Expert Systems With Applications 265 (2025) 125908

Contents lists available at ScienceDirect

Expert Systems With Applications

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

Competition of tribes and cooperation of members algorithm: An

evolutionary computation approach for model free optimization
Zuyan Chen a , Shuai Li a,b ,∗, Ameer Tamoor Khan c , Seyedali Mirjalili d,e
a Faculty of Information Technology and Electrical Engineering, University of Oulu, 90570, Oulu, Finland
b
VTT-Technical Research Centre of Finland, 90590, Oulu, Finland
c
Department of Plant and Environmental Science, University of Copenhagen, Copenhagen, 1172, Denmark
d
Centre for Artificial Intelligence Research and Optimization, Torrens University Australia, Brisbane, QLD 4006, Australia
e
University Research and Innovation Center, Obuda University, Budapest, 1034, Hungary

ARTICLE INFO ABSTRACT

Keywords: Metaheuristic algorithms solve optimization problems mostly by imitating behaviors observed in nature. Over
Evolutionary computation time, these algorithms have proven to be very effective in solving complex optimization problems. Due to
Competition of tribes and cooperation of the rising complexity and scale of practical engineering problems, numerous metaheuristic algorithms have
members algorithm
been developed recently and applied in various fields. In response to this need, researchers continue to
Meta-heuristics
explore novel approaches inspired by natural and social phenomena. Inspired by the competition among
Swarm intelligence
Engineering optimization
ancient tribes and their cooperative behavior, this paper proposes a meta-heuristic called the Competition
Optimization of Tribes and Cooperation of Members Algorithm (CTCM). Experiments are conducted on 23 benchmark
Algorithm test functions and comprehensively compared with other state-of-the-art algorithms, including particle swarm
optimization (PSO), grey wolf optimizer (GWO), sparrow search algorithm (SSA), egret swarm optimization
(ESOA), beetle antennae search (BAS) and whale optimization (WOA). The standard deviation and average, as
well as statistical tests are utilized to compare the performance of each algorithm, which demonstrates that
CTCM is superior in the majority of problems. In addition, the results of Wilcoxon and Friedman rank tests
show that the CTCM achieves the first place in all categories of problems. The results indicate that CTCM
possesses strong global optimization search capability and stability, and has faster convergence speed. The
paper also considers solving practical engineering optimization problems as proof-of-concept case studies, in
which CTCM achieves all the optimal solutions for each engineering problem.

1. Introduction Optimization (PSO) algorithm to simulate the behavior of a flock of

birds foraging for food (Kennedy & Eberhart, 2002), which became
Swarm Intelligence (SI) is a distributed computational and opti- one of the most popular swarm intelligence algorithms and is widely
mization methodology inspired by the collective behavior of natural used in various optimization problems. Although Genetic Algorithms
systems, such as bird flocks, fish schools, and ant colonies. It simulates was proposed by Holland as early as 1975 (Holland, 1992), it has
these behaviors to solve complex problems. In the 1980s, biologists been studied and applied together with other population intelligence
began to study group behavior, such as the collective movements of algorithms in the development of population intelligence, forming a se-
flocks of birds and swarms of fish. In 1986, Reynolds proposed the Boids ries of optimization methods based on biological evolution and natural
model (Reynolds, 1987), which simulated the collective behavior of selection.
flocks of birds, which laid the groundwork for the study of group intelli-
From the 2000s to the 2010s, swarm intelligence algorithms were
gence. In 1991, Dorigo proposed the Ant Colony Optimization (Dorigo,
further extended and diversified. In 2005, Karaboga proposed the Ar-
Birattari, & Stutzle, 2006), ACO, to simulate the behavior of ants
tificial Bee Colony (ABC) algorithm, which simulated the behavior of
by marking paths with pheromones while searching for food. This is
honeybees for foraging (Karaboga & Basturk, 2008). This algorithm is
an important milestone in the application of colony intelligence to
widely used in function optimization and multi-objective optimization
optimization problems.
During the 1990s and 2000s, swarm intelligence was rapidly de- problems. In 2008, Yang proposed Firefly Algorithm, which simulates
veloped. In 1995, Kennedy and Eberhart proposed the Particle Swarm the flashing behavior of fireflies, and is used to solve both continuous

∗ Corresponding author at: Faculty of Information Technology and Electrical Engineering, University of Oulu, 90570, Oulu, Finland.
E-mail addresses: [email protected] (Z. Chen), [email protected] (S. Li).

https://doi.org/10.1016/j.eswa.2024.125908
Received 1 August 2024; Received in revised form 20 October 2024; Accepted 22 November 2024
Available online 30 November 2024
0957-4174/© 2024 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

Table 1
Algorithm abbreviations utilized in this paper.
Algorithm Abbreviation
Competition of Tribes and Cooperation of Members Algorithm CTCM
Particle Swarm Optimization PSO (Kennedy & Eberhart, 2002)
Grey Wolf Optimizer GWO (Mirjalili, Mirjalili, & Lewis, 2014)
Sparrow Search Algorithm SSA (Xue & Shen, 2020)
Egret Swarm Optimization Algorithm ESOA (Chen et al., 2022)
Beetle Antennae Search BAS (Jiang & Li, 2017)
Whale Optimization Algorithm WOA (Mirjalili & Lewis, 2016)

and discrete optimization problems (Yang, 2010). In 2010, Yang pro-

posed Bat Algorithm, which is based on the echolocation behavior of
bats based on the echolocation behavior of bats, which is applied to
global optimization problems (Yang & Hossein Gandomi, 2012).
In the 2010s to the present, group intelligence has seen a new
phase of modern development. Researchers have begun to focus on
multi-intelligent systems and hybrid algorithms that combine multi-
ple population intelligence algorithms to improve the performance
and adaptability of the algorithms (Rostami, Berahmand, Nasiri, &
Forouzandeh, 2021; Tang, Liu, & Pan, 2021). With the development
of big data and machine learning, swarm intelligence algorithms are
applied to feature selection, classification, clustering and other fields,
showing powerful optimization capabilities (Tang, Duan, & Lao, 2023).
For instance, particle swarm optimization algorithms are used for fea-
Fig. 1. The principle of CTCM. The living environment on the earth was extremely
ture selection to improve the accuracy and training speed of machine tough thousands years ago. Humans were organized in tribes to struggle for survivals,
learning models, while artificial bee colony algorithms are used for data i.e., searching for the most favorable place to settle down for living. CTCM mimic
clustering to efficiently handle large-scale datasets (Chuang, Chang, important rules in these human early behaviors, namely, Cooperation: Inside a tribe,
they cooperate; Competition: Between tribes, they fight; Migration: Tribes migrate from
Tu, & Yang, 2008; Huang & Dun, 2008; Karaboga & Ozturk, 2011;
place to place; Conflict: When two tribes move to the same place, conflicts happen.
Zhang & Dolg, 2015). Through these practical applications, population Weaker tribes escape from the place after conflicts while stronger one is not affected.
intelligence algorithms demonstrate their great potential in dealing The above behaviors help humans find the best living places.
with complex system optimization problems (see Table 1).
Intelligent algorithms have shown significant advantages in solving
complex optimization problems. Meta heuristic algorithms are also
developed continuously for different types of problems. To address
widely utilized in applications with particularly complicate real-world the above mentioned limitations of current intelligent algorithms, this
situations, such as nonlinear control systems, object detection, power paper proposes an intelligent algorithm, Competition of Tribes and
grid network, as well as imbalanced data problems, due to their Cooperation of Members(CTCM) Algorithm, inspired by ancient human
excellent optimal-seeking performance (Altbawi, Khalid et al., 2023; tribal competitive and cooperative behaviors. In experimental results,
Altbawi et al., 2024, 2021; Rajwar, Deep, & Das, 2023). In single-stage the proposed algorithm presents excellent global optimization ability
modular PV-Grid-Driven systems, gradient optimization algorithm is and stability with fast convergence speed.
used to regulate the hybrid DC voltage in order to reduce the energy The following paper is organized as follows; Section 2 will present
losses (Altbawi, Mokhtar et al., 2023). In optimizing the charging the algorithmic background, mathematical model and framework of
efficacy of the charging station, african vulture optimization algorithm algorithm. Section 3 is the experimental part where the proposed
is utilized to regulate the parameters of its PID and TID controllers to algorithm will be compared in detail with other state-of-the-art algo-
keep the power system running stably (Altbawi et al., 2023). The fraud rithms. The proposed algorithm will be applied to two real engineering
detection problem is considered to be complex with high-dimensional optimization problems in Section 4. Finally, the summary and future
non-linearity and uneven data, but it could be well solved with the work will be illustrated.
combination of machine learning and heuristic algorithms (Pan, Hu,
Snášel, & Chu, 2023; Yi et al., 2023). 2. Competition of Tribes and Cooperation of Members algorithm
However, these algorithms also have inherent limitations accord- (CTCM)
ing to the No-Free-Lunch (NFL) theorem, which states that for any
optimization algorithm, its performance is, on average, the same in The inspiration of CTCM will be discussed in this section. Mean-
all possible problem spaces; in other words, no algorithm performs while, the mathematical model of CTCM will be established in follow-
best on all problems. . This theory reveals the limitations of intelli- ing (see Fig. 1).
gent algorithms, including problem-specific dependencies, parameter
sensitivity, local optimal traps, and computational complexity (Wolpert 2.1. Anthropological characteristics
& Macready, 1997). Intelligent algorithms typically perform well on
specific types of problems, but may not perform well on other types In ancient early societies, numerous tribes migrated continuously
of problems. In addition, although intelligent algorithms have some in search of the most resource-rich lands, seeking areas most suitable
global search capability, they may still be trapped in local optima and for survival (Lindner, 1982). Each tribe typically had a chieftain, the
difficult to jump out in complex multi-peak optimization problems. most authoritative individual within the group, whose position could
As the problem size increases, the computational complexity of the change dynamically based on the individual’s capabilities (Cameron,
intelligent algorithm also increases significantly, leading to a significant 2013). This mechanism is akin to how individuals in human societies
extension of the computation time, which is especially prominent when rise to leadership roles due to their resource acquisition skills and
dealing with large-scale data or real-time applications. Therefore, in adaptability (Mathew, 2024).
order to improve the optimization efficiency and the adaptation of The competition among tribes was intense, with the strength of a
various situation, new intelligent algorithms need to be proposed and chieftain directly reflecting the tribe’s power (Kradin, 2018). Stronger

2
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

tribes did not need to submit to weaker ones. When a weaker tribe
encountered a stronger one in conflict, it was often forced to migrate
due to the repulsion exerted by the stronger tribe, which remained
unaffected (Lindner, 1982). These tribal conflicts occurred randomly,
with opposing tribes selected at random (Cameron, 2013). Within a
tribe, members cooperated to pursue common goals and directions,
acting primarily under the chieftain’s leadership and their own ex-
periences (Mathew, 2024). Additionally, weaker tribes required high
flexibility to quickly distance themselves from stronger tribes during
conflicts (Puurtinen & Mappes, 2009).
This migration behavior, driven by resource scarcity and tribal con-
flicts, was particularly evident among nomadic peoples (Kradin, 2018;
Lindner, 1982). The dynamic selection of chieftains based on their
abilities also mirrors leadership selection processes in reality (Cameron,
2013; Efferson, Bernhard, Fischbacher, & Fehr, 2024). Tribal competi- Fig. 2. The principle of cooperation of members. There is usually a chief in a tribe
tion and conflict interactions reflect theories of inter-group conflict and who is responsible for managing and planning the subsequent development of the tribe.
In this process, tribal members usually follow the chief’s planning and instructions.
cooperation, extensively validated in sociological studies (Puurtinen &
But each member also has his own individual mind, and so the loyalty to the chief
Mappes, 2009). The internal cooperation among tribe members repre- changes over time. The loyalty of each member is chaotic over time and becomes more
sents the collaborative behaviors observed in human societies, driven pronounced as the number of tribes grows.
by common goals and social cohesion (Efferson et al., 2024).

2.2. Mathematical model and algorithm in which 𝑑 presents the dimension of solution space. Then the fitness
value 𝐹𝑋 of the whole primitive society can be expressed as,
According to previous study on the cooperative and competitive ⎡𝐹1 ⎤ ⎡ 𝐹𝑛,1 ⎤
relationships among tribes and members, we could construct a math- ⎢ ⎥ ⎢ ⎥
⎢𝐹2 ⎥ 𝑝×1 ⎢ 𝐹𝑛,2 ⎥ 𝑚×1
ematical model of competition of tribes and cooperation of members 𝐹 = ⎢ ⎥ ∈ R , 𝐹𝑛 = ⎢ ⎥∈R (2)
algorithm (CTCM). ⎢ ... ⎥ ⎢ ... ⎥
⎢ ⎥ ⎢ ⎥
⎣ 𝐹𝑛 ⎦ ⎣𝐹𝑛,𝑚 ⎦
• In order to survive, members of primitive human societies would
Each individual possesses a velocity vector that can determine
randomly cooperate with other members, and in turn gather to
where the individual goes next. The overall velocity matrix 𝑉 of a
form a tribe. Each tribe was located in a different place, occupied
primitive human society is represented as follows.
different resources, and often migrated due to the demand of
obtaining more resources. ⎡𝑉1 ⎤
⎢ ⎥
• In every tribe there exists a chief who occupies the most rich ⎢𝑉2 ⎥
resource part of the tribe. When tribes migrate, the chief plays 𝑉 = ⎢ ⎥ ∈ R𝑝×𝑑 ,
⎢ ... ⎥
a key role and largely influences the direction of the tribe’s ⎢ ⎥
migration. ⎣ 𝑉𝑛 ⎦
(3)
• Each tribe member will explore to varying degrees based on their ⎡ 𝑉𝑛,1,1 𝑉𝑛,1,2 ... 𝑉𝑛,1,𝑑 ⎤
⎢ ⎥
previous experience, as well as at the command of the chief, in ⎢ 𝑉𝑛,2,1 𝑉𝑛,2,2 ... 𝑉𝑛,2,𝑑 ⎥
𝑉𝑛 = ⎢ 𝑚×𝑑
order to find land that is more fertile in terms of resources. ⎥∈R
⎢ ... ... ... ... ⎥
• Each tribal member, as a separate individual, has the ability to ⎢ ⎥
think independently. This is reflected in the fact that when a tribal ⎣𝑉𝑛,𝑚,1 𝑉𝑛,𝑚,2 ... 𝑉𝑛,𝑚,𝑑 ⎦
member undertakes a migration, she or he is more likely to defer
2.3. Cooperation of members (exploitation)
to her or his own experience or to the instruction of the chief.
• Although the members of the tribe work closely together, they
can replace the current chief once they find more fertile land. The principle of cooperation of members is illustrated in Fig. 2.
• Tribes are competitive and conflicts between two tribes occur In a tribe, the chief is typically responsible for managing and plan-
randomly. In order to avoid greater losses to the tribe, the weaker ning the future development of the tribe. Most members follow the
chief’s plans and instructions, but each member also has their own
tribe will flee in the opposite direction.
individual thoughts, causing their loyalty to fluctuate over time. This
Based on the above characteristics of ancient tribes searching for loyalty exhibits chaotic behavior over time, and as the number of
habitats, the mathematical model of CTCM is constructed as follows. tribes increases, this chaotic nature becomes more pronounced. In this
Suppose the quantity of humans in a primitive society is 𝑝, the number context, the concept of sine chaotic mapping is utilized to characterize
if tribes is 𝑛, and there are 𝑚 members in one tribe. Then the initial the variation of members’ loyalties, while acknowledging the influence
position 𝑋 of the whole society can be written as, of tribal numbers 𝑛 on the chaos state, and the loyalty of individual
⎡𝑋1 ⎤ member 𝑟𝑡 could be described as,
⎢ ⎥ {
⎢𝑋2 ⎥ 𝑈 (0, 1) 𝑖𝑓 𝑡 = 𝑘 ⋅ 𝑚, 𝑘 ∈ Z
𝑋 = ⎢ ⎥ ∈ R𝑝×𝑑 , 𝑟𝑡 = (4)
sin( 𝜋2 𝑟𝑡−1 ) 𝑒𝑙𝑠𝑒
⎢ ... ⎥
⎢ ⎥ The two components of random reassignment and chaotic evolution
⎣𝑋𝑛 ⎦
(1) are combined in Eq. (4) to form a mixed model of loyalty that reflects
⎡ 𝑋𝑛,1,1 𝑋𝑛,1,2 ... 𝑋𝑛,1,𝑑 ⎤
⎢ ⎥ displayed social behavior. The chaotic sinusoidal mapping captures the
⎢ 𝑋𝑛,2,1 𝑋𝑛,2,2 ... 𝑋𝑛,2,𝑑 ⎥ 𝑚×𝑑 unpredictable nature of how individual members adjust their loyalties
𝑋𝑛 = ⎢ ⎥∈R
⎢ ... ... ... ... ⎥ in response to prior values, while the random reassignments represent
⎢ ⎥ significant, abrupt shifts in attitudes that may occur at intervals. As
⎣𝑋𝑛,𝑚,1 𝑋𝑛,𝑚,2 ... 𝑋𝑛,𝑚,𝑑 ⎦

