Adaptive Chaotic Marine Predators Hill Climbing Algorithm For Large-Scale Design Optimizations

Received 23 March 2023, accepted 4 April 2023, date of publication 13 April 2023, date of current version 25 April 2023.
Digital Object Identifier 10.1109/ACCESS.2023.3266991
Adaptive Chaotic Marine Predators Hill Climbing

Algorithm for Large-Scale Design Optimizations
AMIN ABDOLLAHI DEHKORDI 1 , BAHAREH ETAATI2 , MEHDI NESHAT 3,4 ,
AND SEYEDALI MIRJALILI3,5,6 , (Senior Member, IEEE)

1 Department of Computer Engineering, Islamic Azad University, Najafabad Branch, Najafabad 8514143131, Iran
2 JosefRessel Center for Symbolic Regression, University of Applied Sciences Upper Austria, 4232 Hagenberg, Austria
3 Center for Artificial Intelligence Research and Optimization, Torrens University Australia, Brisbane, QLD 4006, Australia
4 Faculty of Engineering and Information Technology, University of Technology Sydney, Ultimo, NSW 2007, Australia
5 Yonsei Frontier Lab, Yonsei University, Seoul 03722, South Korea
6 University Research and Innovation Center, Obuda University, 1034 Budapest, Hungary
Corresponding authors: Amin Abdollahi Dehkordi ([email protected]) and

Seyedali Mirjalili ([email protected])
ABSTRACT Meta-heuristic algorithms have been effectively employed to tackle a wide range of
optimisation issues, including structural engineering challenges. The optimisation of the shape and size
of large-scale truss structures is complicated due to the nonlinear interplay between the cross-sectional
and nodal coordinate pressures of structures. Recently, it was demonstrated that the newly proposed
Marine Predator Algorithm (MPA) performs very well on mathematical challenges. The MPA is a meta-
heuristic that simulates the essential hunting habits of natural marine predators. However, this algorithm
has some disadvantages, such as becoming locked in locally optimal solutions and not exhibiting high
exploratory behaviour. This paper proposes two hybrid marine predator algorithms, Nonlinear Marine
Predator (HNMPA) and Nonlinear-Chaotic Marine Predator Algorithm (HNCMPA), as improved variations
of the marine predator algorithm paired with a hill-climbing (HC) technique for truss optimisation on form
and size. The major advantage of these techniques is that they seek to overcome the MPA’s disadvantages
by using nonlinear values and prolonging the exploration phase with chaotic values; also, the HC algorithm
has been used to avoid locally optimum solutions. In terms of truss optimisation performance, the proposed
algorithm is compared to fourteen well-known meta-heuristics, including the Dragonfly Algorithm (DA),
Henry Gas Solubility optimisation (HGSO), Arithmetic optimisation Algorithm (AOA), Generalized Normal
Distribution Optimisation (GNDO), Salp Swarm Algorithm (SSA), Marine Predators Algorithm (MPA),
Neural Network Algorithm (NNA), Water Cycle Algorithm (WCA), Artificial Gorilla Troops Optimiser
(GTO), Gray Wolf Optimiser (GWO), Moth Flame Optimiser (MFO), Multi-Verse Optimiser (MVO),
Equilibrium Optimiser (EO), and Cheetah Optimiser (CO). Furthermore, seven algorithms were chosen
to test HNCMPA performance on benchmark optimisation sets, including MPA, MVO, PSO, MFO, SSA,
GWO, and WOA. The experiment results demonstrate that the optimisation techniques surpass previously
established meta-heuristics in the field of optimisation, encompassing both traditional and CEC problems,
by a margin of over 95% in terms of attaining a superior ultimate solution. Additionally, with regards
to solving truss optimisation difficulties as a large-scale real-world engineering challenge, the outcomes
indicate a performance boost of over 65% in obtaining significantly better solutions for problems involving
260-bar and 314-bar; conversely, in the case of 340-bar issues, the improvement rate is slightly lower at
almost 25%.
INDEX TERMS Bio-inspired algorithms, hybrid algorithms, Marine predators algorithm, meta-heuristics,
optimization method, truss optimization.
I. INTRODUCTION
The associate editor coordinating the review of this manuscript and The development of various meta-heuristic optimisation tech-
approving it for publication was Zhenzhou Tang . niques over the last few decades has resulted in a significant
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 11, 2023 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 39269
A. A. Dehkordi et al.: Adaptive Chaotic Marine Predators Hill Climbing Algorithm
increase in their application in a wide range of engineering Algorithm (CS) [21], and Dynamic Arithmetic optimisation
problems. These techniques explore the search area in an Algorithm (DAOA) [22].
optimisation problem without the need for a gradient, quasi- Natural frequency is a critical criterion in the issue of
randomly, while keeping some fundamental principles for truss optimisation because it has a significant impact on the
finding the optimal solution in mind. Their popularity in performance of a structure. This criterion is based on the
a variety of fields can be attributed to characteristics such understanding of structural dynamics, a field of study that
as improved performance and efficiency, lower computing provides critical information about the dynamic behaviour
capacity, and easier implementation than fixed algorithms. of structures. Over the last decade, researchers have paid
These algorithms rely on simple principles to produce the close attention to the problem of truss optimisation based
best results, and the search is conducted to find the best on frequency constraints. A significant practical challenge
results in a variety of fields. The use of random elements for the truss problem is to improve its dynamic behaviour
in the structure of meta-heuristic algorithms allows these by taking into account its natural frequencies. Better control
algorithms to explore the entire search area to find the best of this criterion improves the structure’s performance and
solution and reduces the risk of these algorithms becoming prevents the phenomenon of resonance.
trapped in local optimisation. The design and construction of lightweight structures is
One of their notable shortcomings is the low convergence a difficult and critical issue in engineering. The conflict
rate of meta-heuristic algorithms in the face of direct and between weight minimization and frequency constraints com-
straightforward problems. In other words, the use of gradi- plicates the truss optimisation problem even more. As a result
ent information in specific descending gradient algorithms of the perceived need for effective optimisation techniques to
is far more helpful in dealing with these issues than these design these structures and deal with their inherent limita-
algorithms [1]. tions, an active research field has emerged in recent years.
Human, natural, physical, and artistic phenomena are the Researchers are expanding their knowledge and understand-
four categories of meta-heuristic algorithms used in optimi- ing of the subject as research expands. Some of these cases
sation problems. However, because many optimisation prob- are discussed further below. Bellagamba and Yang [23] were
lems have an infinite and limitless solution space, using pioneers in the field of truss optimisation with frequency
basic versions of meta-heuristic algorithms to explore the constraints. Lin et al. [24] developed a bi-factor algorithm for
solution space of these problems may fail to find optimal these structures. Wei et al. [25] proposed a parallel genetic
solutions. algorithm. Charged system search and enhanced CSS [26],
Some meta-heuristic optimisation algorithms inspired by democratic particle swarm optimisation (DPSO) [27], and tug
animal social behaviour or human characteristics that have of war optimisation (TWO) [28] were proposed by Kaveh
been used in various fields include Genetic Algorithm and Zolghadr. Pholdee and Bureerat [29] experimented
(GA) [2], Particle Swarm optimisation (PSO) [3], Ant Colony with various metaheuristic algorithms for this problem.
optimisation (ACO) [4], Imperialist Competitive Algorithm Tejani et al. [19] proposed the search for symbiotic organisms
(ICA) [5], Gray Wolf Optimiser (GWO) [6], Whale opti- (ISOS) in truss structures with a low frequency of occurrence.
misation Algorithm (WOA) [7], and Human Mental Search A learning algorithm based on multi-class training was used
(HMS) [8]. for truss structures with frequency constraints. In a recent
The Gravitational Search Algorithm (GSA) [9], Charged study, the efficiency of the generalized normal distribution
System Search (CSS) [10], Stochastic Paint Optimiser optimisation (GNDO) algorithm [30] was evaluated and com-
(SPO) [11], Simulated Annealing (SA) [12], Harmony Search pared with seven metaheuristic algorithms. The experimental
(HS) [13], and Colliding Bodies optimisation (CBO) [14] results demonstrated that GNDO performed better than other
were all motivated by art, physical, and natural phenomena. latest algorithms regarding convergence speed and optimal
When faced with optimisation problems, meta-heuristic algo- solutions. However, the authors did not test the GNDO’s per-
rithms approximate the optimal solution by either randomly formance for structures larger than 200 bar. In order to over-
identifying the expected optimal solution region or by creat- come the shortcomings of the original seagull optimisation
ing better solutions to these problems with fewer constraints algorithm (SOA), a strategy combined with mutualism and
and fewer computational resources than descending gradient commensalism was proposed [31]. A wide range of numerical
methods [15]. Due to its inherent and challenging difficulty, benchmarks and engineering problems were used to evaluate
truss optimisation has become a difficult issue in structural the improved SOA algorithm’s performance, outperforming
engineering in recent decades. Many researchers have exper- other methods.
imented with meta-heuristic techniques to optimise the size Research in this field has confirmed the effectiveness of
and arrangement of structures. Some of these techniques meta-heuristic optimisation algorithms in dealing with and
are: Genetic Algorithm (GA) [16], Particle Swarm optimi- managing many problems when solving a structural design
sation (PSO) [17], School-Based optimisation (SBO) [18], problem. Given the No Free Launch (NFL) Theorem [32],
Symbiotic Bodies optimisation (SBO) [19], Dynamic no single optimisation technique can solve all optimisation
Water Strider Algorithm (DWSA) [20], Cuckoo Search problems. As a result, developing a new modified algorithm
39270 VOLUME 11, 2023

