DF9 0.5932±6.34e-03 0.5954±1.42e-01 2.2331±5.79e-01 0.2155±3.39e-02
DF10 0.1892±4.71e-04 0.1814±1.70e-02 1.3771±3.14e-01 0.1317±8.44e-03
DF11 10.3910±3.06e-02 0.2670±4.09e-02 1.0163±2.24e-01 0.2108±1.73e-02
DF12 0.5650±3.03e-03 4.4655±5.09e+00 0.7770±6.64e-02 0.2290±2.53e-02
DF13 0.5854±2.49e-03 0.7456±2.24e-01 0.6401±1.10e-01 0.2603±1.30e-02
DF14 0.2267±5.72e-04 0.8972±5.40e-01 0.5049±1.23e-01 0.1907±3.05e-03

investigated, including 10% : 80% : 10%, 15% : 70% : 15% wise by:
and 20% : 60% : 20%, where the corresponding (LT, HT ) fi,j (t) − fi,j (t − 1)
takes value from {(0.01, 0.97), (0.03, 0.78), (0.05, 0.68)}, Scd = max , i = 1, 2, ..., m (12)
fi,j (t − 1) + µ
respectively. Using IGD as the evaluation metric, algorithm
performances with different settings of (LT, HT ) are shown where j stands for the sensor individual.
in Fig. 3, where the dynamic parameters of all 14 benchmark
functions are fixed at τt = 10, nt = 10.
Accordingly, if it satisfies that Scd ≥ Ecd , it can be regarded
As is shown, when LT and HT are set to 0.03 and 0.78, the
that the corresponding individual is greatly influenced by
algorithm presents a more stable performance. To be specific,
the fluctuation in environment, which is not quite a reliable
poor convergence is found on problems DF3 and DF6 when
solution in the new environment. By calculating the proportion
LT = 0.01, HT = 0.97. It may because that the two thresh-
of those unreliable sensors to the whole sensor set, severity
olds are too marginalized to enable the hierarchical response
of the changes can be reflected. Then, considering the three
system timely adapt to the varying environments such as the
optional response modes, the interval (0, 1) is uniformly
concave/convex change on a certain objective function. On the
divided into three sub-intervals, thereby realizing the adaptive
contrary, if LT and HT are set to 0.05 and 0.68, the designed
selection of different response modes. (1) If more than two
response system becomes highly sensitive to the environmental
thirds of sensors are unreliable, then the change is deemed
changes, which can be reflected from the sharp variation
severe and the re-initialization strategy is applied; (2) in
tendency of the IGD value on problems DF12 and DF13.
case of fewer than one third of the sensors are unreliable,
Under this circumstance, once a disconnected time-varying PF
the refinement mode is adopted; (3) otherwise, the TL-based
occurs, the algorithm is likely to treat it as a dramatic change
response is taken.
and accordingly take the re-initialization strategy, which may
cause adverse effects on the convergence. Therefore, it is also
inappropriate to set too centralized thresholds, which cannot Benchmark valuation results obtained by the HRS with
take full advantages of the hierarchical response system. As fixed thresholds (LT, HT ) = (0.03, 0.78) and the HRS
a result, LT = 0.03 and HT = 0.78 are recommended and with adaptive thresholds (denoted as HRS*) are displayed in
they are adopted in other reported experiments in this study. Table IX, where it is found that the adaptive thresholds also
In addition, we also make another attempt to adaptively enable the proposed HRS to effectively handle the dynamic
select the response modes (rather than based on the fixed behaviors with satisfactory convergence. Moreover, according
thresholds), which benefits further strengthening the con- to the rank sum test results, HRS* presents equivalent per-
nections between the response manner and the severity of formance as compared to HRS in six cases, and the former
environmental changes. To be specific, by taking the average even outperforms the latter on DF3, which is a promising
value of CD (in Eq. (5)) regarding to the cardinality of finding that motivates us to explore other adaptive threshold
sensor set S, an environmental changing degree is obtained setting manners and, by doing so, the subjectivity in deter-
1
as Ecd = |S| max{CDi } (i = 1, 2, ..., m), which is deemed mining response modes can be effectively eliminated so as to
as the objective-wise influences of the current change to all establish a highly generalized model that can adapt to diverse
sensors in average level. Meanwhile, one can also obtain the changing situations by connecting the change severity with the
maximum changing degree for each sensor (Scd ) in objective- responses.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 14

DF1 DF2 DF3

0.45 0.16 1.4
(LT,HT)=(0.01,0.97) (LT,HT)=(0.01,0.97) (LT,HT)=(0.01,0.97)
0.4 (LT,HT)=(0.03,0.78) (LT,HT)=(0.03,0.78) (LT,HT)=(0.03,0.78)
(LT,HT)=(0.05,0.68) 0.14 (LT,HT)=(0.05,0.68) 1.2 (LT,HT)=(0.05,0.68)

0.35
0.12 1
0.3

0.25 0.1 0.8

IGD
IGD

IGD
0.2 0.08 0.6

0.15
0.06 0.4
0.1

0.04 0.2
0.05

0 0.02 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Environment Environment Environment

