Swarm and Evolutionary Computation 91 (2024) 101729

Contents lists available at ScienceDirect

Swarm and Evolutionary Computation

journal homepage: www.elsevier.com/locate/swevo

A dimensionality reduction assisted evolutionary algorithm for

high-dimensional expensive multi/many-objective optimization
Zeyuan Yan a , Yuren Zhou a ,∗, Wei Zheng b , Chupeng Su c , Weigang Wu a
a
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, 510006, China
b
Xi’an University of Science and Technology, College of Computer Science and Technology, Xi’an, 250358, China
c
Shien-Ming Wu School of Intelligent Engineering, South China University of Technology, Guangzhou, 511442, China

ARTICLE INFO ABSTRACT

Keywords: Surrogate-assisted multi/many-objective evolutionary algorithms (SA-MOEAs) have shown significant progress
Surrogate model in tackling expensive optimization problems. However, existing research primarily focuses on low-dimensional
Feature extraction optimization problems. The main reason lies in the fact that some surrogate techniques used in SA-MOEAs, such
Expensive optimization
as the Kriging model, are not applicable for exploring high-dimensional decision space. This paper introduces
High-dimensional problems
a surrogate-assisted multi-objective evolutionary algorithm with dimensionality reduction to address high-
dimensional expensive optimization problems. The proposed algorithm includes two key insights. Firstly, we
propose a dimensionality reduction framework containing three different feature extraction algorithms and a
feature drift strategy to map the high-dimensional decision space into a low-dimensional decision space; this
strategy helps to improve the robustness of surrogates. Secondly, we propose a sub-region search strategy to
define a series of promising sub-regions in the high-dimensional decision space; this strategy helps to improve
the exploration ability of the proposed SA-MOEA. Experimental results demonstrate the effectiveness of our
proposed algorithm in comparison to several state-of-the-art algorithms.

1. Introduction when dealing with high-dimensional (with 30 and higher dimension)

EMOPs/EMaOPs [5]. Take the Kriging model as an example. While
Multi-objective optimization problems (MOPs) and many-objective Kriging-based MOEAs are generally effective for low-dimensional
optimization problems (MaOPs) [1] with high economic or physical EMOPs/EMaOPs, they encounter significant challenges in high-dimen-
simulation costs [2] are collectively known as expensive multi/many- sional settings. This is primarily due to the exponential increase in
objective optimization problems (EMOPs/EMaOPs) [3,4]. For such ex- computational complexity associated with training Kriging models as
pensive problems, once function evaluation (FE) may cost hours, days, the number of decision variables grows. Moreover, maintaining predic-
and even weeks [5]. Therefore, conventional multi-objective evolution- tion accuracy in high-dimensional problems necessitates a substantial
ary algorithms (MOEAs) are unsuitable for solving EMOPs/EMaOPs amount of training data, which significantly increases the required
because they must conduct many such FEs before achieving satisfactory number of function evaluations (FEs) [19,20]. For tackling the high-
results. Surrogate-assisted multi-objective evolutionary algorithms (SA- dimensional EMOPs/EMaOP, some successful attempts include [21,22],
MOEAs) [6–12] have been proposed as an effective method to solve
and [23]. The basic ideas of these algorithms can be roughly classified
these problems, which use the solutions in decision space to train surro-
into two types: 1. Designing more robust surrogates and 2. dimension-
gate models, then use surrogate evaluations to replace computationally
ality reduction (DR). The former aims to build more robust surrogates
expensive real FEs to reduce the actual computational cost and assisting
through distinctive model management strategies, such as dropout
MOEAs more efficiently in selecting suitable candidate solutions.
neural network, ensemble model, relationship model, and multi-model.
Despite various effective surrogates, such as support vector ma-
Guo et al. [24] proposed EDN-ARMOEA, which utilizes a computa-
chine (SVM) [13], k-Nearest Neighbor [14], artificial neural network
(ANN) [15], Kriging [16], random forest [17], and radial basis function tionally efficient dropout neural network to approximate the objective
network [18], has been used by SA-MOEAs and these algorithms function. The dropout operation is beneficial for preventing overfitting.
have obtained promising results on low-dimensional (fewer than 30 The effectiveness of EDN-ARMOEA has been demonstrated on a series
dimension) EMOPs/EMaOPs, the curse of dimensionality still exists of testing suites with up to 100 decision variables. HSMEA [25]

∗ Corresponding author.
E-mail address: [email protected] (Y. Zhou).

https://doi.org/10.1016/j.swevo.2024.101729
Received 10 April 2024; Received in revised form 17 August 2024; Accepted 5 September 2024
Available online 16 September 2024
2210-6502/© 2024 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

utilizes multiple surrogates to build ensemble models to approximate Although the above algorithms have shown promising results on
objective functions. Ensemble models offer a compelling advantage high-dimensional EMOPs, significant challenges remain. One key issue
in preventing overfitting, primarily by reducing variance, enhancing is information loss; DR techniques often result in the loss of crucial
generalization, mitigating overfitting risks, and improving robustness to information from the original decision space, making it difficult for
outliers and noise. Li et al. [26] proposed LSA-MoEO to tackle the high- SA-MOEAs to construct effective surrogates. Most SA-MOEAs attempt
dimensional problems, which employs multiple evolutionary operators to train multiple surrogates with different feature sets and construct an
to generate a series of offspring and builds a lightweight local surrogate ensemble model to hedge the information loss issue [21,22]. However,
model to preselect the best candidate as the final offspring. A model- it bypasses enriching the representation capability of dimensional-
free evolutionary optimization rule is also employed by LSA-MoEO, reduced features and does not tackle the information loss issue. Besides,
which is important in preventing the population from stagnation. Hao multi-surrogates also bring about time-consuming issues. The key in-
et al. proposed REMO [27] for tackling high-dimensional problems. sight presented is that, rather than training multiple surrogates for
The classifier of REMO is based on the relationship between pairs of different feature sets, a single feature set with strong representational
solutions instead of its fitness value to predict the quality of solutions. capability can be obtained through an effective feature extraction
REMO has obtained promising results on DTLZ 1-7 [28] up to 50 process. This allows SA-MOEAs to achieve robust performance without
decision variables. Takumi et al. [23] proposed MCEA/D, a multi- increasing the number of surrogates. The second challenge is surrogate
objective evolutionary algorithm based on decomposition (MOEA/D) effectiveness. The core of SA-MOEAs is to use surrogates to preselect
that incorporates SVM for enhanced performance [29]. For each sub- the best candidate solutions. However, these surrogates are not always
problem, MCEA/D constructs an SVM model using a set of historical effective, particularly when dealing with complex EMOPs that have lim-
best individuals as positive examples and the remaining individuals as ited training data. Problems like DTLZ1 and DTLZ3 [28] are difficult to
negative examples. When optimizing each sub-problem, MCEA/D gen- approximate, even in low-dimensional scenarios. Consequently, MOEAs
erates multiple offspring solutions and evaluates them using the SVM. without surrogates may outperform SA-MOEAs in solving such com-
Finally, the most suitable offspring solution is selected for the actual plex EMOPs/EMaOPs. To tackle this issue, some successful attempts
evaluation. MCEA/D performs well on EMOPs/EMaOPs problems with have been obtained by model-free rule [26,34] in the single-objective
50–150 dimensions. optimization area. However, it is still not well-scaled to MOPs/MaOPs.
Although these SA-MOEAs have shown promising results, they still Following the above idea, we propose a surrogate-assisted MOEA/D
grapple with the challenge of surrogate sensitivity to high-dimensional with a feature extraction based DR framework, namely MOEA/D-FEF,
optimization problems. Specifically, as problem dimensionality in- to address high-dimensional EMOPs/EMaOPs; the proposed algorithm
creases, the accuracy of surrogates substantially decreases, causing designs a feature extraction framework that balances the contributions
SA-MOEAs to perform poorly in tackling high-dimensional problems, of linear and nonlinear information, which can provide a more compre-
and in some cases, to fail completely [5]. As a result, many researchers hensive understanding of the data thereby increasing the robustness of
are now focusing on DR techniques to address this challenge. surrogate models. On the other hand, inspired by the model-free evo-
The DR operation aims to map a high-dimensional decision space lutionary optimization rule, this article proposed a sub-region search
into a low-dimensional one, enabling SA-MOEAs to build surrogates (SRS) operation to enhance the search ability in solving complexity
in the latter to alleviate issues related to dimensional sensitivity. ESF- problems, which can local a promising sub-region in decision space
RVEA [30] randomly divides the decision space into a set of sub-spaces without surrogate or real evaluation. The paper makes the following
based on the divide-and-conquer operation and builds a set of surrogate key contributions:
models for those sub-spaces, where surrogate models are combined
into an ensemble model using a weighted method to hedge the infor- 1. The contribution about the DR process. This article proposes
mation loss. Although the divide and conquer strategy is an effective a feature extraction framework to map the high-dimensional
DR method, it may destroy the correlation of the original features. decision space into a series of low-dimensional decision spaces
Therefore, many researchers pay attention to using feature extraction based on multiple feature extraction techniques, including linear
to execute the DR operation. Principal component analysis (PCA) is an or nonlinear methods. Then, MOEA/D-FEF selects a series of fea-
effective feature extraction method that is widely used by SA-MOEAs. tures from those low-dimensional decision spaces based on the
Guo et al. [22] introduced HeE-MOEA, which employs an ensemble variance explained ratio to preserve the linear information of the
approach by integrating nine independent surrogate models. The en- original decision space. On the other hand, this article proposes
semble is trained using three distinct feature groups: PCA-based feature a nonlinear feature adjustment method, namely feature drift
extraction, feature selection, and global features. The impressive per- strategy, to adjust the relative positioning of the dimensionality-
formance of HeE-MOEA in high-dimensional spaces can be attributed reduced data to preserve the nonlinear information.
to its practical feature-handling approach. SA-RVEA-PCA [31] builds a 2. The contribution about the optimizer. MOEA/D-FEF employs
Gaussian process model with PCA to improve the model accuracy for the MOEA/D as its optimizer. To enhance exploration capabil-
each objective function. SA-RVEA-PCA has proven effective in solving ities, a sub-region search operation is used to transform the
problems with up to 160 decision variables. Some approaches attempt original decision space into a discrete decision space; each so-
to enhance the performance of SA-MOEAs through adaptive dropout lution in the discrete decision space is denoted as a sub-region
mechanisms. ADSAPSO [32] proposed a dropout method, a linear DR in the original decision space. This operation is used to identify
technique for model training; it takes advantage of the statistical dif- a series of promising sub-regions through continuous iterations,
ferences between different solution sets in the decision space to guide thereby enhancing the exploration capabilities. In addition, an
the selection of some crucial decision variables. As an extension of adaptive switch strategy is applied to balance the exploration
MCEA/D, Horaguchi et al. [33] proposed a DR mechanism to boost the and exploitation.
accuracy of classification models, which employs an adaptive dropout
ANN for dimension reduction and builds classification models in low- 2. Preliminaries
dimensional decision space to select promising solutions. To overcome
the information loss issue caused by the DR process, Liu et al. [21] 2.1. A multi-objective evolutionary algorithm based on decomposition
proposed a Kriging model-assisted MOEA/D, namely GPEME, to tackle
expensive problems with up to 50 dimensions; GPEME employs both Zhang and Li et al. proposed MOEA/D [29], a typical decomp-
the PCA and Sammon mapping to extract the linear and nonlinear osition-based MOEA to solve MOPs. MOEA/D decomposes an MOP into
information of the original decision space. a series of single objective subproblems based on a set of uniformly

2
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

distributed weight vectors and optimizes these subproblems simulta- 1. Construction neighborhood graph: If 𝒙𝒊 is one point of the k-
neously in one run. It is based on specific decomposition methods, nearest neighbors of 𝒙𝒋 or the two points are closer than 𝜖, the
such as the penalty-based boundary intersection approach, weighted 𝒙𝒋 are connected to 𝒙𝒊 . This process is repeated for all points to
sum approach, and Tchebycheff approach. The Tchebycheff approach construct a neighborhood graph. The edge lengths 𝑑𝑥 (𝑖, 𝑗) are set
is defined as follows. equal to the distance between the two endpoints.
Suppose 𝝀 = {𝜆1 , …, 𝜆𝑁 } is a set of 𝑁 uniformly distributed weight 2. Shortest path computation: 𝑑𝐺 (𝑖, 𝑗) is set as 𝑑𝑥 (𝑖, 𝑗) if 𝒙𝒊 is on
vectors, in which each weight vector 𝝀𝑖 = (𝜆𝑖1 , … , 𝜆𝑖𝑚 ), 𝑖 ∈ (1, … , 𝑁) point of the k-nearest neighbors of 𝒙𝒋 otherwise 𝑑𝐺 (𝑖, 𝑗) is set to
∑
satisfies the constraint 𝑚 𝑖 𝑖
𝑗=1 𝜆𝑗 = 1, 𝜆𝑗 > 0, 𝑚 is the number of objectives
∞. Then, the shortest path algorithm is used to find the shortest
of the current problem. Initialize ideal point 𝒛∗ = (𝑧∗1 , … , 𝑧∗𝑚 ), where distances between all pairs of points in 𝐆.
𝒛∗𝑖 = 𝑚𝑖𝑛{𝑓𝑖 (𝑥)|𝑥 ∈ 𝑃 }, where 𝑃 is denoted as the population of 3. Construction of lower dimensional embedding: MDS is used to
MOEA/D. MOEA/D divides the MOP into 𝑁 subproblems based on the construct lower dimensional embedding.
Tchebycheff technique:
(3) LPP aims to preserve local linear relationships, not just global
𝑔(𝑥|𝜆𝑖 , 𝑧∗ ) = 𝑚𝑎𝑥 𝜆𝑖𝑗 ‖𝑓𝑗 (𝑥) − 𝑧∗𝑗 ‖ (1) variance. It considers the local relationships between data points and
1≤𝑗≤𝑚
uses a reconstruction weight matrix to reduce dimensionality. The core
MOEA/D takes 𝑔(𝒙|𝝀𝑖 , 𝒛∗ ) as the fitness function value of individual 𝒙. steps of LPP include the computation of the reconstruction weight
Due to the weight vector being uniformly distributed, the 𝑔(𝒙|𝝀𝑖 , 𝒛∗ ) of matrix and eigenvalue decomposition.
𝒙 can measure its diversity and convergence. Firstly, LPP constructs a neighborhood graph, which is the same
| 𝒙𝒊 −𝒙𝒋 | 2
as Isomap. If 𝒙𝒊 is connected to 𝒙𝒋 , 𝑤𝑖,𝑗 = 𝑒𝑥𝑝− 𝑡 , otherwise
2.2. Kriging model
𝑊𝑖,𝑗 = 0. LPP then computes a lower-dimensional representation 𝐗′ ,
which minimizes the reconstruction error while preserving the local
The Kriging model [16], or Gaussian process regression, is a widely relationships.
used technique for spatial interpolation and prediction in geostatistics.
∑𝑛 ‖ ∑ ‖2
It assumes that the underlying spatial process is stationary and can be ‖ ′ ‖
min ‖𝐱 − 𝑤𝑖𝑗 𝐱𝑗′ ‖ (4)
described by a mean function and a covariance function. Kriging can ‖ 𝑖 ‖
𝐗′
‖
𝑖=1 ‖ ‖
be used to estimate the value of a variable at an unsampled location
𝑗∈neighbors(𝑖) ‖2
by incorporating spatial correlation information among nearby sampled
3. Proposed algorithm
areas. The algorithm principles are shown below.
Let 𝑍(𝐬) be a spatial process representing the variable of interest
3.1. Algorithm framework
at location 𝐬. As shown in Eq (2), Kriging assumes that 𝑍(𝐬) follows
a Gaussian distribution with a mean function 𝜇(𝐬) and a covariance
function 𝛾(𝐬, 𝐬′ ): Fig. 1 presents the algorithm framework of MOEA/D-FEF, which
employs MOEA/D as the optimizer, decomposing expensive EMOPs/
𝑍(𝑠) ∽  (𝜇(𝐬), 𝛾(𝐬, 𝐬′ )) (2) EMaOPs into a series of subproblems based on the Tchebycheff tech-
nique. This article uses the ‘‘′ ’’ symbol to comprehend the dimen-
The mean function 𝜇(𝐬) represents the global trend of the spatial pro-
sionality-reduced data for easy understanding. MOEA/D-FEF can be
cess, which can be estimated from the available data. The covariance
divided into six steps.
function 𝛾(𝐬, 𝐬′ ) describes the spatial correlation between the values of
(1) Initialization. MOEA/D-FEF initializes a population, 𝐩𝐨𝐩, in this
the variable at two locations 𝐬 and 𝐬′ , which depends on the distance
step. All solutions were under real FE.
and direction between the two locations.
(2) Offspring generation process. This step generates a series of
promising offspring for surrogate evaluation. Firstly, to enhance the
2.3. Feature extraction exploration capabilities, MOEA/D-FEF constructs a discrete decision
space Y𝐷 ⊂ R𝐷 based on SRS operation. Each solution in Y𝐷 de-
Feature extraction techniques aim to map the higher dimensional notes a sub-region in R𝐷 , and a sub-population 𝐄𝐏 is built to local a
data set to the lower dimensional data set. In past decades, many promising sub-region with continuous iterations; meanwhile, the other
unsupervised and discriminant methods of feature extraction tech- sub-population 𝐋𝐏 is built to search promising offspring within the
niques have been proposed, such as PCA and Multi-Dimensional Scal- defined sub-region to enhance the exploitation capabilities. Finally, this
ing (MDS) [35] are linear methods widely applied in data visualiza- step returns an offspring set generated by 𝐄𝐏 or 𝐋𝐏.
tion, making them appealing in revealing the linear relation between (3) Feature extraction and surrogate construction. This step con-
two sets of features. The nonlinear methods include Isometric Feature structs a feature extraction framework (FEF) to map both the training
Mapping (Isomap) [36], Locality Preserving Projection (LPP) [37], set 𝐓 and offspring set 𝐎 into their low-dimensional counterparts, 𝐓′
and Laplacian Eigenmap [38] etc; those methods are robust enough and 𝐎′ , respectively. Then, the dimensionality-reduced training data is
to handle nonlinear data. This paper employs three classical feature used to train a series of Kriging models to evaluate the offspring set.
extraction techniques, including PCA, LPP, and Isomap, to map the (4) Updating sub-population. This step uses the surrogate evaluated
higher-dimensional decision space into a series of lower-dimensional offspring set to update 𝐄𝐏 or 𝐋𝐏, respectively, then repeat steps (2)-(4)
decision spaces. until the predefined 𝑅′𝑚𝑎𝑥 iterations run out.
(1) PCA analyzes the distribution of the training set, considering the (5) Adaptive switching strategy and population update process. The
variable 𝒙 in R𝐷 , PCA exact a solution 𝒙′ ∈ R𝑑 based on the mapping optimization process is divided into two sampling cases: (1) requir-
matrix 𝐌 ∈ R𝑑×𝐷 . ing more convergence and (2) requiring more diversity. An adaptive
switching strategy dynamically selects the appropriate case based on
𝒙′ = (𝒙 − 𝒙𝑚𝑒𝑎𝑛 )𝐌 (3)
the current optimization process. MOEA/D-FEF then generates off-
where 𝒙𝑚𝑒𝑎𝑛 is the mean vector of the training set, 𝐌 is the eigenvector spring based on the selected case, which is used to execute the real FE.
matrix. Finally, these offspring are utilized to update the current population
(2) Isomap aims to find an optimal subspace for preserving the and are added to the training set.
original space. The algorithm of Isomap involves three steps: (6) Repeat steps (2)-(6) until the computation resource runs out.

3
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

Fig. 1. The algorithm framework of MOEA/D-FEF.

3.2. Offspring generation and sub-region search operation

When SA-MOEAs tackle high-dimensional EMOPs/EMaOPs, the

complex decision space makes it hard to construct accurate surrogate
models and execute efficient global searches. Ideally, training a promis-
ing surrogate requires ample training data evenly spread across the
decision space. However, this is unfeasible due to limited FEs and
data sparsity in high-dimensional spaces. A more practical approach
is to ensure training data is as evenly distributed as possible near the
Pareto set (PS). This strategy helps in training reliable local models that
accurately recognize points near the PS, aiding SA-MOEAs in filtering
promising offspring.

𝑌 (𝑥) ∶= arg min ‖𝑥 − 𝑐‖2 (5)

𝑐∈Y𝐷

We introduced the sub-region selection operation (SRS) to achieve

this goal. Inspired by the K-means clustering principle, the SRS opera-
tion generates (𝑆 +1)𝐷 sub-regions that are uniformly distributed in the
original decision space and gets the center points of each sub-region,
where 𝐷 represents the number of dimensionalities. The SRS operation
considers the fitness value of the center point as the fitness for each Fig. 2. The narrowed decision space under SRS operation. The red line is the PS, 𝑌 (𝑥)
is the center point, 𝑥 is a solution in the sub-region of 𝑌 (𝑥), and the orange zone
sub-region, thereby mapping the offspring set 𝑶 ⊂ R𝐷 to the discrete indicates the narrowed decision space under SRS operation.
decision space 𝑌 (𝑶) ⊂ Y𝐷 ⊂ R𝐷 . Here, Y𝐷 consists of the (𝑆+1)𝐷 center
points. The mapping relationship from R𝐷 to Y𝐷 is expressed in Eq. (5),
which finds the center point 𝑐 ∈ Y𝐷 closest to the point 𝑥 and replaces 𝑥
𝐋𝐏, which searches for promising solutions within the identified sub-
with 𝑐. Based on the fundamental principles of evolutionary algorithms,
regions. This search process occurs in the original decision space R𝐷 ,
we postulate that solutions with higher fitness levels are more closely
ensuring center points do not replace generated offspring. Therefore,
related to the PS. Consequently, by searching the better center points this article proposes an adaptive switching and population update
within Y𝐷 , MOEA/D-FEF can locate a series of promising sub-regions strategy to balance exploration and exploitation.
in R𝐷 . Fig. 2 shows the sub-region searched by the SRS operation in The adaptive switching strategy, as
⌊ ⌋ illustrated in Fig. 1, divides all
2-D decision space, and the red line represents the PS. The promising function evaluations into 𝛿 =
𝑅max −𝑁
parts, with each part assigned 𝐿
𝐿
sub-region corresponding to 𝑌 (𝑥) is local in the orange zone, and it is function evaluations. MOEA/D-FEF selects promising offspring for real
found by mapping the 𝑥(0.28, 0.58) to Y𝐷 . function evaluations for each part based on different sampling cases.
The SRS operation is designed to accelerate convergence by identi-
fying promising sub-regions to approximate the PS. While center points 1. Adaptive switching process: This process helps decide whether
can closely approach the PS, they only partially represent it; replacing to focus on diversity or convergence
⌊ ⌋ during the current optimiza-
the population (𝐩𝐨𝐩) with center points can reduce population diver- tion stage. A total of 𝛾 = 𝐵2 function evaluations are reserved
sity. To address this issue, the algorithm introduces a subpopulation, for this decision, where 𝐵 is the neighbor size. These evaluations

4
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

are split equally: half are used for the diversity-demand case, and Algorithm 1: Feature extraction framework
the other half for the convergence-demand case. After comparing Input: 𝐎 (offspring set), 𝐓 (training set), 𝐾 (the maximum number of
the results, the case that yields the better mean fitness value is features), 𝐿𝑅 (learning rate), 𝑅 (shift ratio)
′ ′
selected to guide the remaining 𝐿 − 𝛾 function evaluations. Output: 𝐎 (dimensionally reduced offspring set), 𝐓 (dimensionally
reduced training set)
2. Diversity-demand case: This case focuses on maintaining diver- 1 𝐉 = [𝐓; 𝐎];
′ ′ ′
sity. It searches for promising offspring based on crossover and 2 𝐉1 , 𝐉2 , 𝐉3 ← Apply PCA, Isomap, and LPP to process 𝐉;
mutation, then selects the best offspring in 𝐋𝐏 for real FE. 3 𝑖𝑛𝑑 ← Calculate the feature matrix using Eq. (6);
′ ′

3. Convergence-demand case: This case emphasizes convergence. 4 𝐉 = 𝐉𝑖𝑛𝑑 ;

5 𝐄𝐽 ← Calculate the distance matrix of 𝐉 by Eqs (7) and (8);
It searches for promising sub-regions using the SRS operation ′
6 Initialize the shift matrix 𝐐 as 𝐐 ← 𝐉 ;
and then selects the best center points in 𝐄𝐏 for real FE. 7 for 𝑖𝑡𝑒𝑟 = 1 ∶ 𝐼𝑚𝑎𝑥 do
8 𝐄𝑄 ← Calculate the distance matrix of 𝐐;
Based on the current case, an offspring set is selected from 𝐄𝐏 or 9  ← Calculate loss based on Eq. (9);
𝐋𝐏 for real function evaluation. This selected offspring set is then used 10 ∇ ← Calculate gradient based on ;
11 𝐐 = 𝐐 − 𝐿𝑅 ∗ ∇ ;
to update the population 𝐩𝐨𝐩 and the reference point 𝑧∗ , after which 12 end
the evaluated offspring are added to the overall training data. 13
′ ′
𝐉 = 𝐉 + 𝑅 ∗ (𝐉 − 𝐐);
′

′ ′
[∶,1]),…,𝑣𝑎𝑟(𝐉 [∶,𝐷])
14 𝐕𝐄𝐑 = 𝑠𝑜𝑟𝑡( 𝑣𝑎𝑟(𝐉 𝑣𝑎𝑟(𝐉)
) // Descending order;
3.3. Feature extraction framework 15 𝑑 = 𝐶𝑜𝑢𝑛𝑡(𝐕𝐄𝐑 > 𝐾1 );
∑𝑑
16 if 𝑖=1 𝐕𝐄𝐑(𝑖) > 𝛽 and 𝑑 < 𝐾 then
After generating an offspring set, MOEA/D-FEF trains a series of sur- 17
′ ′
𝐉 = 𝐉 [∶, 1 ∶ 𝑑];
rogates to evaluate these offspring. However, as discussed in Section 1, 18 end
constructing an effective surrogate in high-dimensional problems poses 19 else
′ ′
20 𝐉 = 𝐉 [∶, 1 ∶ 𝐾];
a significant challenge for SA-MOEAs due to the dimensionality sen- 21 end
sitivity of surrogates. To address this issue, this article proposes the 22
′ ′
𝐓 = 𝐉 [1 ∶ 𝑙𝑒𝑛𝑔𝑡ℎ(𝐓), ∶];
′ ′
feature extraction framework (FEF) to map both the training set 𝐓 and 23 𝐎 = 𝐉 [𝑙𝑒𝑛𝑔𝑡ℎ(𝐓) + 1 ∶ 𝑙𝑒𝑛𝑔𝑡ℎ(𝐓) + 𝑁, ∶];
′ ′
offspring set 𝐎 into their low-dimensional counterparts, 𝐓′ and 𝐎′ , 24 return 𝐎 , 𝐓 ;
respectively. Then MOEA/D-FEF trains surrogate by 𝐓′ to evaluate 𝐎′
for hedging the dimensionality sensitivity issue. Two key issues must
be carefully considered in this process.

