The Journal of Supercomputing

https://doi.org/10.1007/s11227-023-05362-5

A fast convergence EO‑based multi‑objective optimization

algorithm using archive evolution path and its application
to engineering design problems

Majid Ilchi Ghazaan1 · Pedram Ghaderi1 · Amirali Rezaeizadeh1

Accepted: 28 April 2023

© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature
2023

Abstract
Time-consuming objective functions are inevitable in practical engineering optimi-
zation problems. This kind of function makes the implementation of metaheuristic
methodologies challenging since the designer must compromise between the qual-
ity of the final solution and the overall runtime of the optimization procedure. This
paper proposes a novel multi-objective optimization algorithm called FC-MOEO/
AEP appropriate for highly time-consuming objective functions. It is based on an
equilibrium optimizer equipped with an archive evolution path (AEP) mechanism.
The AEP mechanism considers the evolutionary trajectory of the decision space
and anticipates the potential regions for optimal solutions. Besides having a high
convergence rate, the newly proposed approach also possesses an intelligent bal-
ance between exploration and exploitation capabilities, enabling the algorithm to
effectively avoid getting stuck in the local Pareto. To assess the efficacy of the FC-
MOEO/AEP, a range of mathematical optimization problems and three real-world
structural design problems were employed. In order to gauge the method’s per-
formance against other approaches, a novel performance indicator known as con-
vergence speed was introduced and utilized, in addition to standard metrics. The
numerical findings demonstrate the robust and consistent performance of the FC-
MOEO/AEP in tackling complex multi-objective problems.

Keywords Multi-objective optimization · Metaheuristic algorithm · Evolution path ·

Structural optimization · Performance metrics · Fast convergence

* Majid Ilchi Ghazaan

[email protected]
Pedram Ghaderi
[email protected]
Amirali Rezaeizadeh
[email protected]
1
School of Civil Engineering, Iran University of Science and Technology,
P.O. Box 16846‑13114, Tehran, Iran

13
Vol.:(0123456789)
 M. Ilchi Ghazaan et al.

1 Introduction

Many engineering applications require optimizing several conflicting objective

functions [1–4]. These optimization problems are called multi-objective optimiza-
tion problems (MOPs). Unlike single-objective optimization [5–7], the goal of solv-
ing MOPs is usually to find a set of trade-off solutions, termed Pareto set, in the
decision space and a set of corresponding trade-off objectives, termed Pareto front,
in the objective space.
There are generally three main approaches for solving multi-objective problems
[8]: a priori, interactive, and a posterior. In the first approach, the multi-objective
problem is converted to a single-objective one by aggregating the objectives. This is
done using a set of weights usually defined by an expert and dictates the importance
of each objective. After aggregating the objectives, a single objective optimizer can
be employed to find the optimal solution [9–11].
In interactive multi-objective optimization, decision-making is done during the
optimization process. This means that human input is fetched to guide the search
process to the regions that are the interests of decision-makers. In this method, users
interact with the optimization algorithm to improve the efficiency of the algorithm,
enrich the optimization model, or evaluate the quality of the solutions obtained [12].
In the last category, the problem’s multi-objective formulation is maintained and
optimized simultaneously. Multi-objective optimization for minimization problem
aims to find design variables minimizing the objectives under the constraints of
equalities and inequalities and can be generally formulated as

(1)
( ) { ( ) ( ) ( )}
Minimize F x⃗ = f1 x⃗ , f2 x⃗ , … , fm x⃗

(2)
( )
subject to: gi x⃗ ≤ 0, i = 1, 2, … , k

(3)
( )
hi x⃗ = 0, i = 1, 2, … , l

Li ≤ xi ≤ Ui , i = 1, 2, … , n (4)
where n is the number of variables, m is the number of objective functions, k is the
number of inequality constraints, l is the number of equality constraints, gi is the
ith inequality constraints, hi indicates the ith equality constraints, and Li , Ui are the
boundaries of the ith variable.
In a posteriori method, Pareto optimal dominance is used to compare the solu-
tions; therefore, a posteriori optimization algorithm needs to store non-dominated
solutions as the best solutions for the problem during the optimization process.
Since the advantages of a posteriori optimization are more than its drawbacks in
comparison to the other methods, in the past two decades, many multi-objective
evolutionary algorithms (MOEAs) have been developed for tackling MOPs. Pos-
teriori MOEAs aim to obtain a set of Pareto-optimal solutions in a single run.
They can be classified into three groups. The first group is Pareto-based MOEAs