and improving its performance can improve the algorithm’s described in the literature. To boost MPA’s performance, two
ability to handle a set of problems compared to existing new features were successfully added to the basic version
algorithms. Simultaneously, the method for dealing with opti- in the first phase (Nonlinear and chaotic parameters), and
misation problems in these methods will not change. This the hill-climbing method was integrated with the suggested
is the motivation for our efforts to improve the efficiency of algorithm in the second step to tackle truss optimisation prob-
the MPA optimisation algorithm, which has recently demon- lems more effectively. The optimisation of trusses consid-
strated excellent performance in optimisation problems, and ered a complex and practical engineering problem, presents
to adapt it to structural design problems better. a significant challenge for meta-heuristic algorithms due to
Faramarzi et al. [33] recently proposed a marine predator extensive local optima caused by the problems’ inherent com-
optimisation algorithm based on the natural behaviour of plexity. As a result, merely having efficient exploration and
marine predators for constrained and unconstrained optimi- exploitation approaches is inadequate to avoid getting trapped
sation problems. Despite the thriving performance of MPO, in these local optima. This study proposes hybrid methods
there are some technical areas for improvement in optimising that integrate hill-climbing, a renowned local search method,
real engineering problems. For example, the prey particles to address this issue and enable the algorithms to move
yield their route and converge on the predator particles in beyond the local optima and approach the global optimum
some multi-modal cases [34]. Shaheen et al. [34] proposed an point. HNMPA and HNCMPA continue to apply to various
improved MPO method to provide a better chance to prevent groups of optimisation issues because no additional fine-
falling into local optimum by determining a random oppor- tuning of MPA parameters is necessary to connect with these
tunity to combine three search stages. In another study [35], new versions.
applying a differential operator is recommended to enhance To make a long story short, the primary contributions of
the exploration phase of the standard MPA. To improve the this study are as follows:
performance of MPA in solving a nonlinear problem, the • Proposing a new adaptive MPA algorithm employs an
tracking the global MPP of shaded PV ystems [36], a com- efficient nonlinear control parameter to achieve a good
bination of the Opposition Based Learning (OBL) method balance between exploration and exploitation strategies,
with Grey Wolf Optimiser (GWO) was proposed to control as well as chaotic values to improve the exploration
the searchability of MPA and acquire more rapid convergence phase by providing the proposed algorithm’s population
rate. with a sufficient degree of diversity, resulting in excel-
Although multi-solution meta-heuristics have some intrin- lent accuracy and rapid convergence.
sic limitations, the literature suggests that these algorithms • Enhancing the performance of the proposed technique
are now the primary method for solving optimisation issues. by employing a Hill Climbing (HC) local search algo-
This research area is among the most popular in computa- rithm to accelerate convergence and avoid trapping at
tional intelligence and has lately presented several methods. local optimal points, hence resolving truss optimisation
Having such a potential possibility to approach a solution concerns.
indicated by a computationally effective optimisation algo- • The optimisation results of several architectures with
rithm capable of finding the best solution somewhat and nonlinear and dynamic constraints are thoroughly ana-
efficiently has encouraged the development of an upgraded lyzed and reviewed to assess the performance of the
hybrid version of the MPA method. proposed approach. To address these issues, the weight
It must be mentioned that related works in the meta- of the structure with specific constraints on the problem
heuristics literature were thoroughly evaluated and described is employed as a target function, together with distinct
in the first part of this section. However, several research and continuous areas of design variables.
works, such as HHO, SSA, MVO, WOA, and other well-
known meta-heuristics, are not natively developed and aim The paper’s structure is as follows. Section II represents
at large-scale and complex problems. For the first time in the the formulation of two large-scale truss problems. Section IV
literature, we customized and adjusted the MPA technique for explains the basic MPA initially, and then it describes the
large-scale challenges, particularly truss problems. proposed IMPA in more detail. The experimental results are
When the form and bar size parameters are combined, shown in Section V. Finally, Section VI draws a conclusion
a multi-modal solution space with dynamic restrictions is from the research findings and results.
created, resulting in a time-consuming optimisation process.
As a result, the majority of actual truss problems are large and II. FORMULATION OF TRUSS PROBLEMS
highly constrained, and scaling issues occur as the problem The goal of shape and sizing optimisation of truss structures
size increases. As a result, this paper employs the proposed is to find the minimal structural weight by simultaneously
optimisation methods (HNMPA, HNCMPA), an extended optimising nodal coordinates and cross-sectional areas. Addi-
population-based meta-heuristic, to tackle structural design tionally, the optimisation problem is supposed to satisfy nat-
challenges. The aim of this research is to apply the pro- ural frequencies, stress, and nodal displacement constraints
posed methods to build the optimal weight of truss structures while minimizing the structural weight. In this problem, the
VOLUME 11, 2023 39271

structural topology is assumed to be fixed and unalterable. positive y-direction, and negative z-direction, respectively.
Thus, the optimisation problem can be formulated as follows: Additionally, all nodes’ displacements in the x, y, and z
directions are confined to a maximum value of ±50 mm.
Find: X = {A1 , A2 , . . . , Am , C1 , C2 , . . . , Cn }
In this optimisation problem, cross-sectional areas and nodal
mn N
X X coordinates variables can take only integer values within
Min: F(X ) = Ai ρi Li + bj range [1, 37] and [9000, 20000], respectively.
i=1 j=1
Where: Amin
i ≤ Ai ≤ Amax
i
Cimin ≤ Ci ≤ Cimax (1)
where X is the structural design variable vector, including Ai
and Ci as structural sizing and shape design variables, which
are confined to the lower and upper bound [Amin i , Ai
max ] and
[Ci , Ci ], respectively. n and m represent the number of

min max
structural cross-sectional areas and nodal coordinates. Also,

in the second formula, ρi , Li , bj , and F(X ) are the ith element’s
mass density, the ith element’s length, the jth node’s mass, FIGURE 1. 314 bars and 84 nodes construct a steel truss structure. The
and the objective function, respectively. Moreover, structural optimisation variables incorporate elements of 14 shapes (C ) and
constraints can be formulated as follows; 314 sizings (A).
|σi | − σimax ≤ 0 2) 260-BAR TRUSS

|δj | − δjmax ≤ 0 (2) The second case study shown in Figure 2 (Part b) is a
three-dimensional steel truss structure including 260 bars and
where σi , σimax ,
δj , and δimax
are the ith element’s stress,
ten nodes. In the structural design, three independent load
ith element’s maximal stress value, the jth nodal displace-
cases are considered to apply the loads to all unsupported
ment value, and the jthe maximal nodal displacement value,
nodes. They include horizontal loads of 5 kN, horizontal
respectively.
loads of 1 kN, and vertical loads of 5 kN applied to positive
Furthermore, in these problems, a penalty function is nec-
x-direction, positive y-direction, and negative z-direction,
essary to convert the constrained problem into an uncon-
respectively. Additionally, all nodes’ displacements in x, y,
strained problem. The penalty function we used in this paper
and z directions are confined to a maximum value of ±25 mm.
is as follows:
In this optimisation problem, cross-sectional areas and nodal
K
X coordinates variables can take only integer values within
FPenalty (X ) = F(X ) + (σVioi + δVioi )Pf (3) range [1, 37] and [-25000, 3500], respectively.
i=1
where Pf and FPenalty (X ) are the penalty factor and penalty
function, respectively. Also, σVioi and δVioi are the stress and
nodal displacement violation values for the design variable
vector.
III. CASE STUDIES

In this section, we explain two case studies proposed by
Bright Optimiser called ISCSO 2018 and ISCSO 2019 [37].
The optimisation problems include concurrent structural
shape and sizing optimisation, while structural topology is
immutable and predetermined. The goal is to minimize the
structural weight while satisfying stress and nodal displace-
ment constraints. In both case studies, for all truss elements,
density, elastic modulus, and yield stress are equivalent to
7.85 ton/m3 , 200 GPa, and 248.2 MPa, respectively.
1) 314-BAR TRUSS
FIGURE 2. 260 bars and 76 nodes construct a steel truss structure. The
The first case study shown in Figure 1 (Part a) is a three- optimisation variables incorporate elements of 10 shapes (C ) and
dimensional steel truss structure including 314 bars and 260 sizings (A).
84 nodes. The structural design is subject to three indepen-
dent load cases. The loads applied to all unsupported nodes 3) 345-BAR TRUSS
include horizontal loads of 12 kN, horizontal loads of 6 kN, The third case study is a 3D steel truss structure with
and vertical loads of 48 kN applied to positive x-direction, 345 elements and 105 nodes with fixed topology, and just
39272 VOLUME 11, 2023

sizing optimisation is considered. The applied independent animals and nature, leading to the survival of the predator’s
loads in this structure are i) in the positive X-coordination, generation. In MPA, the animal’s next position is mathe-
a horizontal load is imposed at 20 kN. ii) the second load is matically predicted based on its current position and the
similar to the first load size but in the positive Y-coordination. probability of moving to the next position.
iii) 24 kN vertical load applied in the Z-coordination The ratio of prey to predator speed is critical in MPA for
inversely. The total feasible displacements for all nodes are simulating and transferring the optimisation process from one
restricted to ±5 mm. The yield stress of the truss mate- step to the next. The high speed of the prey in comparison
rial, elastic modulus, and material density are 248.2 MPa, to the predators is vital in the first phase of the algorithm,
200 GPa, and 7.85 ton/m3 , respectively. Figure 3 shows a whereas in the later stages of the algorithm, the unit ratio and
feasible design of a 345-bar truss structure with the same low speed are significant considerations in the optimisation
element size at 32. process. The MPA has a low number of adjustable variables,
a simple design, a low computational load, and no reliance on
the gradient.
The following is a synopsis of the optimal search, food
search, interaction, and memory processes used by wild
marine predators to find prey:
• Marine predators employ the Levy strategy in areas with
low hunting densities and the Brownian strategy in areas
with multiple preys.
• Marine predators use the same percentage of levy and
brownie movements when crossing different habitats.
• Marine predators remember food search locations as
well as hunting areas.
FIGURE 3. 345 bars and 105 nodes (including 15 supported nodes) 1) MPA’s OPTIMISATION PHASES
construct a steel truss structure. The optimisation variables incorporate This process is divided into three significant stages in MPA,
elements of 345 bars.
inspired by the natural life of predators and prey. These
procedures are as follows:
IV. METHODOLOGY Phase I: The prey is faster than the predator. When the prey
A. MARINE PREDATORS ALGORITHM (MPA) is moving faster than the predator, the predator’s best strategy
Faramarzi et al. [38] developed a new algorithm inspired by is to remain still. The algorithm is currently in the exploration
the marine predator food search strategy to solve optimi- phase. In MPA, this step is represented as follows:
sation problems. This algorithm’s structure is based on the 1
predators’ Lévy and Brownian movements and the optimal While: Iter < MaxIter
3
collision process in the biological interaction between preda- −
→ −−→
⃗
S l = RB ⊗ Elitel − R ⃗B ⊗ − −−→
Prayl
tor and prey. This algorithm has the advantage of simulating
the recall mechanism and memorising the successful hunting −−→ −−→ ⃗ ⊗−→
Prayl = Prayl + P.R Sl (4)
locations of marine predators. MPA, in other words, remem-
bers the optimisation results as well as potential solution The RB vector has random numbers and is based on the
points. In MPA, finding the best solution requires fewer normal distribution of Brownian motion. This vector in Prey
iterations. Therefore, using this algorithm to solve the truss is multiplied to model prey movement. P is a constant equal
optimisation problem is advantageous. to 0.5, and R is a vector of uniform random numbers in the
The two stages of marine predator search and the extent to range [0, 1]. This step occurs in 31 of iterations when the
which predators and prey interact in a marine ecosystem form movement speed is high to allow for high levels of exploration
the core structure of an MPA. Predators and prey are trying (Iter mentions the current iteration, Maxiter is the maximum
to hunt each other down and find natural food. They are pur- one).
suing a phenomenon known as the strongest survival, which Phase II: Both the prey and the predator move at the same
increases hunters’ chances of finding prey. MPA simulates rate. This MPA step occurs when the algorithm is moved
its optimisation phase using Lévy and Brownian strategies. from the exploration phase to the exploitation phase. Explo-
In this algorithm, these two strategies accurately depict the ration and exploitation occur concurrently at this stage. Prey
behaviour of marine predators. engages in exploitation, whereas predators explore.
Random walks in nature are an effective strategy in many 1 2
animals’ food search patterns. This optimal strategy allows while MaxIter < Iter < MaxIter . (5)
3 3
hunters to improve their chances of survival by increasing the −
→ ⃗L ⊗ −
−→ ⃗L ⊗ − −−→
Sl=R Elitel − R Prayl (6)
frequency of prey encounters in the wild. During predators’
lives, this strategy evolved as an inherent process in these −−→ −−→ −
→
⃗ ⊗ S l.
Prayl = Prayl + P.R (7)
VOLUME 11, 2023 39273
Prey motion is simulated for the first half of the popula- as follows:
tion using the RL random number vector based on the Levy
if r < FADs h

distribution. The prey movement in nature is simulated by 

 −−→
adding a step size to the prey position. The steps in the −−→

 Pray l + CF Z⃗min + R ⊗ (Z⃗max − ⃗
Z ⃗
min ] ⊗ U
Prayl =
Levy distribution are typically small. The MPA defines the
if

 r > FADs
− −→
following for the second half of the population: −−→ −−→
Prayl + [FADs (1 − r) + r] Prayr1 − Prayr2