1. The DR process inevitably results in information loss, which

⎛ 0 𝑑1,2 ⋯ 𝑑1,𝑁 ⎞
can significantly affect the performance of surrogate models. ⎜ ⎟
𝑑 0 ⋯ 𝑑2,𝑁 ⎟
Therefore, developing a method that preserves a diverse range 𝐄𝐽 = ⎜ 2,1 (8)
of information from the original feature space, including both ⎜ ⋮ ⋮ ⋱ ⋮ ⎟
⎜𝑑 𝑑𝑁,2 ⋯ 0 ⎟⎠
linear and nonlinear components, is crucial. ⎝ 𝑁,1
2. Balancing different types of information and determining the ∑
𝑁 ∑
𝑁
𝐽 𝑄 2
optimal number of final features are essential tasks. These factors = (𝐸𝑖,𝑗 − 𝐸𝑖,𝑗 ) (9)
critically influence both the representational capability of the 𝑖=1 𝑗=1

features and the accuracy of the surrogate models. Besides linear information, nonlinear aspects, such as the global
spatial relationship, significantly influence the effectiveness of feature
This paper tackles the first challenge with a dual-focused ap-
extraction. Ideally, the relative positions of points in the dimensionally-
proach: first, by maximizing the variance explained ratio (VER) of
the reduced-dimensional data to retain linear information, and sec- reduced data should closely resemble those in the original space. This
ond, by preserving global spatial relationships to capture nonlinear article proposes a feature drift strategy for preserving the relative
information effectively. Specifically, our framework utilizes PCA (a distance information. This process is shown in Algorithm 1 lines 5–13.
linear method) alongside Isomap and LPP (nonlinear methods) to map Firstly, for obtaining the global relative location of the original data,
the high-dimensional decision space into three lower-dimensional sub- line 5 computes the distance matrix 𝐄𝐽 of the original data. Eqs (7) and
spaces. To ensure maximum retention of linear information by these (8) show this process, if 𝐉 is a 𝑁 × 𝐷 matrix, which can be transformed
feature extraction methods, MOEA/D-FEF employs Eq. (6) to identify into a 𝑁 × 𝑁 distance matrix 𝐄𝐽 , where 𝑎(𝑖,∶) and 𝑎(𝑗,∶) are two different
𝐽
solutions of 𝐉, 𝑑(𝑖,𝑗) represents the Euclidean distance between 𝑎(𝑖,∶)
the feature matrix with the highest VER, where VER represents the
proportion of total variance explained by each feature. This process and 𝑎(𝑗,∶) ; line 6 initializes the drift matrix 𝐐 as the dimensionality-
is outlined in Algorithm 1, lines 2–4, with 𝐽 denoting the combined reduced matrix 𝐉′ ; line 8 computes and updates the distance matrix
training and offspring set for streamlined processing. of 𝐐; lines 7–12 optimize the 𝐐 based on gradient descent process;
Line 9 defines the loss function  based on Eq. (9), which measures
𝑣𝑎𝑟(𝐉′1 ) 𝑣𝑎𝑟(𝐉′2 ) 𝑣𝑎𝑟(𝐉′3 )
𝑖𝑛𝑑 = 𝑎𝑟𝑔𝑚𝑎𝑥( , , ) (6) the deviation between 𝐄𝑄 and 𝐄𝐽 . A smaller value of  indicates that
𝑣𝑎𝑟(𝐉) 𝑣𝑎𝑟(𝐉) 𝑣𝑎𝑟(𝐉) the distributions of 𝐐 and 𝐉 are more closely aligned; line 13 executes
where the 𝑐𝑜𝑣(𝐉′1 ), 𝑐𝑜𝑣(𝐉′2 ), and 𝑐𝑜𝑣(𝐉′3 ) are the variance of the feature the feature drift operation, as the result, 𝐉′ is moved towards the 𝐐
matrices extracted by PCA, Isomap, and LPP, respectively. 𝑖𝑛𝑑 is the to preserve the global spatial information. To balance the linear and
feature matrices with the largest VER. As shown in Algorithm 1, line 2 nonlinear information, the feature shift ratio 𝑅 is used to adjust the
employs these methods to handle 𝐉 and generates three feature matrices move distance of the feature drift operation.
𝐉′1 , 𝐉′2 , and 𝐉′3 , lines 3–4 select the feature matrix with largest VER, and Any DR method inherently leads to information loss, and the num-
denote it as 𝐉′ . ber of retained features impacts the amount of preserved information
⎛ 𝑎1,1 𝑎1,2 ⋯ 𝑎1,𝐷 ⎞ and the accuracy of the surrogate model. Typically, more features
⎜ 𝑎 𝑎2,2 ⋯ 𝑎2,𝐷 ⎟ retain more information about the original decision space. This article
⎜ 2,1 ⎟ adaptively selects the number of retained features. Lines 16–22 show
𝐉=⎜ ⋮ ⋮ ⋱ ⋮ ⎟ (7)
⎜𝑎(𝑁−1),1 𝑎(𝑁−1),2 ⋯ 𝑎(𝑁−1),𝐷 ⎟ the feature selection process. Line 14 computes the VER of each feature
⎜ ⎟ of the 𝐉′ and sorts it based on descending order; line 15 computes the
⎝ 𝑎𝑁,1 𝑎𝑁,2 ⋯ 𝑎𝑁,𝐷 ⎠

5
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

Table 1
IGD statistical results on all test instances with D = 50.
Problem M ADSAPSO CPS-MOEA K-RVEA KTA2 MCEA/D HeE-MOEA MOEA/D-FEF
3 1.2448e+3 (5.29e+1) − 1.0755e+3 (8.79e+1) − 1.2342e+3 (8.56e+1) − 1.1215e+3 (9.54e+1) − 9.7339e+2 (1.46e+2) − 1.1373e+3 (6.01e+1) − 1.1491e+2 (6.24e+1)
DTLZ1 7 8.9710e+2 (5.75e+1) − 7.8940e+2 (8.93e+1) − 8.8484e+2 (8.09e+1) − 8.4172e+2 (8.74e+1) − 7.7388e+2 (1.10e+2) − 8.8967e+2 (7.43e+1) − 8.5972e+1 (3.50e+1)
10 7.8371e+2 (6.93e+1) − 7.4738e+2 (6.94e+1) − 8.0198e+2 (9.17e+1) − 7.4630e+2 (6.84e+1) − 6.2972e+2 (8.85e+1) − 8.1166e+2 (5.92e+1) − 1.2387e+2 (4.48e+1)
3 2.4557e+0 (3.28e−1) − 2.4403e+0 (5.78e−1) − 2.9583e+0 (1.75e−1) − 2.6588e+0 (2.03e−1) − 9.3426e−1 (1.23e−1) = 5.8833e−1 (2.46e−2) + 1.0608e+0 (2.80e−1)
DTLZ2 7 2.8469e+0 (2.07e−1) − 2.5996e+0 (4.60e−1) − 3.0612e+0 (1.35e−1) − 2.8647e+0 (1.68e−1) − 1.4939e+0 (1.65e−1) − 9.5691e−1 (2.77e−2) + 1.1350e+0 (1.48e−1)
10 2.7991e+0 (2.61e−1) − 2.7458e+0 (3.55e−1) − 2.8534e+0 (1.90e−1) − 2.7489e+0 (1.53e−1) − 1.3783e+0 (1.24e−1) = 1.0548e+0 (2.57e−2) + 1.3925e+0 (1.90e−1)
3 3.5075e+3 (3.29e+2) − 3.2535e+3 (1.84e+2) − 3.7589e+3 (1.60e+2) − 3.5186e+3 (2.44e+2) − 2.2691e+3 (5.78e+2) − 3.6157e+3 (1.97e+2) − 3.6774e+2 (1.79e+2)
DTLZ3 7 3.3348e+3 (2.29e+2) − 2.9247e+3 (2.20e+2) − 3.2535e+3 (2.16e+2) − 3.4505e+3 (1.98e+2) − 2.3497e+3 (4.49e+2) − 3.2990e+3 (1.88e+2) − 3.2462e+2 (1.15e+2)
10 3.1974e+3 (2.79e+2) − 2.7140e+3 (1.61e+2) − 3.0732e+3 (1.87e+2) − 3.0755e+3 (1.81e+2) − 1.9549e+3 (3.75e+2) − 3.0713e+3 (1.60e+2) − 3.4272e+2 (1.46e+2)
3 2.8926e+0 (3.25e−1) − 2.9189e+0 (5.22e−1) − 3.2859e+0 (2.04e−1) − 3.2100e+0 (1.79e−1) − 1.2133e+0 (1.17e−1) = 1.1056e+0 (2.24e−2) + 1.4334e+0 (3.56e−1)
DTLZ4 7 2.9815e+0 (2.05e−1) − 2.8138e+0 (3.41e−1) − 3.1808e+0 (2.04e−1) − 3.1785e+0 (1.19e−1) − 1.5058e+0 (9.55e−2) = 1.3235e+0 (2.70e−2) + 1.5931e+0 (1.90e−1)
10 2.9418e+0 (1.61e−1) − 2.6504e+0 (3.12e−1) − 3.0200e+0 (1.43e−1) − 3.0888e+0 (1.43e−1) − 1.3907e+0 (3.23e−2) + 1.3552e+0 (2.44e−2) + 1.8703e+0 (3.81e−1)
3 2.4917e+0 (4.05e−1) − 2.3772e+0 (4.76e−1) − 2.9314e+0 (1.38e−1) − 2.6491e+0 (2.08e−1) − 9.1754e−1 (2.14e−1) − 4.5229e−1 (3.00e−2) + 8.9960e−1 (4.51e−1)
DTLZ5 7 2.5910e+0 (1.86e−1) − 2.2250e+0 (4.41e−1) − 2.6333e+0 (1.36e−1) − 2.5986e+0 (1.29e−1) − 9.8774e−1 (2.32e−1) − 4.0286e−1 (3.05e−2) + 7.8772e−1 (2.08e−1)
10 2.3671e+0 (1.43e−1) − 2.3158e+0 (4.02e−1) − 2.3841e+0 (1.50e−1) − 2.3863e+0 (1.21e−1) − 7.2664e−1 (1.33e−1) + 3.7385e−1 (3.38e−2) + 9.1497e−1 (3.09e−1)
3 3.8207e+1 (2.79e+0) − 3.5927e+1 (1.23e+0) − 3.9215e+1 (6.76e−1) − 4.1083e+1 (6.48e−1) − 3.0413e+1 (2.78e+0) − 4.2225e+1 (2.72e−1) − 2.2852e+1 (3.30e+0)
DTLZ6 7 3.7043e+1 (1.25e+0) − 3.8605e+1 (3.69e−1) − 3.8214e+1 (6.77e−1) − 3.8726e+1 (1.96e−1) − 3.4401e+1 (2.15e+0) − 3.8612e+1 (2.54e−1) − 1.9703e+1 (2.86e+0)
10 3.2972e+1 (2.03e+0) − 3.0137e+1 (1.00e+0) − 3.4519e+1 (5.17e−1) − 3.5563e+1 (4.31e−1) − 2.6254e+1 (1.70e+0) − 3.6031e+1 (2.22e−1) − 1.7420e+1 (3.75e+0)
3 8.0748e+0 (1.11e+0) = 9.9390e+0 (7.73e−1) − 1.1905e+0 (2.73e−1) + 8.8337e+0 (7.77e−1) − 9.6669e+0 (7.88e−1) − 1.0228e+1 (6.41e−1) − 7.1689e+0 (1.85e+0)
DTLZ7 7 2.3730e+1 (1.55e+0) − 2.3485e+1 (1.03e+0) − 4.4501e+0 (1.12e+0) + 2.3404e+1 (1.77e+0) − 2.3388e+1 (1.67e+0) − 2.4517e+1 (1.33e+0) − 7.1689e+0 (1.85e+0)
10 3.2543e+1 (1.58e+0) − 3.3722e+1 (1.52e+0) − 8.4443e+0 (1.84e+0) + 3.3304e+1 (1.59e+0) − 3.2090e+1 (2.30e+0) − 3.4582e+1 (1.75e+0) − 2.4072e+1 (4.08e+0)
3 2.3539e+0 (5.50e−2) = 2.3537e+0 (5.63e−2) = 2.0293e+0 (8.09e−2) + 2.1907e+0 (1.07e−1) + 2.3300e+0 (4.38e−2) = 2.3778e+0 (6.35e−2) = 2.2998e+0 (8.37e−2)
WFG1 7 2.9140e+0 (9.15e−2) + 2.9928e+0 (4.01e−2) = 2.8239e+0 (4.60e−2) + 2.9969e+0 (5.58e−2) = 2.9972e+0 (3.17e−2) = 3.0047e+0 (3.41e−2) = 2.9857e+0 (6.06e−2)
10 3.4000e+0 (5.01e−2) + 3.4521e+0 (3.00e−2) = 3.3499e+0 (3.97e−2) = 3.4116e+0 (3.61e−2) = 3.4515e+0 (2.26e−2) = 3.4683e+0 (3.28e−2) = 3.4526e+0 (4.25e−2)
3 1.0032e+0 (6.20e−2) = 9.1941e−1 (4.22e−2) − 9.4234e−1 (3.45e−2) − 8.8748e−1 (7.11e−2) = 8.5603e−1 (7.35e−2) + 9.7428e−1 (1.02e−1) − 8.8251e−1 (1.10e−1)
WFG2 7 2.8702e+0 (4.83e−1) − 2.3305e+0 (3.07e−1) = 1.9739e+0 (3.24e−1) + 2.2929e+0 (4.51e−1) + 2.4482e+0 (4.46e−1) = 2.8282e+0 (5.50e−1) − 2.5046e+0 (5.05e−1)
10 4.5151e+0 (8.55e−1) − 3.7913e+0 (7.13e−1) = 3.1180e+0 (9.24e−1) + 3.1430e+0 (7.29e−1) + 3.9752e+0 (5.40e−1) = 4.1623e+0 (6.89e−1) − 3.5287e+0 (7.41e−1)
3 8.1567e−1 (1.89e−2) − 8.3639e−1 (2.54e−2) − 8.2716e−1 (1.33e−2) − 8.0912e−1 (1.69e−2) − 6.4756e−1 (3.58e−2) + 5.5225e−1 (3.18e−2) + 7.2103e−1 (6.98e−2)
WFG3 7 1.4451e+0 (3.73e−2) − 1.4835e+0 (4.57e−2) − 1.4376e+0 (3.18e−2) − 1.4204e+0 (2.32e−2) − 1.1879e+0 (6.40e−2) + 9.8302e−1 (5.52e−2) + 1.3969e+0 (5.76e−2)
10 1.8391e+0 (7.42e−2) = 1.8802e+0 (7.99e−2) = 1.8508e+0 (5.32e−2) = 1.8554e+0 (3.36e−2) = 1.6238e+0 (9.21e−2) + 1.2596e+0 (7.42e−2) + 1.8205e+0 (5.50e−2)
3 7.9149e−1 (6.38e−2) − 6.5041e−1 (2.55e−2) = 6.2451e−1 (1.76e−2) + 6.1516e−1 (2.11e−2) + 6.3152e−1 (5.18e−2) + 8.0298e−1 (5.64e−2) − 6.7866e−1 (3.61e−2)
WFG4 7 5.4845e+0 (3.12e−1) − 4.1941e+0 (3.35e−1) − 3.6081e+0 (4.32e−1) = 3.9560e+0 (4.90e−1) = 4.3868e+0 (4.20e−1) − 5.5413e+0 (3.28e−1) − 3.8745e+0 (2.34e−1)
10 1.0063e+1 (5.71e−1) − 8.4048e+0 (4.20e−1) − 8.5859e+0 (9.48e−1) − 8.2067e+0 (7.02e−1) − 8.5901e+0 (9.11e−1) − 1.0326e+1 (4.92e−1) − 6.8022e+0 (3.31e−1)
3 8.3343e−1 (2.59e−2) − 7.7745e−1 (1.79e−2) − 7.1711e−1 (2.29e−2) + 7.8522e−1 (1.15e−2) − 7.2904e−1 (3.42e−2) = 8.5483e−1 (1.55e−2) − 7.3825e−1 (5.86e−2)
WFG5 7 4.1202e+0 (2.06e−1) − 3.8196e+0 (1.61e−1) − 3.3168e+0 (2.54e−1) + 3.2708e+0 (1.27e−1) + 3.7393e+0 (2.16e−1) = 3.5130e+0 (6.73e−2) = 3.6464e+0 (3.04e−1)
10 7.9971e+0 (3.41e−1) − 7.3873e+0 (3.58e−1) − 6.8806e+0 (5.24e−1) = 6.4303e+0 (3.61e−1) + 7.2242e+0 (2.74e−1) − 6.4059e+0 (1.55e−1) + 6.7322e+0 (5.61e−1)
3 9.7838e−1 (2.92e−2) + 1.0231e+0 (3.04e−2) = 9.4497e−1 (1.74e−2) + 9.5880e−1 (1.88e−2) + 9.2852e−1 (4.95e−2) + 9.1840e−1 (2.63e−2) + 1.0299e+0 (1.12e−1)
WFG6 7 4.2975e+0 (2.88e−1) − 4.1860e+0 (2.09e−1) − 3.5281e+0 (1.89e−1) + 3.5235e+0 (2.33e−1) + 4.1052e+0 (2.81e−1) − 4.4870e+0 (2.27e−1) − 3.8571e+0 (2.33e−1)
10 8.3128e+0 (5.32e−1) − 7.3873e+0 (3.58e−1) − 6.9006e+0 (5.24e−1) + 6.4303e+0 (3.61e−1) + 7.2242e+0 (2.74e−1) − 6.4059e+0 (1.55e−1) + 7.3689e+0 (6.45e−1)
3 7.7964e−1 (1.95e−2) − 7.9127e−1 (3.77e−2) − 7.8405e−1 (1.32e−2) − 7.6527e−1 (2.62e−2) = 7.1283e−1 (3.10e−2) + 7.4290e−1 (3.46e−2) + 7.7085e−1 (2.43e−2)
WFG7 7 4.4216e+0 (3.47e−1) − 4.2114e+0 (2.92e−1) − 4.1689e+0 (4.87e−1) − 3.7078e+0 (3.13e−1) = 4.2805e+0 (3.17e−1) − 4.5874e+0 (3.35e−1) − 3.8974e+0 (3.05e−1)
10 8.3813e+0 (4.20e−1) − 8.0535e+0 (4.33e−1) − 8.5872e+0 (7.41e−1) − 7.5214e+0 (9.11e−1) = 8.1850e+0 (4.67e−1) − 8.9298e+0 (5.39e−1) − 7.5524e+0 (5.61e−1)
3 8.7976e−1 (2.42e−2) = 9.0623e−1 (2.17e−2) = 8.1947e−1 (2.14e−2) + 8.3064e−1 (2.60e−2) + 8.4057e−1 (4.58e−2) + 8.6148e−1 (3.10e−2) + 8.9886e−1 (2.69e−2)
WFG8 7 4.4263e+0 (1.91e−1) − 4.1596e+0 (2.15e−1) − 3.7887e+0 (2.33e−1) = 3.5076e+0 (2.03e−1) + 4.1871e+0 (2.36e−1) − 4.5445e+0 (1.77e−1) − 3.8617e+0 (2.68e−1)
10 8.3903e+0 (3.41e−1) − 7.9704e+0 (5.26e−1) − 7.8362e+0 (5.47e−1) − 7.2734e+0 (4.78e−1) − 8.1575e+0 (4.16e−1) − 8.3273e+0 (3.74e−1) − 7.0212e+0 (6.21e−1)
3 1.0308e+0 (3.26e−2) = 1.0647e+0 (3.35e−2) = 1.0879e+0 (3.20e−2) = 1.0275e+0 (4.85e−2) = 9.0272e−1 (5.18e−2) + 8.0032e−1 (4.14e−2) + 1.0042e+0 (8.39e−2)
WFG9 7 4.6005e+0 (2.94e−1) − 4.4624e+0 (2.57e−1) = 4.3086e+0 (3.61e−1) = 3.9602e+0 (3.12e−1) + 4.1961e+0 (2.57e−1) = 3.4061e+0 (8.37e−2) + 4.2282e+0 (3.99e−1)
10 8.1882e+0 (3.98e−1) − 8.2517e+0 (3.62e−1) − 8.0480e+0 (5.67e−1) − 7.4394e+0 (5.19e−1) = 7.7722e+0 (4.77e−1) = 6.1275e+0 (1.56e−1) + 7.6076e+0 (6.12e−1)
3 2.9079e+0 (3.57e−1) − 2.4986e+0 (5.64e−1) − 3.4217e+0 (2.03e−1) − 3.0198e+0 (2.14e−1) − 7.9250e−1 (1.41e−1) = 4.7206e−1 (1.82e−1) + 7.8680e−1 (2.51e−1)
MaF1 7 4.6141e+0 (1.02e+0) − 4.6698e+0 (7.53e−1) − 5.6127e+0 (3.99e−1) − 4.7990e+0 (6.40e−1) − 1.2380e+0 (2.03e−1) + 9.1380e−1 (7.21e−2) + 2.0739e+0 (5.26e−1)
10 5.0529e+0 (7.05e−1) − 5.6205e+0 (1.18e+0) − 5.9477e+0 (7.05e−1) − 5.0629e+0 (7.46e−1) − 1.6102e+0 (2.97e−1) + 9.5052e−1 (9.29e−2) + 2.4904e+0 (7.84e−1)
3 2.3892e−1 (1.02e−2) − 2.5838e−1 (1.63e−2) − 2.6807e−1 (7.51e−3) − 2.4566e−1 (7.74e−3) − 1.6128e−1 (9.11e−3) + 2.4080e−1 (4.93e−3) − 1.8810e−1 (2.66e−2)
MaF2 7 2.6888e−1 (1.70e−2) + 2.9110e−1 (1.88e−2) = 2.9612e−1 (1.86e−2) = 2.7691e−1 (2.50e−2) + 3.0258e−1 (1.86e−2) = 3.0546e−1 (1.48e−2) = 3.0077e−1 (2.74e−2)
10 3.3474e−1 (2.39e−2) = 3.5588e−1 (2.61e−2) − 3.5370e−1 (2.04e−2) − 3.2870e−1 (3.99e−2) = 3.5763e−1 (1.59e−2) − 3.5458e−1 (3.03e−2) − 3.2607e−1 (2.64e−2)
3 2.0920e+7 (4.80e+6) − 1.3621e+7 (3.55e+6) − 1.8033e+7 (2.69e+6) − 2.8367e+7 (5.50e+6) − 5.4875e+6 (3.24e+6) − 2.1660e+7 (2.63e+6) − 2.8011e+4 (2.62e+4)
MaF3 7 2.2897e+7 (6.09e+6) − 1.2728e+7 (2.86e+6) − 1.9976e+7 (5.94e+6) − 2.6175e+7 (7.96e+6) − 7.9798e+6 (3.01e+6) − 2.5273e+7 (6.29e+6) − 3.8947e+4 (3.79e+4)
10 2.0852e+7 (7.02e+6) − 9.9729e+6 (1.85e+6) − 1.5355e+7 (2.62e+6) − 1.8708e+7 (4.30e+6) − 5.5094e+6 (2.31e+6) − 2.0917e+7 (5.48e+6) − 4.4292e+5 (1.35e+6)
3 1.2187e+4 (1.18e+3) − 1.0684e+4 (9.25e+2) − 1.2429e+4 (8.77e+2) − 1.2471e+4 (8.04e+2) − 9.1033e+3 (1.85e+3) − 1.2076e+4 (4.36e+2) − 9.0697e+2 (5.75e+2)
MaF4 7 1.9965e+5 (1.07e+4) − 1.7509e+5 (1.18e+4) − 1.9629e+5 (1.73e+4) − 2.0022e+5 (1.31e+4) − 1.6668e+5 (2.07e+4) − 1.8276e+5 (8.19e+3) − 2.2889e+4 (1.32e+4)
10 1.4775e+6 (1.16e+5) − 1.2734e+6 (1.37e+5) − 1.4243e+6 (8.10e+4) − 1.4791e+6 (1.02e+5) − 1.2778e+6 (1.69e+5) − 1.3714e+6 (7.19e+4) − 1.7212e+5 (8.05e+4)
3 1.2314e+1 (5.05e+0) = 1.0005e+1 (7.40e−1) − 9.4632e+0 (6.76e−1) = 1.0657e+1 (3.04e+0) = 7.6003e+0 (8.89e−1) + 9.8870e+0 (8.10e−1) − 8.9352e+0 (1.17e+0)
MaF5 7 4.7263e+1 (1.49e+1) = 4.3866e+1 (6.18e+0) + 3.2032e+1 (6.17e+0) + 5.0461e+1 (2.55e+1) = 4.1816e+1 (6.39e+0) + 8.9055e+1 (2.78e+1) − 5.1548e+1 (1.45e+1)
10 2.1665e+2 (9.14e+1) + 2.1404e+2 (3.12e+1) + 1.8720e+2 (4.26e+1) + 2.3691e+2 (1.16e+2) + 2.4604e+2 (4.42e+1) = 6.0717e+2 (2.62e+2) − 2.5121e+2 (3.70e+1)
3 2.3751e+2 (2.13e+1) − 2.2030e+2 (5.82e+1) − 2.7388e+2 (1.76e+1) − 2.3162e+2 (1.98e+1) − 5.2552e+1 (1.74e+1) + 2.4786e+2 (1.98e+1) − 7.9346e+1 (3.74e+1)
MaF6 7 2.3738e+2 (2.33e+1) − 1.9095e+2 (5.15e+1) − 2.4173e+2 (2.33e+1) − 2.2971e+2 (1.77e+1) − 6.6605e+1 (2.13e+1) = 2.3443e+1 (4.63e+0) + 6.2429e+1 (2.53e+1)
10 2.1036e+2 (2.22e+1) − 1.9450e+2 (3.71e+1) − 2.1347e+2 (1.43e+1) − 2.1608e+2 (1.02e+1) − 6.4642e+1 (1.57e+1) = 1.9224e+1 (3.98e+0) + 6.3420e+1 (2.30e+1)
3 8.3447e+0 (1.08e+0) = 9.8729e+0 (6.84e−1) − 1.1755e+0 (4.18e−1) + 8.5171e+0 (9.69e−1) = 9.6200e+0 (5.52e−1) − 9.3431e+0 (5.54e−1) − 8.3408e+0 (1.52e+0)
MaF7 7 2.3067e+1 (1.69e+0) − 2.4286e+1 (1.06e+0) − 3.9206e+0 (1.04e+0) + 2.2744e+1 (2.05e+0) − 2.3345e+1 (2.06e+0) − 2.3787e+1 (1.13e+0) − 1.6076e+1 (3.19e+0)
10 3.2985e+1 (2.00e+0) − 3.3710e+1 (1.77e+0) − 8.0140e+0 (1.71e+0) + 3.3935e+1 (1.45e+0) − 3.3023e+1 (1.87e+0) − 3.4131e+1 (2.16e+0) − 2.2729e+1 (4.73e+0)
+/-/= 5/54/10 2/55/12 19/42/9 14/41/14 17/34/18 25/39/5
Avg Rank 3.8261 3.5217 3.1594 2.8841 2.4782 2.8261 2.1542
𝑝 value 1.6654e−9 3.9060e−5 8.7449e−4 1.7000e−2 1.0000e+0 2.0000e−2

VER of the first 𝑑 features, where 𝐾 is a parameter used to define 4. Experimental studies
the maximum number of selected features; 𝛽 is a parameter used to
measure the proportion of total variance information retained after DR This section uses a series of EMOPs/EMaOPs to evaluate the perfor-
process, it is set as 99.9% in our algorithm. The feature selection pro- mance of MOEA/D-FEF. 23 test suites are selected from DTLZ 1-7 [28],
∑ WFG 1-9 [39], and MaF 1-7 [40], with decision variables set to 50,
cess prioritizes features with VER greater than 0.999
𝐾
. If 𝑑𝑖=1 𝐕𝐄𝐑(𝑖) > 𝛽,
100, and 150, respectively. For all test suites, the number of objectives
and 𝑑 < 𝐾, MOEA/D-FEF will use the first 𝑑 features to training
is set to 3, 7, and 10. The performance of the compared algorithms is
surrogates; if 𝑑 > 𝐾, this algorithm will use the first 𝐾 features to
∑ measured by IGD [41] and 𝛥𝑝 [42]. To statistically analyze the results,
training surrogates; finally, if 𝐾𝑖=1 𝐕𝐄𝐑(𝑖) > 𝛽, the algorithm uses the the Friedman rank test is conducted at a 0.05 significance level to
first 𝐾 features to limit the number of retained features. Lines 22–23 compare the differences between the algorithms. The notations ‘‘+/-/=’’
restore the 𝐉′ into a training set 𝐓′ and an offspring set 𝐎′ . 𝐓′ is used indicate whether the compared SA-MOEAs perform better than, worse
to train promising surrogates to preselect promising offspring from 𝐎′ than, or similarly to the MOEA/D-FEF, respectively. The best results are
to update 𝐄𝐏 or 𝐋𝐏. highlighted with a gray background. ‘‘Avg Rank’’ represents the average