3
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

𝐼𝑟𝑖𝑣𝑎𝑙 = 𝑟𝑎𝑛𝑑(𝑛) denotes a random conflict with a tribe from among

𝑛 tribes. And the velocity will update as,
𝑅𝑡𝑛,𝑚,𝑑 = 𝑐3 𝑟𝑡3,𝑛,𝑚,𝑑 (𝑋𝑛,𝑚,𝑑
𝑡 𝑏𝑒𝑠𝑡
− 𝑋𝑟𝑖𝑣𝑎𝑙 )
𝑡+1
𝑉𝑛,𝑚,𝑑
{ 𝑡+1
(6)
𝑉𝑛,𝑚,𝑑 − 𝑅𝑡𝑛,𝑚,𝑑 , 𝑖𝑓 𝑏𝑒𝑠𝑡
𝐹𝑛𝑏𝑒𝑠𝑡 < 𝐹𝑟𝑖𝑣𝑎𝑙
= 𝑡+1
𝑉𝑛,𝑚,𝑑 , 𝑒𝑙𝑠𝑒

𝑏𝑒𝑠𝑡 is the best position found by rival. 𝐹 𝑏𝑒𝑠𝑡 is the best fitness
where 𝑋𝑟𝑖𝑣𝑎𝑙 𝑛
𝑏𝑒𝑠𝑡 presents the best fitness of rival. 𝑐 indicates
of 𝑛th tribe while 𝐹𝑟𝑖𝑣𝑎𝑙 3
the tribe escape coefficient and 𝑟𝑡3,𝑛,𝑚,𝑑 is a chaotic random factor
that reveals the retreat speed. Meanwhile the locations of the tribal
members are updated as follows.
𝑡+1 𝑡 𝑡+1
𝑋𝑛,𝑚,𝑑 = 𝑋𝑛,𝑚,𝑑 + 𝑉𝑛,𝑚,𝑑 (7)

𝑡+1
If the updated position of a tribal member 𝑋𝑛,𝑚,𝑑 exceeds the feasible
domain range [𝑋𝑚𝑖𝑛 , 𝑋𝑚𝑎𝑥 ], where 𝑋𝑚𝑖𝑛 is the minimal value of domain
and 𝑋𝑚𝑎𝑥 means the maximum value of supposed domain. For this
instance, a mirror bounce will occurs, and the velocity is reversed and
Fig. 3. The concept of competition of tribes. Conflicts usually break out between
its position information is corrected.
tribes due to competition for resources, and the weaker tribes usually retreat elsewhere
because they cannot compete. In the case of defeat, the retreat is usually chaotic, with 𝑡+1 𝑡+1
𝑉𝑛,𝑚,𝑑 = −𝑉𝑛,𝑚,𝑑 (8)
some tribesmen retreating faster due to psychological fear, and some retreating slower.
The larger the number of tribes, the more unpredictable this retreat can be, as the
meanwhile the position of the tribe member will be corrected. The
defeated tribe will fear conflict with the rest tribes.
mathematical model of CTCM is constructed and its pseudo code is
shown in Algorithm 1.
the number of tribes 𝑛 increases, the complexity and connections of
different tribes can amplify the chaotic behavior of loyalty. More Algorithm 1 The algorithm of CTCM, MATLAB Code (Chen, 2024)
interactions and connections mean more opportunities for the value of Input:
𝑟𝑡 to deviate or fluctuate, making the whole system more sensitive to 𝐺: the maximum iterations
initial conditions. 𝑝: the number of humans
Each individual would communicate with the chief in the tribe to 𝑛: the quantity of tribes
get his or her instructions on what to do next. It will also incorporate 𝑐1 : The experience factor
its own experiences to provide additional sources of information for its 𝑐2 : The obey factor
next move. The specific form of updating the velocity matrix is similar 𝑐3 : The escape factor
to the particle swarm algorithm (Kennedy & Eberhart, 1995). There Initialize a population of 𝑝 humans and assign the relevant
is a certain amount of randomness because each individual possesses parameters
a different mind, with different procedures for listening to the chief’s Output: 𝑋𝑏𝑒𝑠𝑡 , 𝐹𝑏𝑒𝑠𝑡
instructions and following his or her own thoughts, which could be while t<G do
expressed as. for Each human do
𝑡+1 3 𝑡
𝑉𝑛,𝑚,𝑑 = 5
⋅ 𝑉𝑛,𝑚,𝑑 +𝑐1 𝑟𝑡1,𝑛,𝑚,𝑑 (𝑋𝑛,𝑚,𝑑
𝑏𝑒𝑠𝑡 − 𝑋 𝑡
𝑛,𝑚,𝑑
) update loyalty factor 𝑟1,𝑛,𝑚,𝑑 , 𝑟2,𝑛,𝑚,𝑑 by Eq. (4)
(5) 𝑡
update velocity 𝑉𝑛,𝑚,𝑑 by Eq. (5)
+ 𝑐2 𝑟𝑡2,𝑛,𝑚,𝑑 (𝑋𝑡𝑟𝑖𝑏𝑒
𝑏𝑒𝑠𝑡 − 𝑋 𝑡
𝑛,𝑚,𝑑
)
3 𝑡 𝑏𝑒𝑠𝑡
5
⋅ 𝑉𝑛,𝑚,𝑑 + 𝑐1 𝑟1,𝑛,𝑚,𝑑 (𝑋𝑛,𝑚,𝑑
𝑡+1
in which 𝑉𝑛,𝑚,𝑑 means the velocity of 𝑚th member from 𝑛th tribe at 𝑡+ 1 𝑡
−𝑋𝑛,𝑚,𝑑 𝑏𝑒𝑠𝑡 − 𝑋 𝑡
) + 𝑐2 𝑟2,𝑛,𝑚,𝑑 (𝑋𝑡𝑟𝑖𝑏𝑒 )
𝑛,𝑚,𝑑
times, and the constant 35 is the inertia factor while 𝑉𝑛,𝑚,𝑑
𝑡 is the velocity randomly select rival and update 𝑉𝑛,𝑚,𝑑 𝑡 by Eq. (6)
at 𝑡 times. 𝑋𝑛,𝑚,𝑑𝑏𝑒𝑠𝑡 indicates the position of best fitness found by this
𝑏𝑒𝑠𝑡 then
if 𝐹𝑛𝑏𝑒𝑠𝑡 < 𝐹𝑟𝑖𝑣𝑎𝑙
member in the whole period, and 𝑋𝑛,𝑚,𝑑 𝑡 presents the position at 𝑡 times. update retreat factor by Eq. (4)
𝑏𝑒𝑠𝑡
𝑋𝑡𝑟𝑖𝑏𝑒 is the position with best fitness found by the tribe in the whole 𝑡+1 𝑡+1 𝑡 𝑏𝑒𝑠𝑡 )
𝑉𝑛,𝑚,𝑑 = 𝑉𝑛,𝑚,𝑑 − 𝑐3 𝑟3,𝑛,𝑚,𝑑 (𝑋𝑛,𝑚,𝑑 − 𝑋𝑟𝑖𝑣𝑎𝑙
period. 𝑐1 and 𝑐2 represent procedures for tribal members to follow 𝑡+1
their own experiences and to obey the chief’s instructions respectively. update human position 𝑋𝑛,𝑚,𝑑 by Eq. (7)
𝑡+1 𝑡 𝑡+1
𝑟𝑡1,𝑛,𝑚,𝑑 and 𝑟𝑡2,𝑛,𝑚,𝑑 are the chaotic loyalty of each individual member. 𝑋𝑛,𝑚,𝑑 = 𝑋𝑛,𝑚,𝑑 + 𝑉𝑛,𝑚,𝑑
𝑡+1
if 𝑋𝑛,𝑚,𝑑 is out of bound then
2.4. Competition of tribes (exploration) 𝑡+1
𝑉𝑛,𝑚,𝑑 𝑡+1
= −𝑉𝑛,𝑚,𝑑
compute the fitness 𝐹
The concept of competition of tribes is indicated in Fig. 3. Conflicts 𝑏𝑒𝑠𝑡 then
if if 𝐹 < 𝐹𝑡𝑟𝑖𝑏𝑒
between tribes often arise due to competition for resources, and weaker 𝑏𝑒𝑠𝑡 = 𝑋
𝑋𝑡𝑟𝑖𝑏𝑒
tribes are typically forced to retreat as they struggle to compete. When 𝑏𝑒𝑠𝑡 = 𝐹
𝐹𝑡𝑟𝑖𝑏𝑒
faced with defeat, the retreat tends to be chaotic: some members of
tribe flee faster due to psychological fear, while others retreat more if if 𝐹 < 𝐹𝑏𝑒𝑠𝑡 then
slowly. As the number of tribes increases, the unpredictability of these 𝑋 𝑏𝑒𝑠𝑡 = 𝑋
retreats also grows, as the defeated tribe fears further conflict with 𝐹 𝑏𝑒𝑠𝑡 = 𝐹
other remaining tribes. Retrieve the current position
Random clashes can occur between tribes, at which point the 𝑡=𝑡+1
return 𝑋𝑏𝑒𝑠𝑡 , 𝐹𝑏𝑒𝑠𝑡
weaker tribe will flee, while the stronger tribe will be unaffected.

4
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

Fig. 4. The flowchart of CTCM.

And the flow chart of CTCM is described in Fig. 4. the Eq. (9) could be rewritten as,
3
2.5. Computational complexity 𝑉 𝑘+1 = ⋅ 𝑉 𝑘 − (𝑐1 𝑟1,𝑘 + 𝑐2 𝑟2,𝑘 − 𝑐3 𝑟3,𝑘 )𝑠𝑘
5 (12)
3
Computational complexity is a convenient and effective tool for es- = ⋅ 𝑉 𝑘 − 𝑐𝑘 𝑠𝑘
5
timating the consumption of computational resources of an algorithm,
Where 𝑐𝑘 is defined as 𝑐1 𝑟1,𝑘 + 𝑐2 𝑟2,𝑘 − 𝑐3 𝑟3,𝑘 . To eliminate the term of
and it could assist to guide the application of algorithms. The time
velocity, let
complexity of CTCM is first discussed here, which is generally divided
into three procedures, namely initialization, fitness evaluation and state 𝑠𝑘+1 − 𝑠𝑘 = 𝑥𝑘+1 − 𝑥𝑘
(13)
update. Assuming 𝑃 as the number of the whole population of CTCM, = 𝑉 𝑘+1
𝐷 as the dimensions of the problem, and 𝑇 as the maximum iteration of
substituting Eq. (13) into Eq. (12), then
the whole process. The complexity of initialization, fitness evaluation,
and state update are 𝑂(𝑃 × 𝐷), 𝑂(𝑃 ), and 𝑂(𝑃 × 𝐷) respectively. If 3
𝑠𝑘+1 = ⋅ (𝑠𝑘 − 𝑠𝑘−1 ) + (1 − 𝑐𝑘 )𝑠𝑘
the maximum number of iterations is considered, the overall time 5 (14)
complexity of CTCM is 𝑂(𝑇 × 𝑃 × 𝐷). The space complexity refers = (1.6 − 𝑐 𝑘 )𝑠𝑘 − 0.6 ⋅ 𝑠𝑘−1
to the computer storage space occupied by the estimation algorithm hence the characteristic root equation of the difference equation is
during the computation process. As each member of the CTCM needs obtained,
to store the individual optimal position, speed and position, which
occupy space complexity 𝑂(𝐷), 𝑂(𝑃 × 𝐷), and, 𝑂(𝑃 × 𝐷) respectively. 𝜆2 − (1.6 − 𝑐𝑘 )𝜆 + 0.6 = 0 (15)
Thurs, the total space complexity of CTCM is 𝑂(𝑃 × 𝐷), which means
and assuming 𝑐𝑘 to be constant number, then the solution of character-
the relationship between space complexity and population as well as
istic root equation should be,
problem dimension is linear. √
(1.6 − 𝑐𝑘 ) ± (1.6 − 𝑐𝑘 )2 − 2.4
2.6. Convergence analysis 𝜆1,2 = (16)
2
in order to ensure that the CTCM can be in a steady state, the mode of
In order to verify the convergence of the CTCM algorithm, the 2
the characteristic root
√ must be less than 1, so that there is, (1.6 − 𝑐𝑘 ) <
theoretical convergence analysis is described in detail. Meanwhile, the
2.4 and |1.6 − 𝑐𝑘 | < 2.4, hence the 𝑐𝑘 ∈ (0, 3.15). Because 𝑟𝑗 ,𝑘 ∈ [0, 1],
effective theoretical analysis can further guide the design of application
let us consider only the extreme cases. If 𝑟1,𝑘 = 1, 𝑟2,𝑘 = 1 and
parameters of CTCM. Assuming 𝑝∗ as the best solution of individual, 𝑔 ∗
𝑟3,𝑘 = 0 then 0 < 𝑐1 + 𝑐2 < 3.15. 𝑐3 is here an instability perturbation
as the best solution of population and 𝑥∗ as the global optimal point.
term; if 𝑐3 is greater than 0, the system will exhibit instability at some
And in order to be able to carry out the analysis, it is assumed that
𝑝∗ = 𝑔 ∗ = 𝑥∗ during the convergence process. Hence the CTCM speed point, which in turn helps the algorithm to jump out of the local traps.
update equation could be written as, As just mentioned, if 𝑐3 > (𝑐1 𝑟1 + 𝑐2 𝑟2 − 3.15)∕𝑟3 , then the system
increases the perturbation at this stage. So the parameter tuning should
3
𝑉 𝑘+1 = ⋅ 𝑉 𝑘 + 𝑐1 𝑟1,𝑘 (𝑝∗ − 𝑥𝑘 ) + 𝑐2 𝑟2,𝑘 (𝑔 ∗ − 𝑥𝑘 ) be done in such a way that 𝑐1 and 𝑐2 satisfy the condition first of being
5
less than 3.15, and then 𝑐3 starts from a very small positive number
+ 𝑐3 𝑟3,𝑘 (𝑥𝑘 − 𝑥∗ ) (9)
and increases gradually to obtain the most desired global optimization
3
= ⋅ 𝑉 𝑘 + (𝑐1 𝑟1,𝑘 + 𝑐2 𝑟2,𝑘 − 𝑐3 𝑟3,𝑘 ) ∗ (𝑥∗ − 𝑥𝑘 ) seeking capability for different applications.
5
where 𝑉 𝑘+1 describe a individual speed at 𝑘 + 1 iteration, and 2.7. Sensitivity analysis
𝜋
𝑟𝑗 ,𝑘 = sin( 𝑟𝑗 ,𝑘−1 ), 𝑗 ∈ {1, 2, 3} (10)
2 Since there are several important parameters in CTCM, it is essential
Define 𝑠𝑘 as the position offset between the individual and optimal to be tuned before comparing with other algorithms. In order to adapt
solution at 𝑘 iteration, then the CTCM to as various scenarios as possible, seven different cate-
𝑠𝑘 = 𝑥𝑘 − 𝑥∗ (11) gories of benchmark functions (𝑓3 , 𝑓5 , 𝑓7 , 𝑓9 , 𝑓11 , 𝑓18 , 𝑓23 ) are utilized