DF4 DF5 DF6

1.2 1.2 12
(LT,HT)=(0.01,0.97) (LT,HT)=(0.01,0.97) (LT,HT)=(0.01,0.97)
(LT,HT)=(0.03,0.78) (LT,HT)=(0.03,0.78) (LT,HT)=(0.03,0.78)
(LT,HT)=(0.05,0.68) (LT,HT)=(0.05,0.68) 10 (LT,HT)=(0.05,0.68)
1 1

0.8 0.8 8

IGD
IGD

IGD

0.6 0.6 6

0.4 0.4 4

0.2 0.2 2

0 0 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Environment Environment Environment

DF7 DF8 DF9

4.5
(LT,HT)=(0.01,0.97) 0.22 (LT,HT)=(0.01,0.97) (LT,HT)=(0.01,0.97)
4 (LT,HT)=(0.03,0.78) (LT,HT)=(0.03,0.78) (LT,HT)=(0.03,0.78)
(LT,HT)=(0.05,0.68) 0.2 (LT,HT)=(0.05,0.68) (LT,HT)=(0.05,0.68)
0.25
3.5
0.18

3 0.16
0.2
2.5 0.14
IGD

IGD

IGD
0.12
2
0.1 0.15
1.5
0.08
1
0.06 0.1
0.5 0.04

0 0.02
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Environment Environment Environment

DF10 DF11 DF12

1 0.26 1.2
(LT,HT)=(0.01,0.97) (LT,HT)=(0.01,0.97) (LT,HT)=(0.01,0.97)
0.9 (LT,HT)=(0.03,0.78) 0.24 (LT,HT)=(0.03,0.78) (LT,HT)=(0.03,0.78)
(LT,HT)=(0.05,0.68) (LT,HT)=(0.05,0.68) (LT,HT)=(0.05,0.68)
1
0.8 0.22

0.7 0.2
0.8
0.6 0.18
IGD

IGD
IGD

0.5 0.16 0.6

0.4 0.14
0.4
0.3 0.12

0.2 0.1
0.2
0.1 0.08

0 0.06 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Environment Environment Environment

DF13 DF14
0.5 0.3
(LT,HT)=(0.01,0.97) (LT,HT)=(0.01,0.97)
(LT,HT)=(0.03,0.78) (LT,HT)=(0.03,0.78)
0.45 (LT,HT)=(0.05,0.68) (LT,HT)=(0.05,0.68)
0.25

0.4
0.2

0.35
IGD

IGD

0.15
0.3

0.1
0.25

0.05
0.2

0.15 0
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Environment Environment

Fig. 3: Influences of different LT and HT on IGD values (τt = 10, nt = 10).