−
→ ⃗B ⊗ R

⃗B ⊗ − −→ −−→
Sl=R Elitel − Preyl (8) (14)
−−→ −−→ −
→
Prayl = Elitel + P.CF ⊗ S l (9) r is a random number between 0 and 1. Whereas Zmax and

2 Iter Zmin are vectors representing the boundaries of the lower and
Iter MaxIter
CF = 1 − (10) upper dimensions. FADs = 0.2 to apply the effect of FADs
MaxIter
on the optimisation process in MPA. A binary vector U is
The step size of the predators’ movement is controlled by made up of two arrays, zero and one, which are defined by
an adaptive parameter, CF, in MPA. The predator’s move- generating a vector of random values between [0, 1].
ment is simulated using the Brownian method and multiplied Although the performance of the standard MPA is
by RB . Elite simulates prey movement in MPA. Because the considerable in solving most of the fundamental engineer-
prey position is updated in relation to the predators’ move- ing problems, there are some reports of low performance
ment in Brownian motion. of this meta-heuristic in some case studies. The first and
Phase III: The predator is faster than the prey. This MPA foremost shortage of MPA is the low efficiency in local
step contributes to excessive exploitation. search [36]. As the step control parameter (CF) signifi-
At this step, the Predator’s best strategy is to follow the cantly impacts obtaining a proper balance of exploration
Levy strategy. In the Levy strategy, predator motion is simu- and exploitation, some studies [39] proposed modified CF
lated by multiplying RL . The position of the prey is updated in techniques to improve the general performance of MPA.
response to predator movement. In addition, the Elite position
is simulated by increasing the step size. MPA defines this step B. HYBRID NONLINEAR MARINE PREDATORS
as follows: ALGORITHM (HNMPA)
2 The MPA has inspired predator and prey movement in accor-
while Iter > Maxiter −→ (11)
3 dance with the rules and points of multiple research and
−
→ ⃗L ⊗ R
−− → −−−
⃗ L ⊗ Elitel − Prayl→
Sl=R (12) observable behaviours in nature. Despite having satisfac-
−−→ −−→ −
→ tory exploration and exploitation rates, the MPA remains
Prayl = Elitel + P.CF ⊗ S l (13) stuck in optimal local solutions rather than reaching the
MPA divides the process of finding an optimal solution into global optimum solution. The main goal of HNMPA is to
three steps and a limited number of iterations. These steps are improve MPA exploration and exploitation by trying to adjust
inspired by natural predator behaviour and prey and simulate the predator phase’s size towards the prey and trying to
the predator’s step size and movement during the hunting balance the algorithm’s exploration and exploitation phases
process to catch prey. A predator’s percentage of Levy and using the presented control parameter. The MPA algorithm’s
Brownian motion is fixed in MPA. The predator is stationary second phase consists of two main stages: exploration and
and motionless in the early stages of the algorithm. The preda- exploitation.
tor, however, uses Brownian motion to change its position Modifications in this phase will improve the MPA’s effi-
in the second stage and Levy motion in the final stage. This ciency and effectiveness. This phase has been modified as fol-
scenario occurs due to the possibility of prey being predators lows: The nonlinear parameter is used as a control parameter
in nature. The prey uses a Brownian movement in the first by HNMPA to adjust the exploration and exploitation phases.
stage and a levy movement in the second stage to change its This variable is as follows:
position.
s !
Iter
The behaviour of marine predators may change throughout ω = 2 × exp −0.99 × (15)
their lives as a result of environmental issues in their habi- MaxIter
tat. Fish Aggregating Devices (FADs), or eddy formation in In the range [2, 0], ω decreases non-linearly. As a result,
nature, are one of the most well-known examples of these Eqs. 6 and 8 have been updated as follows:
effects. Sharks, as one of the largest marine predators, spend
more than 80% of their time near these FADs and the rest of 1 2
while MaxIter < Iter < MaxIter . (16)
their time searching in different directions and dimensions for 3 3
−
→ ⃗L ⊗ −
−→ ⃗L ⊗ −
−−→
environments with different prey. These FADs are considered Sl=R Elitel − R Prayl
local optima in MPA, and they prevent the algorithm from
× (l = 1, . . . , n/2) (17)
becoming trapped in the local optimal points during the opti-
−−→ −−→ ⃗⊗ − →
misation process. In the algorithm, this model is represented Prayl = ω × Prayl + P.R Sl (18)
39274 VOLUME 11, 2023

Algorithm 1 B. Hybrid Nonlinear Marine Predators Algorithm 2 Hill Climbing Local Search
Algorithm 1: procedure Hill Climbing Local Search(HC)
1: Initialise the control parameters FADs, P, Npop , and ω 2: Initialization
2: Initialise the first swarm (Prey), Zi i ∈ {1, 2, . . . , Npop } 3: Initialize σimax , δjmax ▷ Initialize maximum allowable
3: Evaluate population Prey stress and nodal displacement in the ith element
4: while iter ≤ Maxiter do 4: Steps = (Mins + Maxs )/10, Stepc = (Maxc −
5: Compute Elite matrix Minc )/10 ▷ Compute the step size for GS
6: Calculate CF based on Eq.(10) 5: Trussiter ={S, C} ▷ Read both sections (S) and
7: for each Prey ∈ Zi do coordinate (C) values
8: if Iter < 13 MaxIter then 6: (Weightb , σi , δj )=Eval(Trussiter ) ▷ Evaluate the
9: Update Prey based on Eq.(4) design
10: else if 13 Max_Iter < Iter <√23 Max_Iter then 7: for iter ≤ Maxiter do
11: ω = 2 × exp(−0.99 × Iter/MaxIter ) 8: while t ≤ N + M do
12: if i ≤ n/2 then 9: Temp = Trussiter
13: Update Prey based on Eq.(17) 10: if t ≤ N then
14: else 11: Tempt = Tempt ± Steps ▷ Neighborhood
15: Update Prey based on Eq.(19) search
16: end if 12: else
17: else 13: Tempt = Tempt ± Stepc
18: Update Prey based on Eq.(11) 14: end if
19: end if 15: (Weightt , σi , δj )=Eval(Tempt )
P ,M
20: end for 16: if N i=j=1 (σi + δj ) ̸ = 0 then ▷ design is not
21: Evaluate population Prey feasible
22: Update Prey using FADs based on Eq.(14) 17: Apply the penalty function
23: Apply marine memory saving and update the best 18: end if
solution 19: t =t +1
24: P Compute improvement rate based on ImIter = 20: end while
10
i=1 (fIter −fIter−i ) 21: Trussiter =Min(Weightt ) ▷ Select the best feasible
10
25: if ImIter < µ then solution and update the truss
iter
26: Run Hill Climbing (HC) Algorithm 2 22: Steps = Steps − ( Max iter
Steps ) + 1 ▷ Stepc and
27: if HCBestsol < NCMPABestsol then Steps linearly decreased
28: Update the population based on the best- 23: end for
found solution 24: return Trussiter
29: end if 25: end procedure
30: end if
31: end while
C. HYBRID NONLINEAR-CHAOTIC MARINE PREDATORS
−
→ ALGORITHM (HNCMPA)
⃗B ⊗ −−→ −−→

⃗B ⊗ R
Sl=R Elitel − Preyl 1) CHAOTIC SEARCH BEHAVIORS
× (l = n/2, . . . n) (19) In optimisation algorithms, chaos theory is frequently used
−−→ −−→ −
→ to improve the diversity of initialised solutions. Population
Prayl = ω × Elitel + P.CF ⊗ S l (20) diversity describes potential solutions, parts of solutions,
or structures that can be effectively changed into solutions.
Algorithm 1 shows the technical details of the implementa- According to the literature review, this optimisation algo-
tion of the proposed boosted MPA. The hill climb algorithm rithm is a population-based algorithm, which means it solves
is used (Algorithm 2) in MPA’s second change to prevent the problem by starting with a random solution and then
it from becoming trapped in local optimal points. If the evaluating it based on specified criteria. To solve the prob-
proposed algorithm fails to improve the value of the cost lems in the meta-heuristic, we need to use a search agent
function during the optimisation process and several consec- (in this work, the search agents are predator and prey). The
utive iterations (for example, in a quarter of the algorithm’s original MPA’s search agents start at a random point and
iterations), the HC algorithm is used, and it begins to try to generate random solutions, referred to as population distri-
improve the best answer and position obtained so far by the butions, contributing to a population diversity issue. In this
algorithm. It should be noted that this will continue until the paper, we use the Sine Chaotic Map function to set the
cost function’s value improves. position of these search agents and to improve the algorithm’s
VOLUME 11, 2023 39275

exploration phase. As demonstrated by the definition of In the range [2, 0], w decreases non-linearly. As a result,
chaotic maps and their movement behaviour, there are more Eqs. (12) and (13) have been updated as follows:
possible scenarios for the areas affected and the number of
1 2
times they keep moving. Therefore, a specific map improves while MaxIter < Iter < MaxIter . (24)
an optimisation technique’s exploration of behaviour pat- 3 3
−
→ ⃗
−−→
⃗L ⊗ −
−−→
terns. It is also worth noting that these maps were taken into S l = RL ⊗ Elitel − R Prayl
account because they display a variety of behaviours when
trying to generate chaotic values and have been shown to be × (l = 1, . . . , n/2) (25)
efficient in experiments conducted [40], [41]. Using chaotic −−→ −−→ ⃗ ⊗ −→
Prayl = ω × Prayl + P.CM Sl (26)
maps to enhance initialised solutions improves the efficiency −
→ ⃗B ⊗ − −→ −−→