5
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

for the sensitivity analysis, covering Unimodal, Multimodal and Fixed- 3.1. Unimodal test function
Dimension situations. The sensitivity result is indicated in Table 2,
and the parameters that could be better adapted to each situation are Unimodal test functions are primarily used to evaluate the con-
𝑚 = 20, 𝑐1 = 2.0, 𝑐2 = 1.0 and 𝑐3 = 0.1. In this case, the CTCM is more vergence properties and exploitation capabilities of optimization al-
capable of solving Unimodal functions as well as complex situations, gorithms. These functions, characterized by a single global optimum
and also responds better to Multimodal problems. without any local optima, are valuable for assessing an algorithm’s ef-
In detail, for instance, when 𝑐1 = 2 and 𝑐2 = 1, when 𝑚 = 5, it ficiency in converging to the global optimum. They force the algorithm
to focus on exploitation, measuring its ability to accurately locate and
can be found that the performance of CTCM to solve the problem is
refine the search around the global optimum. By using unimodal test
improved if 𝑐3 is gradually increased, however, the performance on
functions, the stability, reliability, and convergence speed of an algo-
𝑓9 is first weakened and then strengthened. However, when 𝑚 = 10,
rithm under various conditions could be observed, providing insights
the performance of CTCM on 𝑓1 and 𝑓9 gradually weakened, while it
into its internal mechanisms and potential limitations.
gradually got better on the remaining problems. When 𝑚 = 20, CTCM’s
performance is more balanced, and when 𝑐3 is progressively larger, 3.1.1. Analysis of the convergence accuracy
CTCM’s performance is more average at 𝑐3 = 0.1, and its performance is Test functions 1-7 are unimodal test functions and the performance
best on the Unimodal and fixed-dimension problems, and more average of each algorithm can be found in Table 20. It can be noticed that on
on the multimodal. The sensitivity analysis reveals that CTCM performs the 𝐹5 test function, SSA finds the better solution. And on the 𝐹7 test
a bit better on the whole problem when the number of tribal members 𝑚 function PSO finds a better solution, but CTCM comes in second place in
is boosted, which indicates that the more the number of tribal members terms of performance. However for most of the unimodal test functions
is, the better the convergence of CTCM is. However, in the case of CTCM has the best performance, which illustrates that CTCM has good
𝑐1 + 𝑐2 = 4, CTCM will show greater instability, which is consistent with exploitation for unimodal cases.
the convergence analysis results in the previous subsection. And from
the convergence analysis, it can be concluded that 𝑐3 is an unstable 3.1.2. Analysis of the stability
perturbation term, and in the sensitivity analysis table, it can be seen As can be found in Table 20, the standard deviations of the CTCM
that 𝑐3 undergoes a bifurcation when it goes from 0.05 to 0.15. This on the unimodal test functions are all minimal, which shows that CTCM
shows that the convergence of the CTCM increases and then decreases, finds closer optimal solutions with less fluctuation in each run, and
which is also consistent with the results of the convergence analysis. also confirms that it is more stable than other algorithms. In particular,
CTCM has a standard deviation of 1.03𝐸 − 13 on the 𝐹1 test function,
while GWO has the worst performance on it at 3.98𝐸 + 03. In addition,
3. Experimental analysis CTCM has the best average performance on all unimodal test functions.
It is worth mentioning that for the 𝐹7 function performance, PSO finds
In this section, 23 standard test functions as a benchmark test set the optimal solution of 0.0061 and CTCM of 0.0072, but the average
will be utilized. Subsequently, the performance of seven algorithms, optimal solution is 0.0256 for PSO and 0.0115 for CTCM. This also
supports the fact that CTCM can guarantee the stable performance and
CTCM, PSO (Kennedy & Eberhart, 2002), GWO (Mirjalili et al., 2014),
achieve the optimal average performance during the operation.
SSA (Xue & Shen, 2020), ESOA (Chen et al., 2022), BAS (Jiang & Li,
2017), and WOA (Mirjalili & Lewis, 2016) on the benchmark would
3.1.3. Analysis of the convergence speed
be compared in detail. The benchmark consists of seven unimodal test
In order to better compare the performance of the algorithms in
functions, six multimodal test functions as well as fixed-dimension test
terms of convergence speed on the unimodal test function, the con-
functions (Mirjalili et al., 2014; Yang, 2016). The specific details of vergence curves of the individual algorithms are presented in Fig. 6.
the unimodal test functions are indicated in Table 4, where all test From the figure, it can be noticed that CTCM has essentially the fastest
functions’ dimensions are 30, while the optimal value is 0. The details convergence speed on the unimodal test function. On the 𝐹1 , 𝐹2 , 𝐹5 , 𝐹6
of multimodal and fixed-dimension test functions are shown in Tables 5 test functions, the convergence speed of SSA is comparable to that of
and 6 respectively. The Matlab code for experiments in this paper can CTCM, but the final result is not as good as CTCM. On 𝐹4 , CTCM is
be download at Chen (2024). always in the lead, and as it approaches 800 iterations, WOA suddenly
Fig. 5 illustrate the 2D and 3D views of the five different test func- starts to decline and overtakes SSA at 900 iterations. On 𝐹7 , CTCM
tions and the optimal values found by the CTCM on there benchmarks, outperforms the other algorithms at the very beginning and stays in
respectively. For instance, it can be found that although some members the lead. This indicates that CTCM converges faster and works better
of the CTCM will fall into the local optimum, there are still some in the unimodal case compared to other algorithms.
members that can arrive at the global optimum. In summary, compared with other algorithms mentioned in the
In order to better compare the performance of CTCM with other experiments, CTCM illustrates obvious advantages in terms of both
global optimization seeking ability, algorithm stability and convergence
algorithms on these test functions, each algorithm will be run indepen-
speed.
dently 30 times on each test function and the optimal value, mean,
and standard deviation will be tallied. All experiments take a maximum
3.2. Multimodal test function
number of iterations of 1000 and the population size of each algorithm
is 40, and the independent variables will be shifted negatively by 100
Multimodal test functions are characterized by having multiple
in the experiment. The control parameters of each algorithm in the local optima and at least one global optimum, presenting a complex
experiment are presented in Table 3. The tribes size of CTCM is set search space with numerous local extremities, making the optimization
to 20, and the parameters 𝑐1 , 𝑐2 , 𝑐3 are 2.0, 1.0 and 0.1. The parameters process more challenging. These functions exhibit diverse and complex
𝑐1 , 𝑐2 of PSO are 2,2 while the parameters 𝛽1 , 𝛽2 of ESOA are 0.9, 0.99 search spaces, including high dimensionality and nonlinearity, causing
and the parameters 𝑆 𝑇 of SSA is 0.8. The parameters 𝛼̂ of GWO is optimization algorithms to potentially get trapped in local optima
decreased from 2 to 0 while 𝑟1 , 𝑟2 are a random values from [0, 1]. The rather than finding the global optimum. Utilizing multimodal test
parameters 𝑑0 , 𝑑1 of BAS are 0.001, 3, and the parameters 𝛼1 , 𝛼2 of WOA functions can demonstrate and evaluate an optimization algorithm’s
are decreased from 2 to 0 and −1 to −2 respectively. The comprehensive global search capability, robustness, and balance between exploration
performance results of seven algorithms are indicated in Table 20. and exploitation abilities.

6
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

Table 2
Sensitivity analysis for CTCM parameters under various scenarios.
𝑚 𝑐1 𝑐2 𝑐3 𝑓3 𝑓5 𝑓7 𝑓9 𝑓11 𝑓18 𝑓23
5 1 1 0.05 593.8693 140.5015 1.82E−01 6.54E+01 2.45E−01 3 −10.5364
5 1 1 0.1 302.7942 75.9552 1.46E−01 5.94E+01 2.24E−02 3 −10.5364
5 1 1 0.15 621.0863 121.1018 9.26E−02 57.55 8.33E−02 3 −10.5364
5 1 2 0.05 22.3558 23.8596 7.78E−02 4.29E+01 0.00E+00 3 −10.5364
5 1 2 0.1 22.1328 25.2746 3.15E−02 4.12E+01 0.00E+00 3 −10.5364
5 1 2 0.15 25.6227 23.5241 2.08E−02 4.76E+01 0.00E+00 3 −10.5364
5 1.5 1 0.05 260.0564 95.395 1.24E−01 5.73E+01 6.91E−02 3 −10.5364
5 1.5 1 0.1 266.8199 45.8184 6.61E−02 5.04E+01 2.90E−03 3 −10.5364
5 1.5 1 0.15 431.3918 75.8527 6.67E−02 6.00E+01 2.96E−02 3 −10.5364
5 1.5 2 0.05 57.1872 22.7277 2.41E−02 4.99E+01 0.00E+00 3 −10.5364
5 1.5 2 0.1 41.6134 9.7525 1.89E−02 3.66E+01 0.00E+00 3 −10.5364
5 1.5 2 0.15 36.1386 25.1488 6.30E−03 4.88E+01 0.00E+00 3 −10.5364
5 2 1 0.05 301.9079 49.758 8.42E−02 4.89E+01 2.30E−03 3 −10.5364
5 2 1 0.1 130.4581 27.4624 3.70E−02 5.68E+01 0.00E+00 3 −10.5364
5 2 1 0.15 102.1445 19.6839 3.01E−02 4.52E+01 0.00E+00 3 −10.5364
5 2 2 0.05 229.4515 24.2998 1.14E−02 3.99E+01 1.90E−03 3 −10.5364
5 2 2 0.1 228.366 21.9906 1.78E−02 4.28E+01 0.00E+00 3 −10.5364
5 2 2 0.15 306.7427 24.0272 9.70E−03 4.28E+01 0.00E+00 3 −10.5364
10 1 1 0.05 30.0838 13.1369 1.02E−01 32.8415 0.00E+00 3 −10.5364
10 1 1 0.1 156.2986 72.2322 5.10E−02 5.02E+01 0.00E+00 3 −10.5364
10 1 1 0.15 126.951 10.1087 4.61E−02 3.69E+01 2.21E−02 3 −10.5364
10 1 2 0.05 0.7595 12.4042 3.77E−02 3.68E+01 0.00E+00 3 −10.5364
10 1 2 0.1 1.0964 14.7152 1.89E−02 2.79E+01 0.00E+00 3 −10.5364
10 1 2 0.15 0.5628 22.9949 9.60E−03 4.78E+01 0.00E+00 3 −10.5364
10 1.5 1 0.05 18.304 68.8491 1.25E−01 3.75E+01 0.00E+00 3 −10.5364
10 1.5 1 0.1 39.4778 15.8353 8.81E−02 3.53E+01 0.00E+00 3 −10.5364
10 1.5 1 0.15 102.1958 13.7413 5.03E−02 3.92E+01 3.47E−02 3 −10.5364
10 1.5 2 0.05 10.7687 16.9444 1.55E−02 4.48E+01 0.00E+00 3 −10.5364
10 1.5 2 0.1 4.5931 9.2575 9.20E−03 2.98E+01 0.00E+00 3 −10.5364
10 1.5 2 0.15 9.1261 13.9371 8.10E−03 3.28E+01 0.00E+00 3 −10.5364
10 2 1 0.05 1.5571 20.192 7.10E−02 2.99E+01 1.30E−03 3 −10.5364
10 2 1 0.1 4.6366 22.9397 2.57E−02 3.64E+01 0.00E+00 3 −10.5364
10 2 1 0.15 6.7303 14.1616 1.47E−02 4.08E+01 0.00E+00 3 −10.5364
10 2 2 0.05 132.7684 24.7551 1.27E−02 3.28E+01 4.50E−03 3 −10.5364
10 2 2 0.1 263.2481 27.0868 1.37E−02 2.09E+01 1.00E−04 3 −10.5364
10 2 2 0.15 228.4485 23.111 1.26E−02 3.98E+01 0.00E+00 3 −10.5364
20 1 1 0.05 8.0944 22.0098 1.05E−01 2.49E+01 0.00E+00 3 −10.5364
20 1 1 0.1 29.8462 2.9076 1.02E−01 29.8524 0.00E+00 3 −10.5364
20 1 1 0.15 322.2293 15.8966 7.79E−02 3.58E+01 1.00E−04 3 −10.5364
20 1 2 0.05 0.0938 18.7382 2.74E−02 3.38E+01 0.00E+00 3 −10.5364
20 1 2 0.1 0.0849 19.0109 2.33E−02 2.98E+01 0.00E+00 3 −10.5364
20 1 2 0.15 0.0531 15.5078 2.14E−02 3.98E+01 0.00E+00 3 −10.5364
20 1.5 1 0.05 15.6552 21.0959 7.64E−02 30.847 0.00E+00 3 −10.5364
20 1.5 1 0.1 2.4184 12.0854 4.75E−02 34.8243 0.00E+00 3 −10.5364
20 1.5 1 0.15 2.933 19.9459 6.22E−02 30.8437 0.00E+00 3 −10.5364
20 1.5 2 0.05 4.2875 19.9341 1.12E−02 3.18E+01 0.00E+00 3 −10.5364
20 1.5 2 0.1 2.3695 19.4694 9.60E−03 23.879 0.00E+00 3 −10.5364
20 1.5 2 0.15 6.5009 20.1646 9.40E−03 38.8033 0.00E+00 3 −10.5364
20 2 1 0.05 0.2406 15.042 2.62E−02 2.49E+01 0.00E+00 3 −10.5364
20 2 1 0.1 0.0412 0.4961 3.44E−02 3.88E+01 0.00E+00 3 −10.5364
20 2 1 0.15 0.1983 7.1649 2.09E−02 3.48E+01 0.00E+00 3 −10.5364
20 2 2 0.05 140.7373 18.3864 8.40E−03 2.79E+01 1.00E−04 3 −10.5364
20 2 2 0.1 146.1615 26.6298 4.60E−03 27.8588 1.00E−04 3 −10.5364
20 2 2 0.15 135.3297 25.7838 1.05E−02 3.08E+01 0.00E+00 3 −10.5364