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 15

TABLE IX: Fixed thresholds vs. adaptive thresholds in terms DF1, DF2, DF3 and DF12, the proposed HRS-DMOA spends
of MIGD where (nt , τt , τ ) = (10, 10, 200) less time and achieves better performance than the variant SR-
Problems HRS(fixed thresholds) HRS*(adaptive thresholds) DMOA, whose reason lies in that if the environment only has
DF1 0.0778±4.95e-02 0.0889±4.85e-02(=) slight changes, taking TL-based response will increase the
DF2 0.0562±1.85e-02 0.0586±2.80e-02(=) computational burden with little performance enhancement.
DF3 0.5146±1.88e-01 0.2561±7.48e-02(-) On the contrary, if the TL method is used in case of dramatic
DF4 0.3584±8.47e-02 0.6212±2.35e-01(+) changes, the negative transfer phenomenon may occur that can
DF5 0.0417±8.90e-03 0.0438±2.09e-02(=) lead to poor convergence. Consequently, it can be concluded
DF6 1.3116±7.17e-01 2.2021±1.63e+00(=) that the designed HRS can enable flexible responses to the
DF7 0.4297±5.99e-02 0.5761±2.01e-01(+) varying environments, which effectively combines the merits
DF8 0.1268±6.17e-02 0.1423±1.14e-01(=) of different mainstream strategies to pursue comprehensive
DF9 0.2216±4.22e-02 0.2687±5.79e-02(+) performance improvement.
DF10 0.2659±5.77e-02 0.3966±8.04e-02(+)
DF11 0.1586±1.44e-02 0.5913±4.04e-02(+)
DF12 0.3370±1.22e-01 0.2974±4.21e-03(=) G. Ablation Study on HRS
DF13 0.2506±2.24e-02 0.2905±1.55e-02(+) In this subsection, extensive ablation studies are carried
DF14 0.1043±9.09e-03 0.2141±4.88e-03(+) out to validate the effectiveness of the proposed HRS. In
+/-/= \ 7/1/6 particular, five bi- (DF1-DF5) and tri-objective (DF10-DF14)
problems are selected for performance evaluation in this group
of experiments, where the dynamic parameters are set as
F. Investigation on Single TL-based Response Mode (nt , τt ) = (5, 10), and each algorithm is run 20 times to report
Recently, developing the TL-based DMOAs has become the MIGD results in average level.
the novel frontier of solving DMOPs [51], which can realize 1) Influences of the Quantification-based Hierarchy:
the seamless combination of the memory- and prediction- Firstly, in the proposed HRS, selection criterion of the different
based response strategy. In this regard, the proposed HRS response modes is based on the quantification of environ-
further integrates the merits of diversity enhancement when mental changes. To validate whether the quantification-based
making responses. To validate whether such designing is response-making is effective, another variant of the proposed
useful, an algorithm variant with single response namely SR- algorithm is designed, which is named RRS-DMOA that takes
DMOA is investigated, where only the TL-based response is the same three response modes with random probability.
adopted regardless of the change degree of environment. The
comparison results in terms of both the MIGD indicator and TABLE XI: Effectiveness of the quantification-based
the average runtime are reported in Table X, where dynamic hierarchical response modes in terms of MIGD
parameters are set to τt = 10, nt = 10.
Algorithms
Problems
TABLE X: Performance comparison against single TL-based HRS-DMOA RRS-DMOA
response in terms of MIGD and average runtime DF1 0.0657±8.28e-03 0.8467±4.68e-01