of the algorithm. Furthermore, chaos theory can explore the Sl=R ⃗B ⊗ R Elitel − Preyl
solution space more extensively than random search [42].
× (l = n/2, . . . n) (27)
However, it is critical to leverage solution space to try and
−−→ −−→ −
→
make the initial population as efficient as possible. To help Prayl = ω × Elitel + P.CF ⊗ S l (28)
enhance diversity in the population, this work employs Chaos
theory’s Sine Map (SM) to initialise the HCCMPA. This The two proposed algorithms stated so far will eventually be
is due to the fact that chaotic maps aid the optimisation combined by the Hill climb algorithm. Notably, the two pro-
technique’s exploration of the search space. In other words, posed components can be utilized independently as distinct
apart from a probability-based search, the exploration can methodologies or in a combined approach that dynamically
be structured rather than random [41]. Furthermore, chaotic selects the algorithmic component most suited for resolving
maps can help optimisation algorithms prevent optimal solu- the challenges and alternating between them. Nonetheless,
tions while also boosting convergence [43], [44], [45]. The for the purposes of this investigation, they were applied as
CM mathematical formulation is computed as in Eq 21. distinct modalities. Algorithm 2 shows the technical details of
the implementation of the proposed HNCMPA. The hill climb
α algorithm is used (Algorithm 2) in MPA’s second change to
ChaosSineMap = Xn+1 = sin(πXn ), α = 4 (21)
4 prevent it from becoming trapped in local optimal points.
If the proposed algorithm fails to improve the value of the cost
2) HYBRIDISATION function during the optimisation process and several consec-
The MPA has inspired predator and prey movement in accor- utive iterations (for example, in a quarter of the algorithm’s
dance with the rules and points of considerable research and iterations), the HC algorithm is used, and it begins to try to
observable behaviours in nature. Despite having satisfactory improve the best answer and position obtained so far by the
exploration and exploitation rates, the MPA remains stuck in algorithm. It should be noted that this will continue until the
optimal local solutions rather than reaching the global opti- cost function’s value improves.
mum solution. The primary goal of HCCMPA is to improve The proposed hybrid method is able to keep a considerable
MPA exploration and exploitation by attempting to balance balance between searchability and convergence rate in multi-
the algorithm’s exploration and exploitation phases using the modal search space. Furthermore, integrating hill climbing
presented control parameter and chaos values rather than ran- as a robust local search improves the MPA and speeds up the
dom ones. Therefore, the first phase of the MPA is modified convergence rate in the unimodal search space.
as follows.
V. EXPERIMENTAL RESULTS
1
While: Iter < MaxIter A. LANDSCAPE ANALYSIS
3 Recently, the applications of landscape analysis have been
−
→ ⃗B ⊗ −
−→
⃗B ⊗ −
−−→
Sl=R Elitel − R Prayl broadly expanded [46] for a better insight into complicated
−−→ −−→ optimisation problems and clarifying behaviours of algo-
Prayl = Prayl + P.CM ⃗ ⊗− →
Sl (22) rithms, predicting the performance of the optimisation meth-
ods, and automated configurations and selections. In this
The MPA algorithm’s second phase consists of two main section, we performed some experiments for visualising a
stages: exploration and exploitation. Modifications in this violation landscape as further insight into fitness landscapes
phase will improve the MPA’s efficiency and effectiveness. for the truss optimisation problem with a constrained search
This phase has been modified as follows. The nonlinear space.
parameter is used as a control parameter by HCCMPA to A violation landscape is restricted, utilising the identical
adjust the exploration and exploitation phases. This variable features as a fitness landscape. Still, the fitness function is
is as follows: substituted using a violation function [47] that quantifies how
a solution violates the constraints represented on the prob-
Iter 3
ω = 2 × exp −(4 × ( ) ) (23) lem. Therefore, a violation landscape is characterised beyond
MaxIter the decision variable space and supplies a further landscape
39276 VOLUME 11, 2023

perspective to the fitness landscape. The features of violation

landscapes can be studied concerning fitness landscapes to
understand constrained optimisation benchmarks better.
In order to provide a technical landscape analysis, we used
a grid search method. As the number of decision variables is
enormous for both case studies, we assume the same value
for all shape and sizing variables. Figure 4 demonstrates the
fitness plus violation landscape analysis of the 260-bar truss
problem. The whole decision variables include 10 shapes
(−25 × 103 ≤ C ≤ 3500) and 260 sizings (1 ≤ A ≤ 37)
elements. The colour bar highlights the sum of the structure’s
weight and the penalty factor. It is noted that the landscape is
multi-modal, and sizing variables less than 15 make a sharp
violation consisting of stress and displacement. Moreover,
we can see that the best range of shape variables is between
0 and −15000.
FIGURE 5. 3D stress violation landscape analysis of truss structure with

260 bars and 78 nodes. The stress violation of the structure is highlighted
by the colour bar. Dark bluePin the zoomed figure shows the best-found
designs (feasible solution, |δj | − δimax = 0) based on the grid search
analysis.
TABLE 1. The algorithm’s parameter settings for numerical benchmarks.
FIGURE 4. 3D landscape analysis of truss structure with 260 bars and

78 nodes. The whole decision variables include 10 shapes
(−25 × 103 ≤ C ≤ 3500) and 260 sizings (1 ≤ A ≤ 37) elements. The sum
of the structure’s weight and the penalty factor is highlighted by the
colour bar. Dark blue shows the best-found designs based on the grid
search analysis.
To have a more systematic insight into the violation land-

scape of the 260-bar truss, we re-performed the landscape
analysis focused on the stress violation levels. Figure 5 shows
the stress violations of different designs of 260-bar truss,
and we can see the feasible areas highlighted by dark blue.
Obviously, the fitness search space is multi-modal with non- capability to explore, exploit, and escape from local minima.
linear constraints; however, the fitness landscape is simplified Unimodal test functions (F1-F6) are intended to test an
by the assumption of the same shape and sizing values. algorithm’s exploitation ability, whereas multimodal test
Furthermore, the impact of shape values on the stress vio- functions (F7-F13) are used to experiment with the algo-
lation is meaningful. rithm’s exploration phase effectiveness. These two classes of
functions are examined in 300 dimensions. The composition
B. NUMERICAL BENCHMARKS functions (F14-F20) are intended to assess the general perfor-
This study evaluated the proposed method’s performance mance of the proposed algorithm. However, their complexity
using various test functions. Two classes of well-known is similar to real and challenging optimisation problems
benchmarks are used for mathematical optimisation prob- because they have too many local minima. Wu et al. [48] con-
lems. The benchmark contains unimodal, multimodal, and tains more information on these test functions. Tables 1 and 2
composition functions to assess the HNCMPA and HNMPA’s outline the parameters for meta-heuristic algorithms
VOLUME 11, 2023 39277

TABLE 2. The statistical performance criteria for the proposed hybrid methods and other eight optimisation algorithms for unimodal functions.
TABLE 3. The statistical performance criteria for the proposed hybrid methods and eight other optimisation algorithms for multimodal functions.
TABLE 4. The statistical performance criteria for the proposed hybrid methods and eight other optimisation algorithms for Composition functions.
for numerical benchmark and truss optimisation issues, parameters was derived from the ones originally employed
respectively. Notably, the Matlab R2020a platform was uti- by the authors of each algorithm or, alternatively, from
lized to execute the proposed approach and other bench- those frequently utilized by numerous scholars in the current
marked algorithms on a computer equipped with Windows field of study. Furthermore, all iterations and populations
8.1 64-bit and 6 GB of RAM. The determination of for the optimisation algorithms were set at 2000 and 50,
39278 VOLUME 11, 2023

FIGURE 6. The search history, convergence performance of (a) MPA, (b) HNMPA, and (c) HNCMPA on unimodal F1 function.
Using HNCMPA and other methods, the test functions are

solved using a maximum of 100,000 function evaluations and
2000 iterations (Npop = 50). To achieve significant statistical
results, this study ran all optimisation methods 30 times and
presented the results, which included the average and stan-
dard deviation values of the best-so-far solutions observed
in each run. To illustrate the superior performance of the
proposed method over other methods, the test is conducted
for PSO [3], MFO [49], WOA [7], SSA [50], GWO [6],
MVO [51], and MPA [38]. Table 1 summarises the parameter
settings for other techniques. These parameters are either
highly suggested by their developers or fall within the ref-
erence values in order to achieve the ideal performance for
each algorithm [52].
FIGURE 7. The search history, convergence performance of (a) MPA,
(b) HNMPA, and (c) HNCMPA on unimodal F2 function. 1) EXPLOITATION PHASE ANALYSIS
Due to the unimodal functions’ definition, they contain
respectively. Table 1 enumerates the primary parameter con- only one global optimum. Therefore, they can evaluate the
figuration for each algorithm implemented in the present exploitation ability of an algorithm. Table 2 provides the aver-
study. It is worth noting that the hill-climbing technique age and standard deviation values for HNCMPA and other
for classical and CEC benchmark functions, for which techniques on unimodal test functions (F1-F7). The results
the suggested methods are capable of attaining the global show that HNCMPA outperformed most methods in almost
optima in the majority of these functions, was not employed. all test functions. These results demonstrate HNCMPA’s
Source codes of Proposed methods are publicly avail- exploitation capabilities, which can help HNCMPA converge
able at https://github.com/AminDehkordi/Adaptive-Chaotic- toward the global optimum and exploit it efficiently. The
Marine-Predators-Hill-Climbing-Algorithm-for-Large-scale- characterised nonlinear control parameter and small steps of
Design-Optimisations.git. Lévy movements contribute to this ability.
VOLUME 11, 2023 39279

3) ANALYSIS OF HNCMPA LOCAL MINIMA EVASION

The composition test functions are obtained by shifting, rotat-
ing, and hybridising some primitive unimodal and multi-
modal functions (F14-F20). These functions are intended to
test algorithms’ ability to evade local optima while capable
of exploration and exploitation. The efficiency of HNCMPA
and other techniques on these kinds of functions is shown in
Table 4. HNCMPA’s results are very competitive in all func-
tions of this type. These findings indicate that HNCMPA has
finely tuned capabilities between exploration and exploita-
tion and superb performance in escaping from local optima.
The latter is due to the nonlinear control parameter that has
been defined, as well as the more extensive displacement
associated with the chaotic value. HNCMPA’s exploitation,
exploration, and local minima evasion abilities were tested
(b) HNMPA, and (c) HNCMPA on unimodal F10 function. in this section. HNCMPA demonstrated its capacity to com-
prehensively explore the search space and exploit the best
solution while seeking to prevent local minimal stagnation.
Therefore, the hypothesis that chaotic values and a nonlinear
control parameter are used is proven, which improves the
MPA performance.
4) HNMPA AND HNCMPA’s CONVERGENCE ANALYSIS