Table 3 Table 4
Control parameters for each algorithm. Unimodal test functions (Dim = 30).
Algorithm Parameters Values Function Initial range 𝐹𝑚𝑖𝑛
∑𝐷
CTCM 𝑚, 𝑐1 , 𝑐2 , 𝑐3 [20, 2.0, 1.0, 0.1] 𝑓1 (𝑥) = 𝑖=1 𝑥2𝑖 [−100, 100] 0
PSO 𝑐1 , 𝑐2 [2.0, 2.0] ∑𝐷 ∏𝐷
𝑓2 (𝑥) = 𝑖=1 |𝑥𝑖 | + 𝑖=1 |𝑥𝑖 | [−10, 10] 0
GWO 𝛼̂ 2 to 0 ( )2
∑𝐷 ∑𝑖
SSA 𝑆𝑇 0.8 𝑓3 (𝑥) = 𝑖=1 𝑥
𝑗=1 𝑗
[−100, 100] 0
ESOA 𝛽1 , 𝛽2 [0.9, 0.99]
𝑓4 (𝑥) = max𝑖 {|𝑥𝑖 |} [−100, 100] 0
BAS 𝑑0 , 𝑑1 [0.001, 3]
∑𝐷−1 [ ]
WOA 𝛼1 , 𝛼2 2 to 0, −1 to −2 𝑓5 (𝑥) = 𝑖=1 100(𝑥𝑖+1 − 𝑥2𝑖 )2 + (𝑥𝑖 − 1)2 [−30, 30] 0
∑𝐷 ( )2
𝑓6 (𝑥) = 𝑖=1 ⌊𝑥𝑖 + 0.5⌋ [−100, 100] 0
∑𝐷
𝑓7 (𝑥) = 𝑖=1 𝑖𝑥4𝑖 + random[0, 1) [−1.28, 1.28] 0

7
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

Table 5
Multimodal test functions (Dim = 300).
Function Initial range 𝐹𝑚𝑖𝑛
∑𝐷 √
𝑓8 (𝑥) = 𝑖=1 𝑥𝑖 sin( |𝑥𝑖 |) [−500, 500] −418.9829𝐷
∑𝐷 [ 2 ]
𝑓9 (𝑥) = 𝑖=1 𝑥𝑖 − 10 cos(2𝜋 𝑥𝑖 ) + 10 [−5.12, 5.12] 0
( √ ∑ ) ( ∑ )
𝐷 𝐷
𝑓10 (𝑥) = −20 exp −0.2 𝐷1 𝑖=1 𝑥2𝑖 − exp 𝐷1 𝑖=1 cos(2𝜋 𝑥𝑖 ) + 20 + 𝑒 [−32, 32] 0
∑𝐷 ∏𝐷 ( )
𝑥2𝑖 𝑥
𝑓11 (𝑥) = 𝑖=1 4000
− 𝑖=1
cos √𝑖 +1 [−600, 600] 0
𝑖
{ }
𝜋 ∑
𝐷−1
[ ]
𝑓12 (𝑥) = 10 sin2 (𝜋 𝑦1 ) + (𝑦𝑖 − 1)2 1 + 10 sin2 (𝜋 𝑦𝑖+1 ) + (𝑦𝐷 − 1)2
𝐷 𝑖=1

∑
𝐷
+ 𝑢(𝑥𝑖 , 10, 100, 4)
𝑖=1
𝑥𝑖 + 1 [−50, 50] 0
𝑦𝑖 = 1 +
4
⎧𝑘(𝑥𝑖 − 𝑎)𝑚 𝑥𝑖 > 𝑎
⎪
𝑢(𝑥𝑖 , 𝑎, 𝑘, 𝑚) = ⎨0 − 𝑎 ≤ 𝑥𝑖 ≤ 𝑎
⎪
⎩𝑘(−𝑥𝑖 − 𝑎)𝑚 𝑥𝑖 < −𝑎
{ } ∑ 𝐷
𝑓13 (𝑥) = 0.1 sin2 (3𝜋 𝑥1 ) + 𝑔(𝑥) + 𝑢(𝑥𝑖 , 5, 100, 4)
𝑖=1
[−50, 50] 0
∑
𝐷−1
[ ] [ ]
𝑔(𝑥) = (𝑥𝑖 − 1) 1 + sin2 (3𝜋 𝑥𝑖+1 ) + (𝑥𝐷 − 1)2 1 + sin2 (2𝜋 𝑥𝐷 )
2

𝑖=1

Table 6
Fixed-dimension test functions.
Function Dim Initial range 𝐹𝑚𝑖𝑛
∑
𝐷
[ 2 ]
𝑓14 (𝑥) = 𝑦𝑖 − 10 cos(2𝜋 𝑦𝑖 ) + 10
𝑖=1
{ 2 [−65.536, 65.536] 0.998
𝑥𝑖 if |𝑥𝑖 | < 0.5
𝑦𝑖 =
round(2𝑥𝑖 )∕2 otherwise
∑
𝐷
√
𝑓15 (𝑥) = 418.9829𝐷 − 𝑦𝑖 sin( |𝑦𝑖 |)
𝑖=1
{ 2 [−5, 5] 0
𝑥𝑖 if |𝑥𝑖 | < 0.5
𝑦𝑖 =
round(2𝑥𝑖 )∕2 otherwise
[ ]10∕𝐷
∑𝐷
1 ∑ 32
|2𝑗 𝑥𝑖 − round(2𝑗 𝑥𝑖 )|
𝑓16 (𝑥) = 106 1+ 𝑗
− 106 2 [−5, 5] −1.0316
𝑖=1
𝐷 𝑗=1
2
[𝐷 ]
∑ ∑
𝐷
𝑓17 (𝑥) = min (𝑥𝑖 − 𝜇1 )2 , 𝐷𝑠2 + (𝑥𝑖 − 𝜇2 )2
𝑖=1 𝑖=1

∑
𝐷
[ ( )] 2 [−5, 0], [10, 15] 0.3979
+ 10 1 − cos(2𝜋(𝑥𝑖 − 𝜇1 ))
𝑖=1

𝜇1 = 2.5, 𝜇2 = −2.5, 𝑠 = 1
∑[
𝐷−1
]2
𝑓18 (𝑥) = 100(𝑥𝑖+1 − 𝑥2𝑖 )2 + (𝑥𝑖 − 1)2 ∕4000
𝑖=1
( ) 2 [−2, 2] 3
∏
𝐷−1
100(𝑥𝑖+1 − 𝑥2𝑖 )2 + (𝑥𝑖 − 1)2
− cos √ +1
𝑖=1 𝑖
√
⎡ 2 2 2 ⎤
∑⎢
𝐷−1 sin ( 𝑥𝑖 + 𝑥𝑖+1 ) − 0.5 ⎥
𝑓19 (𝑥) = ⎢0.5 + 2 2 ⎥ 3 [0, 1] 3.8628
𝑖=1 ⎢ (1 + 0.001(𝑥𝑖 + 𝑥𝑖+1 ))2 ⎥
⎣ ⎦
[ ( )]
∑𝑁
1∑
𝑛
𝑓20 (𝑥) = 𝜔𝑖 𝑓 (𝑥 ) +𝜖 6 [0, 1] −3.322
𝑖=1
𝑛 𝑗=1 𝑏𝑎𝑠𝑖𝑐 𝑗

∑ ((
10
) ( )𝑇 )−1
𝑓21 (𝑥) = − 𝑥 − 𝑎𝑆 𝐻𝑖 ⋅ 𝑥 − 𝑎𝑆 𝐻𝑖 + 𝑐 𝑆 𝐻𝑖 4 [0, 10] −10.5363
𝑖=1
∑ ((
7
) ( )𝑇 )−1
𝑓22 (𝑥) = − 𝑥 − 𝑎𝑆 𝐻𝑖 ⋅ 𝑥 − 𝑎𝑆 𝐻𝑖 + 𝑐 𝑆 𝐻𝑖 4 [0, 10] −10.4028
𝑖=1
∑ ((
5
) ( )𝑇 )−1
𝑓23 (𝑥) = − 𝑥 − 𝑎𝑆 𝐻𝑖 ⋅ 𝑥 − 𝑎𝑆 𝐻𝑖 + 𝑐 𝑆 𝐻𝑖 4 [0, 10] −10.1532
𝑖=1

3.2.1. Analysis of the convergence accuracy close to the global optimal solution of 0. On 𝐹11 , SSA finds an optimal
As indicated in Table 20, although CTCM performs poorly on 𝐹8 solution of 1.11𝐸 − 16, while CTCM exhibits an optimal solution of
and 𝐹9 , the solutions found on 𝐹12 and 𝐹13 are optimal among all algo- 1.79𝐸−14, followed by PSO, which has an optimal solution of 4.89𝐸−10.
rithms. Next on 𝐹10 and 𝐹11 , although CTCM is not the best performer,
it is in the second position and is not far from the first place SSA. For 3.2.2. Analysis of the stability
instance, on 𝐹10 , SSA finds an optimal solution of 9.40𝐸 − 09, while From the performance of each algorithm on multimodal, CTCM has
CTCM exhibits an optimal solution of 2.59𝐸 − 08, both of which are the smallest standard deviation on 𝐹9 , 𝐹10 , 𝐹11 , and 𝐹13 , reflecting its

8
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

Fig. 5. The results of CTCM tested on the 2 dimension test functions. As observed, the members of the proposed CTCM represented by the red points found their global optimal
solutions in each test function.

good stability in the face of such multimodal problems. It is worth times that of CTCM. This shows that the stability of CTCM in the
mentioning that on 𝐹11 , the standard deviation of CTCM is 0.0069, multimodal case is relatively good.
while the standard deviation of the second place SSA is 0.0143, which
is almost half of SSA. In terms of the average performance of each 3.2.3. Analysis of the convergence speed
algorithm, CTCM performs best on 𝐹10 , 𝐹11 , and 𝐹13 . Among them, on The convergence curves of each algorithm on the multimodal test
𝐹13 , the average performance of CTCM is 0.0011, while the average function are shown in Fig. 7. On 𝐹8 , WOA converges the fastest and
performance of the second-place PSO is 0.0054, which is almost five CTCM looks a bit unsatisfactory in comparison, while on 𝐹9 , CTCM

9
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

Fig. 6. The convergence performance of algorithms on 𝐹1 to 𝐹9 . As observed, the convergence speed and accuracy of proposed CTCM performs other algorithms in most cases.

Fig. 7. The convergence performance of algorithms on 𝐹10 to 𝐹15 . As observed, the convergence speed and accuracy of proposed CTCM performs other algorithms in 𝐹11 , 𝐹12 , 𝐹13 , 𝐹15 .

10
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

Fig. 8. The convergence performance of algorithms on 𝐹16 to 𝐹23 . As observed, the convergence speed and accuracy of proposed CTCM performs other algorithms in most cases.

Table 7 Table 9
Statistic comparison between proposed CTCM, PSO, GWO, SSA, ESOA, BAS, WOA based Statistic comparison between proposed CTCM, PSO, GWO, SSA, ESOA, BAS, WOA based
on Wilcoxon test for Unimodal and Multimodal functions on 30 dimensions. on Wilcoxon test for Unimodal and Multimodal functions on 50 dimensions.
Algorihtms Certira Better Equal Worse 𝑝-value Dec. Algorihtms Certira Better Equal Worse 𝑝-value Dec.
Best 11 0 2 0.1465 ≈ Best 12 0 1 0.0171 ≈
CTCM vs. PSO CTCM vs. PSO
Average 12 0 1 0.0171 + Average 12 0 1 0.0134 +
Best 13 0 0 0.0002 + Best 12 0 1 0.0012 +
CTCM vs. GWO CTCM vs. GWO
Average 13 0 0 0.0002 + Average 12 0 1 0.0012 +
Best 9 0 4 0.2734 ≈ Best 9 0 4 0.2163 ≈
CTCM vs. SSA CTCM vs. SSA
Average 12 0 1 0.0215 + Average 11 0 2 0.0327 +
Best 12 0 1 0.0171 + Best 12 0 1 0.0134 +
CTCM vs. ESOA CTCM vs. ESOA
Average 12 0 1 0.0171 + Average 12 0 1 0.0134 +
Best 13 0 0 0.0002 + Best 13 0 0 0.0002 +
CTCM vs. BAS CTCM vs. BAS
Average 13 0 0 0.0002 + Average 13 0 0 0.0002 +
Best 10 0 3 0.3396 ≈ Best 10 0 3 0.3396 ≈
CTCM vs. WOA CTCM vs. WOA
Average 12 0 1 0.0171 + Average 10 0 3 0.2163 ≈

Table 8 Table 10
Statistic comparison between proposed CTCM, PSO, GWO, SSA, ESOA, BAS, WOA based Statistic comparison between proposed CTCM, PSO, GWO, SSA, ESOA, BAS, WOA based
on Friedman rank test for Unimodal and Multimodal functions on 30 dimensions. on Friedman rank test for Unimodal and Multimodal functions on 50 dimensions.
Algorithms Rank Algorithms Rank
Best Average Best Average
CTCM 1.7692 1.3077 CTCM 1.7692 1.6154
PSO 2.7692 2.8462 PSO 2.9231 3.2308
GWO 4.8462 4.9231 GWO 4.9231 4.7692
SSA 2.6154 3.0000 SSA 2.6923 2.9231
ESOA 5.1538 4.6923 ESOA 5.0769 4.8462
BAS 7.0000 7.0000 BAS 7.0000 7.0000
WOA 3.8462 4.2308 WOA 3.6154 3.6154