(τt = 10, nt = 10) DF2 0.0474±1.55e-02 0.5827±1.72e-01
DF3 0.2795±1.32e-01 0.6866±2.29e-01
MIGD Average runtime (seconds) DF4 0.6621±3.52e-02 1.2031±1.48e-01
Problems
HRS-DMOA SR-DMOA HRS-DMOA SR-DMOA DF5 0.0639±1.78e-02 1.8872±4.68e-01
DF1 0.1336 0.1653 2.4201 4.2164 DF10 0.1820±9.42e-03 0.2170±2.25e-02
DF2 0.0597 0.1148 1.2067 3.025 DF11 0.1486±2.84e-03 0.1899±1.10e-02
DF3 0.1871 0.1888 1.5432 2.4524 DF12 0.4715±1.15e-01 0.5050±3.26e-02
DF4 0.6758 0.9439 2.192 1.9164 DF13 0.2710±1.16e-02 2.0003±3.20e-01
DF5 0.3324 0.2464 2.4921 3.3526 DF14 0.1342±3.63e-03 1.2288±2.74e-01
DF6 1.6638 0.9891 3.352 3.482
DF7 31.4148 28.246 2.4963 2.728 According to the results displayed in Table XI, it is found
DF8 0.8603 1.3746 3.9365 3.6527 that the proposed HRS-DMOA outperforms the variant RRS-
DF9 1.521 1.0455 2.708 3.064 DMOA on all of the adopted test problems with overwhelming
DF10 0.2062 0.2168 6.2245 5.4015 advantages. In particular, only on the four problems of DF3
DF11 0.1722 0.2034 6.0723 5.6432 and DF10-DF12 can random response modes obtain similar
DF12 0.8925 0.9125 2.4647 2.7542 results as the hierarchical ones. In other cases, the proposed
DF13 0.2813 0.2832 7.5568 5.7128 HRS presents significant convergence improvement, which
DF14 0.1367 0.1613 7.2126 5.7605 may due to that although the RRS can also exploit merits of
different response strategies, the randomness-based response-
According to Table X, SR-DMOA only achieves 4 best making lacks analysis on the specific environment. Hence,
MIGD results out of 14 problems, and in none of the cases that it can lead to the inappropriate responses, like to transfer
SR-DMOA can outperform the original HRS-DMOA in terms previous knowledge in the extremely fluctuating situations.
of both MIGD and runtime. It is noticeable that on problems Particularly, on problem DF5, an improvement on MIGD by
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 16

TABLE XII: Ablation study on different response modes in terms of MIGD

Algorithms
HRS HRS v1 HRS v2 HRS v3
Refinement ! % ! !
Response
TL ! ! % !
modes
Re-initialization ! ! ! %
DF1 0.0657±8.28e-03 0.0844±2.34e-02 0.2940±6.52e-02 0.0872±1.78e-02
DF2 0.0474±1.55e-02 0.0577±1.57e-02 0.2801±1.29e-01 0.0553±1.83e-02
DF3 0.2795±1.32e-01 0.1732±1.78e-02 0.8096±1.45e-01 0.2659±1.27e-01
DF4 0.6621±3.52e-02 0.7027±4.44e-02 0.6396±1.91e-01 0.7253±9.31e-02
DF5 0.0639±1.78e-02 0.0897±3.92e-02 2.1753±8.03e-02 0.0732±2.88e-02
Problems
DF10 0.1820±9.42e-03 0.1939±8.98e-03 0.2442±2.83e-02 0.2355±1.32e-01
DF11 0.1486±2.84e-03 0.1508±2.34e-03 0.1817±9.70e-03 0.1507±3.41e-03
DF12 0.4715±1.15e-01 0.4171±4.10e-03 0.5271±2.42e-02 0.5421±1.35e-01
DF13 0.2710±1.16e-02 0.2876±1.64e-02 4.1604±1.79e-01 0.2806±9.69e-03
DF14 0.1342±3.63e-03 0.1372±4.26e-03 1.6557±4.68e-02 0.1360±4.02e-03

28.53% is achieved by the quantification-based hierarchical DF4 are illustrated in Fig. 5, and the PF obtained by the four
responses, which indicates that it is necessary to analyze the algorithms is visualized in Fig. 6.
environment so as to make flexible and efficient responses.
In addition, from the scatter plot illustrated in Fig. 4, it can
be observed that the proposed HRS-DMOA can track the
time-varying PF better than RRS-DMOA, which demonstrates
the superiority of the quantification-based response-making
manner.