This section discusses HNCMPA’s numerical and com-
putational convergence assessments. The convergence of
HNCMPA is reviewed using qualitative metrics such as vari-
ability and trajectories in the experimental procedure. The
analytical section demonstrates that HNCMPA eventually
converges to a stationary point, implying that HNCMPA has
a reasonable convergence rate. Figures 6–10 depict qualita-
(b) HNMPA, and (c) HNCMPA on unimodal F12 function. tive indicators for HNCMPA convergence and performance
evaluation in the mathematical function testing platform com-
2) EXPLORATION PHASE ANALYSIS pared to MPA and HNMPA. The figure’s first column illus-
Multi-modal test functions have many local optima due trates the structure of the functions in a two-dimensional
to their inherent structure, which increases tremendously perspective to provide an overview of the search space topol-
with the number of function design variables. Having more ogy. The search history is represented in the figure’s second
than one optimum is beneficial when evaluating an algo- column as the first criterion to be explained here. This figure
rithm’s exploration capability. High and fixed (low) dimen- depicts the collective exploration of agents and how the inter-
sional multimodal functions are found in F8 through F13. action and collective behaviour patterns of the predators and
Tables 3 demonstrate the results of implementing HNCMPA prey result in the structure of this process in MPA.
and various algorithms to these functions. Table 3 indicates In unimodal functions, Figure 6 illustrates more accumula-
that HNCMPA has superior exploration ability when com- tion of agents around the optimum points and more dispersed
pared to other techniques, particularly in comparison to other attitudes in multimodal and composition functions [38]. The
HNCMPA surpasses all algorithms on most high-dimensional first feature of the template is characterized as assisting in
multimodal functions, particularly MPA and HNMPA, and exploiting the results, which is favourable in unimodal func-
the results are competitive with high-performance optimis- tions, and the latter is described as exploring the domain,
ers for the remaining functions. HNCMPA has achieved which assists MPA in multimodal and composition functions
the global optimum in most problems for fixed-dimensional in searching the entire space [38]. The second criterion is
functions with a high degree of precision (Std) comparable to the convergence curve, which represents the best solution
that of high-performance optimisers. HNCMPA’s exploration discovered thus far. Each function type has its own conver-
is due to its various optimisation steps, chaotic value effect, gence curve model, as shown in Figure 6 and Figure 7. This
and predator Brownian motion. Furthermore, the proposed pattern is fairly smooth in unimodal functions and demon-
method employs a nonlinear control parameter, which results strates improved performance in results with the number
in a proper balance between the exploration and exploitation of iterations, but in multimodal and composition functions,
phases. this pattern changes into a step-by-step behaviour, which is
39280 VOLUME 11, 2023

FIGURE 10. The search history, convergence performance of (a) MPA, (b) and (c) HNCMPA’s on multimodal F13 function.
TABLE 5. The statistical performance criteria for the proposed hybrid methods and other eight optimisation algorithms for CEC-BC-2017 (Unimodal (F1
and F2) and multimodal functions (F4-F10)).
anticipated in these functions [38]. Another observation in the From each type, it is essential to recognise that in unimodal
search history is that the final hybrid shows high exploration functions, HNCMPA can identify and enclose the optimal
that also causes particles being overshoot the search space solution at initial iterations and seek to improve the solu-
(black hots that form a rectangle). tions as iterations continue (see Figure 6 and 7), whereas in
VOLUME 11, 2023 39281

FIGURE 11. Convergence curves of HNCMPA compared with other optimisation methods on various classical test functions.
39282 VOLUME 11, 2023

FIGURE 12. Convergence curves of HNCMPA compared with other optimisation methods on various mathematical CEC2017 test
functions.
VOLUME 11, 2023 39283

TABLE 6. The statistical performance criteria for the proposed hybrid methods and other eight optimisation algorithms for CEC-2017 Hybrid
functions (F11-F20).
TABLE 7. The statistical performance criteria for the proposed hybrid methods and other eight optimisation algorithms for CEC-2017 Composition
functions (F21-F30).
multimodal and composition functions (see Figure 8 and 9)), Suppose we take into account the agents (predators and prey)
the agents attempt to globally browse the solution space even to be collaborators. In that case, the convergence curve
in the final iterations in an endeavour to find an optimal depicts the behaviour patterns of the best player in attaining
solution still. Therefore, despite the enactment of iterations global optimum. Still, it provides no information about the
in some multimodal functions, no advancement in results overall team’s performance [38]. Therefore, we used another
is caused, eventually leading to the stride template in these criterion, ‘‘average fitness history,’’ to assess HNCMPA’s per-
curves. The proposed method’s exploratory and descrip- formance during optimisation. This criterion’s whole method
tive attitude is attributed to long moves deduced from the is similar to the convergence curve, but it reinforces how
agents’ Levy movements, eddy and FAD impacts, and chaotic this interactive attitude improves the outcomes over the initial
values that enhanced the HNCMPA’s exploration ability. random population. Some stride characteristics are visible in
39284 VOLUME 11, 2023

TABLE 8. The statistical performance criteria for the proposed hybrid methods and other eight optimisation algorithms for CEC-2022 benchmarks.
TABLE 9. Results of T-test (p-value) overall runs on CEC-2017.
TABLE 10. Results of T-test (p-value) overall runs on CEC-2022.
these functions in the average fitness history. This happened The trajectory of the agents depicted in column 3 of the
due to improved performance in all agents’ fitness due to Figures (6–10) is another criterion. This criterion illustrates
a phase transition in the algorithm, which led to overall an agent’s spatial variance from the start to the end of the
good agent efficiency. There are slight hillsides in unimodal optimisation procedure. Due to the agents’ shifting in many
functions but steep patterns in multimodal and composition directions in search space, we chose only the first dimen-
functions. sion of an agent to reveal its trajectory in order to ensure
VOLUME 11, 2023 39285

TABLE 11. The configuration details of optimisation methods applied to the truss shape and sizing problem. Npop is the initial population size.
TABLE 12. The statistical performance criteria for the proposed hybrid MPA methods compared with other 14 optimisation algorithms for 260-bar truss
case study.
its exact trajectory. These metrics of figures display sudden second-to-last column, this leads to the proposed algorithm
high amplitude and duration changes in the initial iterations for boosting the performance of the population during the
(exploration phase), which will disappear in the final iter- optimisation procedure. The convergence of this algorithm is
ations (exploitation phase). This pattern proves the algo- also visible in the last column, which shows the fewest values
rithm’s exploratory phase in initial iterations while shifting when compared to MPA.
to exploitation in the last ones, ensuring that an algorithm Figure 11 depicts the convergence curves on some of the
can eventually converge to an optimum global point [53]. test functions for MFO, PSO, GWO, SSA, MVO, WOA,
Due to the essence of multimodal functions, the intensity and MPA, and HNCMPA. The results of this figure indicate
frequency of these shifts are more significant than for uni- that HNCMPA has a couple of unique, readily identifiable
modal functions, and they usually last longer. Figures 6–10 behavioural trends when optimising the test functions. These
directly compare HNCMPA to MPA and on two unimodal tendencies are primarily the result of various optimisation
and three multimodal functions. The search histories (second steps. The first attitude illustrates a rapid convergence toward
column) in all functions display that HNCMPA exploration the near-global optimum point in the first phase and suc-
is more extensive than other methods, as several areas of cessive, slight advancements in the second and third phases,
the search space are coated. Nevertheless, as shown in the proving that a problem can be solved with just one phase.
39286 VOLUME 11, 2023

FIGURE 13. Convergence rate of 14 optimisation algorithms used to minimize the weight of 314-bar truss. The Maximum number of
evaluations is 105 .
FIGURE 14. Convergence rate of 14 optimisation algorithms used to minimize the weight of 260-bar truss case study. The Maximum number
of evaluations is 105 .
The first pattern was demonstrated by F1, F2, and F3. To help According to the functions’ structure demonstrated in
readers comprehend these behavioural patterns, we will use Figure 11, the optimisation process minimises the function’s
F1 as an example for other functions; F2-F4. delivered value with a range of parameters (The solution with
VOLUME 11, 2023 39287

case study.
the lowest value is the best). When we applied this function

to the group of optimisation techniques, all of them found
the best solution close to zero, except for our HNCMPA
algorithm, which could go even lower. The use of a nonlin-
ear control parameter, which results in nonlinear behaviour,
combined with the chaotic values used by the HNCMPA,
may provide the algorithm with more opportunities to diverge
from the optimal local solution and converge to the optimal
global solution. This has enabled HNCMPA to investigate
other areas of the function’s environment that may lead to
more encouraging minimum solutions as iterations progress.
HNCMPA could compete with the MPA and algorithms,
precisely F12–13 and CF1–CF7 (Figure 12). Since the struc-
ture of these functions is rather complicated and includes
many locally optimal solutions, the majority of the proposed
algorithms work even harder to reach the global optima.
Nonetheless, with all of these functions but apart from F12,
our algorithm sustained a satisfactory convergence rate at an
initial stage. In contrast, MPA and WOA algorithms could
only find relatively improved optimal solutions at the final
moment of iterations. However, compared to our proposed FIGURE 15. Convergence rate of 14 optimisation algorithms used to
minimize the weight of 345-bar truss case study. The maximum number
HNCMPA algorithm, these methods demonstrate no viable of evaluations is 105 .
convergence. The slow convergence rate may be quite compu-
tationally costly for some applications that desire an accurate
solution. The newly presented modifications to the MPA by the non-
To summarise this section, we can address the rationale for linear controller and some of the chaotic behaviour patterns
some of the key performance indicators obtained by NMPA prompted by the chaotic map set of equations led to improved
in contrast to other cutting-edge algorithms. The essence of algorithm performance. These new changes are in Section IV
the MPA algorithm and the fact that natural marine predators have made evident the efficiency of our proposed HNCMPA
can memorise their prey’s regions with broader and deeper algorithm in contrast to other benchmark methods.
food sources are the primary reasons we can emphasise here.
Furthermore, the inclusion of Levy and Brownian motions 5) HNMPA AND HNCMPA’s PERFORMANCE ANALYSIS ON
aids the algorithm’s performance in various optimisation THE CEC-BC-2017 AND CEC-BC-2022 TEST
problems. MPA can hold and use this memory to recall some BENCHMARK FUNCTIONS
of the sense of satisfaction in the possible potential area To further indicate the proposed method’s effectiveness,
for a globally optimal solution to be investigated or further we chose among the most subsequent and complicated bench-
exploited, depending on such a significant characteristic that mark test functions from the CEC-BC-2017 [48], and CEC-
other algorithms neglect. This phenomenon has aided our BC-2022 [54] Numerical optimisation competition, which
proposed algorithm, which has a suitable mechanism for includes 30 functions, at least half of which are complex
discovering, on average, superior best solutions to different hybrid and composition functions in CEC-BC-2017 and
algorithms. 12 hybrid and composition functions in the second dataset.
39288 VOLUME 11, 2023