11
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

Table 11 In summary, CTCM still shows a very competitive global optimal

Statistic comparison between proposed CTCM, PSO, GWO, SSA, ESOA, BAS, WOA based
seaking capability in the multimodal case, although sometimes the
on Wilcoxon test for Unimodal and Multimodal functions on 100 dimensions.
solution found is not the best. And it is significantly better than other
Algorihtms Certira Better Equal Worse 𝑝-value Dec.
algorithms in terms of stability and convergence speed.
Best 12 0 1 0.0134 +
CTCM vs. PSO
Average 12 0 1 0.0134 +
Best 13 0 0 0.0002 +
3.3. Fixed-dimension test function
CTCM vs. GWO
Average 13 0 0 0.0002 +
Best 10 0 3 0.0327 + Fixed dimension test functions are a class of test sets describing
CTCM vs. SSA
Average 10 0 3 0.0327 + specific situations, which can be used to evaluate the performance of
Best 12 0 1 0.0046 +
CTCM vs. ESOA an algorithm in a number of complex situations. The fixed-dimension
Average 12 0 1 0.0034 +
Best 13 0 0 0.0002 + test function is closer to the actual engineering optimization problem
CTCM vs. BAS
Average 13 0 0 0.0002 + and can be utilized to verify the actual optimization performance of
CTCM vs. WOA
Best 6 0 7 0.8926 ≈ the algorithm, and to demonstrate the characteristics of the algorithm
Average 8 0 5 0.5879 ≈ during the optimization process.
And the parameters of 𝐹21 , 𝐹22 , 𝐹23 are,
Table 12 𝑇
⎡4 1 8 6 3 2 5 8 6 7⎤
Statistic comparison between proposed CTCM, PSO, GWO, SSA, ESOA, BAS, WOA based ⎢ ⎥
4 1 8 6 7 9 5 1 2 3.6⎥
on Friedman rank test for Unimodal and Multimodal functions on 100 dimensions.
𝑎𝑆 𝐻 = ⎢
Algorithms Rank ⎢4 1 8 6 3 2 3 8 6 7⎥
⎢4 1 8 6 7 9 3 1 2 3.6⎥⎦ (17)
Best Average ⎣
CTCM 1.9231 1.7692 𝑐 𝑆 𝐻 =[0.1 0.2 0.2 0.4 0.4
PSO 3.3846 3.8462
0.6 0.3 0.7 0.5 0.5]
GWO 4.9231 5.1538
SSA 2.9231 2.6154
ESOA 5.3077 5.0769 3.3.1. Analysis of the convergence accuracy
BAS 7.0000 7.0000 CTCM performs best under all complex problems, which means
WOA 2.5385 2.5385 that it is able to find the global optimal solution of the problem with
excellent global optimization finding capability. For instance, on 𝐹20 ,
Table 13 CTCM finds an optimal solution of −3.322, while GWO is −3.1652 and
Statistic comparison between proposed CTCM, PSO, GWO, SSA, ESOA, BAS, WOA based WOA is −3.2439. It can be found that CTCM’s global optimization
on Wilcoxon test for fixed dimension functions. capability performs well in complicate situations.
Algorihtms Certira Better Equal Worse 𝑝-value Dec.
Best 1 9 0 1 ≈ 3.3.2. Analysis of the stability
CTCM vs. PSO
Average 3 4 3 0.4375 ≈ From the standard deviation of each algorithm under different test
Best 10 0 0 0.002 + functions, it can be found that CTCM has the best stability from 𝐹15 to
CTCM vs. GWO
Average 7 0 3 1 ≈
Best 1 9 0 1 ≈
𝐹20 and is the one that almost always finds the global optimal solution.
CTCM vs. SSA On 𝐹14 , CTCM is slightly less stable than the other algorithms, and the
Average 6 3 1 0.4375 ≈
Best 10 0 0 0.002 + best performer is PSO. While on 𝐹21 -𝐹23 , the best performer is ESOA.
CTCM vs. ESOA
Average 5 0 5 0.2324 ≈ the average performance for the algorithms is similar. Although CTCM
Best 10 0 0 0.002 +
CTCM vs. BAS is slightly less stable than the other algorithms in some moments, it still
Average 10 0 0 0.002 +
Best 10 0 0 0.002 + performs the best for most complex cases, which shows that CTCM has
CTCM vs. WOA high stability in the face of complex situations.
Average 6 0 4 0.8457 ≈

3.3.3. Analysis of the convergence speed

Table 14
Statistic comparison between proposed CTCM, PSO, GWO, SSA, ESOA, BAS, WOA based
The convergence curves of each algorithm on complex problems are
on Friedman rank test for fixed dimension functions. shown in Fig. 8. On 𝐹14 , CTCM performs slightly worse, while on 𝐹15 ,
Algorithms Rank CTCM leads all other algorithms from the beginning and achieves better
Best Average
optimization results. This situation is similar on 𝐹16 -𝐹20 , where CTCM
converges quickly and achieves better results than the other algorithms.
CTCM 1.9000 2.9500
PSO 2.0000 2.9000 For 𝐹21 , CTCM converges almost to the optimal solution with the initial
GWO 5.9000 4.4000 20 iterations, while PSO converges close to 200. On 𝐹22 , CTCM also
SSA 2.1000 3.5500 reaches the optimal solution near 20 iterations, while PSO converges
ESOA 4.3000 2.9000 after close to 100 iterations, and ESOA converges close to around 300.
BAS 7.0000 7.0000
The case of 𝐹23 is similar to that of 𝐹22 . This shows that CTCM is the
WOA 4.8000 4.3000
fastest in convergence for most complex problems.
Combined with the above analyses, CTCM has excellent global op-
timization capability, good stability and fast convergence speed when
outperforms all the other algorithms and reaches the vicinity of the facing real complex problems.
optimal solution after nearly 200 iterations. On 𝐹10 , CTCM converges
3.4. Statistical analysis
slightly less quickly than SSA, but also outperforms the other algo-
rithms. On 𝐹11 , CTCM lags behind SSA at the beginning, but surpasses
In order to reflect the performance of CTCM with other algorithms
SSA near 380 iterations and stays in the lead thereafter. The situation on the benchmark functions, statistical analyses are performed here
on 𝐹12 and 𝐹13 is similar to that on 𝐹11 , where CTCM stays ahead of to better indicate the strengths and weaknesses of each algorithm in
the other algorithms after 300 iterations. This shows that CTCM has an different situations. Each algorithm will be run 30 times under various
advantage over the other algorithms in terms of convergence speed in scenarios while Wilcoxon test and Friedman rank test are utilized to
the multimodal case. determine the performance of individual algorithms.

12
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

3.4.1. Unimodal and multimodal functions on 30d 4. Application

The results of the Wilcoxon test are shown in Table 7, and in most
cases the CTCM results are much better than the other algorithms, both In this section, two real-world problems will be utilized to test
in terms of best and average values. In particular, compared to GWO the practical effectiveness of the proposed algorithm. Since there are
and BAS, CTCM outperforms them on all test functions. In order to constraints in both the real world problems, constraint relaxation will
indicate the gap between different algorithms more visually, Friedman be done here by applying penalty functions. Suppose the real fitness
rank test was used for algorithm ranking. As illustrated in Table 8, of the function is 𝑓 (𝑥) while the transformed fitness and constraints
CTCM obtained a ranking of 1.7692 on the best result and 1.3077 on function are 𝑓̂(𝑥) and 𝑔(𝑥).
the average result, both in the first position. ∑
𝑛
𝑓̂(𝑥) = 𝑓 (𝑥) + 𝜙 |𝑔(𝑥)| (18)
𝑖=1

3.4.2. Unimodal and Multimodal functions on 50D in which the penalty weight 𝜙 here is 106 and 𝑛 means the quantity
Table 9 indicates the results of Wilcoxon test between CTCM and of penalty function. The Matlab code for applications experimented in
each other rival algorithm’s performance of Unimodal and Multimodal this paper can be download at Chen (2024).
functions on 50D. From the table, the proposed CTCM is significantly
better than other algorithms, which is confirmed by 𝑝-value in each 4.1. Himmelblau’s nonlinear optimization problem
pair. From the Friedman rank test results in Table 10, the CTCM obtains
ranking of 1.7692 on best result as well as ranking of 1.6154 on the The Himmelblau’s nonlinear optimization problem is a famous
average result, which outperforms all of other algorithms. practical problem for testing the optimization performance of algo-
rithms (Himmelblau, 1972). It is an optimization problem containing 5
variables and 3 constraints. This test problem has multiple local optimal
3.4.3. Unimodal and Multimodal functions on 100D
solutions and is ideal for testing the convergence ability of optimization
To indicate the performance of individual algorithms in the high di- algorithms and the capacity of escaping local optimal solutions. The
mensional case, Wilcoxon test was utilized to examine the performance specific form could be expressed as,
of the algorithms in 100 dimensions. As it is shown in Table 11, CTCM
𝑓 (𝑋) = 5.3578547𝑥23 + 0.8356891𝑥1 𝑥5
is better in most cases, especially compared with GWO and BAS. The
result of Friedman rank test is illustrated in Table 12, CTCM obtains + 37.293239𝑥1 − 40792.141
the best place in terms of best and average result. subject to
𝑔1 (𝑥) = 85.334407 + 0.0056858𝑥2 𝑥5
3.4.4. Fixed-dimension functions + 0.0006262𝑥1 𝑥4 − 0.0022053𝑥3 𝑥5
Since Fixed dimension functions usually have a fixed optimal so- 𝑔2 (𝑥) = 80.51249 + 0.0071317𝑥2 𝑥5
lution, there may be more cases where the results are equal when + 0.0021813𝑥1 𝑥2 + 0.0021813𝑥23
comparisons are made. Therefore statistically analysis the performance 𝑔3 (𝑥) = 9.300961 + 0.0047026𝑥3 𝑥5
of each algorithm can provide a more intuitive representation of its
+ 0.0012547𝑥1 𝑥3 + 0.0019085𝑥3 𝑥4
strengths and weaknesses. Table 13 presents the Wilcoxon test result
0 ≤ 𝑔1 (𝑥) ≤ 92
between CTCM and each algorithm, it is clear that in most cases, the
90 ≤ 𝑔2 (𝑥) ≤ 110
CTCM could retrieve more better results and fewer worse results than
others. From the Friedman rank test result in Table 14, CTCM receives 20 ≤ 𝑔3 (𝑥) ≤ 25
the first place in best results and the second place in average results, 78 ≤ 𝑥1 ≤ 102
which indicates that CTCM could solve the fixed dimension problem 33 ≤ 𝑥2 ≤ 45
more accuracy with enough stability than other algorithms.
27 ≤ 𝑥3 ≤ 45
27 ≤ 𝑥4 ≤ 45
3.5. Discussion 27 ≤ 𝑥5 ≤ 45
In the experiment the population size of CTCM is set to 40 and the
From the exhaustive experimental results above, it could be found maximum number of iterations is 1000, the convergence and statistical
that CTCM performs much better than the other algorithms in most performance is shown in Fig. 9. The CTCM, although converging a bit
Unimodal and Multimodal cases, and occupies a great advantage in slower than SSA and PSO in the very first few iterations, converged
to the global optimal solution near 100 times iterations. From the
convergence accuracy, speed, and stability, and the advantage of CTCM
statistical results graph, it can be found that CTCM, SSA, PSO, and
does not reduce in term of increasing problem dimensions. As for the
GWO perform better and are capable of converging stably to the global
experiments on the fixed-dimension problems, it is clear that CTCM is
optimal solution.
basically in the leading position in terms of performance and is capable
The statistical results of each algorithms after 30 runs are illus-
of finding the optimal value in all problems. However, it is also to
trated in Table 15. The best result found by the CTCM each time is
indicate that CTCM does not perform as well as algorithms in some −30665.5387, from which its statistical standard deviation is 0. The
cases, for instance, in 𝐹8 test function, CTCM seems to fall into a local best solution is 𝑥 = [78, 33, 29.9953, 45, 36.7758], and its constraint is
optimal solution very swiftly and does not jump out of this solution 𝑔(𝑥) = [92, 98.8841, 20]. In the comparison with other algorithms, it
in the subsequent iterations. It indicates that the current CTCM may can be seen that CTCM searched the global optimal solution in all 30
have some limitations in addressing a specific type of multimodal prob- times experiments, leading the other algorithms in terms of best, worst,
lems. Consequently, further enhancements and improvements for such average, and standard deviation cases. It clearly indicates that the
problems are needed in subsequent research to ensure comprehensive CTCM is able to maintain a good global optimization search capability
coverage and the ability to deal with all potential scenarios. as well as stability in practical problems.

13
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

Fig. 9. The convergence and statistical performance of algorithms on Himmelblau’s nonlinear optimization problem. As observed, the CTCM retrieved optimal value at around 100
iterations, and performs excellent stability during the 30 times repeat experiments.

Table 15
The comparison of each algorithm on Himmelblau’s nonlinear optimization problem.
Metrics CTCM PSO GWO SSA ESOA BAS WOA
Best −3.0666E+04 −3.0666E+04 −3.0665E+04 −3.0666E+04 −3.0663E+04 −2.9184E+04 −3.0553E+04
Wrost −3.0666E+04 −3.0666E+04 −3.0655E+04 −3.0666E+04 −3.0508E+04 −2.8597E+04 −2.9652E+04
Mean −3.0666E+04 −3.0666E+04 −3.0663E+04 −3.0666E+04 −3.0591E+04 −2.8845E+04 −3.0216E+04
Std 0 1.0935E−11 1.8867E+00 4.1297E−11 3.8621E+01 2.0815E+02 1.8989E+02
𝑔1 9.2000E+01 9.2000E+01 9.2000E+01 9.2000E+01 9.2000E+01 9.2000E+01 9.2000E+01
𝑔2 9.8841E+01 9.8841E+01 9.8841E+01 9.8841E+01 9.8841E+01 9.8841E+01 9.8841E+01
𝑔3 2.0000E+01 2.0000E+01 2.0000E+01 2.0000E+01 2.0000E+01 2.0000E+01 2.0000E+01

Fig. 10. Speed reducer design problem.

4.2. Speed reducer design the optimal design to minimize the weight of the speed reducer by
optimizing these design parameters. At the same time, several nonlinear
Speed reducer design problem is a classical mechanical structure op- constraints need to be satisfied, which represent practical requirements
timization problem usually utilized to test and validate the performance in physical and engineering design, such as stresses, material strengths,
of optimization algorithms. The objective of this problem is to design spatial constraints, and so on.
a reducer to minimize the total weight of the reducer while satisfying The details are illustrated in Fig. 10, the design variables involved
several design constraints and performance metrics. A speed reducer is in this problem consist of 7 parameters that define the geometry of the
a common component in mechanical design that is used to reduce the reducer. 𝑏 means the gear face width, 𝑚 is the teeth number if small
rotational speed and increase the torque in a machine, and is applied in gear while 𝑧 means the ratio of teeth number between small and large
a wide range of industrial equipment, automation drivelines, and other gear. 𝑑1 and 𝑑2 mean the diameter of the large and small shaft while 𝑙1
mechanical devices. In this optimization problem, the design variables and 𝑙2 are the length of large and small shaft. The design problem has
are related to the dimensions, gear characteristics, bearings, and other several nonlinear constraints that cover the different practical physical
physical properties of the speed reducer, and the objective is to obtain requirements and performance metrics involved in the design. The

14
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

Fig. 11. The convergence and statistical performance of algorithms on Speed reducer design problem. As observed, the CTCM retrieved optimal value at around 20 iterations, and
performs excellent stability during the 30 times repeat experiments.