6
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

Table 2
𝛥𝑃 statistical results on all test instances with D = 50.
Problem M ADSAPSO CPS-MOEA K-RVEA KTA2 MCEA/D HeE-MOEA MOEA/D-FEF
3 1.7555e+3 (6.88e+1) − 1.5664e+3 (3.26e+1) − 1.7677e+3 (3.08e+1) − 1.7659e+3 (4.88e+1) − 1.5324e+3 (9.72e+1) − 1.7240e+3 (3.61e+1) − 6.6609e+2 (5.46e+2)
DTLZ1 7 1.4811e+3 (4.90e+1) − 1.3364e+3 (3.12e+1) = 1.5589e+3 (5.17e+1) − 1.4901e+3 (6.38e+1) − 1.2898e+3 (4.22e+1) = 1.4847e+3 (2.98e+1) − 9.5946e+2 (5.15e+2)
10 1.3590e+3 (6.16e+1) = 1.2485e+3 (3.39e+1) = 1.4512e+3 (4.34e+1) − 1.3994e+3 (3.93e+1) − 1.1776e+3 (5.12e+1) = 1.4104e+3 (2.13e+1) − 1.1142e+3 (4.52e+2)
3 3.1962e+0 (4.92e−1) = 3.9032e+0 (1.29e−1) − 3.9554e+0 (1.12e−1) − 3.2062e+0 (1.81e−1) = 2.8422e+0 (4.65e−1) = 1.9269e+0 (2.41e−1) + 2.9073e+0 (7.25e−1)
DTLZ2 7 3.5585e+0 (1.71e−1) − 3.9608e+0 (1.07e−1) − 3.6578e+0 (9.45e−2) − 3.5095e+0 (7.79e−2) − 3.5175e+0 (3.08e−1) − 2.0741e+0 (1.61e−1) + 2.8366e+0 (4.62e−1)
10 3.3939e+0 (1.60e−1) − 3.7558e+0 (6.11e−2) − 3.2908e+0 (7.87e−2) − 3.2801e+0 (7.39e−2) − 3.0438e+0 (2.06e−1) − 2.1480e+0 (9.82e−2) + 2.8200e+0 (2.06e−1)
3 4.2717e+3 (4.17e+2) − 4.1418e+3 (9.95e+1) − 4.5502e+3 (1.16e+2) − 4.4662e+3 (1.91e+2) − 3.7775e+3 (4.18e+2) − 4.6938e+3 (6.21e+1) − 1.2714e+3 (1.07e+3)
DTLZ3 7 4.5380e+3 (9.98e+1) − 4.1223e+3 (7.53e+1) − 4.3783e+3 (1.26e+2) − 4.5578e+3 (5.27e+1) − 3.9036e+3 (1.69e+2) − 4.4734e+3 (4.98e+1) − 1.9293e+3 (1.27e+3)
10 4.3230e+3 (1.07e+2) − 3.8796e+3 (8.82e+1) − 4.1385e+3 (7.41e+1) − 4.2016e+3 (6.89e+1) − 3.4434e+3 (2.54e+2) − 4.1927e+3 (3.79e+1) − 2.5804e+3 (1.07e+3)
3 3.4737e+0 (2.87e−1) − 3.8299e+0 (1.45e−1) − 3.8950e+0 (1.94e−1) − 3.5845e+0 (2.62e−1) − 2.4192e+0 (3.32e−1) = 2.1034e+0 (3.99e−1) = 2.4589e+0 (6.92e−1)
DTLZ4 7 3.5191e+0 (9.08e−2) − 4.0869e+0 (9.54e−2) − 3.6760e+0 (9.79e−2) − 3.4615e+0 (1.14e−1) − 3.0393e+0 (1.82e−1) − 2.2137e+0 (7.85e−2) + 2.8655e+0 (3.49e−1)
10 3.4048e+0 (8.77e−2) − 3.8675e+0 (8.18e−2) − 3.3833e+0 (5.29e−2) − 3.3991e+0 (1.16e−1) − 2.1740e+0 (9.21e−2) + 2.2540e+0 (3.51e−2) + 3.1708e+0 (2.39e−1)
3 3.3225e+0 (2.78e−1) − 3.9851e+0 (1.36e−1) − 3.9477e+0 (1.22e−1) − 3.2894e+0 (2.20e−1) − 2.7281e+0 (3.44e−1) = 1.7952e+0 (2.23e−1) + 2.4915e+0 (5.95e−1)
DTLZ5 7 3.6922e+0 (1.77e−1) − 3.9629e+0 (8.86e−2) − 3.7277e+0 (9.04e−2) − 3.5925e+0 (8.81e−2) − 3.5058e+0 (3.16e−1) − 2.1566e+0 (2.28e−1) + 2.9311e+0 (4.82e−1)
10 3.4890e+0 (6.28e−2) − 3.7668e+0 (7.30e−2) − 3.3793e+0 (5.82e−2) − 3.3988e+0 (9.14e−2) − 3.1159e+0 (3.25e−1) = 2.3741e+0 (1.79e−1) + 2.8798e+0 (3.40e−1)
3 4.2891e+1 (8.65e−1) − 4.0522e+1 (3.73e−1) − 4.2486e+1 (2.27e−1) − 4.3286e+1 (1.14e−1) − 3.9440e+1 (1.92e+0) = 4.4044e+1 (2.61e−2) − 3.7599e+1 (3.10e+0)
DTLZ6 7 4.0068e+1 (1.42e−1) − 4.0043e+1 (3.62e−2) − 3.9978e+1 (6.39e−2) − 4.0069e+1 (1.07e−2) − 3.9855e+1 (1.18e−1) − 4.0061e+1 (2.05e−2) − 3.9480e+1 (3.31e−1)
10 3.7233e+1 (2.27e−1) − 3.5852e+1 (1.66e−1) − 3.6972e+1 (1.78e−1) − 3.7352e+1 (7.92e−2) − 3.4606e+1 (6.78e−1) − 3.7662e+1 (1.39e−2) − 3.0729e+1 (2.54e+0)
3 8.3095e+0 (1.23e+0) + 1.1040e+1 (4.72e−1) − 1.3742e+0 (3.46e−1) + 1.0072e+1 (7.49e−1) = 1.0762e+1 (5.32e−1) = 1.1379e+1 (3.02e−1) − 9.9458e+0 (1.14e+0)
DTLZ7 7 2.7128e+1 (5.50e−1) − 2.7540e+1 (3.42e−1) − 6.4093e+0 (8.06e−1) + 2.7075e+1 (5.52e−1) − 2.7449e+1 (3.83e−1) − 2.8172e+1 (2.67e−1) − 2.4562e+1 (1.25e+0)
10 3.9238e+1 (6.63e−1) − 3.9476e+1 (5.10e−1) − 2.3015e+1 (1.94e+0) + 3.9080e+1 (5.23e−1) − 3.8835e+1 (6.18e−1) − 4.0610e+1 (1.57e−1) − 3.5425e+1 (1.24e+0)
3 2.3539e+0 (5.50e−2) = 2.3460e+0 (5.72e−2) = 1.9917e+0 (7.33e−2) + 2.1048e+0 (9.33e−2) + 2.3357e+0 (4.37e−2) = 2.2712e+0 (3.78e−2) + 2.3314e+0 (5.68e−2)
WFG1 7 2.9140e+0 (9.15e−2) + 2.9987e+0 (4.15e−2) = 2.8203e+0 (4.61e−2) + 2.8784e+0 (4.03e−2) + 2.9978e+0 (3.19e−2) = 3.0057e+0 (3.55e−2) = 3.0199e+0 (5.61e−2)
10 3.4000e+0 (5.01e−2) + 3.4511e+0 (3.08e−2) = 3.3572e+0 (3.59e−2) + 3.4108e+0 (3.90e−2) + 3.4506e+0 (2.30e−2) = 3.4718e+0 (3.34e−2) = 3.4528e+0 (3.20e−2)
3 1.0032e+0 (6.20e−2) = 9.2058e−1 (4.62e−2) = 9.2566e−1 (2.95e−2) = 8.9335e−1 (8.25e−2) = 8.4803e−1 (7.79e−2) + 8.4914e−1 (3.77e−2) + 9.3612e−1 (1.31e−1)
WFG2 7 2.8702e+0 (4.83e−1) − 2.4100e+0 (2.90e−1) = 2.0186e+0 (3.41e−1) + 2.2991e+0 (4.95e−1) = 2.4908e+0 (4.77e−1) = 2.8694e+0 (5.07e−1) − 2.4182e+0 (6.42e−1)
10 4.5151e+0 (8.55e−1) − 3.8060e+0 (6.55e−1) = 3.2254e+0 (9.87e−1) + 3.1122e+0 (6.94e−1) = 3.8767e+0 (5.47e−1) + 4.1681e+0 (6.89e−1) = 3.7798e+0 (8.81e−1)
3 9.8669e−1 (2.12e−2) + 1.0485e+0 (1.87e−2) = 1.0321e+0 (1.56e−2) = 9.8899e−1 (1.92e−2) + 9.5383e−1 (5.46e−2) + 9.7590e−1 (1.06e−2) + 1.0682e+0 (6.06e−2)
WFG3 7 2.1268e+0 (1.17e−1) + 2.2504e+0 (1.10e−1) + 2.0642e+0 (5.97e−2) + 2.0256e+0 (7.08e−2) + 1.9457e+0 (7.87e−2) + 1.3957e+0 (3.51e−2) + 2.5757e+0 (3.01e−1)
10 2.8941e+0 (1.25e−1) + 3.0781e+0 (8.32e−2) + 2.7585e+0 (8.51e−2) + 2.8052e+0 (1.14e−1) + 2.6926e+0 (1.33e−1) + 1.9196e+0 (2.12e−2) + 3.5910e+0 (2.37e−1)
3 7.9149e−1 (6.38e−2) − 6.5275e−1 (2.74e−2) + 6.1941e+1 (2.38e−2) + 5.9415e−1 (2.18e−2) + 6.3633e−1 (5.61e−2) + 6.0789e−1 (1.89e−2) + 7.0945e−1 (6.48e−2)
WFG4 7 5.4845e+0 (3.12e−1) − 4.1853e+0 (3.54e−1) − 3.5538e+0 (4.64e−1) = 4.0065e+0 (5.29e−1) = 4.4589e+0 (4.39e−1) − 5.5162e+0 (3.53e−1) − 3.8696e+0 (5.03e−1)
10 1.0063e+1 (5.71e−1) − 8.3790e+0 (3.80e−1) − 8.6656e+0 (9.41e−1) − 8.1257e+0 (6.57e−1) − 8.6373e+0 (9.25e−1) − 1.0392e+1 (4.30e−1) − 6.9432e+0 (7.14e−1)
3 8.3343e−1 (2.59e−2) − 7.7335e−1 (1.59e−2) − 7.1647e−1 (1.83e−2) + 7.8582e−1 (2.11e−2) − 7.3852e−1 (2.11e−2) = 7.7092e−1 (4.02e−3) − 7.5196e−1 (3.64e−2)
WFG5 7 4.1202e+0 (2.06e−1) − 3.8158e+0 (1.73e−1) = 3.3097e+0 (2.74e−1) + 3.2615e+0 (1.17e−1) + 3.7464e+0 (2.10e−1) = 3.5190e+0 (6.57e−2) = 3.7263e+0 (2.61e−1)
10 7.9971e+0 (3.41e−1) − 7.4203e+0 (3.90e−1) − 6.8726e+0 (5.73e−1) = 6.4218e+0 (3.25e−1) = 7.2527e+0 (2.84e−1) = 6.4088e+0 (1.60e−1) = 6.8730e+0 (7.26e−1)
3 9.9217e−1 (1.60e−2) + 1.0237e+0 (1.44e−2) = 9.6949e−1 (1.18e−2) + 9.5451e−1 (8.19e−3) + 9.3946e−1 (5.33e−2) + 9.7658e−1 (5.57e−3) + 1.0372e+0 (7.25e−2)
WFG6 7 4.2975e+0 (2.88e−1) − 4.2220e+0 (1.91e−1) − 3.4946e+0 (1.79e−1) + 3.5292e+0 (2.35e−1) + 4.0958e+0 (2.93e−1) − 4.4741e+0 (2.14e−1) − 3.8828e+0 (2.34e−1)
10 8.3128e+0 (5.32e−1) − 7.9101e+0 (3.42e−1) − 7.6593e+0 (3.28e−1) − 6.9006e+0 (6.36e−1) = 7.9151e+0 (4.08e−1) − 8.3322e+0 (3.59e−1) − 7.1089e+0 (5.51e−1)
3 7.7964e−1 (1.95e−2) − 7.8712e−1 (3.27e−2) − 7.8576e−1 (1.77e−2) − 7.4799e−1 (1.36e−2) = 7.1574e−1 (2.98e−2) + 7.2686e−1 (6.93e−3) + 7.5908e−1 (2.43e−2)
WFG7 7 4.4216e+0 (3.47e−1) − 4.1884e+0 (3.13e−1) = 4.1456e+0 (4.48e−1) = 3.7571e+0 (3.08e−1) + 4.2611e+0 (3.28e−1) = 4.5254e+0 (3.36e−1) − 4.0514e+0 (4.63e−1)
10 8.3813e+0 (4.20e−1) − 7.9895e+0 (4.45e−1) − 8.5322e+0 (8.11e−1) − 7.5800e+0 (9.37e−1) = 8.2498e+0 (4.28e−1) − 9.0036e+0 (5.55e−1) − 7.4023e+0 (8.03e−1)
3 8.7976e−1 (2.42e−2) = 9.0952e−1 (2.18e−2) − 8.2432e−1 (1.57e−2) + 8.3291e−1 (2.87e−2) + 8.3809e−1 (4.58e−2) + 8.0749e−1 (7.94e−3) + 8.8881e−1 (4.11e−2)
WFG8 7 4.4263e+0 (1.91e−1) − 4.1880e+0 (2.12e−1) − 3.7387e+0 (2.60e−1) + 3.4845e+0 (2.13e−1) + 4.1481e+0 (2.26e−1) − 4.5172e+0 (1.71e−1) − 3.9190e+0 (2.14e−1)
10 8.3903e+0 (3.41e−1) − 8.0331e+0 (5.56e−1) − 7.9069e+0 (4.63e−1) − 7.3963e+0 (4.47e−1) = 8.1588e+0 (3.97e−1) − 8.2834e+0 (3.99e−1) − 7.3897e+0 (7.08e−1)
3 1.0315e+0 (3.20e−2) = 1.0960e+0 (2.34e−2) − 1.1092e+0 (2.68e−2) − 1.0179e+0 (3.38e−2) = 9.2791e−1 (5.73e−2) + 1.0113e+0 (1.63e−2) = 1.0181e+0 (1.09e−1)
WFG9 7 4.6005e+0 (2.94e−1) − 4.5094e+0 (2.57e−1) − 4.3265e+0 (3.48e−1) = 3.9583e+0 (3.20e−1) = 4.1851e+0 (2.63e−1) = 3.4096e+0 (8.14e−2) + 4.2217e+0 (3.57e−1)
10 8.1882e+0 (3.98e−1) − 8.2607e+0 (3.70e−1) − 7.9165e+0 (4.98e−1) = 7.2673e+0 (4.13e−1) + 7.7500e+0 (5.29e−1) = 6.1422e+0 (1.62e−1) + 7.8059e+0 (6.73e−1)
3 3.9457e+0 (3.56e−1) − 4.6179e+0 (1.62e−1) − 4.3006e+0 (1.23e−1) − 4.1428e+0 (1.87e−1) − 3.0146e+0 (5.13e−1) − 4.0832e+0 (6.18e−2) − 2.4180e+0 (8.30e−1)
MaF1 7 6.8295e+0 (6.63e−1) − 7.8689e+0 (3.13e−1) − 6.9139e+0 (5.02e−1) − 6.8543e+0 (5.09e−1) − 4.8064e+0 (8.47e−1) + 2.8623e+0 (6.52e−1) + 5.9818e+0 (8.68e−1)
10 7.2313e+0 (8.76e−1) − 8.9513e+0 (3.73e−1) − 7.5818e+0 (5.36e−1) − 7.9148e+0 (3.68e−1) − 6.1502e+0 (1.04e+0) = 2.9044e+0 (7.32e−1) + 5.9738e+0 (1.51e+0)
3 2.7844e−1 (1.42e−2) = 3.2698e−1 (4.30e−3) − 3.2236e−1 (5.69e−3) − 2.8542e−1 (8.45e−3) − 2.2791e−1 (1.85e−2) = 2.9193e−1 (4.16e−3) − 2.5393e−1 (3.98e−2)
MaF2 7 2.6888e−1 (1.70e−2) + 2.9110e−1 (1.88e−2) = 2.9612e−1 (1.86e−2) = 2.7691e−1 (2.50e−2) + 3.0258e−1 (1.86e−2) = 3.0546e−1 (1.48e−2) = 3.0077e−1 (2.74e−2)
10 3.3474e−1 (2.39e−2) = 3.5588e−1 (2.61e−2) − 3.5370e−1 (2.04e−2) − 3.2970e−1 (3.99e−2) = 3.5763e−1 (1.59e−2) − 3.5458e−1 (3.03e−2) − 3.2907e−1 (2.64e−2)
3 1.4876e+14 (7.75e+13) − 1.4181e+14 (3.30e+13) − 2.2977e+14 (3.71e+13) − 2.1334e+14 (3.92e+13) − 2.6552e+14 (6.91e+13) − 2.1594e+14 (3.24e+13) − 2.5744e+13 (4.48e+13)
MaF3 7 1.2738e+14 (2.31e+13) − 9.4626e+13 (1.57e+13) − 1.0761e+14 (2.00e+13) − 1.2322e+14 (1.59e+13) − 1.0281e+14 (2.48e+13) − 9.9833e+13 (1.01e+13) − 1.9408e+13 (1.79e+13)
10 9.6066e+13 (1.46e+13) − 6.6621e+13 (9.49e+12) − 8.6601e+13 (1.13e+13) − 9.6018e+13 (1.17e+13) − 9.1647e+13 (1.84e+13) − 7.9529e+13 (9.54e+12) − 1.0728e+13 (9.60e+12)
3 2.2816e+4 (1.12e+3) − 2.1999e+4 (1.41e+3) − 2.4124e+4 (1.08e+3) − 2.3005e+4 (8.47e+2) − 2.2463e+4 (2.26e+3) − 2.3330e+4 (7.12e+2) − 1.1015e+4 (6.15e+3)
MaF4 7 4.0489e+5 (1.90e+4) − 3.6244e+5 (1.48e+4) − 3.7932e+5 (1.17e+4) − 3.9201e+5 (1.61e+4) − 3.6375e+5 (1.68e+4) − 3.5181e+5 (9.30e+3) − 2.7324e+5 (8.99e+4)
10 3.0036e+6 (1.57e+5) − 2.6667e+6 (1.41e+5) − 2.8915e+6 (9.24e+4) − 2.8900e+6 (1.56e+5) − 2.6894e+6 (1.36e+5) − 2.6406e+6 (6.12e+4) − 1.7266e+6 (7.11e+5)
3 2.3441e+1 (2.66e+0) − 2.5523e+1 (1.87e+0) − 2.3208e+1 (1.92e+0) − 2.3327e+1 (1.82e+0) − 1.7696e+1 (2.08e+0) = 2.4844e+1 (8.98e−1) − 1.9759e+1 (3.36e+0)
MaF5 7 3.6247e+2 (1.86e+1) = 3.0291e+2 (1.99e+1) + 3.2794e+2 (2.58e+1) = 3.4992e+2 (2.22e+1) = 3.2956e+2 (2.14e+1) = 2.7056e+2 (1.81e+1) + 3.4533e+2 (3.35e+1)
10 2.6138e+3 (1.51e+2) = 1.7637e+3 (1.18e+2) + 2.5309e+3 (1.42e+2) = 2.6733e+3 (1.48e+2) − 2.5710e+3 (7.28e+1) = 2.2038e+3 (5.32e+1) + 2.5495e+3 (2.11e+2)
3 3.2450e+2 (2.96e+1) − 3.8594e+2 (1.13e+1) − 3.8702e+2 (1.13e+1) − 3.0904e+2 (2.20e+1) − 2.9741e+2 (5.18e+1) = 3.4012e+2 (6.64e+0) − 2.6574e+2 (9.88e+1)
MaF6 7 3.4513e+2 (1.21e+1) − 3.7965e+2 (9.19e+0) − 3.6100e+2 (8.65e+0) − 3.4141e+2 (9.90e+0) − 3.4506e+2 (3.74e+1) − 2.5229e+2 (5.22e+1) + 2.9392e+2 (3.31e+1)
10 3.2634e+2 (1.68e+1) − 3.5659e+2 (7.17e+0) − 3.1089e+2 (1.05e+1) − 3.2086e+2 (9.22e+0) − 3.0166e+2 (2.40e+1) = 2.8782e+2 (3.03e+1) = 2.8407e+2 (2.92e+1)
3 8.7946e+0 (1.15e+0) = 1.1171e+1 (3.13e−1) − 1.3678e+0 (4.35e−1) + 1.0549e+1 (4.37e−1) = 1.0567e+1 (5.57e−1) = 1.1052e+1 (2.03e−1) − 9.5558e+0 (1.75e+0)
MaF7 7 2.7035e+1 (5.82e−1) − 2.7608e+1 (4.07e−1) − 6.0701e+0 (1.00e+0) + 2.6966e+1 (4.49e−1) − 2.7344e+1 (4.48e−1) − 2.8203e+1 (2.27e−1) − 2.4509e+1 (1.48e+0)
10 3.9120e+1 (4.45e−1) − 3.9533e+1 (4.74e−1) − 2.1704e+1 (2.63e+0) + 3.9080e+1 (3.35e−1) − 3.8883e+1 (3.92e−1) − 4.0610e+1 (1.66e−1) − 3.5327e+1 (1.73e+0)
+/-/= 7/51/11 5/51/13 20/39/10 16/37/16 12/30/27 24/36/9
Avg Rank 3.8697 3.5072 3.2463 2.5371 2.3188 2.5797 2.1159
𝑝 value 4.6926e−9 5.6230e−8 2.2500e−2 1.9530e−1 1.0000e+0 4.4400e−2

ranking of all algorithms according to the Friedman test, and the test EMOPs with 𝐷 > 50; the extra FEs are not enough to update the
results are further adjusted using the Bonferroni correction, with the entire population; on the other hand, it is also extravagant in real-world
corresponding ‘‘𝑝-Value’’ presented in the last row of each table. EMOPs to give 11×𝐷−1 solutions as the initial population when 𝐷 > 50.
𝑅𝑚𝑎𝑥 = 300 is used by MCEA/D [23] and KTA2 [23]; this standard sets
4.1. Parameter the initial population as 100 and gives 200 extra FEs for SA-MOEAs
to update the population, a common idea of this standard is to give
4.1.1. A discussion about the maximum number of real function evaluations two generations to ensure convergence or the EMOPs of the real world
There are a series of standards used by different SA-MOEAs, such as can offer enough FEs more than 200. However, these are subjective
𝑅𝑚𝑎𝑥 = 1000 is used by GPEME [21]; Gu er al. [43] adopt 𝑅𝑚𝑎𝑥 = 11×𝐷 opinions; how many FEs can ensure convergence is determined by the
to tackle the large-scale EMOPs; 𝑅𝑚𝑎𝑥 = 11 × 𝐷 + 119 is a popular set for problems and algorithms, and the FEs offered by real-world EMOPs are
tackling EMOPs, it is used by many researcher, such as KRVEA [11], also various. This article considers that SA-MOEAs must be tested in
HeE-MOEA [22], EDN-AMOEA [24], ParEGO [6], and ESF-RVEA [30] limited population size and extra FEs in a laboratory environment to
et al.; 𝑅𝑚𝑎𝑥 = 11 × 𝐷 + 119 was first used to solve the EMOPs with ensure their robustness. We set the population size as 𝑁 = 91 because
𝐷 = 10, the population size of this standard is set as 11 × 𝐷 − 1 = 109, we employ the MOEA/D as an optimizer; its population size should
the extra FEs is set as 120, this setting ensures that the population be the same as the number of uniformly distributed weights, and the
can be updated for one generation. However, it is not suitable for the maximum number of FEs are set as 𝑅𝑚𝑎𝑥 = 182.