1.4
True PF(t=21)
1.2 HRS-DMOA(t=21)
RRS-DMOA(t=21)
(a) DF1 (b) DF2
True PF(t=37)
1
HRS-DMOA(t=37)
RRS-DMOA(t=37)
0.8

0.6

0.4

0.2

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(c) DF3 (d) DF4

Fig. 4: Scatter plots on DF1 of the two algorithms Fig. 5: Box plots on four benchmark problems of the four
(HRS-DMOA vs. RRS-DMOA) at the 21- and 37-th algorithms, where the red circles refer to some outliers, the
environments. red solid and blue dashed lines represent the median and
average value, respectively.
2) Effectiveness of the Response Modes: Secondly, in the
proposed HRS, three response modes are adopted, and to According to the above presented results, we have following
verify the effectiveness of them, three variants of the proposed findings:
algorithm are designed by removing each mode from the 1) On most benchmark problems, the original HRS yields
HRS respectively, which are termed as HRS v1 (without the best convergence performance, which validates that the
refinement), HRS v2 (without TL), and HRS v3 (without proposed HRS can flexibly handle different dynamic behav-
iors, and some variances in 10−3 level also present a stable
! %
re-initialization). The comparison results are reported in Ta-
ble XII, where “ ” and “ ” denote whether a certain performance.
response mode is contained in corresponding algorithm or not. 2) When the TL-based response is removed, the HRS v2
Moreover, the box plots regarding to MIGD of problems DF1- presents the most severe performance declination in many
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 17

In future, we aim to 1) propose other advanced response

1.4
True PF
strategies to the environmental changes; 2) seek potential
1.2
HRS integration of some state-of-the-art DMOAs for further per-
HRS_v1
HRS_v2 formance enhancement; 3) develop a learnable DMOP op-
1 HRS_v3 timizer that can realize adaptive parameter configuration; 4)
employ other advanced population-based heuristic algorithms
0.8
as the static optimizer [29], [47], [63], which can improve
0.6
the efficiency of obtaining the Pareto solutions; 5) investigate
more comprehensive quantification manners of environmental
0.4 changes so as to provide in-depth and thorough insights on
dynamic behaviors, which is also beneficial to the research on
0.2 other dynamic systems [19], [65]; 6) improve and apply our al-
gorithm to some real-world optimization problems such as the
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 complex system modeling [30] and the influence maximization
of complex networks [41] so as to validate and enhance the
Fig. 6: Scatter plots on DF1 of the four hierarchical response engineering practicality of the proposed HRS-DMOA.
systems at the 21-th environment.

V. C ONCLUSIONS
cases such as on problem DF1, where the MIGD is as nearly
five times as that obtained by HRS, which indicates the
In this paper, a novel DMOA has been proposed based on
importance of the TL-based response mode.
a hierarchical response system, whose main idea is to realize
3) On tri-objective problems DF10 and DF11, the four
rational and efficient integration of the mainstream methods so
algorithms obtain similar results, while slight convergence
as to enhance comprehensive performance. The environmental
improvement can be achieved by the original HRS, which
changes have been quantified as three levels and, by doing
implies that the complementary performance is realized in the
so, different response modes can be adopted accordingly to
proposed algorithm.
flexibly handle various dynamic behaviors in DMOPs.
4) The response made by the proposed HRS may not
always be the best choice, as can be seen that on DF3 and Benchmark evaluations have been carried out on 14
DF4, the variants HRS v1 and HRS v2 obtain the best result, DMOPs, and the results show that the proposed HRS-DMOA
respectively. The reason may be that the thresholds of HRS outperforms other four popular baseline DMOAs in terms
are manually set. of both convergence and diversity. In addition, extensive
ablation studies have been carried out to validate the su-
5) Without the re-initialization mode, none of the best
periority of the proposed HRS. On one hand, as compared
MIGD is obtained by the HRS v3, while in some cases like
with randomly taking different responses, the quantification-
DF2 and DF14, the HRS v3 can slightly outperform the
based hierarchical response-making has shown overwhelming
HRS v1 and HRS v2. Hence, it can be concluded that the
advantages, which proves the necessity of in-depth analysis
diversity-based response can still be helpful.
on the dynamic behaviors in DMOPs. On the other hand,
6) From the results on DF4, it can be inferred that the
performance declination to different extents has been observed
TL-based response may even cause adverse effects due to the
when some certain response mode is removed from the HRS,
negative transfer phenomenon, which simultaneously implies
which demonstrates the effectiveness of integrating different
the superiority of the rational integration of diverse strategies
mainstream methods.
as none of the single response can always be the best.
Based on above discussions, it can be concluded that the Although some satisfactory results have been obtained,
proposed HRS does realize the systematical integration of several important issues of the proposed HRS-DMOA deserve
different mainstream response strategies, which is effective in further attention, which mainly include (1) insufficient con-
handling various dynamic behaviors in DMOPs. Moreover, cerns on integrating novel state-of-the-art response strategies;
the quantification of environmental changes is also proven (2) manual intervention on the threshold parameters of HRS.
necessary to some extents, which motivates us to develop In future, we tend to propose some novel response methods
other novel schemes to comprehensively quantify the dynamic with other advanced techniques, and to develop a learnable
behaviors in our future work. optimizer that can avoid the manual parameterization is also a
feasible and innovative direction. Then, how to comprehen-
sively quantify the dynamic behaviors from the aspects of
H. Outlook for Future Work both changing severity and frequency can provide us with a
In addition to above satisfactory results, it is worth pointing more thorough understanding on the essence of the DMOPs.
out that there are still spaces for improvements on the proposed In addition, we are prone to integrate the proposed HRS
HRS-DMOA, where the most important issue is that some to other swarm-intelligence-based static optimizer, and it is
state-of-the-art novel strategies are not well considered in the also promising to apply the HRS-DMOA to more real-world
proposed HRS. optimization scenes to validate its practicality.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 18