FIGURE 16. a) A box and whisker plot indicates the optimisation summary of 15 meta-heuristics for a 260-bar truss, and the
minimum, first quartile, median, third quartile, and maximum findings can be seen. Each box shows the first quartile to the third
quartile, and in the following, the vertical line reaches via the box at the median. b) a zoom version of plot a for the last six
optimisation methods.
FIGURE 17. A box and whisker plot indicate the optimisation summary of
15 meta-heuristics for 314-bar truss, and the minimum, first quartile, FIGURE 18. A box and whisker plot indicate the optimisation summary of
median, third quartile, and maximum findings can be seen. Each box 14 meta-heuristics for 345-bar truss, and the minimum, first quartile,
shows the first quartile to the third quartile, and in the following, the median, third quartile, and maximum findings can be seen.
vertical line reaches via the box at the median.
Appendix A contains the characteristics of the functions. proposed method still performs properly on these test suits.
This function’s complicated numerical formulation is acces- The HNCMPA could outperform other methods and compete
sible [48]. We ran the HNCMPA against all these functions with MPA and HNMPA in the majority of test functions. As a
with 100 dimensions, and the results were compared to the result, the hypothesis of using a nonlinear control parameter
most state-of-the-art techniques in the literature. For all func- and chaotic values to improve the performance of the pro-
tions in this dataset, the dimension is fixed to 10. Tables 5 posed method is proven, as demonstrated by the results.
and 6, as in the preceding part, illustrates the average and The findings presented in Table 8 regarding the CEC-2022
standard deviation. Each technique is performed 30 times functions indicate that HNMPA and HNCMPA exhibit signif-
with a 2000 iteration number and total function assessments icantly better performance than other optimisation methods
of 100,000. As results are reported in Tables 5, 6, and 7, the across all cases evaluated. These two algorithms demon-
proposed method still has its proper performance on these strate strong competitiveness and often outperform others in
test suits. The results in Tables 5, 6, and 7 show that the most functions. This highlights the ability of our proposed
VOLUME 11, 2023 39289

case study.
FIGURE 19. a comparison of the proposed hybrid methods average rank with other meta-heuristic
algorithms computed by the nonparametric tests signrank and ranksum.
techniques to identify optimal solutions by leveraging the the proposed methods. Based on the p-values presented
benefits of nonlinear behaviour with multiple chaotic agents, in Tables 9 and 10, it is evident that both proposed tech-
which can prevent entrapment in local optima. niques, particularly HNCMPA, exhibit significantly better
There is an abundance of statistical techniques available performance than alternative optimisation methods across
in the academic literature for assessing optimisation algo- all evaluated scenarios. The outcomes attained by our pro-
rithms, such as the Wilcoxon rank-sum test, the t-test, and a posed algorithms surpass those of other approaches sig-
newly developed test called paradox-free analysis [55], [56]. nificantly. Notably, neither MPA nor any other strategy
Nevertheless, this study employed the t-test to evaluate could deliver results comparable to those achieved by the
39290 VOLUME 11, 2023

proposed methods. The findings indicate that incorporating

chaotic motion values and a nonlinear control parameter
can enhance the effectiveness of the proposed methods. Sta-
tistical analysis using T-tests, presented in Tables 9 and 10,
suggests that both HNCMPA and HNMPA exhibit superior
performance compared to alternative approaches. Notably,
HNCMPA, followed by HNMPA, consistently achieved
lower global optima than other strategies across different
fitness functions while also exhibiting higher output relative
to MPA.
C. LARGE-SCALE TRUSS STRUCTURES

In this section, we run ten times all optimisation meth-
ods independently to solve each large-scale structural case
study. Two new boosted MPAs and 15 state-of-the-art meta-
heuristics are performed and compared with the same cri-
teria to develop a fair comparative framework. Except for
statistical outcomes, Friedman’s ranking test based on the
ten independent runs is evaluated to ensure that the proposed
optimisation algorithm’s performance best solves these large-
scale real structural problems. The parameters setup for the
aforementioned algorithms has been reported in Table 11.
The average convergence rate of 14 meta-heuristic algo-
rithms for the case study of 314-bar can be seen in Figure 13.
It can be observed that the HNMPA and HNCMPA con-
verged faster than other algorithms in the initial iterations
and handled the dynamic constraints appropriately. More-
over, from Figure 13, most optimisation algorithms faced
premature convergence and struggled with local optimums
such as AOA, MFO, HGSO, SSA, etc. On the flip side, some
algorithms could not converge to a proper solution during
this computational runtime (105 evaluation number) such as
GWO, EO and CO. For the comparative average optimisation
developments in Tables 12, 13 and 14, the best-performed
algorithms based on both mean and STD value are HNMPA
and HNCMPA for three case studies. HNCMPA provided
better solutions than HNMPA in three structures, 260-, 314-,
and 345-bar at 6.48%, 5.41%, and 44.16%, respectively.
Tables 12– 14, MPA is ranked third in effectiveness in solv-
ing truss optimisation problems. In fact, MPA outperforms
other advanced algorithms by a significant margin when it
comes to solving such issues. As a result of these impressive
results on complex multi-dimensional engineering problems, FIGURE 20. Best-found feasible design of 345-bar truss problem
proposed by HNMPA. Total weight plus penalty is 4750.87 kg.
this study’s hypothesis regarding the selection and enhance-
ment of MPA’s performance has been confirmed.
The convergence rate of optimisation algorithms imple- truss problem. For the initial percentage of runtime, GNDO
mented for the 260-bar truss problem can be shown in rapidly converged and surpassed other methods; however, the
Figure 14. Similar to the 314-bar problem, the highest con- best-found solutions were proposed by the HCMPA finally.
vergence rate is related to HNMPA; HNCMPA and MPA The convergence speed of WCA and SSA is considerable in
performed considerably as well. The issue of falling into local this large case study, but both encountered falling into local
optimum can be seen in some of the methods in this case optimum.
study, such as GWO, DA, HGSO, MFO, and EO. Further- In order to compare the performance of the proposed novel
more, The CO and NNA methods need more runtime to find optimisation algorithm with other modern meta-heuristics,
better designs. each method runs ten times, and the best-found solutions
Figure 15 indicates a comparison of 15 optimisation algo- show in Figure 16. This figure demonstrates the distribu-
rithms performed for solving the best design of the 345-bar tion of optimisation methods performance for the 260-bar
VOLUME 11, 2023 39291

case study. The box with smaller distribution represents a dealing with constraint violations. Various penalty variables
more robust optimisation method. From Figure 16, the best- were assessed in order to determine the optimum value in
performed method is HNCMPA, with the lowest total weight the context of best-found solutions. This paper employs three
solutions proposed on average. After HNCMPA, we can see distinct truss issues. They all have a vast structure made
that HNMPA, MPA, and NNA performed well. In both large up of 260, 314, and 345 bars, respectively. It is considered
structures, 314 and 345-bar trusses, the best solutions were that the truss structure is constant and unchangeable. This
proposed by HNMPA, which can be seen in Figure 17 and 18. problem’s main challenge was finding the best truss form and
The second rank is related to HNCMPA and MPA in size parameters by lowering load-bearing capacity (structural
314-bar and 345-bar trusses, respectively. Furthermore, the weight) in proportion to nodal displacement limits, compo-
DA method could handle the dynamic constraints well and nent stress limits, and vibration frequency. This is a compli-
avoid falling into a local optimum. For the comparative aver- cated optimisation challenge due to its defining mathematical
age optimisation developments in Tables 12, 13 and 14, the aspects, which are characterised as large-scale, nonlinear, and
best-performed algorithm based on both mean and STD value multi-modal with dynamic constraints. Since meta-heuristics
is HNMPA for three case studies. HNMPA provided better are renowned as efficient global optimisation techniques,
solutions than HNCMPA in three structures, 260-, 314-, and we will emphasise their applicability to truss optimisation
345-bar, at 6.48%, 5.41%, and 44.16%, respectively. issues in this paper, particularly novel swarm optimisation
Figure 19 shows the average ranking evaluation of 15 opti- approaches. Consequently, we assessed and developed a com-
misation algorithms by a nonparametric test (Friedman rank- parison template for large-scale truss issues using thirteen
sum) for three truss problems. It is clear to observe that distinct swarm optimisation strategies. All control variables
HNMPA obtained the first rank in all case studies and for each optimisation technique were adjusted based on lit-
HNCMPA, MPA, DA, and GNDO received the second to fifth erature guidelines to ensure a meaningful assessment, con-
rank, respectively. Figure 20 shows a 3D and 2D landscape of sidering that no ideal technique exists to achieve the optimal
the best-found design of a 345-bar truss with the minimum parameter settings. Due to rapid and efficient exploration
weight (4750.87 Kg) found by HNMPA. All displacement and exploitation search techniques enhanced by combining
and stress constraints are satisfied, and the sum violation is chaotic map values with a nonlinear parameter, the suggested
zero. approaches outperform existing optimisation methods, par-
ticularly the original MPA employed in this study. In terms
VI. CONCLUSION of the truss optimisation problem, and from an engineering
Two novel versions of the Marine Predator Algorithm (MPA) standpoint, we can conclude that integrating the proposed
are proposed in this study, which employs a series of non- methods with the Hill Climbed approach would increase the
linear control parameters and chaotic values to assist an ade- performance of the proposed methods in dealing with the
quate equilibrium between the exploration and exploitation truss optimisation problem. Based on the results provided by
phases as well as to boost the exploration phase throughout the proposed techniques, the truss structure indicated by the
the optimisation procedure. Our proposed approaches have proposed techniques had the lowest weight, with a substan-
significantly increased efficiency while looking for the best tial difference when compared to the best solutions of other
solutions to large-scale optimisation issues. This enhance- approaches in all three truss case studies. The crucial factor
ment resulted from our proposed approaches’ ability to tran- to remember is that in order to achieve the most outstanding
sition from exploration to exploitation phases by utilising efficiency from the proposed methods, the control parameters
our introduced nonlinearity characteristic. Furthermore, the should be tweaked using these truss issues. The solution space
algorithm’s chaotic behaviour has enabled it to explore a and constraint-handling approach should also be examined.
wider variety of feasible solutions to a specific issue. The In the future, we intend to improve and create collaborative
algorithm may also profit from its nonlinearity characteristic techniques to handle the multi-objective case of large-scale
and devote more emphasis to exploiting promising regions truss challenges.
when necessary. Therefore, our algorithms were able to man- The results reported suggest that both HNCMPA and
age an efficient searching approach for locating an optimal HNMPA exhibited growth rates of 73.31% and 71.94%,
solution to a large-scale specified optimisation issue. The respectively, when compared to MPA’s final solution in
proposed algorithms were developed, and their performance regards to the optimisation of a 260-bar truss. The 314-bar
was compared to that of the original MPA as well as other cur- truss optimisation problem results showed a growth rate of
rent cutting-edge meta-heuristics techniques. Despite minor 66.87% for HNCMPA and 64.78% for HNMPA in terms
differences in obtaining results, both proposed strategies of the final solutions. Regarding the third problem of truss
in this study outperformed other cutting-edge optimisation optimisation, it was observed that HNCMPA and HNMPA
algorithms on mathematical benchmark functions, validating exhibited growth rates of 26.21% and 26.76%, respectively,
them for tackling large-scale challenges. compared to the optimal outcome obtained through MPA.
This study also used thirteen advanced meta-heuristic It can be inferred that integrating the proposed methods
methods to solve the truss form and sizing optimisation issue. with the hill-climbing approach would enhance the perfor-
We defined a penalty function, a typical mechanism for mance of the proposed methods when dealing with the truss
39292 VOLUME 11, 2023