7
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

Table 3
IGD statistical results on all test instances with D = 100.
Problem M ADSAPSO CPS-MOEA K-RVEA KTA2 MCEA/D HeE-MOEA MOEA/D-FEF
3 2.7318e+3 (1.25e+2) − 2.3674e+3 (1.13e+2) − 2.7634e+3 (1.37e+2) − 2.6099e+3 (1.64e+2) − 2.0419e+3 (4.08e+2) − 2.5210e+3 (1.19e+2) − 2.4558e+2 (1.22e+2)
DTLZ1 7 2.0868e+3 (1.28e+2) − 1.8546e+3 (1.12e+2) − 2.0060e+3 (1.26e+2) − 1.9532e+3 (1.73e+2) − 1.7618e+3 (1.94e+2) − 2.0266e+3 (9.27e+1) − 1.6292e+2 (7.59e+1)
10 1.8419e+3 (1.85e+2) − 1.7421e+3 (1.09e+2) − 1.8660e+3 (1.74e+2) − 1.8133e+3 (1.85e+2) − 1.5998e+3 (1.84e+2) − 1.9141e+3 (7.04e+1) − 2.4072e+2 (9.37e+1)
3 6.0319e+0 (5.77e−1) − 5.1725e+0 (1.41e+0) − 6.6645e+0 (2.21e−1) − 6.2021e+0 (3.27e−1) − 1.8821e+0 (4.49e−1) − 8.1207e−1 (4.01e−2) + 1.6815e+0 (5.64e−1)
DTLZ2 7 6.5100e+0 (2.52e−1) − 5.9351e+0 (1.25e+0) − 6.5975e+0 (2.78e−1) − 6.5088e+0 (2.90e−1) − 2.4169e+0 (5.04e−1) − 1.1224e+0 (4.68e−2) + 1.8043e+0 (3.71e−1)
10 6.1667e+0 (3.24e−1) − 5.5735e+0 (1.25e+0) − 6.4703e+0 (2.07e−1) − 6.4003e+0 (3.10e−1) − 2.1951e+0 (3.71e−1) = 1.1940e+0 (3.89e−2) + 2.2938e+0 (5.15e−1)
3 8.0664e+3 (7.10e+2) − 7.4070e+3 (2.88e+2) − 8.5886e+3 (3.15e+2) − 8.2105e+3 (3.28e+2) − 4.7707e+3 (9.05e+2) − 8.2769e+3 (2.51e+2) − 7.8686e+2 (4.13e+2)
DTLZ3 7 8.1219e+3 (3.37e+2) − 7.0605e+3 (2.50e+2) − 8.2750e+3 (3.25e+2) − 8.3358e+3 (3.09e+2) − 5.9045e+3 (7.89e+2) − 7.9332e+3 (2.63e+2) − 6.4808e+2 (3.08e+2)
10 7.7129e+3 (3.52e+2) − 6.6430e+3 (2.57e+2) − 7.9879e+3 (2.27e+2) − 8.0025e+3 (2.98e+2) − 4.7586e+3 (8.58e+2) − 7.6781e+3 (2.93e+2) − 9.3729e+2 (2.62e+2)
3 5.9227e+0 (5.32e−1) − 5.6029e+0 (1.32e+0) − 6.8566e+0 (3.18e−1) − 6.9158e+0 (3.39e−1) − 1.8502e+0 (2.91e−1) = 1.2453e+0 (5.91e−2) + 2.0547e+0 (5.70e−1)
DTLZ4 7 6.3540e+0 (4.45e−1) − 5.3239e+0 (9.75e−1) − 6.8274e+0 (1.92e−1) − 6.7487e+0 (2.28e−1) − 2.1380e+0 (2.22e−1) = 1.4608e+0 (5.03e−2) + 2.4325e+0 (5.30e−1)
10 6.1590e+0 (4.72e−1) − 5.5554e+0 (1.29e+0) − 6.5428e+0 (2.76e−1) − 6.6899e+0 (2.45e−1) − 1.6647e+0 (9.44e−2) + 1.4782e+0 (4.56e−2) + 2.6150e+0 (6.22e−1)
3 5.8179e+0 (4.26e−1) − 4.9029e+0 (1.28e+0) − 6.5389e+0 (3.15e−1) − 6.2359e+0 (2.48e−1) − 1.6595e+0 (5.74e−1) − 6.6950e−1 (4.12e−2) + 1.4828e+0 (3.30e−1)
DTLZ5 7 6.2126e+0 (2.40e−1) − 4.9032e+0 (1.26e+0) − 6.2397e+0 (2.08e−1) − 6.0875e+0 (3.62e−1) − 2.0889e+0 (5.95e−1) − 6.2208e−1 (3.60e−2) + 1.5030e+0 (3.35e−1)
10 5.8805e+0 (2.91e−1) − 5.0808e+0 (1.24e+0) − 5.9418e+0 (1.98e−1) − 5.9158e+0 (2.68e−1) − 1.5987e+0 (4.02e−1) + 5.7147e−1 (4.18e−2) + 2.0412e+0 (4.70e−1)
3 7.7954e+1 (6.30e+0) − 7.4364e+1 (2.84e+0) − 8.4170e+1 (1.00e+0) − 8.6475e+1 (5.22e−1) − 5.9768e+1 (5.43e+0) − 8.7085e+1 (5.83e−1) − 4.7029e+1 (7.16e+0)
DTLZ6 7 7.8933e+1 (4.47e+0) − 7.2704e+1 (2.76e+0) − 8.1416e+1 (7.41e−1) − 8.3049e+1 (8.29e−1) − 6.3841e+1 (6.38e+0) − 8.3631e+1 (2.59e−1) − 4.6232e+1 (5.42e+0)
10 7.4964e+1 (2.86e+0) − 6.8491e+1 (3.25e+0) − 7.9141e+1 (5.95e−1) − 8.0631e+1 (3.63e−1) − 5.8633e+1 (5.99e+0) − 8.0868e+1 (2.40e−1) − 4.4787e+1 (5.24e+0)
3 9.2356e+0 (9.03e−1) = 1.0481e+1 (4.59e−1) − 6.3933e+0 (9.38e−1) + 1.0180e+1 (6.79e−1) − 1.0483e+1 (5.98e−1) − 1.0724e+1 (4.50e−1) − 8.6670e+0 (2.20e+0)
DTLZ7 7 2.5109e+1 (1.18e+0) − 2.5590e+1 (1.05e+0) − 1.5111e+1 (2.07e+0) + 2.5306e+1 (1.27e+0) − 2.5530e+1 (1.38e+0) − 2.6549e+1 (6.06e−1) − 2.0028e+1 (3.20e+0)
10 3.6733e+1 (1.52e+0) − 3.6892e+1 (1.68e+0) − 1.8867e+1 (2.31e+0) + 3.7518e+1 (1.33e+0) − 3.6468e+1 (1.60e+0) − 3.7295e+1 (1.67e+0) − 3.0628e+1 (3.60e+0)
3 2.3960e+0 (3.30e−2) − 2.3281e+0 (4.73e−2) = 2.0782e+0 (9.47e−2) + 2.2269e+0 (1.78e−1) = 2.3232e+0 (4.62e−2) = 2.4038e+0 (4.63e−2) − 2.2430e+0 (9.15e−2)
WFG1 7 2.9789e+0 (3.39e−2) = 3.0115e+0 (3.57e−2) = 2.8906e+0 (5.05e−2) = 2.9484e+0 (5.17e−2) = 3.0053e+0 (3.64e−2) = 3.0161e+0 (4.40e−2) = 2.9968e+0 (7.40e−2)
10 3.4671e+0 (3.49e−2) = 3.4615e+0 (3.33e−2) = 3.4056e+0 (3.93e−2) = 3.4502e+0 (2.82e−2) = 3.4614e+0 (2.96e−2) = 3.4735e+0 (3.53e−2) = 3.4625e+0 (3.26e−2)
3 1.0618e+0 (6.59e−2) − 9.3436e−1 (5.83e−2) − 9.8123e−1 (7.46e−2) − 9.5677e−1 (7.01e−2) − 8.3225e−1 (5.32e−2) = 9.4887e−1 (6.37e−2) − 8.5168e−1 (6.92e−2)
WFG2 7 2.8722e+0 (4.17e−1) = 2.2984e+0 (3.78e−1) = 2.7133e+0 (3.76e−1) − 2.4774e+0 (6.54e−1) − 2.5726e+0 (4.36e−1) − 3.0267e+0 (4.43e−1) − 2.1641e+0 (4.67e−1)
10 4.6232e+0 (9.13e−1) − 3.8201e+0 (6.66e−1) = 4.1329e+0 (9.78e−1) − 3.7644e+0 (7.06e−1) + 4.1774e+0 (9.03e−1) − 5.2295e+0 (6.98e−1) − 4.0277e+0 (6.92e−1)
3 8.5071e−1 (1.15e−2) − 8.5293e−1 (1.79e−2) − 8.4982e−1 (1.13e−2) − 8.5523e−1 (9.06e−3) − 6.7031e−1 (4.39e−2) + 5.7563e−1 (2.64e−2) + 7.2012e−1 (5.93e−2)
WFG3 7 1.4765e+0 (2.81e−2) = 1.5193e+0 (3.90e−2) − 1.4657e+0 (3.28e−2) = 1.4695e+0 (2.34e−2) = 1.2212e+0 (5.42e−2) + 1.0135e+0 (5.70e−2) + 1.4355e+0 (7.27e−2)
10 1.8969e+0 (4.01e−2) = 1.9235e+0 (6.83e−2) = 1.8843e+0 (3.24e−2) = 1.8898e+0 (3.54e−2) = 1.4752e+0 (7.96e−2) + 1.2873e+0 (7.51e−2) + 1.8718e+0 (8.93e−2)
3 8.0149e−1 (6.23e−2) − 6.5041e−1 (2.55e−2) = 6.2451e−1 (1.76e−2) + 6.1516e−1 (2.11e−2) + 6.3152e−1 (5.18e−2) + 8.0298e−1 (5.64e−2) − 6.7434e−1 (3.15e−2)
WFG4 7 5.5229e+0 (2.19e−1) − 4.3611e+0 (3.82e−1) − 4.7395e+0 (4.67e−1) − 4.7932e+0 (5.41e−1) − 4.1555e+0 (5.69e−1) = 5.6220e+0 (2.91e−1) − 3.9846e+0 (4.63e−1)
10 1.0134e+1 (5.45e−1) − 8.4433e+0 (7.49e−1) − 9.2188e+0 (5.41e−1) − 9.3236e+0 (7.34e−1) − 7.9853e+0 (7.23e−1) − 1.0438e+1 (5.18e−1) − 7.0572e+0 (5.15e−1)
3 7.8650e−1 (1.77e−2) − 7.8650e−1 (1.77e−2) − 7.9758e−1 (2.15e−2) − 8.2392e−1 (1.04e−2) − 7.3900e−1 (2.52e−2) = 8.6016e−1 (1.19e−2) − 7.2744e−1 (5.32e−2)
WFG5 7 4.1674e+0 (2.06e−1) − 3.7998e+0 (1.67e−1) − 3.5711e+0 (1.93e−1) = 3.7110e+0 (2.19e−1) = 3.7439e+0 (1.78e−1) − 3.5022e+0 (8.66e−2) = 3.6935e+0 (2.59e−1)
10 8.0503e+0 (3.99e−1) − 8.4433e+0 (7.49e−1) − 9.2188e+0 (5.41e−1) − 9.3236e+0 (7.34e−1) − 7.9853e+0 (7.23e−1) − 1.0438e+1 (5.18e−1) − 6.6962e+0 (5.86e−1)
3 1.0249e+0 (2.56e−2) = 1.0307e+0 (1.60e−2) = 9.8456e−1 (1.06e−2) + 1.0118e+0 (1.43e−2) = 9.4506e−1 (3.00e−2) + 9.1761e−1 (2.93e−2) + 1.0516e+0 (1.10e−1)
WFG6 7 4.4417e+0 (1.81e−1) − 4.1224e+0 (2.60e−1) − 3.7648e+0 (2.68e−1) + 4.1416e+0 (2.28e−1) − 4.2327e+0 (2.22e−1) − 4.4964e+0 (2.08e−1) − 4.0241e+0 (3.06e−1)
10 8.5306e+0 (4.27e−1) − 7.8211e+0 (4.03e−1) − 7.6865e+0 (4.56e−1) = 8.1268e+0 (5.85e−1) − 8.2038e+0 (5.44e−1) − 8.4130e+0 (3.71e−1) − 7.4514e+0 (4.89e−1)
3 8.0814e−1 (2.53e−2) − 7.8806e−1 (1.72e−2) = 7.8207e−1 (1.93e−2) = 7.9276e−1 (2.23e−2) = 7.2051e−1 (2.82e−2) + 7.3225e−1 (3.68e−2) + 7.7125e−1 (3.29e−2)
WFG7 7 4.5431e+0 (2.33e−1) − 4.0539e+0 (3.68e−1) = 3.9254e+0 (2.40e−1) = 4.2007e+0 (2.57e−1) − 4.2503e+0 (3.36e−1) − 4.5695e+0 (2.72e−1) − 4.0368e+0 (5.16e−1)
10 8.5994e+0 (5.15e−1) − 8.0986e+0 (3.89e−1) − 8.3144e+0 (4.01e−1) − 8.2125e+0 (5.51e−1) − 8.4786e+0 (6.51e−1) − 8.7585e+0 (5.48e−1) − 7.5547e+0 (7.11e−1)
3 8.9209e−1 (2.55e−2) = 8.7164e−1 (1.50e−2) = 8.4186e−1 (1.81e−2) + 8.7881e−1 (1.74e−2) = 8.3666e−1 (4.07e−2) + 8.3774e−1 (2.23e−2) + 8.8171e−1 (4.10e−2)
WFG8 7 4.2966e+0 (1.73e−1) − 4.0483e+0 (2.69e−1) = 3.7667e+0 (1.84e−1) = 4.1179e+0 (1.92e−1) = 4.0478e+0 (2.92e−1) = 4.4643e+0 (1.70e−1) − 3.8612e+0 (2.97e−1)
10 8.4579e+0 (4.06e−1) − 7.6768e+0 (4.95e−1) − 7.6086e+0 (3.86e−1) − 7.9858e+0 (5.24e−1) − 8.1712e+0 (5.86e−1) − 8.3292e+0 (3.46e−1) − 7.2652e+0 (7.36e−1)
3 1.0907e+0 (3.87e−2) − 1.0951e+0 (3.22e−2) = 1.0889e+0 (3.12e−2) = 1.0850e+0 (2.97e−2) = 9.3950e−1 (4.99e−2) + 8.0119e−1 (5.22e−2) + 9.9575e−1 (9.03e−2)
WFG9 7 4.5181e+0 (2.57e−1) = 4.4841e+0 (2.05e−1) − 4.2277e+0 (3.48e−1) = 4.1210e+0 (2.08e−1) = 4.1028e+0 (2.30e−1) = 3.2554e+0 (6.38e−2) + 4.1504e+0 (3.03e−1)
10 8.5276e+0 (4.75e−1) − 8.2144e+0 (3.56e−1) − 8.2162e+0 (4.23e−1) − 7.8100e+0 (4.58e−1) = 7.5815e+0 (6.02e−1) = 5.7943e+0 (1.36e−1) + 7.7015e+0 (6.26e−1)
3 7.1787e+0 (5.42e−1) − 6.6970e+0 (1.65e+0) − 7.6768e+0 (3.09e−1) − 7.3485e+0 (4.17e−1) − 1.5588e+0 (4.15e−1) + 1.0166e+0 (1.28e−1) + 2.1789e+0 (7.85e−1)
MaF1 7 1.2555e+1 (1.37e+0) − 1.0108e+1 (3.05e+0) − 1.4252e+1 (4.78e−1) − 1.3662e+1 (8.22e−1) − 2.7074e+0 (5.26e−1) + 1.1919e+0 (1.85e−1) + 5.6299e+0 (1.65e+0)
10 1.4434e+1 (1.86e+0) − 1.1772e+1 (2.90e+0) − 1.6753e+1 (9.40e−1) − 1.6358e+1 (1.05e+0) − 1.6622e+0 (2.40e−1) + 1.3186e+0 (1.47e−1) + 8.3495e+0 (2.64e+0)
3 5.1554e−1 (2.84e−2) − 4.7742e−1 (5.89e−2) − 5.6532e−1 (8.85e−3) − 5.4723e−1 (2.04e−2) − 2.4227e−1 (1.47e−2) + 1.6302e−1 (7.88e−3) + 2.8137e−1 (3.20e−2)
MaF2 7 3.7535e−1 (1.73e−2) = 4.0009e−1 (1.61e−2) − 3.9478e−1 (1.57e−2) = 3.8551e−1 (1.99e−2) = 3.7552e−1 (1.52e−2) = 3.7896e−1 (1.53e−2) = 3.6544e−1 (2.50e−2)
10 3.8724e−1 (3.55e−2) = 4.1347e−1 (2.41e−2) = 4.0310e−1 (2.25e−2) = 4.0338e−1 (2.42e−2) = 4.0966e−1 (2.81e−2) = 4.1344e−1 (2.02e−2) = 3.8323e−1 (3.09e−2)
3 9.6432e+7 (2.22e+7) − 6.3213e+7 (6.70e+6) − 1.2498e+8 (5.05e+7) − 2.4629e+8 (2.33e+8) − 2.6646e+7 (1.03e+7) − 2.3729e+8 (1.67e+8) − 1.1132e+7 (2.75e+7)
MaF3 7 1.5632e+8 (8.45e+7) − 6.4045e+7 (1.09e+7) − 9.7884e+7 (1.63e+7) − 1.7303e+8 (1.31e+8) − 3.6593e+7 (1.29e+7) − 1.3680e+8 (5.23e+7) − 2.2471e+5 (1.72e+5)
10 1.8659e+8 (1.36e+8) − 5.6407e+7 (8.27e+6) − 9.5893e+7 (1.47e+7) − 1.1624e+8 (4.15e+7) − 2.6845e+7 (1.00e+7) − 1.1468e+8 (4.85e+7) − 5.0814e+6 (1.52e+7)
3 2.6699e+4 (1.16e+3) − 2.3424e+4 (1.75e+3) − 2.7229e+4 (1.11e+3) − 2.7994e+4 (8.54e+2) − 2.1306e+4 (2.79e+3) − 2.5566e+4 (1.47e+3) − 3.3046e+3 (1.59e+3)
MaF4 7 4.5661e+5 (2.25e+4) − 3.9162e+5 (2.53e+4) − 4.4890e+5 (1.97e+4) − 4.6962e+5 (1.82e+4) − 3.6238e+5 (4.64e+4) − 4.2915e+5 (1.52e+4) − 5.9455e+4 (3.56e+4)
10 3.5449e+6 (1.78e+5) − 3.0332e+6 (1.55e+5) − 3.5560e+6 (1.11e+5) − 3.6289e+6 (1.26e+5) − 2.7566e+6 (2.47e+5) − 3.2634e+6 (1.18e+5) − 4.3185e+5 (2.79e+5)
3 2.1686e+1 (7.69e+0) − 1.8558e+1 (1.36e+0) − 2.0665e+1 (5.21e+0) − 2.1785e+1 (7.46e+0) − 1.3671e+1 (2.73e+0) = 9.5871e+0 (4.55e−1) + 1.5382e+1 (3.32e+0)
MaF5 7 6.9118e+1 (4.62e+1) = 5.1129e+1 (1.58e+0) = 5.4422e+1 (2.61e+1) = 1.1893e+2 (1.06e+2) = 5.0775e+1 (4.60e+0) = 1.3601e+2 (3.10e+1) − 5.6340e+1 (1.59e+1)
10 2.8025e+2 (8.80e+1) = 2.5270e+2 (3.92e+1) + 2.1830e+2 (3.27e+1) + 3.5487e+2 (1.17e+2) = 2.7719e+2 (2.83e+1) = 9.2740e+2 (3.39e+2) − 4.1683e+2 (5.06e+2)
3 5.7254e+2 (5.02e+1) − 4.6234e+2 (1.31e+2) − 6.2065e+2 (3.79e+1) − 5.9675e+2 (3.35e+1) − 1.3176e+2 (3.56e+1) = 7.5756e+1 (1.69e+1) + 1.4000e+2 (4.39e+1)
MaF6 7 5.9412e+2 (3.56e+1) − 4.4205e+2 (1.18e+2) − 6.0206e+2 (2.09e+1) − 5.9158e+2 (3.87e+1) − 1.9619e+2 (5.00e+1) − 4.3282e+1 (5.20e+0) + 1.4454e+2 (3.66e+1)
10 5.6781e+2 (3.02e+1) − 4.3495e+2 (1.22e+2) − 5.8064e+2 (1.91e+1) − 5.7481e+2 (3.40e+1) − 1.3183e+2 (4.22e+1) + 3.7478e+1 (5.90e+0) + 2.0753e+2 (8.36e+1)
3 9.3118e+0 (8.23e−1) = 1.0599e+1 (4.25e−1) − 6.8245e+0 (8.53e−1) + 1.0111e+1 (8.32e−1) = 1.0338e+1 (6.20e−1) = 1.0738e+1 (4.38e−1) − 9.4898e+0 (1.22e+0)
MaF7 7 2.5222e+1 (1.10e+0) − 2.5746e+1 (8.75e−1) − 1.4362e+1 (2.07e+0) + 2.5869e+1 (8.18e−1) − 2.5596e+1 (9.01e−1) − 2.6233e+1 (7.77e−1) − 2.1646e+1 (2.50e+0)
10 3.7071e+1 (1.38e+0) − 3.7436e+1 (8.30e−1) − 1.9097e+1 (2.53e+0) + 3.8039e+1 (9.31e−1) − 3.6735e+1 (2.30e+0) − 3.7218e+1 (1.33e+0) − 3.2046e+1 (3.63e+0)
+/-/= 0/55/14 1/54/15 12/43/14 2/49/18 15/35/19 26/38/5
Avg Rank 3.8261 3.3913 2.9711 2.8985 2.4782 3.1014 1.7681
𝑝 value 1.1646e−11 1.4000e−2 2.1579e−5 8.5083e−11 1.0000e+0 2.4000e−2

4.1.2. Parameter settings • FEs for per part: 𝐿 = 20

We adopt the recommended settings from the original literature for • Replacement times: 𝑛𝑟 = 2
a fair comparison. The parameters for MOEA/D-FEF are as follows: • Neighbor size: 𝐵 = 10
• Max iterations for surrogate optimization: 𝑅′max = 3
1. Initial Population Size: • Max center points in the 𝑖th decision variable: 𝑆 = 4
• Features for dimensionality reduction and feature fusion:
• For 𝑀 = 3 and 𝑀 = 7: 𝑁 = 91
• For 𝑀 = 10: 𝑁 = 110 𝐾 = 14
• Learning rate: In this study, the learning rate is set to
2. Maximum Number of Real FEs:
the value that enables the fastest convergence of the loss
• 𝑅max = 182 for all test suites. function. We set the 𝐿𝑅 = 0.000001 for the WFG test suites
3. MOEA/D-FEF Parameters: and the 𝐿𝑅 = 0.00007 for the DTLZ/MaF test suites.

8
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

Table 4
𝛥𝑃 statistical results on all test instances with D = 100.
Problem M ADSAPSO CPS-MOEA K-RVEA KTA2 MCEA/D HeE-MOEA MOEA/D-FEF
3 3.6779e+3 (1.68e+2) − 3.3098e+3 (5.67e+1) − 3.7823e+3 (5.54e+1) − 3.7037e+3 (4.97e+1) − 3.1034e+3 (1.75e+2) − 3.5799e+3 (4.75e+1) − 1.6383e+3 (1.20e+3)
DTLZ1 7 3.2317e+3 (5.48e+1) − 2.9248e+3 (6.87e+1) − 3.3857e+3 (1.04e+2) − 3.1841e+3 (8.45e+1) − 2.8370e+3 (1.11e+2) − 3.1647e+3 (3.14e+1) − 1.8316e+3 (1.06e+3)
10 3.0871e+3 (8.08e+1) = 2.8040e+3 (4.93e+1) + 3.2670e+3 (8.06e+1) − 3.1397e+3 (9.30e+1) = 2.6226e+3 (1.10e+2) + 3.1036e+3 (3.07e+1) = 2.8516e+3 (6.72e+2)
3 7.2552e+0 (4.32e−1) − 8.2664e+0 (1.59e−1) − 8.0466e+0 (8.59e−2) − 7.4683e+0 (1.19e−1) − 6.6336e+0 (5.56e−1) − 4.1215e+0 (5.48e−1) + 5.7906e+0 (9.89e−1)
DTLZ2 7 7.8008e+0 (1.47e−1) − 8.5515e+0 (1.78e−1) − 7.9158e+0 (6.23e−2) − 7.6574e+0 (1.11e−1) − 7.6195e+0 (4.62e−1) − 4.9111e+0 (3.91e−1) + 5.7149e+0 (9.46e−1)
10 7.5187e+0 (1.90e−1) − 8.3643e+0 (9.90e−2) − 7.6303e+0 (9.62e−2) − 7.4915e+0 (1.09e−1) − 6.9993e+0 (3.54e−1) − 5.0554e+0 (2.91e−1) + 6.0444e+0 (7.04e−1)
3 9.6901e+3 (3.03e+2) − 8.7864e+3 (1.41e+2) − 1.0078e+4 (1.36e+2) − 9.9652e+3 (1.02e+2) − 7.8309e+3 (6.63e+2) − 9.8515e+3 (9.92e+1) − 3.8784e+3 (2.30e+3)
DTLZ3 7 9.9678e+3 (1.18e+2) − 9.0213e+3 (1.29e+2) − 9.9326e+3 (7.62e+1) − 9.9282e+3 (9.09e+1) − 8.7198e+3 (3.26e+2) − 9.7497e+3 (4.28e+1) − 5.6030e+3 (2.50e+3)
10 9.6446e+3 (1.35e+2) − 8.7480e+3 (1.09e+2) − 9.6186e+3 (9.25e+1) − 9.6821e+3 (8.29e+1) − 7.9655e+3 (4.40e+2) − 9.4841e+3 (3.32e+1) − 5.6789e+3 (2.66e+3)
3 7.2549e+0 (3.28e−1) − 8.1880e+0 (3.64e−1) − 7.9595e+0 (2.35e−1) − 7.6034e+0 (2.36e−1) − 5.4710e+0 (6.26e−1) = 4.1488e+0 (5.92e−1) + 5.2112e+0 (1.43e+0)
DTLZ4 7 7.7299e+0 (1.69e−1) − 8.7735e+0 (1.37e−1) − 7.8776e+0 (9.67e−2) − 7.6859e+0 (1.20e−1) − 6.6322e+0 (2.69e−1) = 4.6487e+0 (2.41e−1) + 6.4435e+0 (6.15e−1)
10 7.5948e+0 (1.38e−1) − 8.6351e+0 (1.42e−1) − 7.5923e+0 (9.45e−2) − 7.5199e+0 (1.06e−1) − 5.2526e+0 (3.15e−1) + 4.9396e+0 (1.33e−1) + 7.1220e+0 (7.03e−1)
3 7.3486e+0 (3.48e−1) − 8.3177e+0 (1.77e−1) − 8.1287e+0 (1.17e−1) − 7.5253e+0 (1.31e−1) − 6.2035e+0 (7.74e−1) = 3.9659e+0 (4.63e−1) + 6.1387e+0 (1.49e+0)
DTLZ5 7 7.8785e+0 (1.30e−1) − 8.5668e+0 (1.44e−1) − 8.0458e+0 (8.81e−2) − 7.8017e+0 (7.37e−2) − 7.6536e+0 (6.75e−1) − 5.5088e+0 (3.38e−1) + 6.3626e+0 (9.37e−1)
10 7.7082e+0 (2.07e−1) − 8.2911e+0 (1.27e−1) − 7.7077e+0 (8.61e−2) − 7.6028e+0 (1.54e−1) − 6.7162e+0 (6.19e−1) = 5.8012e+0 (3.96e−1) + 6.5565e+0 (5.31e−1)
3 8.7648e+1 (1.52e+0) − 8.3185e+1 (5.60e−1) − 8.8256e+1 (2.68e−1) − 8.8929e+1 (7.78e−2) − 8.0908e+1 (3.09e+0) = 8.9981e+1 (4.36e−2) − 7.7818e+1 (7.10e+0)
DTLZ6 7 8.5175e+1 (4.91e−1) − 8.2221e+1 (3.26e−1) − 8.5022e+1 (8.43e−2) − 8.5496e+1 (8.92e−2) − 8.1618e+1 (9.13e−1) − 8.5514e+1 (1.56e−2) − 6.8057e+1 (4.77e+0)
10 8.2521e+1 (3.23e−1) − 7.9573e+1 (2.75e−1) − 8.2398e+1 (1.32e−1) − 8.2821e+1 (8.18e−2) − 7.7683e+1 (1.45e+0) − 8.3576e+1 (2.17e−2) − 6.7353e+1 (3.48e+0)
3 9.6434e+0 (1.16e+0) + 1.1551e+1 (1.60e−1) − 6.9840e+0 (8.97e−1) + 1.1263e+1 (4.71e−1) − 1.1210e+1 (3.96e−1) = 1.1581e+1 (1.97e−1) − 1.0756e+1 (6.64e−1)
DTLZ7 7 2.7462e+1 (3.00e−1) − 2.7775e+1 (3.60e−1) − 1.7284e+1 (1.88e+0) + 2.7457e+1 (3.30e−1) − 2.7622e+1 (2.08e−1) − 2.8344e+1 (2.06e−1) − 2.5622e+1 (1.68e+0)
10 3.9215e+1 (3.54e−1) − 3.9709e+1 (3.81e−1) − 2.9883e+1 (1.93e+0) + 3.9505e+1 (2.67e−1) − 3.9454e+1 (3.43e−1) − 4.0733e+1 (4.07e−1) − 3.6907e+1 (9.92e−1)
3 2.3960e+0 (3.30e−2) − 2.3186e+0 (4.61e−2) = 2.0867e+0 (9.16e−2) + 2.2369e+0 (1.81e−1) = 2.3230e+0 (4.96e−2) = 2.4028e+0 (5.89e−2) − 2.3455e+0 (4.89e−2)
WFG1 7 2.9789e+0 (3.39e−2) = 3.0166e+0 (3.75e−2) = 2.8915e+0 (4.92e−2) + 2.9470e+0 (5.72e−2) + 3.0083e+0 (3.55e−2) = 3.0093e+0 (4.26e−2) = 3.0156e+0 (7.04e−2)
10 3.4671e+0 (3.49e−2) = 3.4601e+0 (3.43e−2) = 3.4125e+0 (3.59e−2) + 3.4503e+0 (2.63e−2) = 3.4619e+0 (2.29e−2) = 3.4764e+0 (3.63e−2) = 3.4737e+0 (4.20e−2)
3 1.0618e+0 (6.59e−2) − 9.2662e−1 (6.08e−2) = 9.7316e−1 (4.84e−2) − 9.5788e−1 (7.80e−2) = 8.3436e−1 (5.90e−2) + 9.6725e−1 (6.27e−2) − 9.1752e−1 (1.18e−1)
WFG2 7 2.8722e+0 (4.17e−1) = 2.2851e+0 (3.88e−1) = 2.7425e+0 (3.37e−1) = 2.4013e+0 (5.83e−1) = 2.5773e+0 (4.50e−1) = 3.0066e+0 (4.85e−1) − 2.5521e+0 (5.16e−1)
10 4.6232e+0 (9.13e−1) − 3.8197e+0 (6.42e−1) = 4.1115e+0 (9.90e−1) = 3.8180e+0 (7.28e−1) = 3.9652e+0 (7.59e−1) = 5.1532e+0 (6.71e−1) − 3.7836e+0 (1.03e+0)
3 1.0078e+0 (1.80e−2) = 1.0619e+0 (2.28e−2) = 1.0312e+0 (1.62e−2) = 1.0056e+0 (1.49e−2) = 9.5963e−1 (5.61e−2) + 6.7738e−1 (2.83e−2) + 1.0437e+0 (5.36e−2)
WFG3 7 2.1479e+0 (1.02e−1) + 2.2831e+0 (1.20e−1) + 2.0680e+0 (5.31e−2) + 2.0708e+0 (9.30e−2) + 1.9658e+0 (1.12e−1) + 1.3903e+0 (2.02e−2) + 2.5188e+0 (2.65e−1)
10 2.9074e+0 (1.64e−1) + 3.2291e+0 (1.37e−1) + 2.7096e+0 (1.04e−1) + 2.8234e+0 (1.25e−1) + 2.6989e+0 (1.98e−1) + 1.9126e+0 (3.52e−2) + 3.6817e+0 (2.79e−1)
3 8.0149e−1 (6.23e−2) − 6.4955e−1 (2.98e−2) + 6.6412e−1 (4.92e−2) + 7.1405e−1 (6.03e−2) = 6.3331e−1 (2.54e−2) + 7.9853e−1 (4.56e−2) − 7.0221e−1 (3.29e−2)
WFG4 7 5.5229e+0 (2.19e−1) − 4.3725e+0 (4.15e−1) − 4.7651e+0 (4.54e−1) − 4.8565e+0 (5.54e−1) − 4.1317e+0 (5.81e−1) = 5.5766e+0 (3.08e−1) − 3.8974e+0 (3.83e−1)
10 1.0134e+1 (5.45e−1) − 8.4869e+0 (8.05e−1) − 9.1935e+0 (5.88e−1) − 9.3895e+0 (7.44e−1) − 7.9784e+0 (7.46e−1) − 1.0496e+1 (4.92e−1) − 7.0002e+0 (3.51e−1)
3 8.3835e−1 (2.39e−2) − 7.8924e−1 (1.78e−2) = 7.9789e−1 (2.32e−2) = 8.2752e−1 (1.32e−2) − 7.4399e−1 (2.01e−2) + 8.6454e−1 (1.28e−2) − 7.8076e−1 (5.65e−2)
WFG5 7 4.1674e+0 (2.06e−1) − 3.8201e+0 (1.70e−1) − 3.5623e+0 (1.94e−1) = 3.7462e+0 (2.25e−1) = 3.7410e+0 (1.86e−1) = 3.5838e+0 (9.22e−2) = 3.5133e+0 (2.78e−1)
10 8.0503e+0 (3.99e−1) − 7.5114e+0 (2.76e−1) − 7.0862e+0 (3.67e−1) − 7.5091e+0 (4.07e−1) − 7.1654e+0 (3.84e−1) − 6.5965e+0 (1.69e−1) = 6.5930e+0 (6.00e−1)
3 1.0277e+0 (2.10e−2) = 1.0356e+0 (1.14e−2) = 9.8863e−1 (6.31e−3) + 1.0104e+0 (2.12e−2) + 9.5757e−1 (2.60e−2) + 8.8929e−1 (2.62e−2) + 1.0999e+0 (1.05e−1)
WFG6 7 4.4417e+0 (1.81e−1) − 4.1660e+0 (2.67e−1) − 3.9491e+0 (2.34e−1) = 4.1605e+0 (2.04e−1) − 4.2113e+0 (1.91e−1) − 4.5243e+0 (2.04e−1) − 3.8472e+0 (2.82e−1)
10 8.5306e+0 (4.27e−1) − 7.8375e+0 (4.12e−1) − 7.7404e+0 (4.55e−1) − 8.0870e+0 (6.31e−1) − 8.2189e+0 (5.67e−1) − 8.4033e+0 (3.84e−1) − 7.0163e+0 (5.15e−1)
3 8.0814e−1 (2.53e−2) − 7.9751e−1 (1.77e−2) − 7.8067e−1 (1.06e−2) = 7.8651e−1 (2.16e−2) = 7.1935e−1 (3.01e−2) + 7.1893e−1 (3.50e−2) + 7.8207e−1 (2.27e−2)
WFG7 7 4.5431e+0 (2.33e−1) − 3.9745e+0 (3.32e−1) = 3.9378e+0 (2.13e−1) = 4.2163e+0 (2.59e−1) − 4.2995e+0 (3.37e−1) − 4.5704e+0 (2.86e−1) − 3.9116e+0 (3.84e−1)
10 8.5994e+0 (5.15e−1) − 8.0946e+0 (4.01e−1) − 8.2261e+0 (3.73e−1) − 8.2595e+0 (5.49e−1) − 8.4492e+0 (6.73e−1) − 8.7054e+0 (5.94e−1) − 7.5760e+0 (6.38e−1)
3 8.9213e−1 (2.54e−2) = 8.8862e−1 (2.34e−2) = 8.4465e−1 (1.58e−2) + 8.6350e−1 (2.34e−2) + 8.3747e−1 (3.60e−2) + 8.1410e−1 (3.03e−2) + 8.8108e−1 (3.10e−2)
WFG8 7 4.2966e+0 (1.73e−1) − 4.0335e+0 (2.91e−1) − 3.7354e+0 (1.70e−1) = 4.1142e+0 (1.94e−1) − 4.1177e+0 (2.64e−1) − 4.4618e+0 (1.79e−1) − 3.7163e+0 (3.13e−1)
10 8.4579e+0 (4.06e−1) − 7.6173e+0 (4.79e−1) − 7.6903e+0 (3.78e−1) − 8.0125e+0 (5.42e−1) − 8.2605e+0 (3.95e−1) − 8.3803e+0 (3.55e−1) − 7.1119e+0 (4.91e−1)
3 1.0914e+0 (3.78e−2) − 1.1344e+0 (2.87e−2) − 1.1078e+0 (2.02e−2) − 1.0748e+0 (2.99e−2) = 9.8950e−1 (6.93e−2) = 7.8261e−1 (3.71e−2) + 1.0289e+0 (7.08e−2)
WFG9 7 4.5181e+0 (2.57e−1) = 4.4756e+0 (2.25e−1) = 4.2286e+0 (3.71e−1) = 4.1402e+0 (1.93e−1) = 4.1051e+0 (2.46e−1) = 3.2593e+0 (7.09e−2) + 4.3291e+0 (3.86e−1)
10 8.5276e+0 (4.75e−1) − 8.2682e+0 (3.66e−1) = 8.1775e+0 (4.49e−1) = 7.7866e+0 (4.76e−1) = 7.6807e+0 (6.30e−1) = 5.7804e+0 (1.46e−1) + 7.8108e+0 (8.06e−1)
3 9.0098e+0 (3.03e−1) − 9.8889e+0 (2.26e−1) − 9.6647e+0 (1.71e−1) − 9.2944e+0 (1.07e−1) − 6.8625e+0 (1.26e+0) = 4.1152e+0 (8.90e−1) + 6.9692e+0 (1.44e+0)
MaF1 7 1.6150e+1 (7.02e−1) − 1.7350e+1 (8.40e−1) − 1.7337e+1 (2.15e−1) − 1.6457e+1 (3.22e−1) − 1.1178e+1 (2.20e+0) = 7.2108e+0 (1.86e+0) + 1.2308e+1 (3.09e+0)
10 1.9100e+1 (1.75e+0) − 2.0898e+1 (6.09e−1) − 2.0600e+1 (3.95e−1) − 2.0356e+1 (5.02e−1) − 1.1495e+1 (3.61e+0) + 8.6829e+0 (2.66e+0) + 1.5585e+1 (3.37e+0)
3 6.0011e−1 (2.15e−2) − 6.7522e−1 (1.69e−2) − 6.6987e−1 (7.97e−3) − 6.2876e−1 (1.25e−2) − 4.0921e−1 (4.74e−2) = 1.8750e−1 (2.08e−2) + 4.1982e−1 (7.11e−2)
MaF2 7 3.7535e−1 (1.73e−2) = 4.0009e−1 (1.61e−2) − 3.9478e−1 (1.57e−2) = 3.8551e−1 (1.99e−2) = 3.7552e−1 (1.52e−2) = 3.7896e−1 (1.53e−2) = 3.6544e−1 (2.50e−2)
10 3.8724e−1 (3.55e−2) = 4.1347e−1 (2.41e−2) = 4.0310e−1 (2.25e−2) = 4.0338e−1 (2.42e−2) = 4.0966e−1 (2.81e−2) = 4.1344e−1 (2.02e−2) = 3.9323e−1 (3.09e−2)
3 3.4077e+15 (8.06e+14) − 2.3160e+15 (4.00e+14) − 3.6920e+15 (3.73e+14) − 4.0970e+15 (4.53e+14) − 4.6984e+15 (9.90e+14) − 3.4238e+15 (3.89e+14) − 1.3865e+15 (1.63e+15)
MaF3 7 2.6048e+15 (2.87e+14) − 2.0018e+15 (2.70e+14) − 2.2691e+15 (2.35e+14) − 2.5828e+15 (3.38e+14) − 2.2974e+15 (2.55e+14) − 1.9279e+15 (1.10e+14) − 3.8848e+14 (4.25e+14)
10 2.2575e+15 (2.25e+14) − 1.7659e+15 (1.75e+14) − 1.7789e+15 (2.21e+14) − 2.2769e+15 (2.58e+14) − 1.5281e+15 (3.71e+14) − 1.6937e+15 (1.18e+14) − 4.0969e+14 (5.47e+14)
3 4.8738e+4 (1.75e+3) − 4.5929e+4 (2.19e+3) − 4.9393e+4 (1.42e+3) − 4.8603e+4 (1.33e+3) − 4.6778e+4 (2.86e+3) − 4.5579e+4 (1.14e+3) − 3.6072e+4 (1.74e+4)
MaF4 7 8.7077e+5 (3.05e+4) − 8.0239e+5 (2.82e+4) − 8.1084e+5 (1.62e+4) − 8.5034e+5 (2.62e+4) − 7.7240e+5 (4.29e+4) − 7.2449e+5 (1.30e+4) − 5.9289e+5 (1.94e+5)
10 6.7233e+6 (2.60e+5) − 6.0250e+6 (2.86e+5) − 6.3734e+6 (1.92e+5) − 6.5203e+6 (1.74e+5) − 5.9191e+6 (3.03e+5) − 5.7393e+6 (1.40e+5) − 4.4042e+6 (1.54e+6)
3 4.9538e+1 (4.45e+0) − 5.2642e+1 (6.11e+0) − 5.4470e+1 (3.15e+0) − 4.7894e+1 (4.58e+0) − 4.0258e+1 (4.17e+0) = 3.3169e+1 (4.84e+0) + 4.1270e+1 (8.70e+0)
MaF5 7 7.9171e+2 (4.28e+1) − 7.1040e+2 (5.07e+1) = 8.5290e+2 (4.00e+1) − 8.0694e+2 (6.07e+1) − 7.2546e+2 (3.10e+1) = 5.8438e+2 (2.43e+1) + 6.9413e+2 (8.54e+1)
10 6.0461e+3 (2.73e+2) = 4.0037e+3 (2.92e+2) + 6.4586e+3 (2.71e+2) − 6.2995e+3 (2.35e+2) − 4.8950e+3 (2.25e+2) + 4.8043e+3 (8.86e+1) + 5.7599e+3 (5.16e+2)
3 7.2639e+2 (3.31e+1) − 8.3703e+2 (1.69e+1) − 8.0067e+2 (1.13e+1) − 7.3975e+2 (1.57e+1) − 6.5515e+2 (6.88e+1) = 4.9315e+2 (8.05e+1) + 6.0317e+2 (9.91e+1)
MaF6 7 7.7008e+2 (1.23e+1) − 8.2834e+2 (1.24e+1) − 7.9155e+2 (9.62e+0) − 7.6100e+2 (1.04e+1) − 7.5000e+2 (6.25e+1) − 6.5844e+2 (7.27e+1) = 6.4865e+2 (7.34e+1)
10 7.4653e+2 (1.09e+1) − 8.1404e+2 (1.03e+1) − 7.5976e+2 (9.15e+0) − 7.4288e+2 (9.16e+0) − 7.1953e+2 (7.94e+1) − 6.7437e+2 (3.76e+1) − 6.2255e+2 (3.93e+1)
3 9.5959e+0 (8.69e−1) = 1.1458e+1 (3.25e−1) − 7.1439e+0 (6.89e−1) + 1.1268e+1 (2.65e−1) − 1.1122e+1 (4.16e−1) − 1.1635e+1 (1.98e−1) − 1.0157e+1 (9.64e−1)
MaF7 7 2.7205e+1 (4.91e−1) − 2.7662e+1 (2.75e−1) − 1.6686e+1 (1.39e+0) + 2.7409e+1 (2.48e−1) − 2.7633e+1 (2.61e−1) − 2.8404e+1 (2.66e−1) − 2.6183e+1 (9.45e−1)
10 3.9502e+1 (4.17e−1) − 3.9687e+1 (4.40e−1) − 3.0296e+1 (2.06e+0) + 3.9558e+1 (2.36e−1) − 3.9381e+1 (4.56e−1) − 4.0819e+1 (1.21e−1) − 3.7201e+1 (9.75e−1)
+/-/= 3/54/12 5/49/15 14/42/13 5/49/15 13/32/24 26/35/8
Avg Rank 4.0289 3.2753 3.0431 2.8841 2.3623 3.1739 1.8261
𝑝 value 4.8571e−11 1.2904e−8 2.1579e−5 7.2509e−7 1.0000e+0 6.5500e−2