Table 16
The comparison of each algorithm on Speed reducer design optimization problem.
Metrics CTCM PSO GWO SSA ESOA BAS WOA
Best 2.2306E+03 2.2306E+03 2.2481E+03 2.2306E+03 2.2313E+03 2.6492E+03 3.3203E+03
Wrost 2.3242E+03 2.2399E+03 4.2186E+03 2.2360E+03 2.2357E+03 2.0167E+04 4.8154E+03
Mean 2.2337E+03 2.2309E+03 2.7999E+03 2.2309E+03 2.2333E+03 1.5055E+04 4.0023E+03
Std 1.7099E+01 1.7041E+00 7.2833E+02 1.3890E+00 1.0841E+00 7.6533E+03 2.8972E+02
𝑔1 −7.3915E−02 −7.3915E−02 −7.3915E−02 −7.3915E−02 −7.3915E−02 −7.3915E−02 −7.3915E−02
𝑔2 −1.9800E−01 −1.9800E−01 −1.9800E−01 −1.9800E−01 −1.9800E−01 −1.9800E−01 −1.9800E−01
𝑔3 −4.5665E−01 −4.5665E−01 −4.5665E−01 −4.5665E−01 −4.5665E−01 −4.5665E−01 −4.5665E−01
𝑔4 −4.1453E−01 −4.1453E−01 −4.1453E−01 −4.1453E−01 −4.1453E−01 −4.1453E−01 −4.1453E−01
𝑔5 −1.1102E−15 −1.1102E−15 −1.1102E−15 −1.1102E−15 −1.1102E−15 −1.1102E−15 −1.7764E−15
𝑔6 −4.7740E−15 −4.7740E−15 −4.7740E−15 −4.7740E−15 −4.7740E−15 −4.7740E−15 −5.6621E−15
𝑔7 −7.0250E−01 −7.0250E−01 −7.0250E−01 −7.0250E−01 −7.0250E−01 −7.0250E−01 −7.0250E−01
𝑔8 0.0000E+00 0.0000E+00 0.0000E+00 0.0000E+00 0.0000E+00 0.0000E+00 0.0000E+00
𝑔9 −5.8333E−01 −5.8333E−01 −5.8333E−01 −5.8333E−01 −5.8333E−01 −5.8333E−01 −5.8333E−01
𝑔10 −5.1326E−02 −5.1326E−02 −5.1326E−02 −5.1326E−02 −5.1326E−02 −5.1326E−02 −5.1326E−02
𝑔11 −2.4732E−01 −2.4732E−01 −2.4732E−01 −2.4732E−01 −2.4732E−01 −2.4732E−01 −2.4732E−01

√
main constraints include: ensuring that the design does not exceed the 𝐻 2 + 157.5 × 106
𝑔6 (𝑥) = − 1 ≤ 0,
stress limits of the material; ensuring that the strength of the gears 85𝑑23
and shafts meets the requirements; ensuring that the dimensions of the 𝑚𝑧
𝑔7 (𝑥) = − 1 ≤ 0,
gearbox do not exceed the limitations of the design space; ensuring 40
5𝑚
that the gearbox is capable of efficiently transmitting the power; and 𝑔8 (𝑥) = − 1 ≤ 0,
𝑏
ensuring that the geometry of the design meets the fabrication and 𝑏
𝑔9 (𝑥) = − 1 ≤ 0,
operational requirements. 12𝑚
1.5𝑑1 + 1.9
Minimize 𝑔10 (𝑥) = − 1 ≤ 0,
𝑙1
𝑓 (𝑏, 𝑚, 𝑧, 𝑙1 , 𝑙2 , 𝑑1 , 𝑑2 ) = 1.1𝑑2 + 1.9
𝑔11 (𝑥) = − 1 ≤ 0,
0.7854𝑏𝑚𝑧2 (3.3333𝑧 + 14.9334) 𝑙2
− 43.0934 − 1.508𝑏(𝑑12 + 𝑑22 ) with bounds
+ 7.4777(𝑑13 + 𝑑23 ) + 0.7854(𝑙1 𝑑12 + 𝑙2 𝑑22 ) 2.6 ≤ 𝑏 ≤ 3.6,
0.7 ≤ 𝑚 ≤ 0.8,
subject to
17 ≤ 𝑧 ≤ 28,
27
𝑔1 (𝑥) = − 1 ≤ 0, 7.3 ≤ 𝑙1 ≤ 8.3,
𝑏𝑚𝑧2
397.5 7.8 ≤ 𝑙2 ≤ 8.3,
𝑔2 (𝑥) = − 1 ≤ 0,
𝑏𝑚𝑧2 𝑧2 2.9 ≤ 𝑑1 ≤ 3.9,
1.93𝑑13 where
𝑔3 (𝑥) = − 1 ≤ 0,
𝑚𝑧𝑑12 745𝑙1
𝑀= ,
1.93𝑑23 𝑚𝑧
𝑔4 (𝑥) = − 1 ≤ 0, 745𝑙2
𝑚𝑧𝑑22 𝐻= .
𝑚𝑧
√
𝑀 2 + 16.9 × 106 In the experiment the population size of CTCM is set to 40 and the
𝑔5 (𝑥) = − 1 ≤ 0, maximum number of iterations is 1000, the convergence and statistical
110𝑑13
performance is shown in Fig. 11. The convergence curve of CTCM drops
rapidly at the beginning, ahead of the other algorithms, and converges

15
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

to near the optimum in close to 20 iterations. The statistical results derived from the statistics, it can be found that CTCM’s box is on the
graph indicates that CTCM’s box is in the lead, and although there is lower side of all the algorithms, which means that the optimal solution
one outlier that is far apart, it does not detract from the fact that in found by CTCM is more advantageous. Although SSA and PSO have
most cases CTCM outperforms the other algorithms. somewhat better averages than CTCM, SSA and PSO have some very
The statistics of its performance are presented in Table 16. The opti- poor outliers, while CTCM is significantly more stable.
mal value retrieved by CTCM is 2230.5739 and the optimal solution is Table 17 indicates in detail the statistical performance of CTCM
𝑥 = [3.5, 0.7, 17, 7.3, 7.8, 3.3502, 3.6099], and the constraints are all under with other algorithms, CTCM, PSO, SSA and WOA find the best optimal
constraints. In the statistical results of 30 experiments, CTCM found the solution. However, in terms of worst-case scenarios, CTCM finds a
optimal solution, SSA found the best solution on average, and ESOA solution of 4735.7, while PSO, SSA and WOA find solutions of 5062.5,
was the best in the worst case as well as in standard deviation. Although 5060.2 and 5724.0, respectively, which is much worse than CTCM’s
CTCM is not as stable as ESOA in this application scenario, CTCM is able performance. ESOA, although a bit better than CTCM in terms of stabil-
to quickly obtain the optimal solution to this problem. This shows that ity, finds the best solution of 3377.2 The best solution found by ESOA is
CTCM is still able to show strong global optimization capability and 3377.2, while the best solution found by CTCM is 3368.4, which is infe-
high stability in the face of complex engineering optimization problems rior to CTCM. The best solution is 𝑥 = [0.9015, 0.1000, 46.7106, 126.7901],
with constraints. and the constraints are 𝑔 = [0, −64.603, 0, −230]. This experimental data
demonstrates that CTCM has outstanding global optimization seeking
4.3. Pressure vessel design problem capability while possessing high stability to run stably on practical
engineering problems.
Pressure Vessel Design Problem is a typical nonlinear constrained
optimization problem in the field of engineering optimization and is 4.4. Welded beam design problem
widely utilized to test various optimization algorithms, and the details
are presented in Fig. 12. The objective of the problem is to design Welded Beam Design Problem is one of the important nonlinear
a pressure vessel that minimizes its total manufacturing cost while constrained optimization problems in the field of engineering optimiza-
satisfying constraints on strength, geometry, pressure, and safety. A tion. The objective of the problem is to minimize the total fabrication
pressure vessel is a closed vessel subjected to internal or external cost of the welded beam by optimizing the design parameters while
pressure and is widely used in industries such as petrochemical, energy, satisfying several constraints in the design such as stress, stiffness, and
and food processing. When designing a pressure vessel, the strength, geometric constraints. Welded beams are widely applied in industrial
safety and manufacturing cost of its structure must be considered. and mechanical design to support and carry heavy loads. Optimizing
The manufacturing cost of a pressure vessel consists of material cost, the design of welded beams is important to ensure adequate strength,
welding cost and assembly cost. The shape of a pressure vessel is usually stiffness, and stability while minimizing manufacturing cost and mate-
a cylindrical shell with hemispherical heads at each end. This geometry rial usage. This involves balancing cost and performance by optimizing
can effectively withstand pressure and is easy to fabricate. The goal the beam size, material usage and weld length.
of the problem is to minimize the total cost of the pressure vessel, As it is illustrated as Fig. 14, there are 4 variables, which are the
including material, fabrication and welding costs. Typically, the design height of the bar (𝑡), the length of the attached bar (𝑙), the thickness
variables include the vessel’s shell thickness, head thickness, internal of the weld(ℎ), and the thickness of the bar (𝑏). The purpose of the
radius of the shell, and length. problem is to minimize the total cost of the welded beam, which
Pressure vessel design problem contains 4 design variables which includes material costs, welding costs, and other fabrication-related
are shell thickness 𝑧1 , head thickness 𝑧2 , inner radius of the packaging costs. The optimization problem has several constraints that cover
𝑧3 as well as length of the packaging 𝑧4 . The problem involves multiple the strength, stiffness, shear stress, and deflection requirements of the
design constraints that are used to ensure that the pressure vessel meets welded beam. The shear stress constraint ensures that the shear stress
safety and manufacturing standards during the design process. The con- does not exceed a limit of 13,600 psi. The bending stress constraint
straints can be expressed as follows: shell thickness and head thickness ensures that the bending stress does not exceed a limit of 30,000 psi.
must not be less than the manufacturing limitations; ensure that the Deflection constraints ensure that the maximum deflection of the beam
shell thickness is large enough to withstand the internal pressure; head does not exceed 0.25 inches. The mathematical modal is, Consider:
thickness is large enough to withstand the internal pressure; and ensure
that the vessel length does not exceed 240 inches. The mathematical 𝑥⃗ = [𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 ] = [ℎ, 𝑙, 𝑡, 𝑏]
modal is as below, Minimize:
Consider 𝑧⃗ = [𝑧1 𝑧2 𝑧3 𝑧4 ] = [𝑇𝑠 𝑇ℎ 𝑅𝐿],
Minimize 𝑥) = 1.10471𝑥21 𝑥2 + 0.04811𝑥3 𝑥4 (14.0 + 𝑥2 )
𝑓 (⃗

𝑓 (⃗
𝑧) = 0.6224𝑧1 𝑧3 𝑧4 + 1.7781𝑧2 Subject to:
𝑧23 + 3.1661𝑧21 𝑧4 + 19.84𝑧21 𝑧3 𝑔1 (⃗ 𝑥) − 𝜏max ≤ 0,
𝑥) = 𝜏(⃗
, 𝑔2 (⃗ 𝑥) − 𝜎max ≤ 0,
𝑥) = 𝜎(⃗
Subject to:
𝑔3 (⃗ 𝑥) − 𝛿max ≤ 0,
𝑥) = 𝛿(⃗
𝑧) = −𝑧1 + 0.0193𝑧3 ≤ 0,
𝑔1 (⃗
𝑥) = 𝑥1 − 𝑥4 ≤ 0,
𝑔4 (⃗
𝑧) = −𝑧3 + 0.00954𝑧3 ≤ 0,
𝑔2 (⃗
𝑔5 (⃗ 𝑥) ≤ 0,
𝑥) = 𝑃 − 𝑃𝑐 (⃗
4
𝑧) = −𝜋 𝑧23 𝑧4 − 𝜋 𝑧33 + 1, 296, 000 ≤ 0,
𝑔3 (⃗ 𝑥) = 0.125 − 𝑥1 ≤ 0,
𝑔6 (⃗
3
𝑧) = 𝑧4 − 240 ≤ 0.
𝑔4 (⃗ 𝑥) = 1.10471𝑥21 + 0.04811𝑥3 𝑥4 (14.0 + 𝑥2 ) − 5.0 ≤ 0
𝑔7 (⃗
The population size of CTCM in this experiment is set to 40, and Variables range:
the maximum iterations is 1000, the convergence and statistical per-
0.1 ≤ 𝑥1 ≤ 2,
formance is presented in Fig. 13. From the convergence graph, CTCM
carries out rapid convergence at the beginning of the iteration and is 0.1 ≤ 𝑥2 ≤ 10,
once ahead of the other algorithms, and converges to the neighborhood 0.1 ≤ 𝑥3 ≤ 10,
of the optimal solution after about 30 iterations. From the box plots 0.1 ≤ 𝑥4 ≤ 2

16
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

Fig. 12. Pressure vessel design problem.

Fig. 13. The convergence and statistical performance of algorithms on Pressure vessel design problem. As observed, the CTCM retrieved optimal value at around 30 iterations,
and performs excellent stability during the 30 times repeat experiments.