R EFERENCES [22] Y. Hu, J. Zheng, S. Jiang, J. Zou and R. Wang, A Mahalanobis distance-
based approach for dynamic multi-objective optimization with stochastic
[1] L. Altin and H. Topcuoglu, Impact of sensor-based change detection changes, IEEE Transactions on Evolutionary Computation, 2023. https:
schemes on the performance of evolutionary dynamic optimization //doi.org/10.1109/TEVC.2023.3253850
techniques, Soft Computing, vol. 22, no. 14, pp. 4741-4762, 2018. [23] S. Jiang and S. Yang, Evolutionary dynamic multiobjective optimization:
[2] A. Anand and A. Singh, Hybrid nature-inspired optimization and benchmarks and algorithm comparisons, IEEE Transactions on Cyber-
encryption-based watermarking for E-healthcare, IEEE Transactions netics, vol. 47, no. 1, pp. 198-211, 2017.
on Computational Social Systems, 2022. https://doi.org/10.1109/TCSS. [24] M. Jiang, W. Hu, W. Qiu, M. Li, H. Ming and K. Tan, Solving
2022.3140862 dynamic multi-objective optimization problems via support vector ma-
[3] R. Azzouz, S. Bechikh and L. Said, A multiple reference point-based chine, Proceedings of the 10th International Conference on Advanced
evolutionary algorithm for dynamic multi-objective optimization with Computational Intelligence, pp. 819-824, 2018.
undetectable changes, 2014 IEEE Congress on Evolutionary Computa- [25] M. Jiang, Z. Huang, L. Qiu, W. Huang and G. Yen, Transfer learning-
tion, pp. 3168-3175, 2014. based dynamic multiobjective optimization algorithms, IEEE Transac-
[4] X. Cai, H. Hu, M. Li, Y. Xiao, H. Ishibuchi and Z. Fan, A grid-based tions on Evolutionary Computation, vol. 22, no. 4, pp. 501-514, 2017.
inverted generational distance for multi/many-objective optimization, [26] M. Jiang, Z. Wang, H. Hong and G. Yen, Knee point-based imbalanced
IEEE Transactions on Evolutionary Computation, vol. 25, no. 1, pp. 21- transfer learning for dynamic multiobjective optimization, IEEE Trans-
34, 2021. actions on Evolutionary Computation, vol. 25, no. 1, pp. 117-129, 2020.
[5] B. Chen, Y. Chen, G. Zeng and Q. She, Fractional-order convolutional [27] S. Jiang, S. Yang, X. Yao, K. Tan, M. Kaiser and N. Krasnogor, Bench-
neural networks with population extremal optimization, Neurocomput- mark problems for CEC2018 competition on dynamic multiobjective
ing, vol. 477, pp. 36-45, 2022. optimisation, 2018 CEC Competition, pp. 1-18, 2018.
[6] D. Chen, F. Zou, R. Lu and X. Wang, A hybrid fuzzy inference [28] Q. Li, S. Yang, J. Zheng and G. Ruan, A predictive strategy based on
prediction strategy for dynamic multi-objective optimization, Swarm and special points for evolutionary dynamic multi-objective optimization,
Evolutionary Computation, vol. 43, pp. 147-165, 2018. Soft Computing, vol. 23, pp. 3723-3739, 2019.
[7] J. Chen, Y. Wu, X. Xu, H. Zheng, Z. Ruan and Q. Xuan, PSO-ANE: [29] H. Li, J. Li, P. Wu, Y. You and N. Zeng, A ranking-system-based
Adaptive network embedding with particle swarm optimization, IEEE switching particle swarm optimizer with dynamic learning strategies,
Transactions on Computational Social Systems, vol. 6, no. 4, pp. 649- Neurocomputing, vol. 494, pp. 356-367, 2022.
659, 2019. [30] H. Li, P. Wu, N. Zeng, Y. Liu and F. E. Alsaadi, A survey on
[8] M. Chen, Y. Guo, Y. Jin, S. Yang, D. Gong and Z. Yu, An environment- parameter identification, state estimation and data analytics for lateral
driven hybrid evolutionary algorithm for dynamic multi-objective op- flow immunoassay: from systems science perspective, International
timization problems, Complex and Intelligent Systems, vol. 9, no. 1, Journal of Systems Science, vol. 53, no. 16, pp. 3556-3576, 2022.
pp. 659-675, 2023. [31] Z. Liang, S. Zheng, Z. Zhu and S. Yang, Hybrid of memory and pre-
[9] H. Cheng, Z. Wang, Z. Wei, L. Ma and X. Liu, On adaptive learning diction strategies for dynamic multiobjective optimization, Information
framework for deep weighted sparse autoencoder: A multiobjective Sciences, vol. 485, pp. 200-218, 2019.
evolutionary algorithm, IEEE Transactions on Cybernetics, vol. 52,
[32] Z. Liang, T. Wu, X. Ma, Z. Zhu and S. Yang, A dynamic multiobjective
no. 5, pp. 3221-3231, 2022.
evolutionary algorithm based on decision variable classification, IEEE
[10] C. Cortes and V. Vapnik, Support-vector networks, Machine Learning, Transactions on Cybernetics, vol. 52, no. 3, pp. 1602-1615, 2022.
vol. 20, no. 3, pp. 273-297, 1995.
[33] Z. Liang, X. Xu, L. Liu, Y. Tu and Z. Zhu, Evolutionary many-task op-
[11] W. Dai, Q. Yang, R. Xue and Y. Yu, Boosting for transfer learning,
timization based on multisource knowledge transfer, IEEE Transactions
International Conference on Machine Learning, 2007.
on Evolutionary Computation, vol. 26, no. 2, pp. 319-333, 2022.