4.2. Comparison over other algorithms suites, respectively, but is outperformed by these algorithms in 7, 5,
20, 16, 12, and 24 test suites, respectively. MOEA/D-FEF ranked first
4.2.1. Comparison results on all test suites with 𝐷 = 50 on these test suites with 𝐷 = 50.
The IGD results for different test suites with 𝐷 = 50 are shown in It is worth noting that K-RVEA and KTA2 are the Kriging-assisted
Table 1. HeE-MOEA performs best in most test suites (23 out of 69). SA-MOEAs; these algorithms perform worse than MOEA/D-FEF in most
MOEA/D-FEF, K-RVEA, KTA2, MCEA/D, and ADSAPSO perform best DTLZ/MaF test suites; however, there is no clear difference in terms
in 18, 16, 6, 5, and 1 test suites, respectively. Specifically, MOEA/D- of ‘‘+/-/=’’ and these algorithms for WFG test suites with 𝐷 = 50.
FEF outperforms ADSAPSO, CPS-MOEA, K-RVEA, KTA2, MCEA/D, and MCEA/D, ADSAPSO, and HeE-MOEA are competitive algorithms that
HeE-MOEA in 54, 55, 42, 41, 34, and 39 test suites, respectively. tackle high-dimensional problems. Similar to MOEA/D-FEF, MCEA/D
However, it is outperformed by these algorithms in 5, 2, 19, 14, 17, uses MOEA/D as its optimizer, while ADSAPSO and HeE-MOEA ap-
and 25 test suites, respectively. The 𝛥𝑝 results for the same test suites ply DR strategies to handle decision variables in high-dimensional
are shown in Table 2. MOEA/D-FEF achieves the best performance in spaces. MOEA/D-FEF significantly outperforms these algorithms on
22 out of 69 test suites, while HeE-MOEA, K-RVEA, KTA2, MCEA/D, most test suites with 𝐷 = 50. Additionally, MCEA/D and CPS-MOEA
CPS-MOEA, and ADSAPSO lead in 18, 13, 6, 8, 1, and 1 test suites, re- are classification-based SA-MOEAs. As discussed by Takumi et al. [44],
spectively. MOEA/D-FEF outperforms ADSAPSO, CPS-MOEA, K-RVEA, classification-based SA-MOEAs generally perform better and reduce
KTA2, MCEA/D, and HeE-MOEA in 51, 51, 39, 37, 30, and 36 test the risk of overfitting compared to approximation-based SA-MOEAs.

9
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

Table 5
IGD statistical results on all test instances with D = 150.
Problem M ADSAPSO CPS-MOEA K-RVEA KTA2 MCEA/D HeE-MOEA MOEA/D-FEF
3 4.2521e+3 (1.69e+2) − 3.7065e+3 (1.28e+2) − 4.1908e+3 (9.69e+1) − 4.1107e+3 (1.83e+2) − 3.2243e+3 (4.53e+2) − 3.9344e+3 (1.04e+2) − 4.4550e+2 (1.87e+2)
DTLZ1 7 3.1096e+3 (1.95e+2) − 2.8569e+3 (1.70e+2) − 3.1894e+3 (1.35e+2) − 3.1560e+3 (1.61e+2) − 2.9125e+3 (2.29e+2) − 3.1621e+3 (1.16e+2) − 2.8844e+2 (1.57e+2)
10 2.9769e+3 (3.08e+2) − 2.7435e+3 (1.61e+2) − 2.9806e+3 (1.85e+2) − 3.0335e+3 (2.26e+2) − 2.5708e+3 (3.92e+2) − 2.9978e+3 (1.15e+2) − 4.3377e+2 (2.35e+2)
3 9.5841e+0 (5.48e−1) − 7.4797e+0 (2.01e+0) − 1.0118e+1 (3.03e−1) − 1.0101e+1 (2.60e−1) − 2.4504e+0 (7.42e−1) = 9.6497e−1 (7.09e−2) + 2.5534e+0 (8.75e−1)
DTLZ2 7 1.0165e+1 (4.52e−1) − 8.4359e+0 (1.97e+0) − 1.0168e+1 (3.46e−1) − 1.0269e+1 (2.59e−1) − 3.5891e+0 (8.99e−1) − 1.2358e+0 (5.48e−2) + 2.8813e+0 (6.41e−1)
10 9.9215e+0 (5.08e−1) − 8.3460e+0 (2.04e+0) − 1.0084e+1 (3.08e−1) − 1.0098e+1 (3.69e−1) − 3.8331e+0 (8.23e−1) = 1.2765e+0 (5.94e−2) + 3.8087e+0 (8.80e−1)
3 1.2714e+4 (1.31e+3) − 1.1386e+4 (5.58e+2) − 1.3633e+4 (4.17e+2) − 1.3623e+4 (3.24e+2) − 8.3336e+3 (9.52e+2) − 1.2999e+4 (3.25e+2) − 1.0857e+3 (4.97e+2)
DTLZ3 7 1.2632e+4 (8.61e+2) − 1.1160e+4 (5.63e+2) − 1.3262e+4 (2.39e+2) − 1.3444e+4 (4.07e+2) − 8.6602e+3 (1.29e+3) − 1.2556e+4 (4.29e+2) − 1.2470e+3 (4.46e+2)
10 1.2634e+4 (4.70e+2) − 1.1155e+4 (3.53e+2) − 1.2872e+4 (3.31e+2) − 1.2962e+4 (4.25e+2) − 8.9297e+3 (9.56e+2) − 1.2327e+4 (4.20e+2) − 1.4971e+3 (2.70e+2)
3 9.3647e+0 (7.33e−1) − 7.7544e+0 (1.84e+0) − 1.0469e+1 (4.13e−1) − 1.0716e+1 (4.08e−1) − 2.6351e+0 (7.06e−1) = 1.3902e+0 (6.94e−2) + 2.8350e+0 (6.33e−1)
DTLZ4 7 9.9873e+0 (5.75e−1) − 8.0386e+0 (2.18e+0) − 1.0452e+1 (2.65e−1) − 1.0455e+1 (3.60e−1) − 2.9405e+0 (4.45e−1) + 1.5874e+0 (7.29e−2) + 3.4982e+0 (1.08e+0)
10 9.2660e+0 (7.41e−1) − 7.7475e+0 (1.91e+0) − 1.0197e+1 (3.07e−1) − 1.0408e+1 (3.12e−1) − 3.0863e+0 (5.66e−1) + 1.6169e+0 (6.31e−2) + 4.5882e+0 (1.83e+0)
3 9.1890e+0 (1.14e+0) − 7.7849e+0 (2.05e+0) − 1.0217e+1 (2.98e−1) − 1.0103e+1 (3.20e−1) − 2.5406e+0 (7.19e−1) = 8.6090e−1 (8.66e−2) + 2.6766e+0 (6.81e−1)
DTLZ5 7 9.8278e+0 (3.19e−1) − 8.6846e+0 (2.16e+0) − 9.8250e+0 (3.31e−1) − 1.0008e+1 (2.34e−1) − 3.2091e+0 (9.51e−1) − 7.7117e−1 (6.81e−2) + 2.8942e+0 (7.09e−1)
10 9.5219e+0 (4.15e−1) − 9.3338e+0 (1.52e+0) − 9.6137e+0 (3.17e−1) − 9.8305e+0 (2.53e−1) − 3.1048e+0 (6.47e−1) = 7.3695e−1 (5.72e−2) + 3.3687e+0 (1.06e+0)
3 1.1844e+2 (6.02e+0) − 1.1373e+2 (3.58e+0) − 1.3027e+2 (9.84e−1) − 1.3181e+2 (4.61e−1) − 9.5562e+1 (1.08e+1) − 1.3200e+2 (5.60e−1) − 7.2371e+1 (1.02e+1)
DTLZ6 7 1.1843e+2 (4.80e+0) − 1.1300e+2 (2.70e+0) − 1.2699e+2 (7.81e−1) − 1.2834e+2 (5.28e−1) − 9.9712e+1 (5.61e+0) − 1.2858e+2 (4.66e−1) − 6.9874e+1 (8.96e+0)
10 1.1741e+2 (4.34e+0) − 1.1012e+2 (2.36e+0) − 1.2475e+2 (6.08e−1) − 1.2562e+2 (5.45e−1) − 9.8352e+1 (9.38e+0) − 1.2592e+2 (3.57e−1) − 7.0173e+1 (4.51e+0)
3 9.5197e+0 (6.16e−1) = 1.0824e+1 (4.31e−1) − 1.0579e+1 (4.33e−1) − 1.0827e+1 (3.42e−1) − 1.0730e+1 (5.37e−1) − 1.0934e+1 (2.73e−1) − 9.8375e+0 (1.12e+0)
DTLZ7 7 2.5789e+1 (1.27e+0) − 2.6324e+1 (6.66e−1) − 2.4477e+1 (1.93e+0) − 2.6646e+1 (8.77e−1) − 2.6079e+1 (8.91e−1) − 2.6579e+1 (1.04e+0) − 2.2697e+1 (2.51e+0)
10 3.7609e+1 (1.62e+0) − 3.8436e+1 (1.13e+0) − 3.3197e+1 (2.21e+0) = 3.8159e+1 (1.54e+0) − 3.7731e+1 (2.22e+0) − 3.8744e+1 (1.09e+0) − 3.4381e+1 (2.91e+0)
3 2.3671e+0 (3.85e−2) = 2.3373e+0 (5.69e−2) = 2.2271e+0 (7.50e−2) + 2.3373e+0 (9.74e−2) = 2.3117e+0 (5.01e−2) = 2.3768e+0 (5.30e−2) = 2.3123e+0 (9.66e−2)
WFG1 7 3.0103e+0 (3.31e−2) = 3.0233e+0 (4.23e−2) = 2.9351e+0 (4.47e−2) = 2.9890e+0 (5.44e−2) = 3.0059e+0 (3.65e−2) = 3.0240e+0 (2.75e−2) = 3.0096e+0 (4.67e−2)
10 3.4577e+0 (3.87e−2) = 3.4490e+0 (3.09e−2) = 3.4118e+0 (4.80e−2) = 3.4524e+0 (4.86e−2) = 3.4919e+0 (3.16e−2) − 3.4689e+0 (3.16e−2) = 3.4608e+0 (4.19e−2)
3 1.0223e+0 (9.29e−2) − 9.3632e−1 (7.96e−2) − 9.7767e−1 (6.11e−2) − 9.9783e−1 (6.23e−2) − 8.5574e−1 (5.57e−2) + 9.4144e−1 (7.27e−2) − 8.9791e−1 (1.36e−1)
WFG2 7 3.2236e+0 (6.34e−1) − 2.3594e+0 (3.47e−1) = 2.8092e+0 (3.97e−1) − 2.9268e+0 (6.96e−1) − 2.5517e+0 (4.71e−1) = 2.9213e+0 (6.04e−1) − 2.3893e+0 (5.70e−1)
10 4.8516e+0 (6.11e−1) − 3.4134e+0 (5.93e−1) + 3.9607e+0 (6.99e−1) = 4.3859e+0 (7.20e−1) − 4.1123e+0 (9.28e−1) = 4.5719e+0 (5.96e−1) − 3.8562e+0 (8.09e−1)
3 8.6588e−1 (1.39e−2) − 8.6445e−1 (2.46e−2) − 8.5673e−1 (6.50e−3) − 8.6439e−1 (9.33e−3) − 6.8088e−1 (3.83e−2) + 5.7836e−1 (2.87e−2) + 7.1277e−1 (4.89e−2)
WFG3 7 1.4922e+0 (3.21e−2) − 1.5391e+0 (5.23e−2) − 1.4640e+0 (1.71e−2) − 1.5010e+0 (2.53e−2) − 1.2084e+0 (4.48e−2) + 9.9388e−1 (5.44e−2) + 1.4233e+0 (6.44e−2)
10 1.9117e+0 (4.40e−2) = 1.9392e+0 (5.39e−2) − 1.8672e+0 (2.25e−2) = 1.9175e+0 (4.95e−2) − 1.4904e+0 (7.17e−2) + 1.2968e+0 (6.75e−2) + 1.8554e+0 (7.66e−2)
3 7.8955e−1 (7.30e−2) − 6.6215e−1 (3.30e−2) + 7.1956e−1 (5.64e−2) − 7.7380e−1 (6.46e−2) − 6.3343e−1 (2.72e−2) + 8.0208e−1 (8.06e−2) − 7.0149e−1 (4.96e−2)
WFG4 7 5.3946e+0 (3.99e−1) − 4.3972e+0 (3.45e−1) − 5.1085e+0 (2.86e−1) − 5.0714e+0 (4.42e−1) − 4.3269e+0 (5.02e−1) − 5.5302e+0 (4.08e−1) − 4.0809e+0 (2.57e−1)
10 1.0129e+1 (4.84e−1) − 8.5075e+0 (5.37e−1) − 9.6950e+0 (5.79e−1) − 9.6019e+0 (5.75e−1) − 8.4744e+0 (9.50e−1) − 1.0333e+1 (5.48e−1) − 7.0940e+0 (5.31e−1)
3 8.4521e−1 (2.22e−2) − 7.8812e−1 (1.02e−2) − 8.2360e−1 (9.27e−3) − 8.4989e−1 (1.75e−2) − 7.4048e−1 (3.89e−2) = 8.6884e−1 (1.13e−2) − 7.3994e−1 (4.97e−2)
WFG5 7 4.1868e+0 (1.83e−1) − 3.8375e+0 (1.89e−1) − 3.7924e+0 (1.80e−1) − 4.0257e+0 (2.20e−1) − 3.7232e+0 (1.53e−1) = 3.5344e+0 (8.21e−2) = 3.6968e+0 (3.54e−1)
10 8.2169e+0 (4.31e−1) − 7.5177e+0 (4.29e−1) − 7.4456e+0 (2.91e−1) − 7.5208e+0 (4.50e−1) − 7.2602e+0 (3.30e−1) − 6.3884e+0 (1.30e−1) + 6.8841e+0 (5.63e−1)
3 1.0300e+0 (1.80e−2) + 1.0359e+0 (1.57e−2) = 1.0090e+0 (1.28e−2) = 1.0343e+0 (2.03e−2) = 9.4128e−1 (1.96e−2) + 9.2147e−1 (2.66e−2) + 1.1087e+0 (9.95e−2)
WFG6 7 4.5108e+0 (1.74e−1) − 4.0704e+0 (2.40e−1) = 4.0575e+0 (1.90e−1) = 4.4359e+0 (2.38e−1) − 4.1658e+0 (2.76e−1) − 4.3984e+0 (2.40e−1) − 3.9129e+0 (2.18e−1)
10 8.4682e+0 (3.81e−1) − 7.6948e+0 (4.25e−1) − 7.9632e+0 (2.82e−1) − 8.0853e+0 (3.70e−1) − 7.8248e+0 (5.16e−1) − 8.1704e+0 (2.98e−1) − 7.1299e+0 (5.94e−1)
3 8.0343e−1 (2.43e−2) = 7.9532e−1 (2.46e−2) − 7.8555e−1 (1.46e−2) − 8.2224e−1 (2.17e−2) − 7.1714e−1 (2.51e−2) + 7.4789e−1 (4.61e−2) = 7.6504e−1 (2.84e−2)
WFG7 7 4.4969e+0 (2.73e−1) − 4.0643e+0 (2.85e−1) = 4.2026e+0 (2.84e−1) − 4.3514e+0 (2.55e−1) − 4.1452e+0 (2.51e−1) − 4.4674e+0 (2.85e−1) − 3.9211e+0 (2.97e−1)
10 8.6308e+0 (6.22e−1) − 8.0353e+0 (4.44e−1) − 8.3861e+0 (4.27e−1) − 8.5369e+0 (4.47e−1) − 8.3687e+0 (5.85e−1) − 8.7269e+0 (3.84e−1) − 7.5106e+0 (7.07e−1)
3 8.8675e−1 (1.76e−2) − 8.8073e−1 (2.71e−2) = 8.5133e−1 (1.96e−2) = 8.9717e−1 (2.46e−2) − 8.2676e−1 (3.67e−2) + 8.2712e−1 (2.65e−2) + 8.6316e−1 (2.64e−2)
WFG8 7 4.3412e+0 (1.89e−1) − 3.9546e+0 (1.85e−1) = 3.9473e+0 (1.81e−1) = 4.2522e+0 (2.26e−1) − 3.9910e+0 (2.80e−1) = 4.4436e+0 (1.84e−1) − 3.8330e+0 (2.37e−1)
10 8.4105e+0 (3.65e−1) − 7.8573e+0 (3.10e−1) − 7.9190e+0 (2.25e−1) − 8.2747e+0 (4.32e−1) − 8.0343e+0 (4.83e−1) − 8.3813e+0 (4.48e−1) − 7.3785e+0 (6.29e−1)
3 1.0951e+0 (4.18e−2) − 1.1106e+0 (3.55e−2) = 1.0980e+0 (2.46e−2) = 1.0967e+0 (3.35e−2) = 9.4312e−1 (5.19e−2) + 7.9249e−1 (3.44e−2) + 1.0147e+0 (6.79e−2)
WFG9 7 4.5592e+0 (2.86e−1) − 4.5189e+0 (2.45e−1) − 4.3350e+0 (1.94e−1) − 4.3783e+0 (2.80e−1) − 4.0524e+0 (2.34e−1) = 3.1985e+0 (7.83e−2) + 4.1695e+0 (3.42e−1)
10 8.3427e+0 (5.64e−1) − 8.2012e+0 (4.59e−1) − 8.1893e+0 (4.05e−1) − 8.1926e+0 (3.68e−1) − 7.4588e+0 (3.40e−1) + 5.6945e+0 (1.32e−1) + 7.8311e+0 (8.04e−1)
3 1.1298e+1 (1.11e+0) − 8.0026e+0 (1.83e+0) − 1.2059e+1 (4.19e−1) − 1.2077e+1 (5.68e−1) − 2.2199e+0 (5.71e−1) + 8.9313e−1 (1.27e−1) + 3.1739e+0 (1.23e+0)
MaF1 7 2.1213e+1 (1.67e+0) − 1.6042e+1 (3.82e+0) − 2.2887e+1 (7.96e−1) − 2.2494e+1 (1.22e+0) − 4.3950e+0 (6.54e−1) + 1.5390e+0 (2.39e−1) + 9.8964e+0 (2.73e+0)
10 2.4208e+1 (3.17e+0) − 2.1014e+1 (6.14e+0) − 2.7627e+1 (7.72e−1) − 2.7563e+1 (1.08e+0) − 2.6819e+0 (3.45e−1) + 1.7479e+0 (2.63e−1) + 1.3674e+1 (5.41e+0)
3 7.8426e−1 (4.02e−2) − 6.7778e−1 (1.41e−1) − 8.5750e−1 (2.14e−2) − 8.5835e−1 (2.23e−2) − 3.2118e−1 (4.82e−2) = 1.8322e−1 (6.95e−3) + 3.2318e−1 (3.76e−2)
MaF2 7 4.9508e−1 (1.44e−2) − 5.2402e−1 (1.76e−2) − 5.0788e−1 (1.02e−2) − 5.1145e−1 (1.83e−2) − 4.4790e−1 (3.09e−2) = 4.4963e−1 (1.14e−2) = 4.3522e−1 (3.16e−2)
10 4.5233e−1 (2.90e−2) = 4.5973e−1 (2.05e−2) = 4.6757e−1 (1.79e−2) = 4.6424e−1 (2.52e−2) = 4.7583e−1 (2.51e−2) = 4.6441e−1 (1.51e−2) = 4.5211e−1 (2.24e−2)
3 4.2236e+8 (3.73e+8) − 1.5980e+8 (1.92e+7) − 2.4575e+8 (3.27e+7) − 7.7313e+8 (8.34e+8) − 7.1499e+7 (2.14e+7) − 9.0611e+8 (8.38e+8) − 5.9283e+5 (5.27e+5)
MaF3 7 4.7622e+8 (5.41e+8) − 1.5824e+8 (1.51e+7) − 2.9051e+8 (1.26e+8) − 7.4477e+8 (7.61e+8) − 7.9479e+7 (2.52e+7) − 5.4972e+8 (4.28e+8) − 1.0255e+7 (1.75e+7)
10 3.9923e+8 (2.52e+8) − 1.5483e+8 (1.79e+7) − 2.8830e+8 (1.02e+8) − 4.7629e+8 (4.46e+8) − 6.9321e+7 (2.45e+7) − 4.0133e+8 (3.84e+8) − 1.6049e+7 (5.63e+7)
3 4.0884e+4 (2.32e+3) − 3.6634e+4 (1.47e+3) − 4.2917e+4 (1.06e+3) − 4.3447e+4 (1.07e+3) − 3.0929e+4 (4.34e+3) − 3.9665e+4 (1.26e+3) − 5.4791e+3 (3.34e+3)
MaF4 7 7.1361e+5 (2.14e+4) − 6.3108e+5 (2.41e+4) − 7.1161e+5 (1.95e+4) − 7.3149e+5 (2.18e+4) − 5.8932e+5 (6.21e+4) − 6.6957e+5 (2.91e+4) − 7.8246e+4 (3.44e+4)
10 5.6261e+6 (1.65e+5) − 4.9068e+6 (2.13e+5) − 5.5765e+6 (1.90e+5) − 5.6425e+6 (2.07e+5) − 4.5066e+6 (3.71e+5) − 5.2527e+6 (1.60e+5) − 6.1909e+5 (1.78e+5)
3 3.6747e+1 (1.65e+1) − 2.7928e+1 (8.36e−1) = 4.0096e+1 (1.33e+1) − 3.9453e+1 (1.30e+1) − 1.8178e+1 (3.07e+0) = 9.2477e+0 (7.61e−1) + 2.4769e+1 (1.12e+1)
MaF5 7 1.4246e+2 (1.39e+2) = 5.6232e+1 (2.33e+0) + 1.0786e+2 (7.28e+1) = 1.2364e+2 (7.36e+1) − 5.5255e+1 (3.98e+0) + 1.4707e+2 (3.65e+1) − 1.0559e+2 (1.32e+2)
10 3.1718e+2 (1.55e+2) = 2.8928e+2 (3.05e+1) + 4.5928e+2 (4.07e+2) = 9.3355e+2 (7.93e+2) − 2.9154e+2 (3.40e+1) + 8.2374e+2 (4.03e+2) − 3.2194e+2 (7.35e+1)
3 9.3768e+2 (9.11e+1) − 6.6824e+2 (1.77e+2) − 9.8693e+2 (3.92e+1) − 1.0024e+3 (3.40e+1) − 2.2818e+2 (8.05e+1) = 6.5198e+1 (7.09e+0) + 2.3387e+2 (7.55e+1)
MaF6 7 9.5496e+2 (4.92e+1) − 7.0091e+2 (1.97e+2) − 9.6048e+2 (3.50e+1) − 9.8264e+2 (2.31e+1) − 3.1679e+2 (6.91e+1) − 5.5472e+1 (9.11e+0) + 2.5847e+2 (6.13e+1)
10 9.3053e+2 (6.63e+1) − 7.5516e+2 (2.14e+2) − 9.4309e+2 (3.63e+1) − 9.5639e+2 (2.15e+1) − 2.5885e+2 (6.28e+1) = 5.3503e+1 (8.31e+0) + 3.1073e+2 (1.27e+2)
3 9.6492e+0 (8.03e−1) = 1.0679e+1 (2.90e−1) − 1.0411e+1 (6.16e−1) = 1.0504e+1 (5.39e−1) = 1.0812e+1 (4.40e−1) − 1.0975e+1 (4.55e−1) − 9.4518e+0 (1.57e+0)
MaF7 7 2.5476e+1 (1.15e+0) − 2.6680e+1 (9.74e−1) − 2.4909e+1 (1.27e+0) = 2.6610e+1 (7.65e−1) − 2.5843e+1 (1.34e+0) − 2.6704e+1 (1.10e+0) − 2.3450e+1 (2.30e+0)
10 3.7637e+1 (1.38e+0) − 3.8386e+1 (8.46e−1) − 3.3516e+1 (2.62e+0) = 3.8118e+1 (2.00e+0) − 3.8582e+1 (1.36e+0) − 3.8701e+1 (1.05e+0) − 3.0425e+1 (4.74e+0)
+/-/= 1/58/10 4/53/12 1/52/16 0/62/7 17/33/19 26/36/7
Avg Rank 3.9275 3.2028 3.5217 3.2608 2.5337 2.5507 1.7971
𝑝 value 4.9673e−14 8.4379e−5 9.8311e−10 1.9606e−22 1.0000e+0 6.2403e−4