Table 17
The comparison of each algorithm on Pressure vessel design problem.
Metrics CTCM PSO GWO SSA ESOA BAS WOA
Best 3.3684E+03 3.3684E+03 3.3732E+03 3.3684E+03 3.3772E+03 9.5102E+05 3.3684E+03
Wrost 4.7357E+03 5.0625E+03 4.9430E+03 5.0602E+03 4.2796E+03 9.9327E+05 5.7240E+03
Mean 4.0477E+03 3.4813E+03 4.0060E+03 3.6651E+03 3.4794E+03 9.6862E+05 4.3182E+03
Std 5.1326E+02 4.2982E+02 4.3009E+02 6.1458E+02 2.0603E+02 1.2057E+04 4.1514E+02
𝑔1 −2.8866E−15 −7.7507E−08 −6.7724E−15 −1.3950E−08 −2.9964E−10 −2.8866E−15 −2.8866E−15
𝑔2 −6.4603E+01 −5.3056E+01 −4.7319E+01 −5.2484E+01 −5.1630E+01 −6.4603E+01 −6.4603E+01
𝑔3 0.0000E+00 −5.9351E+00 0.0000E+00 −2.0284E−06 −6.2774E−02 0.0000E+00 0.0000E+00
𝑔4 −2.3000E+02 −1.6766E+02 −1.2296E+02 −1.6374E+02 −1.5769E+02 −2.3000E+02 −2.3000E+02

where: 𝑃 = 6000 lb, 𝐿 = 14 in., 𝛿max = 0.25 in.,

√
𝑥2
𝑥) =
𝜏(⃗ (𝜏 ′ )2 + 2𝜏 ′ 𝜏 ′′
+ (𝜏 ′′ )2 , 𝐸 = 30 × 106 psi, 𝐺 = 12 × 106 psi,
2𝑅
𝑝 𝑀𝑅 𝜏max = 13600 psi, 𝜎max = 30000 psi
𝜏′ = √ , 𝜏 ′′ = ,
2𝑥1 𝑥2 𝐽
( In this experiment, the population size of CTCM is set to 40 and the
𝑥2 )
𝑀 =𝑃 𝐿+ , maximum iteration is set to 1000, and the convergence and statistical
2
√ performance is presented in Fig. 15. From the convergence graph, it can
( )
𝑥22 𝑥1 + 𝑥3 2
𝑅= + , be noticed that the fitness value of CTCM decreases rapidly at the very
4 2
{ [ 2 ( ) ]} beginning of the iteration, to be much ahead of the other algorithms,
√ 𝑥2 𝑥1 + 𝑥3 2
𝐽 =2 2𝑥1 𝑥2 + , and converges to the neighborhood area of the optimal value after close
4 2
to 40 iterations. From the box plots generated based on the statistical
6𝑃 𝐿 6𝑃 𝐿3 data, it can be seen that CTCM performs much better, which indicates
𝑥) =
𝜎(⃗ , 𝑥) =
𝛿(⃗ ,
𝑥4 𝑥23 𝐸 𝑥23 𝑥4 that CTCM has excellent solving ability and stability on this type of
√
𝑥23 𝑥64 ( √ ) problem.
4.013𝐸 36 𝑥3 𝐸
𝑃𝑐 (⃗
𝑥) = 1− The statistical results are exhibited in Table 18, CTCM is ahead of
𝐿2 2𝐿 4𝐺
the other algorithms in both the best, worst and average cases. In the

17
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

Fig. 14. Welded beam design problem.

Fig. 15. The convergence and statistical performance of algorithms on Welded beam design problem. As observed, the CTCM retrieved optimal value at around 40 iterations, and
performs excellent stability during the 30 times repeat experiments.

Table 18
The comparison of each algorithm on Welded beam design problem.
Metrics CTCM PSO GWO SSA ESOA BAS WOA
Best 1.2783E+00 1.2783E+00 1.3048E+00 1.2783E+00 1.2793E+00 1.9132E+00 1.4689E+00
Wrost 1.2783E+00 1.2783E+00 2.4395E+00 1.2783E+00 1.3139E+00 4.9240E+00 4.4510E+00
Mean 1.2783E+00 1.2783E+00 1.5391E+00 1.2783E+00 1.2839E+00 3.7192E+00 3.3370E+00
Std 0 9.6321E−10 2.0958E−01 6.7005E−07 7.2396E−03 1.2148E+00 1.0781E+00
𝑔1 −2.5466E−11 −2.8795E−09 −8.8585E−10 −5.8935E−10 −2.8194E−10 −1.4188E−10 −3.6380E−12
𝑔2 −4.7294E−11 −1.5571E−09 −1.7390E−09 −7.6398E−11 −5.8888E−08 −2.1828E−11 −8.5736E−08
𝑔3 −5.4000E−02 −5.4000E−02 −5.4000E−02 −5.4000E−02 −5.4000E−02 −5.4000E−02 −5.4000E−02
𝑔4 0.0000E+00 −1.2079E−12 −4.5475E−13 0.0000E+00 −2.8422E−14 −1.2079E−13 −2.1316E−14
𝑔5 −4.5566E−10 −2.7285E−12 −5.6752E−10 −5.4415E−09 −4.4929E−10 −9.5666E−08 −9.0949E−13
𝑔6 −8.0730E−02 −8.0730E−02 −8.0730E−02 −8.0730E−02 −8.0730E−02 −8.0730E−02 −8.0730E−02
𝑔7 −3.6839E+00 −3.6839E+00 −3.6839E+00 −3.6839E+00 −3.6839E+00 −3.6839E+00 −3.6839E+00

best case, CTCM, PSO and SSA all find the global optimal solution to the 4.5. Rolling element bearing design problem
problem for 1.2783, while ESOA finds the best result for just 1.2793.
In the worst case, CTCM still finds the global optimal solution to the
Rolling element bearing design problem is a essential optimization
problem, while ESOA finds the optimal solution for 1.3139. In terms of
problem in the field of mechanical engineering, which aims to optimize
stability, CTCM has a standard deviation of 0, while PSO and SSA have
the geometry and material properties of a bearing so that the bearing
standard deviations of 9.6321𝐸−10 and 6.7005𝐸−07, respectively, which
meets the load, life and performance requirements while minimizing its
are somewhat worse than CTCM. The data illustrates that in such cases,
cost or weight. Rolling bearings are critical components widely used in
CTCM shows a great advantage over other algorithms in terms of both
rotating equipment, so it is important to optimize their design. A rolling
accuracy, optimization seeking ability and stability.
bearing consists of an inner ring, an outer ring, a cage and a rolling

18
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

Fig. 16. Rolling element bearing design problem.

Fig. 17. The convergence and statistical performance of algorithms on Rolling element bearing design problem. As observed, the CTCM retrieved optimal value at around 30
iterations, and performs excellent stability during the 30 times repeat experiments.

Table 19
The comparison of each algorithm on Rolling element bearing design problem.
Metrics CTCM PSO GWO SSA ESOA BAS WOA
Best 1.4614E+04 1.4614E+04 1.4628E+04 1.4614E+04 1.4620E+04 2.0286E+04 1.4668E+04
Wrost 1.4701E+04 1.5497E+04 7.8868E+04 1.5497E+04 1.4665E+04 1.0468E+05 1.9113E+04
Mean 1.4667E+04 1.4724E+04 1.6810E+04 1.4711E+04 1.4638E+04 7.8132E+04 1.5323E+04
Std 3.8351E+01 1.4686E+02 1.1721E+04 2.1407E+02 1.1013E+01 2.8534E+04 9.4875E+02
𝑔1 −2.6390E+00 −2.6289E+00 −2.6357E+00 −2.6357E+00 −2.6212E+00 −2.6212E+00 −2.6390E+00
𝑔2 −1.0000E+00 −2.3461E+00 −7.7726E+00 −8.0000E+00 −1.8249E+00 −2.3018E+00 −1.0000E+00
𝑔3 −6.0000E+00 −7.6453E+00 −1.3000E+01 −1.3000E+01 −6.0000E+00 −6.0000E+00 −6.0000E+00
𝑔4 0.0000E+00 0.0000E+00 0.0000E+00 0.0000E+00 0.0000E+00 0.0000E+00 0.0000E+00
𝑔5 −6.2000E+00 −2.6000E+00 −5.0000E+00 −5.0000E+00 0.0000E+00 0.0000E+00 −6.2000E+00
𝑔6 −1.8800E+01 −1.3017E+01 −1.1369E−13 −1.4211E−13 −2.0800E+01 −5.0000E+00 −1.8800E+01
𝑔7 −6.5725E−14 −1.1102E−13 −6.0000E−01 −6.0000E−01 −1.8627E+00 −3.0995E+00 −6.5725E−14
𝑔8 −8.5000E−02 −8.5000E−02 −8.5000E−02 −8.5000E−02 −8.5000E−02 −8.5000E−02 −8.5000E−02
𝑔9 −8.5000E−02 −8.5000E−02 −8.5000E−02 −8.5000E−02 −8.5000E−02 −8.5000E−02 −8.5000E−02

19
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

Table 20
Performance of each algorithm on test functions.
CTCM PSO GWO SSA ESOA BAS WOA

Best Ave Std Best Ave Std Best Ave Std Best Ave Std Best Ave Std Best Ave Std Best Ave Std

F1 4.92E−15 8.18E−14 1.03E−13 6.14E−11 1.86E−09 3.25E−09 0.3304 3.98E+03 9.64E+03 4.50E−12 5.59E−11 8.43E−11 7.5093 9.9819 1.9367 5.73E+04 6.70E+04 6.18E+03 2.4673 26.1331 16.8817
F2 1.47E−09 6.96E−09 5.92E−09 1.19E−06 1.09E−04 2.36E−04 3.5435 12.1365 10.6959 5.40E−08 2.53E−07 2.84E−07 3.7634 7.9171 2.5145 3.77E+07 3.07E+12 6.69E+12 0.8036 1.7979 0.7519
F3 1.9409 6.6395 3.8029 115.2618 1.48E+03 2.44E+03 6.89E+03 1.70E+04 9.64E+03 58.6546 167.1205 117.4013 6.51E+03 8.95E+03 1.60E+03 8.50E+04 1.45E+05 6.30E+04 4.96E+04 8.30E+04 2.05E+04
F4 0.0783 0.2315 0.1114 2.0379 3.5107 0.8421 0.5359 0.6582 0.1367 12.6571 20.1959 3.8165 15.5412 23.165 4.5988 76.495 85.0754 4.2196 56.9607 73.8318 7.1645
F5 5.6493 20.3366 7.0706 16.3092 91.2572 172.8688 149.8239 590.4524 578.5083 0.0021 36.4291 35.6431 199.07 370.4274 87.5191 1.68E+08 2.21E+08 3.97E+07 0.5777 66.4687 177.3516
F6 8.82E−15 5.57E−14 6.25E−14 9.43E−11 1.45E−09 3.55E−09 0.4756 1.01E+03 3.18E+03 2.89E−12 6.23E−11 8.00E−11 5.918 10.0699 2.3821 4.72E+04 6.31E+04 8.24E+03 9.9062 20.2567 9.1056
F7 0.0072 0.0115 0.003 0.0061 0.0256 0.0094 0.6041 1.3815 0.6298 0.0196 0.0428 0.022 0.3681 0.4645 0.0812 83.4253 117.7015 23.0799 0.0758 0.2107 0.1174
F8 −8.56E+03 −7.10E+03 794.6067 −9.82E+03 −9.04E+03 620.932 −7.09E+03 −6.57E+03 697.3462 −9.33E+03 −8.73E+03 400.3225 −1.05E+04 −9.95E+03 299.6899 −3.47E+03 −2.68E+03 601.7448 −1.26E+04 −1.14E+04 1.84E+03
F9 38.8034 53.5287 11.8637 32.8336 52.8394 18.915 165.77 221.9815 42.9097 95.5158 139.8927 18.9774 72.7989 96.6822 18.8388 422.7171 439.8682 18.6023 0.8381 84.1891 92.8376
F10 2.59E−08 5.10E−08 2.29E−08 3.58E−07 4.02E−04 1.20E−03 20.0193 20.1915 0.0727 9.40E−09 2.75E−06 6.06E−06 19.9631 19.9655 0.0013 20.4777 20.6589 0.1277 0.2267 1.991 1.2318
F11 1.79E−14 0.0043 0.0069 4.89E−10 0.0123 0.0141 0.0762 4.3064 2.7536 1.11E−16 0.0069 0.0143 1.0525 1.0787 0.0152 410.5167 555.6157 78.0869 0.9964 1.0381 0.0312
F12 1.15E−15 4.15E−02 1.31E−01 3.01E−12 2.42E−09 6.20E−09 0.0493 0.1345 0.1349 1.35E−05 0.7446 0.968 2.2093 4.1321 0.9744 3.34E+08 5.39E+08 1.38E+08 0.0553 1.7623 2.9967
F13 2.95E−15 0.0011 0.0035 2.37E−10 0.0054 0.0075 0.0618 0.1333 0.0538 9.89E−11 0.2131 0.636 2.6197 6.2576 2.1522 7.68E+08 1.13E+09 2.06E+08 0.1674 0.7453 0.6685
F14 0.998 3.0698 2.4741 0.998 0.998 0 0.998 0.998 2.56E−10 0.998 0.998 7.40E−17 0.998 1.9075 0.7473 1.2003 32.1047 43.3888 0.998 1.0974 0.3143
F15 3.07E−04 3.07E−04 8.36E−19 3.07E−04 2.60E−03 6.30E−03 0.0012 0.0014 1.40E−04 3.08E−04 0.0026 0.0063 6.68E−04 8.58E−04 1.11E−04 0.0187 0.1089 0.0838 9.77E−04 0.0077 0.0079
F16 −1.0316 −1.0316 0 −1.0316 −1.0316 7.40E−17 −1.0316 −1.0308 5.98E−04 −1.0316 −1.0316 0 −1.0316 −1.0316 5.23E−10 −1.0091 −0.5169 0.5133 −1.0316 −1.0316 8.94E−05
F17 0.3979 0.3979 0 0.3979 0.3979 0 0.3979 0.4027 8.90E−03 0.3979 0.3979 0 0.3979 0.3979 2.33E−10 0.4069 1.2245 1.0299 0.3979 0.3979 3.09E−06
F18 3 3 0 3 3 0 3.0089 3.0407 0.0303 3 3 0.00E+00 3 3 1.19E−08 6.667 34.0777 24.0755 3 5.7825 8.7989
F19 −3.8628 −3.8628 0 −3.8628 −3.8628 9.36E−16 −3.8613 −3.8497 0.0121 −3.8628 −3.8628 7.83E−16 −3.8628 −3.8628 2.09E−10 −3.8312 −3.5905 0.2155 −3.8627 −3.7991 0.1149
F20 −3.322 −3.322 4.68E−16 −3.322 −3.2548 0.0888 −3.1652 −3.0826 0.0668 −3.322 −3.2625 0.0627 −3.322 −3.322 4.97E−07 −2.7966 −2.1473 0.5252 −3.2439 −3.1308 0.0936
F21 −10.1532 −6.8964 3.55E+00 −10.1532 −7.4016 3.6212 −10.11 −7.7497 3.0383 −10.1532 −7.38 3.0146 −10.1531 −10.1526 5.53E−04 −0.9755 −0.6601 0.2043 −10.1133 −10.0388 0.0897
F22 −10.4029 −7.6761 3.5769 −10.4029 −9.6378 2.4195 −10.3595 −8.8006 2.995 −10.4029 −6.8237 3.8633 −10.4029 −10.4023 7.00E−04 −0.721 −0.6238 0.0777 −10.4028 −10.2337 0.1563
F23 −10.5364 −5.6396 3.4629 −10.5364 −8.186 3.7862 −10.4509 −10.3674 0.0978 −10.5364 −6.1245 3.8875 −10.5362 −10.5361 1.29E−04 −1.6397 −1.0303 0.4153 −10.5198 −10.3446 0.2453