[12] K. Deb, N. Rao and S. Karthik, Dynamic multi-objective optimization
[34] J. Lin, H. Liu, K. Tan and F. Gu, An effective knowledge transfer ap-
and decision-making using modified NSGA-II: a case study on hydro-
proach for multiobjective multitasking optimization, IEEE Transactions
thermal power scheduling, International Conference on Evolutionary
on Cybernetics, vol. 51, no. 6, pp. 3238-3248, 2021.
Multi-Criterion Optimization, pp. 803-817, 2007.
[13] J. Derrac, S. Garcia, D. Molina and F. Herrera, A practical tutorial [35] R. Liu, J. Li, C. Mu and L. Jiao, A coevolutionary technique based on
on the use of nonparametric statistical tests as a methodology for multi-swarm particle swarm optimization for dynamic multi-objective
comparing evolutionary and swarm intelligence algorithms, Swarm and optimization, European Journal of Operational Research, vol. 261, no. 3,
Evolutionary Computation, vol. 1, no. 1, pp. 3-18, 2011. pp. 1028-1051, 2017.
[14] J. Eaton, S. Yang and M. Gongora, Ant colony optimization for [36] R. Liu, P. Yang and J. Liu, A dynamic multi-objective optimization
simulated dynamic multi-objective railway junction rescheduling, IEEE evolutionary algorithm for complex environmental changes, Knowledge-
Transactions on Intelligent Transportation Systems, vol. 18, no. 11, Based Systems, vol. 216, article no. 106612, 2021.
pp. 2980-2992, 2017. [37] X. Liu, X. Xu, Z. Zhan, Y. Fang and J. Zhang, Interaction-based
[15] L. Feng, W. Zhou, W. Liu, Y. Ong and K. Tan, Solving dynamic prediction for dynamic multiobjective optimization, IEEE Transactions
multiobjective problem via autoencoding evolutionary search, IEEE on Evolutionary Computation, 2023. doi:10.1109/TEVC.2023.3234113
Transactions on Cybernetics, vol. 52, no. 5, pp. 2649-2662, 2022. [38] X. Liu, Y. Du, M. Jiang and X. Zeng, Multiobjective particle swarm
[16] Y. Guo, H. Yang, M. Chen, J. Cheng and D. Gong, Ensemble prediction- optimization based on network embedding for complex network com-
based dynamic robust multi-objective optimization methods, Swarm and munity detection, IEEE Transactions on Computational Social Systems,
Evolutionary Computation, vol. 48, pp. 156-171, 2019. vol. 7, no. 2, pp. 437-449, 2020.
[17] H. Han, Y. Liu, L. Zhang, H. Liu, H. Yang and J. Qiao, Knowledge [39] X. Liu, Z. Zhan, T. Gu, S. Kwong, Z. Lu, H. Duh and J. Zhang,
reconstruction for dynamic multi-objective particle swarm optimization Neural network-based information transfer for dynamic optimization,
using fuzzy neural network, International Journal of Fuzzy Systems, IEEE Transactions on Neural Networks and Learning Systems, vol. 31,
2023. https://doi.org/10.1007/s40815-023-01477-2 no. 5, pp. 1557-1570, 2020.
[18] M. Helbig and A. Engelbrecht, Performance measures for dynamic [40] Y. Liu, N. Zhu and M. Li, Solving many-objective optimization problems
multi-objective optimisation algorithms, Information Sciences, vol. 250, by a Pareto-based evolutionary algorithm with preprocessing and a
pp. 61-81, 2013. penalty mechanism, IEEE Transactions on Cybernetics, vol. 51, no. 11,
[19] J. Hu, C. Jia, H. Liu, X. Yi and Y. Liu, A survey on state estimation of pp. 5585-5594, 2021.
complex dynamical networks, International Journal of Systems Science, [41] L. Ma, Z. Shao, X. Li, Q. Lin, J. Li, V. Leung and A. Nandi, Influence
vol. 52, no. 16, pp. 3351-3367, 2021. maximization in complex networks by using evolutionary deep rein-
[20] L. Hu, W. Naeem, E. Rajabally, G. Watson, T. Mills, Z. Bhuiyan, forcement learning, IEEE Transactions on Emerging Topics in Computa-
C. Raeburn, I. Salter and C. Pekcan, A multiobjective optimization tional Intelligence, 2021. https://doi.org/10.1109/TETCI.2021.3136643
approach for COLREGs-compliant path planning of autonomous surface [42] J. Ou, M. Li, L. Xing, J. Lv, Y. Hu, N. Dong and G. Zhang, Individual-
vehicles verified on networked bridge simulators, IEEE Transactions on based self-learning prediction method for dynamic multi-objective opti-
Intelligent Transportation Systems, vol. 21, no. 3, pp. 1167-1179, 2020. mization, Information Sciences, vol. 613, pp. 401-418, 2022.
[21] Y. Hu, J. Zheng, S. Jiang, S. Yang and J. Zou, Handling dynamic [43] R. Rambabu, P. Vadakkepat, K. Tan and M. Jiang, A mixture-of-
multiobjective optimization environments via layered prediction and experts prediction framework for evolutionary dynamic multiobjective
subspace-based diversity maintenance, IEEE Transactions on Cybernet- optimization, IEEE Transactions on Cybernetics, vol. 50, pp. 5099-5112,
ics, vol. 53, pp. 2572-2585, 2021. 2019.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final
publication. Citation information: DOI10.1109/TCSS.2023.3293331, IEEE Transactions on Computational Social Systems
FINAL VERSION 19