The superior performance of MOEA/D-FEF over these SA-MOEAs sug- suites are shown in Table 4. MOEA/D-FEF performs best in 29 out of 69
gests that the feature extraction framework in MOEA/D-FEF effectively test suites, while HeE-MOEA, K-RVEA, MCEA/D, and CPS-MOEA lead
mitigates the risk of overfitting in approximation-based SA-MOEAs. in 25, 9, 4, and 2 test suites, respectively. MOEA/D-FEF outperforms
ADSAPSO, CPS-MOEA, K-RVEA, KTA2, MCEA/D, and HeE-MOEA in 54,
4.2.2. Comparison results on all test suites with 𝐷 = 100 49, 42, 49, 32, and 35 test suites, respectively, but is outperformed in
The IGD results for different test suites with 𝐷 = 100 are shown in 3, 5, 14, 5, 13, and 26 test suites, respectively.
Table 3. MOEA/D-FEF achieves the best results in most test suites (25 It is worth noting that the number of ‘‘-’’ for K-RVEA and KTA2
out of 69), while HeE-MOEA, K-RVEA, KTA2, and MCEA/D perform increases compared to 𝐷 = 50 because Kriging models struggle to
best in 25, 13, 2, and 4 test suites, respectively. Specifically, MOEA/D- handle high-dimensional inputs.
FEF outperforms ADSAPSO, CPS-MOEA, K-RVEA, KTA2, MCEA/D, and
HeE-MOEA in 55, 54, 43, 49, 35, and 38 test suites, respectively. 4.2.3. Comparison results on all test suites with 𝐷 = 150
However, these algorithms outperform MOEA/D-FEF in 0, 1, 12, 2, The IGD results for different test suites with 𝐷 = 150 are shown
15, and 26 test suites, respectively. The 𝛥𝑝 results for the same test in Table 5. MOEA/D-FEF achieves the best results in most test suites

10
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

Table 6
𝛥𝑃 statistical results on all test instances with D = 150.
Problem M ADSAPSO CPS-MOEA K-RVEA KTA2 MCEA/D HeE-MOEA MOEA/D-FEF
3 5.6742e+3 (1.26e+2) − 4.9856e+3 (8.11e+1) − 5.7278e+3 (8.75e+1) − 5.7347e+3 (6.34e+1) − 4.8834e+3 (2.58e+2) − 5.5306e+3 (3.30e+1) − 2.5319e+3 (1.72e+3)
DTLZ1 7 4.9299e+3 (1.05e+2) − 4.5190e+3 (9.04e+1) = 5.2921e+3 (5.95e+1) − 5.0050e+3 (1.14e+2) − 4.3659e+3 (1.04e+2) = 4.8773e+3 (2.60e+1) − 3.4388e+3 (1.56e+3)
10 4.8091e+3 (1.28e+2) = 4.3297e+3 (9.12e+1) = 5.1354e+3 (8.54e+1) = 4.9931e+3 (1.08e+2) = 4.4718e+3 (1.08e+2) = 4.8001e+3 (6.65e+1) = 4.1868e+3 (1.67e+3)
3 1.1240e+1 (4.08e−1) − 1.2525e+1 (1.53e−1) − 1.1854e+1 (1.27e−1) − 1.1716e+1 (1.15e−1) − 9.9969e+0 (1.34e+0) = 1.1171e+1 (7.65e−2) − 9.0331e+0 (2.40e+0)
DTLZ2 7 1.2002e+1 (2.29e−1) − 1.3105e+1 (2.53e−1) − 1.1873e+1 (5.03e−2) − 1.1822e+1 (1.11e−1) − 1.1510e+1 (6.44e−1) − 7.7621e+0 (7.05e−1) + 9.7480e+0 (1.36e+0)
10 1.1870e+1 (3.47e−1) − 1.3008e+1 (1.26e−1) − 1.1646e+1 (8.22e−2) − 1.1636e+1 (1.03e−1) − 1.1328e+1 (5.81e−1) − 8.3669e+0 (4.42e−1) + 9.8788e+0 (7.85e−1)
3 1.4642e+4 (1.20e+3) − 1.3439e+4 (1.48e+2) − 1.5529e+4 (7.49e+1) − 1.5434e+4 (1.14e+2) − 1.1979e+4 (1.21e+3) − 1.4986e+4 (4.76e+1) − 6.4872e+3 (4.59e+3)
DTLZ3 7 1.5295e+4 (1.98e+2) − 1.3885e+4 (1.37e+2) − 1.5416e+4 (6.05e+1) − 1.5401e+4 (8.61e+1) − 1.3439e+4 (6.04e+2) − 1.4975e+4 (6.01e+1) − 8.8751e+3 (4.19e+3)
10 1.5107e+4 (1.34e+2) − 1.3675e+4 (9.99e+1) − 1.5107e+4 (5.43e+1) − 1.5110e+4 (7.32e+1) − 1.3727e+4 (3.54e+2) − 1.4769e+4 (3.67e+1) − 1.0758e+4 (3.63e+3)
3 1.1312e+1 (4.12e−1) − 1.2300e+1 (2.93e−1) − 1.1820e+1 (2.72e−1) − 1.1784e+1 (2.69e−1) − 9.3334e+0 (9.65e−1) − 1.1563e+1 (1.33e−1) − 8.1635e+0 (1.73e+0)
DTLZ4 7 1.2110e+1 (3.19e−1) − 1.3460e+1 (1.67e−1) − 1.1882e+1 (5.91e−2) − 1.1819e+1 (1.19e−1) − 1.0382e+1 (2.76e−1) − 7.0284e+0 (2.63e−1) + 9.4610e+0 (1.30e+0)
10 1.1858e+1 (2.20e−1) − 1.3280e+1 (1.78e−1) − 1.1757e+1 (4.39e−2) − 1.1692e+1 (9.08e−2) − 1.0854e+1 (3.24e−1) = 7.6347e+0 (1.24e−1) + 1.0757e+1 (6.62e−1)
3 1.1063e+1 (7.82e−1) − 1.2767e+1 (2.11e−1) − 1.1928e+1 (1.26e−1) − 1.1804e+1 (1.50e−1) − 9.6698e+0 (1.24e+0) = 1.1234e+1 (5.80e−2) − 9.5752e+0 (1.62e+0)
DTLZ5 7 1.2079e+1 (1.94e−1) − 1.3092e+1 (1.85e−1) − 1.1970e+1 (6.56e−2) − 1.1894e+1 (1.03e−1) − 1.1176e+1 (9.22e−1) − 9.4453e+0 (6.12e−1) = 9.3919e+0 (1.67e+0)
10 1.1918e+1 (2.35e−1) − 1.2891e+1 (1.40e−1) − 1.1747e+1 (6.33e−2) − 1.1719e+1 (1.33e−1) − 1.1069e+1 (4.85e−1) − 9.0158e+0 (6.34e−1) + 1.0258e+1 (8.32e−1)
3 1.3253e+2 (2.00e+0) − 1.2605e+2 (6.57e−1) = 1.3425e+2 (5.58e−2) − 1.3446e+2 (8.56e−2) − 1.2503e+2 (3.00e+0) = 1.3428e+2 (2.80e−2) − 1.1376e+2 (1.62e+1)
DTLZ6 7 1.3038e+2 (5.60e−1) − 1.2579e+2 (5.97e−1) − 1.3077e+2 (6.69e−2) − 1.3097e+2 (6.99e−2) − 1.2509e+2 (1.22e+0) − 1.3239e+2 (2.35e−2) − 1.0411e+2 (8.93e+0)
10 1.2775e+2 (5.33e−1) − 1.2341e+2 (5.45e−1) − 1.2811e+2 (7.48e−2) − 1.2825e+2 (9.80e−2) − 1.2370e+2 (1.14e+0) − 1.2947e+2 (1.73e−2) − 1.0475e+2 (6.62e+0)
3 1.0082e+1 (7.70e−1) + 1.1684e+1 (2.73e−1) − 1.0870e+1 (4.14e−1) = 1.1601e+1 (2.32e−1) = 1.1327e+1 (3.94e−1) = 1.1432e+1 (9.20e−2) = 1.1078e+1 (7.07e−1)
DTLZ7 7 2.7592e+1 (4.53e−1) − 2.7867e+1 (2.12e−1) − 2.6393e+1 (7.28e−1) = 2.7644e+1 (2.50e−1) − 2.7704e+1 (1.87e−1) − 2.8395e+1 (3.17e−1) − 2.6042e+1 (5.98e−1)
10 3.9385e+1 (3.72e−1) − 3.9679e+1 (2.66e−1) − 3.8536e+1 (2.56e−1) − 3.9525e+1 (1.71e−1) − 3.9630e+1 (3.24e−1) − 4.0957e+1 (1.57e−1) − 3.7940e+1 (7.17e−1)
3 2.3671e+0 (3.85e−2) = 2.3417e+0 (6.22e−2) = 2.2209e+0 (8.28e−2) + 2.3498e+0 (6.81e−2) = 2.3147e+0 (4.40e−2) = 2.1988e+0 (3.32e−2) + 2.3392e+0 (7.21e−2)
WFG1 7 3.0103e+0 (3.31e−2) = 3.0255e+0 (4.67e−2) = 2.9315e+0 (4.47e−2) + 2.9901e+0 (5.78e−2) = 3.0069e+0 (3.97e−2) = 3.0226e+0 (2.96e−2) = 3.0153e+0 (4.92e−2)
10 3.4577e+0 (3.87e−2) = 3.4547e+0 (3.12e−2) = 3.4163e+0 (5.04e−2) + 3.4430e+0 (4.91e−2) = 3.5010e+0 (2.37e−2) − 3.4689e+0 (2.45e−2) = 3.4719e+0 (3.96e−2)
3 1.0223e+0 (9.29e−2) − 9.3671e−1 (8.60e−2) = 9.6470e−1 (5.37e−2) = 9.9415e−1 (6.23e−2) = 8.5539e−1 (6.14e−2) + 8.0502e−1 (8.37e−3) + 9.4754e−1 (9.06e−2)
WFG2 7 3.2236e+0 (6.34e−1) − 2.3893e+0 (3.56e−1) = 2.7825e+0 (4.24e−1) − 3.0056e+0 (7.40e−1) − 2.6161e+0 (4.64e−1) = 2.9567e+0 (6.52e−1) − 2.3330e+0 (4.34e−1)
10 4.8516e+0 (6.11e−1) − 3.6607e+0 (5.79e−1) = 4.0342e+0 (7.23e−1) = 4.4413e+0 (7.46e−1) − 3.9474e+0 (9.25e−1) = 4.4664e+0 (6.16e−1) − 3.5667e+0 (7.30e−1)
3 1.0039e+0 (1.26e−2) + 1.0657e+0 (2.32e−2) = 1.0129e+0 (8.97e−3) + 1.0099e+0 (1.01e−2) + 9.8596e−1 (6.44e−2) + 9.9212e−1 (5.81e−3) + 1.0609e+0 (5.59e−2)
WFG3 7 2.1688e+0 (7.55e−2) + 2.2636e+0 (1.04e−1) + 1.9351e+0 (2.71e−2) + 2.0449e+0 (6.70e−2) + 1.9606e+0 (1.04e−1) + 1.3599e+0 (1.75e−2) + 2.6452e+0 (1.72e−1)
10 2.9503e+0 (1.33e−1) + 3.1885e+0 (1.76e−1) + 2.5749e+0 (3.96e−2) + 2.8037e+0 (1.05e−1) + 2.6403e+0 (2.08e−1) + 1.8902e+0 (3.21e−2) + 3.6734e+0 (2.92e−1)
3 7.8955e−1 (7.30e−2) − 6.6540e−1 (3.60e−2) + 7.2616e−1 (6.13e−2) = 7.8255e−1 (6.66e−2) − 6.3224e−1 (2.69e−2) + 5.6077e−1 (1.14e−2) + 7.1073e−1 (3.13e−2)
WFG4 7 5.3946e+0 (3.99e−1) − 4.3229e+0 (2.77e−1) − 5.0953e+0 (3.10e−1) − 5.0066e+0 (4.28e−1) − 4.3156e+0 (5.17e−1) = 5.5987e+0 (3.91e−1) − 4.0065e+0 (3.05e−1)
10 1.0129e+1 (4.84e−1) − 8.4119e+0 (5.38e−1) − 9.6883e+0 (5.29e−1) − 9.6999e+0 (5.21e−1) − 8.6809e+0 (8.96e−1) − 1.0374e+1 (5.09e−1) − 7.0485e+0 (4.72e−1)
3 8.4521e−1 (2.22e−2) − 7.8871e−1 (1.09e−2) − 8.2398e−1 (7.82e−3) − 8.5060e−1 (1.94e−2) − 7.4229e−1 (3.40e−2) = 7.8520e−1 (1.83e−3) − 7.6057e−1 (4.09e−2)
WFG5 7 4.1868e+0 (1.83e−1) − 3.8481e+0 (2.09e−1) − 3.8180e+0 (1.85e−1) − 4.0464e+0 (2.34e−1) − 3.7334e+0 (1.44e−1) = 3.5328e+0 (8.53e−2) = 3.6793e+0 (2.96e−1)
10 8.2169e+0 (4.31e−1) − 7.5498e+0 (4.37e−1) − 7.4642e+0 (3.04e−1) − 7.5483e+0 (4.39e−1) − 7.2403e+0 (3.36e−1) = 6.3708e+0 (1.28e−1) + 6.9937e+0 (6.84e−1)
3 1.0303e+0 (1.72e−2) + 1.0413e+0 (1.31e−2) + 1.0133e+0 (8.46e−3) + 1.0366e+0 (2.10e−2) + 9.4458e−1 (1.88e−2) + 1.0008e+0 (1.94e−3) + 1.1393e+0 (1.01e−1)
WFG6 7 4.5108e+0 (1.74e−1) − 4.1062e+0 (2.17e−1) − 4.0539e+0 (2.13e−1) = 4.4291e+0 (2.46e−1) − 4.1722e+0 (3.02e−1) − 4.4031e+0 (2.64e−1) − 3.9381e+0 (2.53e−1)
10 8.4682e+0 (3.81e−1) − 7.6959e+0 (4.36e−1) − 7.9445e+0 (2.90e−1) − 8.0700e+0 (3.72e−1) − 7.7773e+0 (5.14e−1) − 8.1357e+0 (2.96e−1) − 7.0784e+0 (5.80e−1)
3 8.0343e−1 (2.43e−2) = 7.9104e−1 (2.07e−2) = 7.8561e−1 (1.61e−2) = 8.2427e−1 (2.26e−2) − 7.1693e−1 (2.65e−2) + 7.4303e−1 (2.42e−3) + 7.9252e−1 (4.06e−2)
WFG7 7 4.4969e+0 (2.73e−1) − 4.0319e+0 (2.66e−1) − 4.2797e+0 (2.59e−1) − 4.3084e+0 (2.55e−1) − 4.1382e+0 (2.12e−1) − 4.5063e+0 (3.00e−1) − 3.8134e+0 (2.50e−1)
10 8.6308e+0 (6.22e−1) − 7.9965e+0 (4.36e−1) − 8.3155e+0 (4.07e−1) − 8.5374e+0 (4.91e−1) − 8.3395e+0 (6.08e−1) − 8.8021e+0 (3.53e−1) − 7.4179e+0 (6.98e−1)
3 8.8675e−1 (1.76e−2) − 8.8456e−1 (2.70e−2) − 8.5082e−1 (1.65e−2) = 9.0014e−1 (2.55e−2) − 8.2566e−1 (3.99e−2) + 7.8405e−1 (4.09e−3) + 8.6089e−1 (4.00e−2)
WFG8 7 4.3412e+0 (1.89e−1) − 3.9758e+0 (1.77e−1) = 3.9415e+0 (1.97e−1) = 4.2640e+0 (2.47e−1) − 3.9926e+0 (2.56e−1) = 4.4549e+0 (1.73e−1) − 3.8341e+0 (2.26e−1)
10 8.4105e+0 (3.65e−1) − 7.8506e+0 (2.96e−1) − 7.9296e+0 (2.41e−1) − 8.2899e+0 (4.74e−1) − 8.0544e+0 (4.24e−1) − 8.3314e+0 (4.45e−1) − 7.2896e+0 (5.71e−1)
3 1.0953e+0 (4.16e−2) − 1.1367e+0 (1.87e−2) − 1.1062e+0 (1.48e−2) − 1.0906e+0 (2.95e−2) − 9.6652e−1 (6.33e−2) + 1.0108e+0 (5.80e−3) + 1.0602e+0 (4.91e−2)
WFG9 7 4.5592e+0 (2.86e−1) − 4.5244e+0 (2.41e−1) − 4.3083e+0 (1.81e−1) = 4.3658e+0 (2.81e−1) = 4.0867e+0 (2.31e−1) = 3.1991e+0 (8.40e−2) + 4.2816e+0 (4.27e−1)
10 8.3427e+0 (5.64e−1) − 8.2352e+0 (4.72e−1) = 8.1152e+0 (3.24e−1) = 8.1457e+0 (3.87e−1) = 7.4126e+0 (3.07e−1) + 5.6915e+0 (1.35e−1) + 7.9154e+0 (5.55e−1)
3 1.3828e+1 (5.30e−1) − 1.5307e+1 (5.59e−1) − 1.4441e+1 (1.63e−1) − 1.4220e+1 (2.22e−1) − 1.0421e+1 (2.14e+0) = 6.6416e+0 (2.16e+0) + 9.2047e+0 (3.63e+0)
MaF1 7 2.5867e+1 (4.98e−1) − 2.7413e+1 (5.49e−1) − 2.6676e+1 (2.03e−1) − 2.6490e+1 (3.05e−1) − 1.7056e+1 (3.79e+0) = 1.1876e+1 (4.91e+0) + 1.8778e+1 (4.46e+0)
10 3.0981e+1 (1.51e+0) − 3.3032e+1 (8.40e−1) − 3.2327e+1 (3.43e−1) − 3.2392e+1 (4.82e−1) − 1.7621e+1 (5.92e+0) + 1.5034e+1 (3.98e+0) + 2.6486e+1 (4.30e+0)
3 9.2951e−1 (3.77e−2) − 1.0240e+0 (2.02e−2) − 9.8797e−1 (1.09e−2) − 9.6811e−1 (1.55e−2) − 5.0374e−1 (1.01e−1) = 2.1190e−1 (3.43e−2) + 4.8836e−1 (1.33e−1)
MaF2 7 4.9508e−1 (1.44e−2) − 5.2855e−1 (1.19e−2) − 5.0788e−1 (1.02e−2) − 5.1145e−1 (1.83e−2) − 4.6769e−1 (2.21e−2) − 4.5963e−1 (1.14e−2) = 4.3530e−1 (3.15e−2)
10 4.5833e−1 (2.90e−2) = 4.5973e−1 (2.05e−2) = 4.6757e−1 (1.79e−2) = 4.6424e−1 (2.52e−2) = 4.5583e−1 (2.51e−2) = 4.6441e−1 (1.51e−2) = 4.5271e−1 (2.24e−2)
3 1.8013e+16 (4.13e+15) − 1.1870e+16 (1.83e+15) − 2.0669e+16 (1.61e+15) − 2.0622e+16 (2.71e+15) − 2.2208e+16 (4.20e+15) − 1.6534e+16 (1.27e+15) − 6.1973e+15 (7.89e+15)
MaF3 7 1.3618e+16 (1.22e+15) − 1.1936e+16 (1.59e+15) − 1.3740e+16 (1.52e+15) − 1.4233e+16 (1.64e+15) − 1.2528e+16 (2.10e+15) − 9.9032e+15 (5.53e+14) − 1.8551e+15 (2.08e+15)
10 1.2711e+16 (1.76e+15) − 1.0010e+16 (8.18e+14) − 1.2078e+16 (9.56e+14) − 1.2985e+16 (1.42e+15) − 8.5140e+15 (1.70e+15) − 9.2263e+15 (5.46e+14) − 2.4166e+15 (2.10e+15)
3 7.4497e+4 (2.44e+3) − 6.9607e+4 (2.34e+3) = 7.5333e+4 (1.26e+3) − 7.3708e+4 (1.07e+3) − 7.1156e+4 (3.96e+3) = 6.7257e+4 (9.25e+2) = 5.9438e+4 (2.19e+4)
MaF4 7 1.3386e+6 (4.78e+4) − 1.2450e+6 (5.57e+4) − 1.2242e+6 (1.86e+4) − 1.2724e+6 (2.52e+4) − 1.1752e+6 (4.91e+4) = 1.0899e+6 (1.97e+4) = 9.4056e+5 (3.29e+5)
10 1.0429e+7 (5.70e+5) − 9.4818e+6 (4.23e+5) − 9.6704e+6 (1.56e+5) − 9.9650e+6 (2.38e+5) − 9.0217e+6 (3.44e+5) − 8.7474e+6 (1.09e+5) − 7.1681e+6 (2.06e+6)
3 7.7289e+1 (8.58e+0) − 8.3058e+1 (4.49e+0) − 8.5725e+1 (3.60e+0) − 7.8744e+1 (6.03e+0) − 6.0669e+1 (6.69e+0) + 4.5007e+1 (7.18e+0) + 6.7951e+1 (1.15e+1)
MaF5 7 1.2311e+3 (7.74e+1) − 1.0541e+3 (7.92e+1) + 1.4120e+3 (4.26e+1) − 1.2641e+3 (7.10e+1) − 1.1256e+3 (4.32e+1) = 8.8268e+2 (4.28e+1) + 1.1461e+3 (1.02e+2)
10 9.2661e+3 (5.47e+2) = 6.4085e+3 (5.66e+2) + 1.1093e+4 (2.00e+2) − 1.0166e+4 (4.50e+2) = 7.4289e+3 (4.08e+2) + 7.4151e+3 (1.47e+2) + 9.3509e+3 (1.21e+3)
3 1.1319e+3 (7.52e+1) − 1.2479e+3 (2.82e+1) − 1.1813e+3 (1.39e+1) − 1.1665e+3 (1.94e+1) − 9.1627e+2 (1.34e+2) = 7.7340e+2 (1.36e+2) + 9.6856e+2 (1.85e+2)
MaF6 7 1.1992e+3 (3.42e+1) − 1.2942e+3 (1.57e+1) − 1.1849e+3 (6.88e+0) − 1.1815e+3 (9.78e+0) − 1.1627e+3 (9.82e+1) − 1.0529e+3 (7.41e+1) = 9.6240e+2 (1.83e+2)
10 1.1614e+3 (1.69e+1) − 1.2732e+3 (2.18e+1) − 1.1615e+3 (5.57e+0) − 1.1557e+3 (1.10e+1) − 1.1410e+3 (1.25e+2) − 1.0520e+3 (4.65e+1) = 1.0280e+3 (1.27e+2)
3 1.0228e+1 (8.88e−1) = 1.1466e+1 (2.17e−1) − 1.0837e+1 (3.17e−1) = 1.1415e+1 (2.15e−1) − 1.1452e+1 (3.74e−1) − 1.1708e+1 (1.81e−1) − 1.0188e+1 (7.00e−1)
MaF7 7 2.7404e+1 (2.73e−1) − 2.7862e+1 (3.00e−1) − 2.6399e+1 (6.99e−1) = 2.7597e+1 (1.92e−1) − 2.7701e+1 (3.29e−1) − 2.8487e+1 (2.55e−1) − 2.6030e+1 (1.21e+0)
10 3.9337e+1 (4.32e−1) − 3.9637e+1 (2.66e−1) − 3.8490e+1 (2.63e−1) − 3.9543e+1 (3.14e−1) − 3.9849e+1 (4.86e−1) − 4.0812e+1 (4.21e−1) − 3.6564e+1 (1.26e+0)
+/-/= 5/55/8 6/48/15 7/47/15 4/55/10 13/31/25 26/31/12
Avg Rank 3.6811 3.3623 3.3828 3.1159 2.2318 2.4492 1.7681
𝑝 value 6.8995e−13 5.7350e−10 5.3040e−9 4.9265e−14 1.0000e+0 5.7611e−2

(31 out of 69), while HeE-MOEA, K-RVEA, MCEA/D, and CPS-MOEA Those test results can support the effectiveness of MOEA/D-FEF. In
perform best in 26, 4, 5, and 3 test suites, respectively. Specifically, addition, WFG3 and WFG9 are challenging to the MOEA/D-FEF, and
MOEA/D-FEF outperforms ADSAPSO, CPS-MOEA, K-RVEA, KTA2, a more significant number of ‘‘-’’/‘‘=’’ is observed on these test suites,
MCEA/D, and HeE-MOEA in 58, 53, 52, 62, 33, and 36 test suites, which are categorized as disconnected or non-separable problems [39].
respectively. However, these algorithms outperform MOEA/D-FEF in Thus, the performance of MOEA/D-FEF may be challenged in solving
1, 4, 1, 0, 17, and 26 test suites, respectively. The 𝛥𝑝 results for the the problems with disconnected or non-separable PF.
same test suites are shown in Table 6. MOEA/D-FEF achieves the best
performance in 38 out of 69 test suites, while HeE-MOEA, K-RVEA, 4.3. The effects of the SRS operation
MCEA/D, ADSAPSO, and CPS-MOEA lead in 22, 2, 5, 1, and 1 test
suites, respectively. MOEA/D-FEF outperforms ADSAPSO, CPS-MOEA, To evaluate the impact of the SRS operation, we set 𝑆 =
K-RVEA, KTA2, MCEA/D, and HeE-MOEA in 55, 48, 47, 55, 31, and 31 {0, 3, 4, 5, 6, 7, 8, 9}, where a higher value of 𝑆 increases the number
test suites, respectively, but is outperformed in 5, 6, 7, 4, 13, and 26 of center points distributed within the decision space. Here, 𝑆 = 0
test suites, respectively. represents the IGD value of MOEA/D-FEF without the SRS operation.
It is worth noting that the number of ‘‘-’’ of the K-RVEA and the The notations ‘‘+/-/=’’ indicate that the results of MOEA/D-FEF with
KTA2 still increases with the 𝐷. Compared to the MCEA/D and HeE- other 𝑆 values are better than, worse than, or similar to that with 𝑆 = 4,
MOEA, there is no apparent difference in the term ‘‘+/-=’’ with the respectively. The best results are highlighted with a gray background.
increase of 𝐷. All IGD values in Table 7 are obtained from 20 independent runs.