optimisation problem. The outcomes obtained from the pro- [22] N. Khodadadi, V. Snasel, and S. Mirjalili, ‘‘Dynamic arithmetic optimiza-
posed techniques indicate that the truss structure suggested tion algorithm for truss optimization under natural frequency constraints,’’
IEEE Access, vol. 10, pp. 16188–16208, 2022.
by these methods yielded the lowest weight, with a substantial [23] L. Bellagamba and T. Y. Yang, ‘‘Minimum-mass truss structures with
difference compared to other approaches in all three truss case constraints on fundamental natural frequency,’’ AIAA J., vol. 19, no. 11,
studies. It is important to note that in order to achieve opti- pp. 1452–1458, Nov. 1981.
[24] J. H. Lin, W. Y. Che, and Y. S. Yu, ‘‘Structural optimization on geometrical
mal efficiency from these proposed methods, it is necessary
configuration and element sizing with statical and dynamical constraints,’’
to adjust control parameters using these truss issues. Addi- Comput. Struct., vol. 15, no. 5, pp. 507–515, Jan. 1982.
tionally, the solution space and constraint-handling approach [25] L. Wei, T. Tang, X. Xie, and W. Shen, ‘‘Truss optimization on shape
should be analyzed. In future work, we aim to develop collab- and sizing with frequency constraints based on parallel genetic algo-
rithm,’’ Struct. Multidisciplinary Optim., vol. 43, no. 5, pp. 665–682,
orative techniques to address multi-objective cases of large- May 2011.
scale truss challenges and improve existing methodologies. [26] A. Mortazavi, ‘‘Size and layout optimization of truss structures with
dynamic constraints using the interactive fuzzy search algorithm,’’ Eng.
Optim., vol. 53, no. 3, pp. 369–391, Mar. 2021.
REFERENCES
[27] A. Kaveh and A. Zolghadr, ‘‘Democratic PSO for truss layout and size opti-
[1] Z. Beheshti and S. M. H. Shamsuddin, ‘‘A review of population-based mization with frequency constraints,’’ Comput. Struct., vol. 130, pp. 10–21,
meta-heuristic algorithms,’’ Int. J. Adv. Soft. Comput. Appl., vol. 5, no. 1, Jan. 2014.
pp. 1–35, 2013. [28] L. F. F. Miguel and L. F. F. Miguel, ‘‘Shape and size optimization of truss
[2] D. Whitley, ‘‘A genetic algorithm tutorial,’’ Statist. Comput., vol. 4, no. 2, structures considering dynamic constraints through modern metaheuristic
pp. 65–85, 1998. algorithms,’’ Exp. Syst. Appl., vol. 39, no. 10, pp. 9458–9467, Aug. 2012.
[3] R. Eberhart and J. Kennedy, ‘‘A new optimizer using particle swarm [29] N. Pholdee and S. Bureerat, ‘‘Comparative performance of meta-heuristic
theory,’’ in Proc. 6th Int. Symp. Micro Mach. Human Sci., 1995, pp. 39–43. algorithms for mass minimisation of trusses with dynamic constraints,’’
[4] M. Dorigo, M. Birattari, and T. Stutzle, ‘‘Ant colony optimization,’’ IEEE Adv. Eng. Softw., vol. 75, pp. 1–13, Sep. 2014.
Comput. Intell. Mag., vol. 1, no. 4, pp. 28–39, Nov. 2006. [30] N. Khodadadi and S. Mirjalili, ‘‘Truss optimization with natural frequency
[5] E. Atashpaz-Gargari and C. Lucas, ‘‘Imperialist competitive algorithm: constraints using generalized normal distribution optimization,’’ Appl.
An algorithm for optimization inspired by imperialistic competition,’’ in Intell., vol. 52, no. 9, pp. 10384–10397, 2022.
Proc. IEEE Congr. Evol. Comput., Sep. 2007, pp. 4661–4667. [31] Y. Che and D. He, ‘‘An enhanced seagull optimization algorithm for
[6] S. Mirjalili, S. M. Mirjalili, and A. Lewis, ‘‘Grey wolf optimizer,’’ Adv. solving engineering optimization problems,’’ Appl. Intell., vol. 52, no. 11,
Eng. Softw., vol. 69, pp. 46–61, Mar. 2014. pp. 13043–13081, 2022.
[7] S. Mirjalili and A. Lewis, ‘‘The whale optimization algorithm,’’ Adv. Eng. [32] D. H. Wolpert and W. G. Macready, ‘‘No free lunch theorems for optimiza-
Softw., vol. 95, pp. 51–67, May 2016. tion,’’ IEEE Trans. Evol., vol. 1, no. 1, pp. 67–82, Apr. 1997.
[8] S. J. Mousavirad and H. Ebrahimpour-Komleh, ‘‘Human mental search:
[33] A. Faramarzi, M. Heidarinejad, B. Stephens, and S. Mirjalili, ‘‘Equilibrium
A new population-based metaheuristic optimization algorithm,’’ Appl.
optimizer: A novel optimization algorithm,’’ Knowl.-Based Syst., vol. 191,
Intell., vol. 47, no. 3, pp. 850–887, 2017.
Mar. 2020, Art. no. 105190.
[9] E. Rashedi, H. Nezamabadi-Pour, and S. Saryazdi, ‘‘GSA: A gravitational
[34] A. M. Shaheen, A. M. Elsayed, R. A. El-Sehiemy, S. Kamel, and
search algorithm,’’ J. Inf. Sci., vol. 179, no. 13, pp. 2232–2248, 2009.
S. S. M. Ghoneim, ‘‘A modified marine predators optimization algorithm
[10] A. Kaveh and S. Talatahari, ‘‘A novel heuristic optimization method:
for simultaneous network reconfiguration and distributed generator allo-
Charged system search,’’ Acta Mechanica, vol. 213, nos. 3–4, pp. 267–289,
cation in distribution systems under different loading conditions,’’ Eng.
2010.
Optim., vol. 54, no. 4, pp. 687–708, Apr. 2022.
[11] A. Kaveh, S. Talatahari, and N. Khodadadi, ‘‘Stochastic paint optimizer:
[35] M. A. Elaziz, S. B. Thanikanti, I. A. Ibrahim, S. Lu, B. Nastasi,
Theory and application in civil engineering,’’ Eng. Comput., vol. 38,
M. A. Alotaibi, M. A. Hossain, and D. Yousri, ‘‘Enhanced marine
pp. 1921–1952, Oct. 2020.
predators algorithm for identifying static and dynamic photovoltaic
[12] D. Bertsimas and J. Tsitsiklis, ‘‘Simulated annealing,’’ Statist. Sci., vol. 8,
models parameters,’’ Energy Convers. Manag., vol. 236, May 2021,
no. 1, pp. 10–15, 1993.
Art. no. 113971.
[13] Z. W. Geem, J. H. Kim, and G. V. Loganathan, ‘‘A new heuristic opti-
[36] E. H. Houssein, M. A. Mahdy, A. Fathy, and H. Rezk, ‘‘A modified marine
mization algorithm: Harmony search,’’ J. Simul., vol. 76, no. 2, pp. 60–68,
predator algorithm based on opposition based learning for tracking the
Feb. 2001.
global MPP of shaded PV system,’’ Exp. Syst. Appl., vol. 183, Nov. 2021,
[14] A. Kaveh and V. R. Mahdavi, ‘‘Colliding bodies optimization: A novel
Art. no. 115253.
meta-heuristic method,’’ Comput. Struct., vol. 139, pp. 18–27, Jan. 2014.
[37] Bright Optimizer: International Student Competition in Structural
[15] A. Kaveh, N. Khodadadi, B. F. Azar, and S. Talatahari, ‘‘Optimal design
Optimization. Accessed: Dec. 25, 2021. [Online]. Available:
of large-scale frames with an advanced charged system search algorithm
https://www.brightoptimizer.com
using box-shaped sections,’’ Eng. Comput., vol. 37, no. 4, pp. 2521–2541,
Oct. 2021. [38] A. Faramarzi, M. Heidarinejad, S. Mirjalili, and A. H. Gandomi, ‘‘Marine
[16] S. D. Rajan, ‘‘Sizing, shape, and topology design optimization of trusses predators algorithm: A nature-inspired metaheuristic,’’ Exp. Syst. Appl.,
using genetic algorithm,’’ J. Struct. Eng., vol. 121, no. 10, pp. 1480–1487, vol. 152, Aug. 2020, Art. no. 113377.
1995. [39] E. H. Houssein, M. Hassaballah, I. E. Ibrahim, D. S. AbdElminaam, and
[17] G.-C. Luh and C.-Y. Lin, ‘‘Optimal design of truss-structures using particle Y. M. Wazery, ‘‘An automatic arrhythmia classification model based on
swarm optimization,’’ Comput. Struct., vol. 89, nos. 23–24, pp. 2221–2232, improved marine predators algorithm and convolutions neural networks,’’
Dec. 2011. Exp. Syst. Appl., vol. 187, Jan. 2022, Art. no. 115936.
[18] S. O. Degertekin, H. Tutar, and L. Lamberti, ‘‘School-based optimization [40] A. H. Gandomi, G. J. Yun, X.-S. Yang, and S. Talatahari, ‘‘Chaos-enhanced
for performance-based optimum seismic design of steel frames,’’ Eng. accelerated particle swarm optimization,’’ Commun. Nonlinear Sci. Numer.
Comput., vol. 37, no. 4, pp. 3283–3297, Oct. 2021. Simul., vol. 18, no. 2, pp. 327–340, Feb. 2013.
[19] G. G. Tejani, V. J. Savsani, V. K. Patel, and S. Mirjalili, ‘‘Truss optimiza- [41] G.-G. Wang, S. Deb, A. H. Gandomi, Z. Zhang, and A. H. Alavi, ‘‘Chaotic
tion with natural frequency bounds using improved symbiotic organisms cuckoo search,’’ Soft Comput., vol. 20, no. 9, pp. 3349–3362, 2016.
search,’’ Knowl.-Based Syst., vol. 143, pp. 162–178, Mar. 2018. [42] G. I. Sayed, A. Darwish, and A. E. Hassanien, ‘‘A new chaotic whale
[20] A. Kaveh, A. D. Eslamlou, and N. Khodadadi, ‘‘Dynamic water strider optimization algorithm for features selection,’’ J. Classification, vol. 35,
algorithm for optimal design of skeletal structures,’’ Periodica Polytech- no. 2, pp. 300–344, 2018.
nica Civil Eng., vol. 64, no. 3, pp. 904–916, Jun. 2020. [43] A. A. Dehkordi, A. S. Sadiq, S. Mirjalili, and K. Z. Ghafoor,
[21] A. H. Gandomi, S. Talatahari, X.-S. Yang, and S. Deb, ‘‘Design optimiza- ‘‘Nonlinear-based chaotic Harris Hawks Optimizer: Algorithm and Inter-
tion of truss structures using cuckoo search algorithm,’’ Struct. Des. Tall net of Vehicles application,’’ Appl. Soft Comput., vol. 109, Sep. 2021,
Special Buildings, vol. 22, no. 17, pp. 1330–1349, Dec. 2013. Art. no. 107574.
VOLUME 11, 2023 39293