element, and is mainly used to carry radial and axial loads and to where:
[ { ( )1.72
reduce friction in rotary motion. When designing a rolling bearing, it is 1−𝛾
necessary to balance a variety of factors such as load carrying capacity, 𝑓𝑐 = 37.91 1 + 1.04
1+𝛾
stiffness, life, vibration, material, friction, etc. The goal is usually to
optimize certain design parameters to minimize cost, weight or friction ( )0.41 }10∕3 ⎤−0.3
𝑓𝑖 (2𝑓𝑜 − 1) ⎥
losses while satisfying various constraints, for instances, load, life, and ×
𝑓𝑜 (2𝑓𝑖 − 1) ⎥
bearing size (see Fig. 17). ⎦
𝐷𝑏 𝑟𝑖 𝑟𝑜
𝛾= , 𝑓𝑖 = , 𝑓𝑜 = ,
This engineering optimization problem has a total of 10 variables 𝐷𝑚 𝐷𝑏 𝐷𝑏
[ ]2 { }
and aims to satisfy the ability to maximize its dynamic loading under (𝐷 − 𝑑)∕2 − 3(𝑇 ∕4) 𝐷∕2 − 𝑇 ∕4 − 𝐷𝑏 2
𝑥= + ,
geometric constraints, and the details can be found in Fig. 16. 𝑍 𝑇 ∕4 𝑑∕2 + 𝑇 ∕4
means the number of rolling elements, 𝐷𝑏 is the ball diameter, 𝐶𝑑 [ ]{ }
(𝐷 − 𝑑)∕2 − 3(𝑇 ∕4) 𝐷∕2 − 𝑇 ∕4 − 𝐷𝑏
is the dynamic load rating while 𝐷𝑚 presents pitch diameter. 𝐷 is 𝑦=2 ,
𝑇 ∕4 𝑑∕2 + 𝑇 ∕4
the outer diameter of bearing, 𝑑 means bore diameter of bearing, ( )
𝑥
while 𝐵𝑤 presents the width of bearing. The constraints in this prob- 𝜑0 = 2𝛱 − cos−1 ,
𝑦
lem include physical reality constraints, design parameter constraints,
𝑇 = 𝐷 − 𝑑 − 2𝐷𝑏 , 𝐵𝑤 = 30, 𝐷 = 160, 𝑑 = 90,
stress constraints, stability constraints, and geometric constraints. The
mathematical modal is as below, 𝑟𝑖 = 𝑟𝑜 = 11.033, 𝐷𝑏 = 11.033, 𝐷𝑚 = 22.066,
𝜁 = 0.75, 𝐾𝐷 min = 0.4, 𝐾𝐷 max = 0.7, 𝑒 = 0.1
Maximize 𝐶𝑑 defined as: In this experiment, the population size of CTCM is set to 40 and
⎧ 2∕3 1.8 the maximum iteration is set to 1000, and Fig. 15 indicates the con-
⎪𝑓 𝑐 𝑍 𝐷 𝑏 if 𝐷 ≤ 25.4 mm
𝐶𝑑 = ⎨ vergence and statistical performance. The convergence plot of the
⎪3.647𝑓𝑐 𝑍 2∕3 𝐷𝑏1.4 if 𝐷 > 25.4 mm fitness function illustrates that CTCM possesses a superiority over other
⎩
algorithms for this kind of problem, in terms of convergence speed,
CTCM reaches the neighborhood of the global optimal solution at about
30 iterations. The box plot indicates that CTCM outperforms most of the
Subject to: other algorithms both in terms of average performance and stability.
𝜑0
𝑔1 (⃗
𝑧) = − 𝑍 + 1 ≤ 0, ESOA might be better in terms of stability, however CTCM produces
2 sin−1 (𝐷𝑏 ∕𝐷𝑚 ) significantly better optimal solutions than ESOA.
𝑔2 (⃗
𝑧) = 2𝐷𝑏 − 𝐾𝐷 min (𝐷 − 𝑑) > 0, Statistical Table 19 better illustrates the advantages of CTCM, the
optimal value retrieved by CTCM is 14701 while the average value is
𝑧) = 𝐾𝐷 max (𝐷 − 𝑑) − 2𝐷𝑏 ≥ 0,
𝑔3 (⃗ 14667 with the standard deviation 40.0790. The best solution is 𝑥 =
[125.0000, 18.0000, 4.0000, 0.6000, 0.6000, 0.5000, 0.6001, 0.3547, 0.0663,
𝑧) = 𝜁 𝐵𝑤 − 𝐷𝑏 ≤ 0,
𝑔4 (⃗
0.6000] under the whole constraints. In terms of the best results, only
𝑧) = 𝐷𝑚 − 0.5(𝐷 + 𝑑) ≥ 0,
𝑔5 (⃗ CTCM, PSO and SSA find globally optimal solutions. In terms of the
worst results, CTCM finds a solution of 14701, PSO and SSA both find
𝑧) = (0.5 + 𝑒)(𝐷 + 𝑑) − 𝐷𝑚 ≥ 0,
𝑔6 (⃗ a solution of 15 497, which seems to be trapped in a local optimum.
𝑧) = 0.5(𝐷 − 𝐷𝑚 − 𝐷𝑏 ) − 𝑒𝐷𝑏 ≥ 0,
𝑔7 (⃗ In terms of algorithmic stability, ESOA is the most stable, followed
by CTCM, however, CTCM has a stronger ability to find the global
𝑧) = 𝑓𝑖 ≥ 0.515,
𝑔8 (⃗
optimum. The experimental data demonstrates that CTCM balances
𝑧) = 𝑓𝑜 ≥ 0.515
𝑔9 (⃗ stability with excellent optimization capability.

20
Z. Chen et al. Expert Systems With Applications 265 (2025) 125908

5. Conclusion and future work Chen (2024). Competition of tribes and cooperation with members (CTCM).
https://se.mathworks.com/matlabcentral/fileexchange/168661-competition-of-
tribes-and-cooperation-with-members-ctcm. (Accessed 27 June 2024).
In this paper, optimization algorithms, Competition of Tribes and
Chen, Zuyan, Francis, Adam, Li, Shuai, Liao, Bolin, Xiao, Dunhui, Ha, Tran, et
Cooperation of Members Algorithm, which mimic human tribes, for al. (2022). Egret swarm optimization algorithm: An evolutionary computation
the seek of more favorable living environments. The mathematical approach for model free optimization. Biomimetics (Basel), 7(4), 144.
model and system framework of the algorithm were constructed. In Chuang, Li-Yeh, Chang, Hsueh-Wei, Tu, Chung-Jui, & Yang, Cheng-Hong (2008). Im-
proved binary PSO for feature selection using gene expression data. Computational
order to test the performance of the algorithm, CTCM is applied on
Biology and Chemistry, 32(1), 29–38.
23 benchmark functions including unimodal, multimodal and fixed- Dorigo, Marco, Birattari, Mauro, & Stutzle, Thomas (2006). Ant colony optimization.
dimension cases. The performance is compared with other state-of-the- IEEE Computational Intelligence Magazine, 1(4), 28–39.
art algorithms, including PSO, GWO, SSA, ESOA, BAS, WOA, using Efferson, Charles, Bernhard, Helen, Fischbacher, Urs, & Fehr, Ernst (2024).
Matlab coding (Chen, 2024). The results indicate that CTCM possesses Super-additive cooperation. Nature, 626(8001), 1034–1041.
Himmelblau, David M. (1972). Applied nonlinear programming. New York, NY, USA:
a strong global optimization capability and stability on the benchmark
McGraw-Hill.
functions, and has a faster convergence speed compared with other Holland, John H. (1992). Genetic algorithms. Scientific American, 267(1), 66–73.
algorithms. Finally, CTCM demonstrates superior solution capability Huang, Cheng-Lung, & Dun, Jian-Fan (2008). A distributed PSO–SVM hybrid system
and stability on five real engineering optimization problems. with feature selection and parameter optimization. Applied Soft Computing, 8(4),
1381–1391.
Although the CTCM has demonstrated extremely superior perfor-
Jiang, Xiangyuan, & Li, Shuai (2017). BAS: Beetle antennae search algorithm for
mance in numerous situations, it has performed relatively unsatisfac- optimization problems. arXiv:1710.10724.
torily for a number of specific situations. In future work, this part of Karaboga, D., & Basturk, B. (2008). On the performance of artificial bee colony (ABC)
the CTCM’s shortcomings will be targeted for improvement, so that the algorithm. Applied Soft Computing, 8(1), 687–697.
CTCM can be applied to different kinds of situations. In addition, some Karaboga, Dervis, & Ozturk, Celal (2011). A novel clustering approach: Artificial bee
colony (ABC) algorithm. Applied Soft Computing, 11(1), 652–657.
of the existing deficiencies of the CTCM will be addressed to improve
Kennedy, J., & Eberhart, R. (1995). Particle swarm optimization. Vol. 4, In Proceedings
its overall performance. In addition, CTCM will be used in the opti- of iCNN’95 - international conference on neural networks (pp. 1942–1948 vol.4).
mization of more complex engineering problems, such as mechanical http://dx.doi.org/10.1109/ICNN.1995.488968.
structure optimization, production process optimization, sensor struc- Kennedy, J., & Eberhart, R. (2002). Particle swarm optimization. Vol. 4, In Proceedings
of iCNN’95 - international conference on neural networks (pp. 1942–1948 vol.4). IEEE.
ture optimization and so on. Meanwhile, CTCMs for multi-objective
Kradin, Nikolay (2018). Ancient steppe nomad societies. Oxford University Press.
optimization problems will be developed. Lindner, Rudi Paul (1982). What was a nomadic tribe? Comparative Studies in Society
and History, 24(4), 689–711.
CRediT authorship contribution statement Mathew, Sarah (2024). Why reciprocity is common in humans but rare in other animals.
Nature, 626(8001), 955–956.
Mirjalili, Seyedali, & Lewis, Andrew (2016). The whale optimization algorithm.
Zuyan Chen: Paper writing, Software, Data collection. Shuai Advances in Engineering Software, 95, 51–67.
Li: Conceptualization, Methodology, Mathematical modeling. Ameer Mirjalili, Seyedali, Mirjalili, Seyed Mohammad, & Lewis, Andrew (2014). Grey wolf
Tamoor Khan: Algorithm framework, Data correction, Experiment. optimizer. Advances in Engineering Software, 69, 46–61.
Pan, Jeng-Shyang, Hu, Pei, Snášel, Václav, & Chu, Shu-Chuan (2023). A survey
Seyedali Mirjalili: Review, Supervision, Validation.
on binary metaheuristic algorithms and their engineering applications. Artificial
Intelligence Review, 56(7), 6101–6167.
Declaration of competing interest Puurtinen, Mikael, & Mappes, Tapio (2009). Between-group competition and human
cooperation. Proceedings: Biological Sciences, 276(1655), 355–360.
Rajwar, Kanchan, Deep, Kusum, & Das, Swagatam (2023). An exhaustive review of the
The authors declare that they have no known competing finan-
metaheuristic algorithms for search and optimization: taxonomy, applications, and
cial interests or personal relationships that could have appeared to open challenges. Artificial Intelligence Review, 56(11), 13187–13257.
influence the work reported in this paper. Reynolds, Craig W. (1987). Flocks, herds and schools: A distributed behavioral model.
Computer Graphics (ACM), 21(4), 25–34.
Rostami, Mehrdad, Berahmand, Kamal, Nasiri, Elahe, & Forouzandeh, Saman
Data availability (2021). Review of swarm intelligence-based feature selection methods. Engineering
Applications of Artificial Intelligence, 100(104210), Article 104210.
Data will be made available on request. Tang, Jun, Duan, Haibin, & Lao, Songyang (2023). Swarm intelligence algorithms for
multiple unmanned aerial vehicles collaboration: a comprehensive review. Artificial
Intelligence Review, 56(5), 4295–4327.
Tang, Jun, Liu, Gang, & Pan, Qingtao (2021). A review on representative swarm
References
intelligence algorithms for solving optimization problems: Applications and trends.
IEEE/CAA Journal of Automatica Sinica, 8(10), 1627–1643.
Abdallah Altbawi, Saleh Masoud, Abdul Khalid, Saifulnizam, Safawi Mokhtar, Ahmad, Wolpert, D. H., & Macready, W. G. (1997). No free lunch theorems for optimization.
Hamza Alsisi, Rayan, Ahmad Arfeen, Zeeshan, Shareef, Hussain, et al. (2023). IEEE Transactions on Evolutionary Computation, 1(1), 67–82.
Improve power quality of charging station unit using african vulture optimization Xue, Jiankai, & Shen, Bo (2020). A novel swarm intelligence optimization approach:
algorithm. Bulletin of Electrical Engineering and Informatics, 12(5), 2605–2614. sparrow search algorithm. Systems Science & Control Engineering, 8(1), 22–34.
Altbawi, Saleh Masoud Abdallah, Khalid, Saifulnizam Bin Abdul, Mokhtar, Ahmad Yang, Xin She (2010). Firefly algorithm, stochastic test functions and design
Safawi Bin, Shareef, Hussain, Husain, Nusrat, Yahya, Ashraf, et al. (2023). An optimisation. International Journal of Bio-Inspired Computation, 2(2), 78.
improved gradient-based optimization algorithm for solving complex optimization Yang, Xin-She (Ed.), (2016). Cuckoo search and firefly algorithm: Theory and applications.
problems. Processes (Basel), 11(2), 498. Cham, Switzerland: Springer International Publishing.
Altbawi, Saleh Masoud Abdallah, Mokhtar, Ahmad Safawi Bin, Jumani, Touqeer Ahmed, Yang, Xin-She, & Hossein Gandomi, Amir (2012). Bat algorithm: a novel approach
Khan, Ilyas, Hamadneh, Nawaf N, & Khan, Afrasyab (2024). Optimal design of for global engineering optimization. Engineering with Computers (Swansea), 29(5),
fractional order PID controller based automatic voltage regulator system using 464–483.
gradient-based optimization algorithm. Journal of King Saud University - Engineering Yi, Ziwei, Cao, Xinwei, Pu, Xujin, Wu, Yiding, Chen, Zuyan, Khan, Ameer Tamoor, et
Sciences, 36(1), 32–44. al. (2023). Fraud detection in capital markets: A novel machine learning approach.
Altbawi, Saleh Masoud Abdallah, Mokhtar, Ahmad Safawi Bin, Khalid, Saifulnizam Expert Systems with Applications, 231, Article 120760.
Bin Abdul, Husain, Nusrat, Yahya, Ashraf, Haider, Syed Aqeel, et al. (2023). Zhang, Jun, & Dolg, Michael (2015). ABCluster: the artificial bee colony algo-
Optimal control of a single-stage modular PV-grid-driven system using a gradient rithm for cluster global optimization. Physical Chemistry Chemical Physics, 17(37),
optimization algorithm. Energies, 16(3), 1492. 24173–24181.
Altbawi, Saleh Masoud Abdallah, et al. (2021). Enhacement of microgrid technologies
using various algorithms. Turkish Journal of Computer and Mathematics Education
(TURCOMAT), 12(7), 1127–1170.
Cameron, Catherine M. (2013). How people moved among ancient societies: Broadening
the view. American Anthropologist, 115(2), 218–231.

21