11
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

Table 7
The IGD statistical results of different 𝑆 on DTLZ and WFG test suites with 100 dimension decision variable.
Problem S = 0 S = 3 S = 5 S = 6 S = 7 S = 8 S = 9 S = 4
DTLZ1 2.1532e+3 (4.38e+2) − 1.4706e+3 (4.38e+2) − 1.6831e+3 (5.35e+2) − 1.6831e+3 (5.35e+2) − 1.6081e+3 (5.30e+2) − 1.7551e+3 (5.70e+2) − 3.0713e+2 (8.64e+1) − 2.4558e+2 (1.22e+2)
DTLZ2 1.8547e+0 (4.18e−1) − 2.0328e+0 (5.42e−1) − 1.9896e+0 (4.96e−1) − 1.9896e+0 (4.96e−1) − 2.0692e+0 (5.06e−1) − 1.8092e+0 (3.07e−1) = 2.1466e+0 (5.17e−1) − 1.6815e+0 (5.64e−1)
DTLZ3 4.6924e+3 (1.72e+3) − 3.1747e+3 (1.04e+3) − 3.1145e+3 (5.14e+2) − 3.1145e+3 (5.14e+2) − 3.7103e+3 (1.34e+3) − 3.5143e+3 (1.38e+3) − 7.8170e+2 (2.25e+2) = 7.8686e+2 (4.13e+2)
DTLZ4 2.3378e+0 (9.70e−1) − 2.5278e+0 (6.25e−1) − 2.0632e+0 (4.54e−1) = 2.0632e+0 (4.54e−1) = 2.1536e+0 (3.33e−1) = 2.5189e+0 (1.24e+0) − 2.5480e+0 (6.27e−1) − 2.0547e+0 (5.70e−1)
DTLZ5 1.9924e+0 (6.22e−1) − 2.0092e+0 (5.00e−1) − 1.7196e+0 (3.44e−1) = 1.7196e+0 (3.44e−1) = 2.0803e+0 (8.24e−1) − 4.3125e+0 (1.73e+0) − 4.0398e+0 (1.86e+0) − 1.4828e+0 (3.30e−1)
DTLZ6 5.1553e+1 (9.06e+0) = 3.5845e+1 (1.02e+1) + 4.6133e+1 (4.69e+0) = 4.6133e+1 (4.69e+0) = 4.1435e+1 (5.47e+0) + 4.6406e+1 (5.06e+0) = 4.6678e+1 (3.77e+0) = 4.7029e+1 (7.16e+0)
DTLZ7 1.0072e+1 (6.85e−1) − 7.9102e+0 (2.08e+0) + 9.4962e+0 (1.03e+0) − 9.4962e+0 (1.03e+0) − 4.8484e+0 (2.37e+0) = 8.5713e+0 (1.24e+0) = 9.4845e+0 (1.09e+0) − 8.6670e+0 (2.20e+0)
WFG1 2.2544e+0 (9.13e−2) = 2.3056e+0 (1.16e−1) − 2.2770e+0 (9.57e−2) = 2.2734e+0 (6.53e−2) = 2.3072e+0 (7.15e−2) − 2.3042e+0 (7.53e−2) − 2.2911e+0 (1.26e−1) = 2.2430e+0 (9.15e−2)
WFG2 9.1160e−1 (1.04e−1) − 9.7650e−1 (1.49e−1) − 9.4688e−1 (1.17e−1) − 9.3643e−1 (1.35e−1) − 9.5940e−1 (1.37e−1) − 9.3044e−1 (9.07e−2) − 9.4377e−1 (1.17e−1) − 8.5168e−1 (6.92e−2)
WFG3 6.9625e−1 (7.63e−2) = 7.0368e−1 (3.62e−2) = 7.3624e−1 (6.34e−2) = 7.2479e−1 (7.43e−2) = 7.1785e−1 (5.54e−2) = 7.5648e−1 (5.99e−2) = 7.0434e−1 (5.76e−2) = 7.2012e−1 (5.93e−2)
WFG4 7.0431e−1 (4.42e−2) − 6.7451e−1 (2.74e−2) = 6.8068e−1 (3.41e−2) = 6.9559e−1 (4.53e−2) − 6.8669e−1 (5.05e−2) = 7.0954e−1 (6.73e−2) − 6.8285e−1 (4.11e−2) = 6.7434e−1 (3.15e−2)
WFG5 7.8448e−1 (5.43e−2) − 7.7326e−1 (4.66e−2) − 7.2909e−1 (5.82e−2) = 7.4135e−1 (3.98e−2) = 7.5579e−1 (3.71e−2) − 7.3275e−1 (6.88e−2) = 7.5309e−1 (7.15e−2) = 7.2744e−1 (5.32e−2)
WFG6 1.0503e+0 (1.01e−1) = 1.0389e+0 (9.12e−2) = 1.0673e+0 (9.35e−2) = 1.0543e+0 (8.94e−2) = 1.1199e+0 (1.02e−1) = 1.0640e+0 (9.07e−2) = 1.1116e+0 (9.80e−2) = 1.0516e+0 (1.10e−1)
WFG7 7.6486e−1 (3.14e−2) = 7.9960e−1 (2.68e−2) − 7.7180e−1 (4.46e−2) = 7.6481e−1 (3.03e−2) = 7.8472e−1 (2.73e−2) = 7.8156e−1 (3.07e−2) = 7.8599e−1 (3.77e−2) = 7.7125e−1 (3.29e−2)
WFG8 8.9168e−1 (4.06e−2) = 8.6985e−1 (3.49e−2) = 8.7709e−1 (3.32e−2) = 8.7656e−1 (4.02e−2) = 8.7121e−1 (3.59e−2) = 8.8328e−1 (2.45e−2) = 8.8493e−1 (3.70e−2) = 8.8171e−1 (4.10e−2)
WFG9 1.0901e+0 (8.59e−2) = 1.0763e+0 (5.57e−2) = 1.0484e+0 (7.18e−2) = 1.0815e+0 (3.89e−2) = 1.0672e+0 (7.03e−2) = 1.0613e+0 (4.79e−2)= 1.0598e+0 (8.22e−2) = 9.9575e−1 (9.03e−2)
+/-/= 1/6/9 2/5/9 0/4/12 1/6/9 1/6/9 1/4/11 0/3/13

Fig. 3. The average computational time (in seconds) of MOEA/D-FEF with different feature sets in 3-objectives DTLZ/WFG test suites with 𝐷 = 100.

Table 7 shows that the results obtained by MOEA/D-FEF with 𝑆 > 0 • MOEA/D-LPI: This variant utilizes all feature extraction methods
are better than those with 𝑆 = 0 on most test suites. Both DTLZ1 and selects the features with the highest VER value to train the
and DTLZ3, which have multimodal landscapes, present challenges for Kriging model but does not employ the feature drift strategy.
convergence, making it difficult for surrogate models to approximate
their objective functions accurately. The SRS operation enhances con- Table 8 shows the Friedman rank-sum test results and the average
vergence in these cases, demonstrating more robust performance. On ranks obtained from the IGD values of these algorithms. As shown in
the other hand, DTLZ2 and DTLZ5 are EMOPs that are relatively easier Table 8, MOEA/D-K obtains the worst results on all test suites, and
to approximate, and these test suites need more diversity. MOEA/D- MOEA/D-FEF obtains the best results on most test suites; it ranked first.
FEF tends to employ the D part to tackle those test suites, prioritizing Compared to the MOEA/D-LPI, MOEA/D-FEF obtained better results on
exploitation over exploration directly. For the DTLZ4 and WFG1 test six test suites, proving the effectiveness of the feature shift strategy. The
suites, the initial training set with poor diversity poses a challenge for rank of MOEA/D-LPI is lower than MOEA/D-PCA and MOEA/D-Isomap
training a promising surrogate model. WFG6 and WFG9 are catego- but higher than MOEA/D-LPP. In more detail, MOEA/D-LPI obtains
rized as non-separable problems. WFG7 is a separable and unimodal better results than MOEA/D-LPP on DTLZ6-7 and WFG2, which denotes
problem. The MOEA/D-FEF with 𝑆 > 0 performs similarly to that with that the VAR is an effective indicator in these test suites; MOEA/D-LPI
𝑆 = 0 on these test suites. MOEA/D-FEF also obtains the best result obtains worse results than MOEA/D-LPP on DTLZ1, DTLZ3, and WFG3,
on the DTLZ6 problem with degenerate PF and the DTLZ7 problem it is challenge for MOEA/D-LPI to select a promising dimensionality
with a discontinuous PF. However, compared with DTLZ7, it is more reduction method on these test suites. For testing the impact of different
challenging to maintain diverse solutions on WFG3; the MOEA/D-FEF 𝐾, Table 9 summarizes the averaged IGD values of MOEA/D-FEF with
with 𝑆 = 0 performs better than that with 𝑆 > 0 on WFG3. The 𝐾 = {6, 8, 10, 12, 14, 16} on the DTLZ and WFG problems, respectively.
MOEA/D-FEF with 𝑆 = 0 performs worse than those with 𝑆 > 0 on 𝐾 = 14 is used as a default value. In the table, ‘‘+’’, ‘‘-’’, and ‘‘=’’
most test suites, and 𝑆 = 4 can be defined as the best setting in our denote that the IGD value of MOEA/D-FEF with 𝐾 = {6, 8, 10, 12, 16}
experimental setting. Thus, the number of center points for the SRS is significantly better, worse, and competitive to that obtained by
operation is set as 4 in this article.
MOEA/D-FEF with 𝐾 = 14, respectively. The MOEA/D-FEF with 𝐾 <
10 is worse than that with 𝐾 ≥ 10 due to 𝐾 < 10 cannot carry
4.4. The effects of the dimensionality reduction framework
enough information for the surrogates. From the overall results, 𝐾 =
14 can be defined as the best setting in our algorithm, as it derives
To evaluate the effectiveness of the dimensionality reduction frame-
work, we test several variants of MOEA/D-FEF: MOEA/D-K, MOEA/D- competitive performances on all test suites. MOEA/D-FEF employs a
LPP, MOEA/D-PCA, MOEA/D-I, and MOEA/D-LPI. feature extraction to map a high-dimensional decision space into a
low-dimensional decision space. The reduced features used to build
• MOEA/D-K: Does not employ any feature extraction strategy. surrogates could save computational time. Fig. 3 shows the average
• MOEA/D-LPP, MOEA/D-PCA, MOEA/D-I: Each variant utilizes computational time (in seconds) of MOEA/D-FEF with different feature
a single feature extraction method, LPP, PCA, or Isomap, respec- sets in 3-objectives DTLZ/WFG test suites with 𝐷 = 100, the blue data
tively, to handle the original decision variables. denotes the MOEA/D-FEF without the feature extraction framework,

12
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

Table 8
IGD statistical results obtained by different variants of MOEA/D-FEF on DTLZ/WFG test suites with 100 dimension decision variable.
Problem MOEA/D-K LPP PCA Isomap MOEA/D-LPI MOEA/D-FEF
DTLZ1 2.5989e+2 (1.19e+2) = 1.5132e+2 (1.09e+2) + 2.6440e+2 (1.31e+2) = 2.8709e+2 (8.02e+1) = 2.6268e+2 (1.07e+2) = 2.4558e+2 (1.22e+2)
DTLZ2 4.7443e+0 (1.39e+0) − 1.7662e+0 (5.28e−1) − 2.7943e+0 (8.39e−1) − 1.7656e+0 (5.61e−1) − 1.7602e+0 (4.60e−1) − 1.6815e+0 (5.64e−1)
DTLZ3 7.7203e+2 (2.99e+2) = 5.3091e+2 (2.41e+2) + 6.9237e+2 (2.98e+2) + 9.2675e+2 (3.60e+2) − 7.8792e+2 (2.46e+2) = 7.8686e+2 (4.13e+2)
DTLZ4 5.5487e+0 (1.00e+0) − 2.1066e+0 (4.69e−1) = 2.2249e+0 (7.08e−1) = 2.1031e+0 (5.16e−1) = 2.0344e+0 (4.45e−1) = 2.0547e+0 (5.70e−1)
DTLZ5 5.0127e+0 (1.38e+0) − 1.6554e+0 (3.59e−1) − 3.0567e+0 (1.07e+0) − 1.7086e+0 (5.30e−1) − 1.7750e+0 (7.11e−1) − 1.4828e+0 (3.30e−1)
DTLZ6 6.1081e+1 (7.58e+0) − 5.6680e+1 (5.86e+0) − 5.0971e+1 (8.19e+0) = 4.8544e+1 (5.30e+0) = 4.9956e+1 (4.76e+0) = 4.7029e+1 (7.16e+0)
DTLZ7 1.0202e+1 (6.82e−1) − 9.7495e+0 (9.95e−1) − 9.1599e+0 (1.06e+0) − 9.1200e+0 (1.01e+0) − 9.3142e+0 (1.36e+0) − 8.6670e+0 (2.20e+0)
WFG1 2.3118e+0 (6.75e−2) − 2.3148e+0 (7.23e−2) − 2.3082e+0 (8.55e−2) − 2.2771e+0 (6.44e−2) = 2.3450e+0 (5.71e−2) = 2.2430e+0 (9.15e−2)
WFG2 9.9789e−1 (8.74e−2) − 9.1189e−1 (1.33e−1) − 9.7876e−1 (1.12e−1) − 9.1985e−1 (1.05e−1) − 8.9824e−1 (1.10e−1) − 8.5168e−1 (6.92e−2)
WFG3 8.4384e−1 (1.35e−2) − 6.8278e−1 (3.21e−2) = 7.6079e−1 (6.17e−2) − 7.4168e−1 (4.72e−2) = 7.6480e−1 (6.30e−2) − 7.2012e−1 (5.93e−2)
WFG4 7.3104e−1 (5.54e−2) − 6.7771e−1 (4.54e−2) = 6.9703e−1 (5.29e−2) − 6.8448e−1 (4.13e−2) = 6.7876e−1 (3.74e−2) = 6.7434e−1 (3.15e−2)
WFG5 8.0653e−1 (3.05e−2) − 7.6430e−1 (3.23e−2) − 7.7183e−1 (4.94e−2) − 7.4527e−1 (6.49e−2) = 7.6137e−1 (5.08e−2) − 7.2744e−1 (5.32e−2)
WFG6 1.0534e+0 (3.82e−2) = 9.7120e−1 (6.51e−2) = 1.0449e+0 (8.02e−2) = 1.0838e+0 (1.02e−1) = 1.1111e+0 (1.02e−1) = 1.0516e+0 (1.10e−1)
WFG7 8.3606e−1 (3.78e−2) − 7.7512e−1 (4.55e−2) = 7.9074e−1 (3.44e−2) − 7.8436e−1 (3.52e−2) = 7.8454e−1 (2.91e−2) = 7.7125e−1 (3.29e−2)
WFG8 9.2232e−1 (3.63e−2) − 8.5058e−1 (3.97e−2) + 8.8921e−1 (4.41e−2) = 8.7376e−1 (3.29e−2) = 8.8441e−1 (2.90e−2) = 8.8171e−1 (4.10e−2)
WFG9 1.1048e+0 (6.36e−2) = 9.9891e−1 (8.28e−2) = 1.0808e+0 (4.96e−2) = 1.0404e+0 (6.58e−2) = 1.0272e+0 (6.42e−2) = 9.9575e−1 (9.03e−2)
+/-/= 0/12/4 3/7/6 1/9/6 0/5/11 0/6/10
Avg Rank 2.9375 1.6875 2.25 1.9375 1.75 1.3125
𝑝 value 2.1486e−7 1.0000e+0 1.1100e−2 1.2233e−1 2.7300e−2

Table 9
IGD statistical results obtained by MOEA/D-FEF with different 𝐾 on DTLZ/WFG test suites with 𝐷 = 100.
Problem K=6 K=8 K=10 K=12 K=16 K=14
DTLZ1 2.0947e+2 (1.17e+2) = 2.2136e+2 (1.20e+2) = 2.1877e+2 (1.21e+2) = 2.7749e+2 (1.06e+2) = 2.2982e+2 (1.22e+2) = 2.4558e+2 (1.22e+2)
DTLZ2 2.3645e+0 (9.13e−1) − 2.2026e+0 (8.71e−1) − 2.0775e+0 (9.15e−1) − 1.7304e+0 (4.81e−1) = 1.6570e+0 (4.74e−1) = 1.6815e+0 (5.64e−1)
DTLZ3 6.2073e+2 (2.42e+2) + 7.5963e+2 (3.35e+2) = 6.7112e+2 (2.85e+2) + 7.6155e+2 (2.90e+2) = 8.2341e+2 (3.07e+2) = 7.8686e+2 (4.13e+2)
DTLZ4 2.4919e+0 (1.17e+0) − 2.1672e+0 (5.57e−1) = 2.1231e+0 (3.79e−1) = 2.1967e+0 (6.96e−1) = 1.9842e+0 (4.23e−1) = 2.0547e+0 (5.70e−1)
DTLZ5 2.3198e+0 (1.22e+0) − 1.8235e+0 (4.26e−1) − 1.5924e+0 (4.01e−1) = 1.5895e+0 (5.24e−1) − 1.7045e+0 (5.61e−1) − 1.4828e+0 (3.30e−1)
DTLZ6 5.2949e+1 (8.45e+0) = 4.9720e+1 (7.67e+0) = 4.8170e+1 (5.63e+0) = 4.8199e+1 (4.03e+0) = 4.5975e+1 (3.62e+0) = 4.7029e+1 (7.16e+0)
DTLZ7 9.7338e+0 (1.27e+0) − 8.9790e+0 (1.49e+0) = 9.1881e+0 (1.30e+0) = 8.9245e+0 (9.86e−1) − 9.4448e+0 (1.41e+0) − 8.6670e+0 (2.20e+0)
WFG1 2.3467e+0 (6.90e−2) − 2.3047e+0 (9.43e−2) = 2.3034e+0 (6.08e−2) = 2.3040e+0 (7.64e−2) = 2.2952e+0 (7.99e−2) = 2.2430e+0 (9.15e−2)
WFG2 9.1085e−1 (1.22e−1) − 8.9683e−1 (8.65e−2) − 9.2547e−1 (1.26e−1) = 9.7478e−1 (1.39e−1) = 8.8749e−1 (1.01e−1) = 8.5168e−1 (6.92e−2)
WFG3 7.2229e−1 (5.56e−2) = 7.5293e−1 (6.44e−2) = 7.5183e−1 (6.34e−2) = 7.4094e−1 (7.32e−2) = 7.5611e−1 (6.99e−2) = 7.2012e−1 (5.93e−2)
WFG4 6.8420e−1 (4.58e−2) = 7.0434e−1 (3.53e−2) = 6.9196e−1 (3.34e−2) = 7.1279e−1 (6.63e−2) = 7.0789e−1 (4.77e−2) = 6.7434e−1 (3.15e−2)
WFG5 7.7740e−1 (8.26e−2) − 7.5166e−1 (5.64e−2) − 7.3757e−1 (5.80e−2) = 7.4748e−1 (5.18e−2) = 7.3252e−1 (4.65e−2) = 7.2744e−1 (5.32e−2)
WFG6 1.0450e+0 (8.76e−2) = 1.1004e+0 (8.94e−2) = 1.0861e+0 (1.10e−1) = 1.0742e+0 (9.79e−2) = 1.1165e+0 (1.05e−1) = 1.0516e+0 (1.10e−1)
WFG7 7.9751e−1 (5.33e−2) = 7.8516e−1 (3.52e−2) = 7.7849e−1 (3.25e−2) = 7.8108e−1 (3.70e−2) = 7.8683e−1 (3.58e−2) = 7.7125e−1 (3.29e−2)
WFG8 8.7568e−1 (2.92e−2) = 8.7353e−1 (3.76e−2) = 8.7026e−1 (3.42e−2) = 8.7069e−1 (3.83e−2) = 8.7374e−1 (3.75e−2) = 8.8171e−1 (4.10e−2)
WFG9 1.0341e+0 (6.03e−2) = 1.0279e+0 (7.11e−2) = 1.0642e+0 (8.86e−2) = 1.0220e+0 (8.01e−2) = 1.0503e+0 (5.26e−2) = 9.9575e−1 (9.03e−2)
+/-/= 1/7/8 0/4/12 1/1/14 0/2/14 0/2/14

Table 10 values of MOEA/D-FEF with 𝐿 = {10, 20, 30, 40} on the DTLZ and WFG
The IGD statistical results of different 𝐿 on DTLZ test suites with 100 dimension dec-
problems. 𝐿 = 20 is used as the default value. In the table, ‘‘+,’’ ‘‘-,’’ and
ision variable.
Problem 𝐿 = 10 𝐿 = 30 𝐿 = 40 𝐿 = 20
‘‘=’’ denote that the IGD value of MOEA/D-FEF with 𝐿 = {10, 20, 30, 40}
DTLZ1 2.1651e+2 (1.08e+2) = 2.3533e+2 (1.63e+2) = 2.5949e+2 (9.27e+1) = 2.4558e+2 (1.22e+2) is significantly better, worse, or competitive compared to that obtained
DTLZ2 1.6859e+0 (3.90e−1) = 1.8427e+0 (7.69e−1) = 2.1127e+0 (1.16e+0) − 1.6815e+0 (5.64e−1)
DTLZ3 7.9620e+2 (2.91e+2)= 8.1940e+2 (3.55e+2) = 8.0674e+2 (3.04e+2) = 7.8686e+2 (4.13e+2) by MOEA/D-FEF with 𝐿 = 20, respectively.
DTLZ4 2.0314e+0 (4.42e−1) = 2.4425e+0 (1.20e+0) 1.9657e+0 (2.86e−1) = 2.0547e+0 (5.70e−1)
DTLZ5 1.7093e+0 (7.67e−1) = 1.6248e+0 (7.23e−1)
−
= 1.9462e+0 (1.01e+0) − 1.4828e+0 (3.30e−1)
As shown in Table 10, the performance of MOEA/D-FEF with 𝐿 = 20
DTLZ6 4.7520e+1 (4.75e+0) = 4.6306e+1 (4.61e+0) = 4.7517e+1 (4.23e+0) = 4.7029e+1 (7.16e+0) is similar to that with 𝐿 = 10, 𝐿 = 30, and 𝐿 = 40 on most test suites.
DTLZ7 9.1188e+0 (1.50e+0) = 8.6066e+0 (1.76e+0) = 8.8556e+0 (1.51e+0) = 8.6670e+0 (2.20e+0)
WFG1 2.2698e+0 (1.06e−1) = 2.2981e+0 (8.38e−2) = 2.2675e+0 (7.85e−2) = 2.2430e+0 (9.15e−2) MOEA/D-FEF is insensitive to the setting of 𝐿. We set 𝐿 = 20 in our
WFG2 9.2299e−1 (1.20e−1) − 8.9035e−1 (8.58e−2) = 9.1374e−1 (9.90e−2) − 8.5168e−1 (6.92e−2)
WFG3 7.3523e−1 (6.67e−2) = 7.5871e−1 (5.23e−2) = 7.4058e−1 (6.18e−2) = 7.2012e−1 (5.93e−2) algorithm because MOEA/D-FEF with 𝐿 = 20 achieves the best results
WFG4
WFG5
6.7769e−1 (3.17e−2) =
7.3876e−1 (4.10e−2) =
6.9810e−1
7.6039e−1
(4.32e−2)
(5.34e−2)
=
−
6.9933e−1
7.6393e−1
(3.47e−2)
(3.98e−2)
=
−
6.7434e−1
7.2744e−1
(3.15e−2)
(5.32e−2)
on most test suites.
WFG6 1.0764e+0 (9.44e−2) = 1.0806e+0 (8.31e−2) = 1.0958e+0 (1.05e−1) = 1.0516e+0 (1.10e−1)
WFG7 7.8136e−1 (2.64e−2) = 7.8502e−1 (2.62e−2) = 7.9228e−1 (3.69e−2) = 7.7125e−1 (3.29e−2)
WFG8 8.7755e−1 (3.87e−2) = 8.8511e−1 (3.32e−2) = 8.7916e−1 (2.51e−2) = 8.8171e−1 (4.10e−2) 4.6. Impact of maximum repeat time
WFG9 1.0417e+0 (9.15e−2) = 1.0231e+0 (6.88e−2) = 1.0620e+0 (6.68e−2) = 9.9575e−1 (9.03e−2)
+/-/= 0/1/15 0/2/14 0/4/12
We analyze the impact of the maximum repeat time 𝑅′max ; a large
value of 𝑅′max increases the probability of generating promising solu-
tions. Table 12 summarizes the average IGD values of MOEA/D-FEF
the red one denotes that with the proposed framework. The fact that with 𝑅′max = {1, 3, 5, 10} on the DTLZ and WFG problems, respectively.
the MOEA/D-FEF without the feature extraction framework needs more 𝑅′max = 3 is used as a default value. In this table, ‘‘+,’’ ‘‘-,’’ and ‘‘=’’
computational time can denote the effectiveness of our method. denote that the IGD value of MOEA/D-FEF with 𝑅′max = {1, 5, 10} is
significantly better, worse, or competitive compared to that obtained by
4.5. Impact of the length of each prat MOEA/D-FEF with 𝑅′max = 3, respectively. MOEA/D-FEF with 𝑅′max = 1
is worse than that with 𝑅′max > 1 on most test suites because 𝑅′max = 1
For testing the effect of different FEs assigned to each part, this cannot ensure convergence. 𝑅′max > 3 obtains similar results to 𝑅′max =
section sets four different values for 𝐿 = {10, 20, 30, 40}. If 𝐿 > 40, 3. We set 𝑅′max = 3 in our algorithm because it achieves competitive
the evaluation process will be divided into two parts, so 𝐿 is set to performance on both the DTLZ and WFG test suites while consuming
less than 40 in our algorithm. Table 10 summarizes the average IGD less computational time than when 𝑅′max > 3.