[44] A. A. Ewees and M. A. Elaziz, ‘‘Performance analysis of chaotic multi- AMIN ABDOLLAHI DEHKORDI received the
verse Harris Hawks Optimization: A case study on solving engineering M.Sc. degree in artificial intelligence and robotics
problems,’’ Eng. Appl. Artif. Intell., vol. 88, Feb. 2020, Art. no. 103370. from Islamic Azad University (IAU), Najafabad,
[45] Y. Xu, ‘‘An efficient chaotic mutative moth-flame-inspired optimizer Isfahan, Iran, in 2018. He is a Research Assis-
for global optimization tasks,’’ Exp. Syst. Appl., vol. 129, pp. 135–155, tant with an extensive experience in the use of
Sep. 2019. analytical and numerical data science techniques,
[46] K. M. Malan, ‘‘A survey of advances in landscape analysis for optimisa- such as machine learning, deep learning, and opti-
tion,’’ Algorithms, vol. 14, no. 2, p. 40, Jan. 2021. mization. He has engaged in a wide range of
[47] K. M. Malan, J. F. Oberholzer, and A. P. Engelbrecht, ‘‘Characterising
multi-disciplinary and complicated data modeling
constrained continuous optimisation problems,’’ in Proc. IEEE Congr.
projects, such as the IoT, truss optimization, and
Evol. Comput. (CEC), May 2015, pp. 1351–1358.
health systems. He has published multiple research papers in top journals,
[48] G. Wu, R. Mallipeddi, and P. N. Suganthan, ‘‘Problem definitions and
evaluation criteria for the CEC 2017 competition on constrained real- in order to apply his skills to new real-world engineering problems.
parameter optimization,’’ 2016.
[49] S. Mirjalili, ‘‘Moth-flame optimization algorithm: A novel nature- BAHAREH ETAATI received the master’s degree
inspired heuristic paradigm,’’ Knowl.-Based Syst., vol. 89, pp. 228–249, in artificial intelligence from the Amirkabir Uni-
Nov. 2015. versity of Technology, in 2018. She has partici-
[50] S. Mirjalili, A. H. Gandomi, S. Z. Mirjalili, S. Saremi, H. Faris, and pated in several engineering optimization projects
S. M. Mirjalili, ‘‘Salp Swarm Algorithm: A bio-inspired optimizer for with the University of Adelaide, the University
engineering design problems,’’ Adv. Eng. Softw., vol. 114, pp. 163–191, of Technology Sydney, and Torrens University
Dec. 2017. Australia. Recently, she attended a joint human-
[51] S. Mirjalili, S. M. Mirjalili, and A. Hatamlou, ‘‘Multi-verse optimizer:
computer interaction project between the Univer-
A nature-inspired algorithm for global optimization,’’ Neural Comput.
sity of Johannes Kepler Linz and the University of
Appl., vol. 27, no. 2, pp. 495–513, 2016.
[52] M. Jain, V. Singh, and A. Rani, ‘‘A novel nature-inspired algorithm for Applied Sciences Upper Austria as a Ph.D. candi-
optimization: Squirrel search algorithm,’’ Swarm Evol. Comput., vol. 44, date. Her research interests include machine learning, symbolic regression,
pp. 148–175, Feb. 2018. humancomputer collaboration with a focus on interpretability and explain-
[53] F. van den Bergh and A. P. Engelbrecht, ‘‘A study of particle swarm ability of AI, designing novel nature-inspired meta-heuristics, and applying
optimization particle trajectories,’’ Inf. Sci., vol. 176, no. 8, pp. 937–971, them to real-world problems, such as structural optimisation.
2006.
[54] W. Luo, X. Lin, C. Li, S. Yang, and Y. Shi, ‘‘Benchmark functions for CEC MEHDI NESHAT received the Ph.D. degree in
2022 competition on seeking multiple optima in dynamic environments,’’ computer science from the University of Adelaide,
2022, arXiv:2201.00523. Australia, in 2020. From 2020 to 2023, he was
[55] Y. Yan, Q. Liu, and A. Y. Li, ‘‘Paradox-free analysis for comparing the
a Postdoctoral Research Associate in data sci-
performance of optimization algorithms,’’ IEEE Trans. Evol. Comput.,
ence, machine learning, and deep learning with
early access, Aug. 17, 2022, doi: 10.1109/TEVC.2022.3199647.
the University of South Australia, Australia. He is
[56] Q. Liu, W. V. Gehrlein, L. Wang, Y. Yan, Y. Cao, W. Chen, and Y. Li,
‘‘Paradoxes in numerical comparison of optimization algorithms,’’ IEEE an Adjunct Research Fellow with the Center for
Trans. Evol. Comput., vol. 24, no. 4, pp. 777–791, Aug. 2020. Artificial Intelligence Research and Optimization,
[57] S. Mirjalili, ‘‘Dragonfly algorithm: A new meta-heuristic optimiza- Torrens University Australia, Australia. He has
tion technique for solving single-objective, discrete, and multi-objective published more than 100 papers in top interna-
problems,’’ Neural Comput. Appl., vol. 27, no. 4, pp. 1053–1073, tional conferences and journals. His primary research interests include artifi-
2015. cial intelligence, optimization, and machine/deep learning. In 2019 and 2020,
[58] F. A. Hashim, E. H. Houssein, M. S. Mabrouk, W. Al-Atabany, he received two Best Paper Awards from the most prestigious Genetic and
and S. Mirjalili, ‘‘Henry gas solubility optimization: A novel physics- Evolutionary Computation Conference (GECCO).
based algorithm,’’ Future Gener. Comput. Syst., vol. 101, pp. 646–667,
Dec. 2019.
[59] L. Abualigah, A. Diabat, S. Mirjalili, M. A. Elaziz, and A. H. Gandomi, SEYEDALI MIRJALILI (Senior Member, IEEE)
‘‘The arithmetic optimization algorithm,’’ Comput. Methods Appl. Mech. is a Full Professor, a top AI Scientist, and the
Eng., vol. 376, Apr. 2021, Art. no. 113609. Director of the Centre for Artificial Intelligence
[60] Y. Zhang, Z. Jin, and S. Mirjalili, ‘‘Generalized normal distribu- Research and Optimization, Torrens University
tion optimization and its applications in parameter extraction of pho- Australia. He is internationally recognized for
tovoltaic models,’’ Energy Convers. Manag., vol. 224, Nov. 2020, his advances in swarm intelligence and opti-
Art. no. 113301. mization, including the first set of algorithms
[61] A. Sadollah, H. Sayyaadi, and A. Yadav, ‘‘A dynamic metaheuris- from a synthetic intelligence standpoint—a radical
tic optimization model inspired by biological nervous systems: Neu- departure from how natural systems are typically
ral network algorithm,’’ Appl. Soft Comput., vol. 71, pp. 747–782, understood; and a systematic design framework
Oct. 2018. to reliably benchmark, evaluate, and propose computationally cheap robust
[62] H. Eskandar, A. Sadollah, A. Bahreininejad, and M. Hamdi, ‘‘Water optimization algorithms. He has published over 500 publications with over
cycle algorithm—A novel metaheuristic optimization method for solv- 72000 citations and an H-index of 88. As the most cited researcher in robust
ing constrained engineering optimization problems,’’ Comput. Struct.,
optimization, he is in the list of 1% highly-cited researchers and named as
vols. 110–111, pp. 151–166, Nov. 2012.
one of the most influential researchers in the world by the Web of Science.
[63] B. Abdollahzadeh, F. S. Gharehchopogh, and S. Mirjalili, ‘‘Artificial
His research interests include robust optimization, engineering optimization,
gorilla troops optimizer: A new nature-inspired metaheuristic algorithm
for global optimization problems,’’ Int. J. Intell. Syst., vol. 36, no. 10, multi-objective optimization, swarm intelligence, evolutionary algorithms,
pp. 5887–5958, 2021. and artificial neural networks. He is an Editor of several journals, includ-
[64] M. A. Akbari, M. Zare, R. Azizipanah-Abarghooee, S. Mirjalili, and ing Neurocomputing, Applied Soft Computing, Advances in Engineering
M. Deriche, ‘‘The cheetah optimizer: A nature-inspired metaheuristic Software, Applied Intelligence, and Engineering Applications of Artificial
algorithm for large-scale optimization problems,’’ Sci. Rep., vol. 12, no. 1, Intelligence.
pp. 1–20, Jun. 2022.
39294 VOLUME 11, 2023

Adaptive Chaotic Marine Predators Hill Climbing Algorithm For Large-Scale Design Optimizations

Uploaded by

Copyright:

Available Formats

Adaptive Chaotic Marine Predators Hill Climbing Algorithm For Large-Scale Design Optimizations

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Adaptive Chaotic Marine Predators Hill Climbing Algorithm For Large-Scale Design Optimizations

Uploaded by

Copyright:

Available Formats

Received 23 March 2023, accepted 4 April 2023, date of publication 13 April 2023, date of current version 25 April 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3266991

Adaptive Chaotic Marine Predators Hill Climbing

AND SEYEDALI MIRJALILI3,5,6 , (Senior Member, IEEE)

Corresponding authors: Amin Abdollahi Dehkordi ([email protected]) and

39270 VOLUME 11, 2023

VOLUME 11, 2023 39271

[Ci , Ci ], respectively. n and m represent the number of

structural cross-sectional areas and nodal coordinates. Also,

|σi | − σimax ≤ 0 2) 260-BAR TRUSS

III. CASE STUDIES

39272 VOLUME 11, 2023

39274 VOLUME 11, 2023

VOLUME 11, 2023 39275

39276 VOLUME 11, 2023

perspective to the fitness landscape. The features of violation

FIGURE 5. 3D stress violation landscape analysis of truss structure with

TABLE 1. The algorithm’s parameter settings for numerical benchmarks.

FIGURE 4. 3D landscape analysis of truss structure with 260 bars and

To have a more systematic insight into the violation land-

VOLUME 11, 2023 39277

39278 VOLUME 11, 2023

Using HNCMPA and other methods, the test functions are

VOLUME 11, 2023 39279

3) ANALYSIS OF HNCMPA LOCAL MINIMA EVASION

4) HNMPA AND HNCMPA’s CONVERGENCE ANALYSIS

39280 VOLUME 11, 2023

VOLUME 11, 2023 39281

39282 VOLUME 11, 2023

VOLUME 11, 2023 39283

39284 VOLUME 11, 2023

TABLE 9. Results of T-test (p-value) overall runs on CEC-2017.

TABLE 10. Results of T-test (p-value) overall runs on CEC-2022.

VOLUME 11, 2023 39285

39286 VOLUME 11, 2023

VOLUME 11, 2023 39287

the lowest value is the best). When we applied this function

39288 VOLUME 11, 2023

VOLUME 11, 2023 39289

39290 VOLUME 11, 2023

proposed methods. The findings indicate that incorporating

C. LARGE-SCALE TRUSS STRUCTURES

VOLUME 11, 2023 39291

39292 VOLUME 11, 2023

VOLUME 11, 2023 39293

39294 VOLUME 11, 2023

You might also like