[44] S. Sahmoud and H. Topcuoglu, Sensor-based change detection schemes [67] H. Zhang, D. Yue, X. Xie, S. Hu and S. Weng, Multi-elite guide hybrid
for dynamic multi-objective optimization problems, 2016 IEEE Sympo- differential evolution with simulated annealing technique for dynamic
sium Series on Computational Intelligence, pp. 1-8, 2016. economic emission dispatch, Applied Soft Computing, vol. 34, pp. 312-
[45] S. Sahmoud and H. Topcuoglu, A memory-based NSGA-II algorithm for 323, 2015.
dynamic multi-objective optimization problems, European Conference [68] J. Zheng, Z. Peng, J. Zou and R. Shen, A prediction strategy based
on the Applications of Evolutionary Computation, pp. 296-310, 2016. on guide-individual for dynamic multi-objective optimization, Acta
[46] S. Sahmoud and H. Topcuoglu, Exploiting characterization of dynamism Electronica Sinica, vol. 43, no. 9, pp. 1816-1825, 2015.
for enhancing dynamic multi-objective evolutionary algorithms, Applied [69] J. Zheng, F. Zhou, J. Zou, S. Yang and Y. Hu, A dynamic multi-
Soft Computing, vol. 85, article no. 105783, 2019. objective optimization based on a hybrid of pivot points prediction
[47] B. Song, H. Miao and L. Xu, Path planning for coal mine robot via and diversity strategies, Swarm and Evolutionary Computation, vol. 78,
improved ant colony optimization algorithm, Systems Science & Control article no. 101284, 2023.
Engineering, vol. 9, no. 1, pp. 283-289, 2021. [70] J. Zheng, Q. Wu, J. Zou, S. Yang and Y. Hu, A dynamic multi-
[48] H. Sun, A. Cao, Z. Hu, X. Li and Z. Zhao, A novel quantile-guided objective evolutionary algorithm using adaptive reference vector and
dual prediction strategies for dynamic multi-objective optimization, linear prediction, Swarm and Evolutionary Computation, vol. 78, article
Information Sciences, vol. 579, pp. 751-775, 2022. no. 101281, 2023.
[71] A. Zhou, Y. Jin and Q. Zhang, A population prediction strategy for
[49] H. Sun, X. Ma, Z. Hu, J. Yang and H. Cui, A two stages prediction
evolutionary dynamic multi-objective optimization, IEEE Transaction
strategy for evolutionary dynamic multi-objective optimization, Applied
on Cybernetics, vol. 33, no. 1, pp. 40-53, 2012.
Intelligence, vol. 53, pp. 1115-1131, 2023.
[72] J. Zou, Q. Li, H. Bai and J. Zheng, A prediction strategy based on
[50] J. Sun, X. Gan, D. Gong, X. Tang, H. Dai and Z. Zhong, A self-
center points and knee points for evolutionary dynamic multi-objective
evolving fuzzy system online prediction-based dynamic multi-objective
optimization, Applied Soft Computing, vol. 61, pp. 806-818, 2017.
evolutionary algorithm, Information Sciences, vol. 612, pp. 638-654,
2022.
[51] K. Tan, L. Feng and M. Jiang, Evolutionary transfer optimization-a
new frontier in evolutionary computation research, IEEE Computational
Intelligence Magazine, vol. 16, no. 1, pp. 22-33, 2021.
[52] F. Wang, Y. Li, F. Liao and H. Yan, An ensemble learning based
prediction strategy for dynamic multi-objective optimization, Applied
Soft Computing, vol. 96, article no. 106592, 2020.
[53] Y. Wang, K. Li and G. Wang, Combining key-points-based transfer
learning and hybrid prediction strategies for dynamic multi-objective
optimization, Mathematics, vol. 10, no. 12, article no. 2117, 2022.
[54] X. Wang, K. Zhang, J. Wang and Y. Jin, An enhanced competitive swarm
optimizer with strongly convex sparse operator for large-scale multiob-
jective optimization, IEEE Transactions on Evolutionary Computation,
vol. 26, no. 5, pp. 859-871, 2021.
[55] X. Wang, Y. Jin, S. Schmitt and M. Olhofer, Transfer learning based co-
surrogate assisted evolutionary bi-objective optimization for objectives
with non-uniform evaluation times, Evolutionary Computation, vol. 30,
no. 2, pp. 221-251, 2022.
[56] Z. Wang, Z. Zhan, W. Yu, Y. Lin, J. Zhang, T. Gu and J. Zhang, Dynamic
group learning distributed particle swarm optimization for large-scale
optimization and its application in cloud workflow scheduling, IEEE
Transactions on Cybernetics, vol. 50, no. 6, pp. 2715-2729, 2020.
[57] F. Wu, J. Chen and W. Wang, A dynamic multi-objective evolutionary
algorithm based on prediction, Journal of Computational Design and
Engineering, vol. 10, no. 1, pp. 1-15, 2023.
[58] D. Xu, M. Jiang, W. Hu, S. Li, R. Pan and G. Yen, An online prediction
approach based on incremental support vector machine for dynamic
multiobjective optimization, IEEE Transactions on Evolutionary Com-
putation, vol. 26, no. 4, pp. 690-703, 2021.
[59] X. Xu, Y. Tan, W. Zheng and S. Li, Memory-enhanced dynamic multi-
objective evolutionary algorithm based on Lp decomposition, Applied
Sciences, vol. 8, no. 9, pp. 1673-1695, 2018.
[60] F. Yao and G. Wang, Transfer learning based on clustering difference for
dynamic multi-objective optimization, Applied Science, vol. 13, no. 8,
article no. 4795, 2023.
[61] Y. Ye, L. Li, Q. Lin, K. Wong, J. Li and Z. Ming, Knowledge
guided Bayesian classification for dynamic multi-objective optimization,
Knowledge-Based Systems, vol. 250, article no. 109173, 2022.
[62] N. Yu, R. Yang and M. Huang, Deep common spatial pattern based
motor imagery classification with improved objective function, Inter-
national Journal of Network Dynamics and Intelligence, vol. 1, no. 1,
pp. 73-84, 2022.
[63] N. Zeng, Z. Wang, W. Liu, H. Zhang, K. Hone and X. Liu, A dynamic
neighborhood-based switching particle swarm optimization algorithm,
IEEE Transactions on Cybernetics, vol. 52, no. 9, pp. 9290-9301, 2022.
[64] Q. Zhang and H. Li, MOEA/D: A multiobjective evolutionary algorithm
based on decomposition, IEEE Transactions on Evolutionary Computa-
tion, vol. 11, no. 6, pp. 712-731, 2007.
[65] W. Zhang, L. Sheng and M. Gao, Finite-horizon H∞ state estimation for
time-varying complex networks based on the outputs of partial nodes,
Systems Science & Control Engineering, vol. 9, no. s2, pp. 48-59, 2021.
[66] G. Zhao, Y. Li and Q. Xu, From emotion AI to cognitive AI, Inter-
national Journal of Network Dynamics and Intelligence, vol. 1, no. 1,
pp. 65-72, 2022.