13
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

Table 11
The IGD statistical results of different 𝑅 on DTLZ/WFG test suites with 100 dimension decision variable.
Problem 𝑅 = 0.1 𝑅 = 0.5 𝑅 = 0.7 𝑅 = 0.9 𝑅 = 0.3
DTLZ1 2.6329e+2 (9.49e+1) = 2.8440e+2 (1.39e+2) = 2.9693e+2 (1.38e+2) = 2.1167e+2 (1.16e+2) = 2.4558e+2 (1.22e+2)
DTLZ2 1.8264e+0 (6.15e−1) − 1.7531e+0 (5.90e−1) = 1.6650e+0 (5.82e−1) = 1.6654e+0 (3.75e−1) = 1.6815e+0 (5.64e−1)
DTLZ3 8.0383e+2 (2.98e+2) = 7.6194e+2 (3.38e+2) = 7.4355e+2 (3.44e+2) = 7.6090e+2 (3.23e+2) = 7.8686e+2 (4.13e+2)
DTLZ4 2.0670e+0 (4.76e−1) = 2.2011e+0 (5.29e−1) = 1.9359e+0 (4.77e−1) = 2.0622e+0 (8.15e−1) = 2.0547e+0 (5.70e−1)
DTLZ5 1.5582e+0 (5.23e−1) = 1.6662e+0 (4.74e−1) − 1.4501e+0 (3.13e−1) = 1.6086e+0 (4.77e−1) − 1.4828e+0 (3.30e−1)
DTLZ6 4.9810e+1 (4.06e+0) = 4.7006e+1 (7.08e+0) = 4.7890e+1 (6.90e+0) = 4.7711e+1 (5.31e+0) = 4.7029e+1 (7.16e+0)
DTLZ7 8.7141e+0 (1.93e+0) = 8.7489e+0 (1.13e+0) = 9.1968e+0 (1.45e+0) = 9.6666e+0 (1.12e+0) − 8.6670e+0 (2.20e+0)
WFG1 2.2735e+0 (1.14e−1) = 2.2608e+0 (8.49e−2) = 2.2809e+0 (9.41e−2) = 2.2732e+0 (9.41e−2) = 2.2430e+0 (9.15e−2)
WFG2 8.7207e−1 (8.28e−2) = 8.9188e−1 (1.10e−1) − 9.3123e−1 (1.20e−1) − 9.4125e−1 (1.20e−1) − 8.5168e−1 (6.92e−2)
WFG3 7.3594e−1 (7.11e−2) = 7.4907e−1 (7.19e−2) = 7.7704e−1 (5.43e−2) − 7.5224e−1 (5.43e−2) − 7.2012e−1 (5.93e−2)
WFG4 6.9981e−1 (3.57e−2) = 7.0259e−1 (3.78e−2) = 6.9636e−1 (5.66e−2) = 6.8123e−1 (5.66e−2) = 6.7434e−1 (3.15e−2)
WFG5 7.4697e−1 (5.13e−2) = 7.4417e−1 (5.50e−2) = 7.2256e−1 (6.35e−2) = 7.2426e−1 (6.35e−2) = 7.2744e−1 (5.32e−2)
WFG6 1.0554e+0 (8.74e−2) = 1.0795e+0 (1.01e−1) = 1.0714e+0 (1.01e−1) = 1.0967e+0 (1.01e−1) = 1.0516e+0 (1.10e−1)
WFG7 7.8452e−1 (2.58e−2) − 7.8103e−1 (3.00e−2) = 7.8794e−1 (3.53e−2) − 7.9872e−1 (3.53e−2) − 7.7125e−1 (3.29e−2)
WFG8 8.6056e−1 (3.43e−2) = 8.7822e−1 (3.92e−2) = 8.8065e−1 (3.30e−2) = 8.7895e−1 (3.30e−2) = 8.8171e−1 (4.10e−2)
WFG9 1.0547e+0 (7.85e−2) = 1.0355e+0 (7.96e−2) = 1.0543e+0 (5.58e−2) = 1.0362e+0 (5.58e−2) = 9.9575e−1 (9.03e−2)
+/-/= 0/2/14 0/2/14 0/3/13 0/5/11

Table 12
The IGD statistical results of different 𝑅′𝑚𝑎𝑥 on DTLZ/WFG test suites with 100 dime-
nsion decision variable.
Problem 𝑅′𝑚𝑎𝑥 = 1 𝑅′𝑚𝑎𝑥 = 5 𝑅′𝑚𝑎𝑥 = 10 𝑅′𝑚𝑎𝑥 = 3
DTLZ1 1.9150e+2 (1.10e+2) = 3.0482e+2 (1.50e+2) − 2.4820e+2 (9.75e+1) = 2.4558e+2 (1.22e+2)
DTLZ2 2.0293e+0 (4.04e−1) − 1.7583e+0 (3.81e−1) = 1.9171e+0 (6.27e−1) = 1.6815e+0 (5.64e−1)
DTLZ3 5.8540e+2 (3.07e+2) + 7.6716e+2 (1.80e+2) = 8.2481e+2 (3.30e+2) = 7.8686e+2 (4.13e+2)
DTLZ4 2.0314e+0 (4.42e−1) = 2.4425e+0 (1.20e+0) = 1.9657e+0 (2.86e−1) = 2.0547e+0 (5.70e−1)
DTLZ5 1.8382e+0 (4.93e−1) − 1.5470e+0 (5.61e−1) = 1.6062e+0 (7.33e−1) = 1.4828e+0 (3.30e−1)
DTLZ6 5.4204e+1 (6.39e+0) − 4.4989e+1 (5.31e+0) = 4.6044e+1 (4.77e+0) = 4.7029e+1 (7.16e+0)
DTLZ7 9.6920e+0 (1.17e+0) − 8.1726e+0 (1.47e+0) = 7.4928e+0 (2.23e+0) + 8.6670e+0 (2.20e+0)
WFG1 2.3575e+0 (5.49e−2) − 2.2647e+0 (6.62e−2) = 2.2313e+0 (9.35e−2) = 2.2430e+0 (9.15e−2)
WFG2 8.9598e−1 (1.13e−1) − 8.6946e−1 (1.19e−1) = 8.8649e−1 (8.62e−2) = 8.5168e−1 (6.92e−2)
WFG3 7.3420e−1 (3.38e−2) = 7.3410e−1 (7.64e−2) = 7.3825e−1 (8.28e−2) = 7.2012e−1 (5.93e−2)
WFG4 6.9172e−1 (3.37e−2) = 6.8988e−1 (4.00e−2) = 6.8660e−1 (3.92e−2) = 6.7434e−1 (3.15e−2)
WFG5 7.5705e−1 (5.66e−2) − 7.2125e−1 (5.52e−2) = 7.2163e−1 (6.86e−2) = 7.2744e−1 (5.32e−2)
WFG6 1.0159e+0 (7.52e−2) = 1.0934e+0 (1.03e−1) = 1.0910e+0 (8.64e−2) = 1.0516e+0 (1.10e−1)
WFG7 8.0176e−1 (3.64e−2) − 7.8142e−1 (4.21e−2) = 7.5609e−1 (2.79e−2) = 7.7125e−1 (3.29e−2)
WFG8 8.8129e−1 (2.69e−2) = 8.6052e−1 (3.69e−2) = 8.5491e−1 (3.83e−2) = 8.8171e−1 (4.10e−2)
WFG9 1.0300e+0 (6.60e−2) = 1.0309e+0 (7.60e−2) = 1.0332e+0 (6.70e−2) = 9.9575e−1 (9.03e−2)
+/-/= 1/8/7 0/1/15 1/0/15

4.7. Impact of shift ratio

We analyze the shift ratio 𝑅; the value of 𝑅 can significantly

impact the performance of the feature shift strategy. A large value Fig. 4. The solutions obtained by different SA-MOEAs in the 2nd stage of the multi-pass
work roll cooling design problem.
of 𝑅 increases the pressure on the dimensionally reduced data to
obtain a similar relative location to the original data. We set 𝑅 =
{0.1, 0.3, 0.5, 0.7, 0.9}. Table 11 summarizes the average IGD values of
MOEA/D-FEF with 𝑅 = {0.1, 0.3, 0.5, 0.7, 0.9} on the DTLZ and WFG decision variables, with seven variables assigned to each pass. The
problems, respectively. 𝑅 = 0.3 is used as a default value. In this table, nine objectives focus on minimizing the changes in temperature at the
‘‘+,’’ ‘‘-,’’ and ‘‘=’’ denote that the IGD value of MOEA/D-FEF with roll surface (𝛥𝑇𝑠 ) and radial stress at the roll surface (𝑅𝑡𝑠 ). The multi-
𝑅 = {0.1, 0.3, 0.5, 0.7, 0.9} is significantly better, worse, or competitive pass work roll cooling design problem is mathematically defined as
compared to that obtained by MOEA/D-FEF with 𝑅 = 0.3, respectively. follows [45,46]:
MOEA/D-FEF with 𝑅 = 0.3 performs similarly to that with 𝑅 = 0.1 ⎡𝑓1 (𝒙) = 𝛥𝑇𝑠−𝑝1 ⎤
on most test suites and performs better than the latter on two test min 𝑭 (𝒙) = ⎢ ⋮ ⎥ s.t. 𝒙∈𝛺 (10)
𝒙 ⎢ ⎥
suites. MOEA/D-FEF with 𝑅 > 0.5 is worse than that with 𝑅 ≤ 0.5, ⎣ 𝑓9 (𝒙) = 𝑅𝑡𝑠−𝑝5 ⎦
demonstrating that a large 𝑅 may destroy the linear information of the where 𝑭 = (𝑓1 , … , 𝑓9 ) represents nine objectives of multi-pass work roll
dimensionally reduced data. Hence, we set 𝑅 = 0.3 in our algorithm. cooling design problem; 𝛥𝑇𝑠−𝑝𝑖 denotes the changes in temperature at
the roll surface of 𝑖nd pass work roll, 𝑅𝑡𝑠−𝑝𝑖 denotes the radial stress at
4.8. Multi-pass work roll cooling design problem the roll surface of 𝑖nd pass work roll; 𝒙 = (𝑥1 , … , 𝑥35 ) represents the
decision variable, and its range is provided in Table 13.
In the preceding sections, the effectiveness of MOEA/D-FEF has KTA2, MCEA/D, and HeE-MOEA are compared with MOEA/D-FEF
been demonstrated through various benchmark problems. Then, this to demonstrate their performance on this real-world problem. The total
section extends the application of MOEA/D-FEF to a multi-pass work number of simulations is set to 182, with each SA-MOEA executing
roll cooling design (MPWRCD) problem, which is a 9-objective problem ten independent runs. The Hypervolume (HV) metric [47] is used to
with 35 decision variables [45]. Multi-pass work rolling is a structured, assess performance, and the statistical results are provided in Table 14.
multi-stage process in which a metal strip travels through multiple The results show that MOEA/D-FEF consistently outperforms the other
stands sequentially, with the output from one pass serving as the algorithms at all stages. The second pass work roll, a classical two-
input for the next. This problem considers five passes, leading to 35 objective optimization problem commonly used for visualizing decision

14
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

Table 13
Decision variables and their ranges of Eq. (10).
Variable Range Variable Range Variable Range Variable Range Variable Range
𝑥1 [5, 15] 𝑥2 [1230, 1250] 𝑥3 [10, 30] 𝑥4 [15, 50] 𝑥5 [0.14, 0.2]
𝑥6 [40, 80] 𝑥7 [65, 75] 𝑥8 [5, 15] 𝑥9 [1155, 1195] 𝑥10 [10, 30]
𝑥11 [15, 50] 𝑥12 [0.17, 0.29] 𝑥13 [40, 80] 𝑥14 [75.75, 85.75] 𝑥15 [5, 15]
𝑥16 [1080, 1120] 𝑥17 [10, 30] 𝑥18 [15, 50] 𝑥19 [0.32, 0.44] 𝑥20 [40, 80]
𝑥21 [82.25, 92.25] 𝑥22 [5, 15] 𝑥23 [1005, 1045] 𝑥24 [10, 30] 𝑥25 [15, 50]
𝑥26 [0.57, 0.69] 𝑥27 [40, 80] 𝑥28 [86.2, 96.2] 𝑥29 [5, 15] 𝑥30 [950, 970]
𝑥31 [10, 30] 𝑥32 [15, 50] 𝑥33 [0.98, 1.1] 𝑥34 [40, 80] 𝑥35 [88.58, 98.58]

Table 14
HV statistical results of HeE-MOEA, KTA2, MCEA/D, and MOEA/D-FEF on multi-pass work roll cooling design problem.
Problem Stage HeE-MOEA KTA2 MCEA/D MOEA/D-FEF
p 1-2 8.8605e−1 (1.58e−2) − 9.2322e−1 (1.18e−2) − 9.1480e−1 (1.46e−2) − 9.4516e−1 (1.67e−2)
MPWRCD p 1-3 8.8389e−1 (1.40e−2) − 9.0731e−1 (1.12e−2) − 9.1025e−1 (9.52e−3) = 9.2657e−1 (1.99e−2)
p 1-4 8.7123e−1 (1.24e−2) − 8.9970e−1 (1.53e−2) − 8.9580e−1 (1.51e−2) − 9.1490e−1 (2.27e−2)
p 1-5 1.5243e−1 (4.55e−3) = 1.5780e−1 (5.09e−3) = 1.5811e−1 (5.36e−3) = 1.5952e−1 (5.25e−3)
+/-/= 0/3/1 0/3/1 0/2/2

solutions [45], is also employed in this study. The solutions obtained by [3] M.H. Aissa, T. Verstraete, Metamodel-assisted multidisciplinary design opti-
different SA-MOEAs in the second pass work roll are shown in Fig. 4, mization of a radial compressor, Int. J. Turbomach. Propuls. Power 4 (4)
(2019).
highlighting the superior convergence of MOEA/D-FEF compared to the
[4] Y. Jin, Surrogate-assisted evolutionary computation: Recent advances and future
other competitors. challenges, Swarm Evol. Comput. 1 (2) (2011) 61–70, http://dx.doi.org/10.
1016/j.swevo.2011.05.001, URL https://www.sciencedirect.com/science/article/
5. Conclusion pii/S2210650211000198.
[5] M.C. Zhou, M. Cui, D. Xu, S. Zhu, Z. Zhao, A. Abusorrah, Evolutionary opti-
mization methods for high-dimensional expensive problems: A survey, IEEE/CAA
This paper introduces a surrogate-assisted multi-objective evolution- J. Autom. Sin. 11 (5) (2024) 1092–1105, http://dx.doi.org/10.1109/JAS.2024.
ary algorithm with dimensionality reduction to address high- 124320.
dimensional expensive optimization problems. Firstly, we propose a [6] J. Knowles, ParEGO: a hybrid algorithm with on-line landscape approximation
for expensive multiobjective optimization problems, IEEE Trans. Evol. Comput.
dimensionality reduction framework containing three different feature
10 (1) (2006) 50–66, http://dx.doi.org/10.1109/TEVC.2005.851274.
extraction algorithms and a feature drift strategy to map the high- [7] Z. Wang, Q. Zhang, Y.-S. Ong, S. Yao, H. Liu, J. Luo, Choose appropriate
dimensional decision space into a low-dimensional decision space; this subproblems for collaborative modeling in expensive multiobjective optimization,
strategy helps to improve the robustness of surrogates. Secondly, we IEEE Trans. Cybern. 53 (1) (2023) 483–496, http://dx.doi.org/10.1109/TCYB.
2021.3126341.
propose a sub-region search strategy to define a series of promising
[8] Q. Zhang, W. Liu, E. Tsang, B. Virginas, Expensive multiobjective optimization
sub-regions in the high-dimensional decision space; this strategy helps by MOEA/D with Gaussian process model, IEEE Trans. Evol. Comput. 14 (3)
to improve the exploration ability of the proposed SA-MOEA. Experi- (2010) 456–474, http://dx.doi.org/10.1109/TEVC.2009.2033671.
mental results demonstrate the effectiveness of our proposed algorithm [9] G. Li, Q. Zhang, Multiple penalties and multiple local surrogates for expensive
in comparison to several state-of-the-art algorithms. constrained optimization, IEEE Trans. Evol. Comput. 25 (4) (2021) 769–778,
http://dx.doi.org/10.1109/TEVC.2021.3066606.
[10] Z. Song, H. Wang, C. He, Y. Jin, A kriging-assisted two-archive evolutionary
CRediT authorship contribution statement algorithm for expensive many-objective optimization, IEEE Trans. Evol. Comput.
25 (6) (2021) 1013–1027, http://dx.doi.org/10.1109/TEVC.2021.3073648.
[11] T. Chugh, Y. Jin, K. Miettinen, J. Hakanen, K. Sindhya, A surrogate-assisted
Zeyuan Yan: Writing – original draft, Methodology, Funding ac-
reference vector guided evolutionary algorithm for computationally expensive
quisition, Data curation. Yuren Zhou: Writing – original draft, Project many-objective optimization, IEEE Trans. Evol. Comput. 22 (1) (2018) 129–142,
administration. Wei Zheng: Resources, Formal analysis. Chupeng Su: http://dx.doi.org/10.1109/TEVC.2016.2622301.
Supervision, Software, Resources. Weigang Wu: Funding acquisition, [12] G. Li, Z. Wang, Q. Zhang, J. Sun, Offline and online objective reduction via
Gaussian mixture model clustering, IEEE Trans. Evol. Comput. 27 (2) (2023)
Data curation.
341–354, http://dx.doi.org/10.1109/TEVC.2022.3168836.
[13] B.E. Boser, I.M. Guyon, V.N. Vapnik, A training algorithm for optimal margin
Declaration of competing interest classifiers, in: Proceedings of the Fifth Annual Workshop on Computational
Learning Theory, COLT ’92, Association for Computing Machinery, New York,
NY, USA, 1992, pp. 144–152, http://dx.doi.org/10.1145/130385.130401.
The authors declare that they have no known competing finan-
[14] T. Cover, P. Hart, Nearest neighbor pattern classification, IEEE Trans. Inform.
cial interests or personal relationships that could have appeared to Theory 13 (1) (1967) 21–27, http://dx.doi.org/10.1109/TIT.1967.1053964.
influence the work reported in this paper. [15] D. Nelson, J. Wang, Introduction to artificial neural systems, Neurocomputing 4
(6) (1992) 328–330.
[16] D.G. Krige, A statistical approach to some mine valuation and allied problems on
Data availability the Witwatersrand: By DG Krige (Ph.D. thesis), University of the Witwatersrand,
1951.
No data was used for the research described in the article. [17] H. Wang, Y. Jin, A random forest-assisted evolutionary algorithm for data-driven
constrained multiobjective combinatorial optimization of trauma systems, IEEE
Trans. Cybern. 50 (2) (2020) 536–549, http://dx.doi.org/10.1109/TCYB.2018.
References 2869674.
[18] D. Broomhead, D. Lowe, Radial Basis Functions, Multi-Variable Functional
[1] C. Coello Coello, Evolutionary multi-objective optimization: a historical view of Interpolation and Adaptive Networks, Royal Signals and Radar Establishment
the field, IEEE Comput. Intell. Mag. 1 (1) (2006) 28–36, http://dx.doi.org/10. Malvern (United Kingdom), 1988, RSRE-MEMO-4148.
1109/MCI.2006.1597059. [19] Y. Shui, H. Li, J. Sun, Q. Zhang, The combination of MOEA/D and WOF
[2] C. He, Y. Tian, H. Wang, Y. Jin, A repository of real-world datasets for for solving high-dimensional expensive multiobjective optimization problems,
data-driven evolutionary multiobjective optimization, Complex Intell. Syst. (1) in: 2023 IEEE Congress on Evolutionary Computation, CEC, 2023, pp. 1–8,
(2020). http://dx.doi.org/10.1109/CEC53210.2023.10254083.

15
Z. Yan et al. Swarm and Evolutionary Computation 91 (2024) 101729

[20] Y. Liu, J. Liu, Y. Jin, Surrogate-assisted multipopulation particle swarm optimizer [33] Y. Horaguchi, M. Nakata, High-dimensional expensive optimization by
for high-dimensional expensive optimization, IEEE Trans. Syst. Man Cybern Syst. classification-based multiobjective evolutionary algorithm with dimensionality
52 (7) (2022) 4671–4684, http://dx.doi.org/10.1109/TSMC.2021.3102298. reduction, in: 2023 62nd Annual Conference of the Society of Instrument
[21] B. Liu, Q. Zhang, G.G.E. Gielen, A Gaussian process surrogate model assisted and Control Engineers, SICE, 2023, pp. 1535–1542, http://dx.doi.org/10.23919/
evolutionary algorithm for medium scale expensive optimization problems, IEEE SICE59929.2023.10354103.
Trans. Evol. Comput. 18 (2) (2014) 180–192, http://dx.doi.org/10.1109/TEVC. [34] S. Rahnamayan, S.J. Mousavirad, Towards solving large-scale expensive opti-
2013.2248012. mization problems efficiently using coordinate descent algorithm, in: 2020 IEEE
[22] D. Guo, Y. Jin, J. Ding, T. Chai, Heterogeneous ensemble-based infill crite- International Conference on Systems, Man, and Cybernetics, SMC, 2020, pp.
rion for evolutionary multiobjective optimization of expensive problems, IEEE 2506–2513, http://dx.doi.org/10.1109/SMC42975.2020.9283224.
Trans. Cybern. 49 (3) (2019) 1012–1025, http://dx.doi.org/10.1109/TCYB. [35] T. Cox, M. Cox, Multidimensional scaling, second edition, Chapman & Hall/CRC
2018.2794503. Monographs on Statistics & Applied Probability, CRC Press, 2000.
[23] T. Sonoda, M. Nakata, Multiple classifiers-assisted evolutionary algorithm based [36] J.B. Tenenbaum, V. de Silva, J.C. Langford, A global geometric framework for
on decomposition for high-dimensional multi-objective problems, IEEE Trans. nonlinear dimensionality reduction, Science 290 (5500) (2000) 2319–2323.
Evol. Comput. (2022) 1, http://dx.doi.org/10.1109/TEVC.2022.3159000. [37] X. He, P. Niyogi, Locality preserving projections, in: Neural Informa-
[24] D. Guo, X. Wang, K. Gao, Y. Jin, J. Ding, T. Chai, Evolutionary optimization of tion Processing Systems, 2003, URL https://api.semanticscholar.org/CorpusID:
high-dimensional multiobjective and many-objective expensive problems assisted 10726702.
by a dropout neural network, IEEE Trans. Syst. Man Cybern Syst. 52 (4) (2022) [38] M. Belkin, P. Niyogi, Laplacian eigenmaps for dimensionality reduction and data
2084–2097, http://dx.doi.org/10.1109/TSMC.2020.3044418. representation, Neural Comput. 15 (6) (2003) 1373–1396, http://dx.doi.org/10.
[25] A. Habib, H.K. Singh, T. Chugh, T. Ray, K. Miettinen, A multiple surrogate 1162/089976603321780317.
assisted decomposition-based evolutionary algorithm for expensive multi/many- [39] S. Huband, P. Hingston, L. Barone, L. While, A review of multiobjective test
objective optimization, IEEE Trans. Evol. Comput. 23 (6) (2019) 1000–1014, problems and a scalable test problem toolkit, IEEE Trans. Evol. Comput. 10 (5)
http://dx.doi.org/10.1109/TEVC.2019.2899030. (2006) 477–506, http://dx.doi.org/10.1109/TEVC.2005.861417.
[26] G. Li, Z. Wang, M. Gong, Expensive optimization via surrogate-assisted and [40] R. Cheng, et al., A benchmark test suite for evolutionary many-objective
model-free evolutionary optimization, IEEE Trans. Syst. Man Cybern Syst. 53 optimization, Complex Intell. Syst. 3 (1) (2017) 67–81.
(5) (2023) 2758–2769, http://dx.doi.org/10.1109/TSMC.2022.3219080. [41] P. Bosman, D. Thierens, The balance between proximity and diversity in
[27] H. Hao, A. Zhou, H. Qian, H. Zhang, Expensive multiobjective optimization multiobjective evolutionary algorithms, IEEE Trans. Evol. Comput. 7 (2) (2003)
by relation learning and prediction, IEEE Trans. Evol. Comput. 26 (5) (2022) 174–188, http://dx.doi.org/10.1109/TEVC.2003.810761.
1157–1170, http://dx.doi.org/10.1109/TEVC.2022.3152582. [42] O. Schutze, X. Esquivel, A. Lara, C.A.C. Coello, Using the averaged hausdorff
[28] K. Deb, L. Thiele, M. Laumanns, E. Zitzler, Scalable test problems for evolutionary distance as a performance measure in evolutionary multiobjective optimization,
multiobjective optimization, CEC ’02. Proc. 2002 Congr. 1 (2002) 825–830, IEEE Trans. Evol. Comput. 16 (4) (2012) 504–522, http://dx.doi.org/10.1109/
http://dx.doi.org/10.1109/CEC.2002.1007032. TEVC.2011.2161872.
[29] Q. Zhang, H. Li, MOEA/D: a multiobjective evolutionary algorithm based on [43] H. Gu, H. Wang, Y. Jin, Surrogate-assisted differential evolution with adap-
decomposition, IEEE Trans. Evol. Comput. 11 (6) (2007) 712–731, http://dx. tive multisubspace search for large-scale expensive optimization, IEEE Trans.
doi.org/10.1109/TEVC.2007.892759. Evol. Comput. 27 (6) (2023) 1765–1779, http://dx.doi.org/10.1109/TEVC.2022.
[30] Q. Lin, X. Wu, L. Ma, J. Li, M. Gong, C.A.C. Coello, An ensemble surrogate-based 3226837.
framework for expensive multiobjective evolutionary optimization, IEEE Trans. [44] T. Sonoda, M. Nakata, Multiple classifiers-assisted evolutionary algorithm based
Evol. Comput. 26 (4) (2022) 631–645, http://dx.doi.org/10.1109/TEVC.2021. on decomposition for high-dimensional multi-objective problems, IEEE Trans.
3103936. Evol. Comput. (2022) 1, http://dx.doi.org/10.1109/TEVC.2022.3159000.
[31] M. Zhao, K. Zhang, G. Chen, X. Zhao, C. Yao, H. Sun, Z. Huang, J. Yao, [45] Y. Azene, Work Roll System Optimisation Using Thermal Analysis and Genetic
A surrogate-assisted multi-objective evolutionary algorithm with dimension- Algorithm (Ph.D. thesis), Cranfield University, Cranfield, U.K., 2011.
reduction for production optimization, J. Pet. Sci. Eng. 192 (2020) 107192, [46] V. Palakonda, J.-M. Kang, H. Jung, Benchmarking real-world many-objective
http://dx.doi.org/10.1016/j.petrol.2020.107192. problems: A problem suite with baseline results, IEEE Access 12 (2024)
[32] j. Li, C. He, R. Cheng, Adaptive dropout for high-dimensional expensive multiob- 49275–49290, http://dx.doi.org/10.1109/ACCESS.2024.3383916.
jective optimization, in: Complex & Intelligent Systems, (8) 2021, pp. 271–285, [47] E. Zitzler, L. Thiele, Multiobjective evolutionary algorithms: a comparative case
http://dx.doi.org/10.1007/s40747-021-00362-5. study and the strength Pareto approach, IEEE Trans. Evol. Comput. 3 (4) (1999)
257–271, http://dx.doi.org/10.1109/4235.797969.

16