13
A fast convergence EO‑based multi‑objective optimization…

(e.g., Non-dominated Sorting Genetic Algorithm II (NSGA-II) [13], Multi-objec-

tive Salp Swarm Algorithm (MSSA) [14], Multi-objective Cuckoo Search Opti-
mization (MOCSO) [15], Multi-objective Artificial Bee Colony (MOABC) [16]).
The second group is decomposition-based MOEAs that Multi-Objective Evolu-
tionary Algorithm Based on Decomposition (MOEA/D) [17] is delegate in this
category. The third category is indicator-based MOEAs (e.g., Indicator-Based
Evolutionary Algorithm (IBEA) [18], Indicator Based Multi-Objective Evolu-
tionary Algorithm with Reference Point Adaptation (AR-MOEA) [19], Hypervol-
ume Estimation Algorithm for Multi-Objective Optimization (HypE) [20]).
Evolutionary algorithms face two main challenges: (1) achieve an approxima-
tion of the true Pareto, called convergence, and (2) provide the Pareto includ-
ing diverse optimal solutions in objective space, called coverage. Many types
of research have been done to overcome these challenges. Examples of these
researches include using the novel reproduction mechanism [21], improving the
selection mechanism [22], using reference points [23], employing the maxmin
function [24], calculation of crowding distance [13], and objective space grad-
ing [25]. Therefore, researchers have tried to balance exploration and exploitation
ability in the algorithms using different ideas to present the best performance in
Pareto estimating.
The metaheuristic approaches have become very popular today. They are used
to optimize many engineering real-world applications due to their robustness and
easy implementation [26]. The utilization of these algorithms in practical scenar-
ios has presented an additional hurdle. Evolutionary algorithms typically rely on a
trial-and-error approach to identify optimal solutions by generating possible solu-
tions and assessing them based on objective functions and constraints. However,
evaluating these objective functions or constraints can prove to be an arduous and
expensive task in numerous real-world engineering applications, particularly in civil
or mechanical engineering. This is due to the fact that each evaluation of a can-
didate solution in civil engineering problems necessitates structural analysis and
several constraints computations [27–29]. Conducting such computations for large-
scale structural problems, such as optimizing high-rise towers, incurs substantial
computational expenses. Another instance where the evaluation of objective func-
tions or constraints entails a complicated finite element analysis would also result in
significant computational costs [30, 31]. Executing such analyses numerous times,
possibly reaching several thousand evaluations, would prove to be excessively time-
consuming or infeasible. Consequently, the ability of an algorithm to furnish a satis-
factory Pareto front while requiring fewer evaluations constitutes the third challenge
confronting evolutionary algorithms. This attribute of algorithms can be classified
as their convergence speed.
Although the speed of convergence holds significance, it has not garnered con-
siderable research interest. Majority of the studies that introduce novel algorithms
typically assess their efficacy on standard problems using around 100,000 evalua-
tions and compare them against existing algorithms. However, the efficiency of such
algorithms for time-intensive problems remains unknown. Empirical evidence sug-
gests that algorithms that perform better than their competitors in high evaluations

13
 M. Ilchi Ghazaan et al.

may underperform in low evaluations. Thus, it is crucial to focus on the algorithm’s

performance in low evaluations.
One of the successful research experiences in increasing convergence speed has
been paying attention to the evolution path in the decision space. The concept of
decision space pertains to the range of decision variables or design parameters that
define the potential solutions for consideration. This space serves as the arena where
the search for the optimal solution occurs. For instance, in the previously mentioned
scenario, the decision space encompasses all feasible combinations of design param-
eters that could impact the efficiency and cost of a system, such as the dimensions of
components, material selection, and operational conditions. Conversely, the objec-
tive space pertains to the space occupied by the objective functions employed to
evaluate a set of candidate solutions. In a two-objective optimization problem aimed
at minimizing cost and maximizing efficiency, the objective space would be a two-
dimensional space, wherein every point would represent a unique combination of
cost and efficiency values.
By considering the evolutionary process, helpful information can be obtained,
e.g., individuals’ distribution and movement direction in decision space. This infor-
mation can lead the algorithm to the promising region more quickly and increasing
convergence speed. For example, the multi-objective covariance matrix adaptation
evolution strategy, termed MO-CMA-ES, employs the evolution path to update the
individual’s covariance matrix, based on which offspring are generated [32]. He
et al. [33] proposed an evolution path-based reproduction operator, termed EPDE,
for many-objective optimization. Each reference vector is associated with a set of
evolution paths in this operator. Based on these evolution paths, Lagrange poly-
nomial is constructed to depict the individual movement and predict its tendency.
However, this operator can only be embedded into decomposition-based algorithms
by employing a set of pre-defined reference vectors. It should be noted that each
individual has their evolution path, and there is no information exchange between
different individuals both in MO-CMA-ES and EPDE.
In another study, Song et al. [34] proposed a novel path-based reproduction oper-
ator, termed path evolution (PE), to generate potential solutions more effectively
for MOEAs. In PE, the calculation of the evolution path is simple. The size and
the direction of the evolution path are updated online with a simple method, which
keeps the evolution path instructive for guiding the population towards the Pareto
set in the early stages, and along the Pareto set in the later stages of the evolutionary
process. PE is highly effective in generating potential offspring, thus accelerating the
evolutionary process significantly.
In the mentioned studies, the decision space information is mainly used to guide
the searching process to potential regions; however, their research shows that if the
evolutionary path in the decision space is correctly estimated, it can be used directly
to generate very potential solutions.
This paper introduces a new multi-objective optimization algorithm that uses
both information obtained from the decision space and objective space to find opti-
mal solutions.

13
A fast convergence EO‑based multi‑objective optimization…

The proposed algorithm has two mechanisms for generating, namely MOEO
(Multi-Objective equilibrium optimizer) and AEP (archive evolution path). The
combination of these two methods to generate the solutions has created a very
potential population in each iteration and has the following advantages:

• Using information from the evolution path in the decision space and considering
the results obtained in the objective space, no helpful information is ignored.
• Using this idea has created an intelligent balance between exploration and exploita-
tion. This ability allows the algorithm to save itself easily from the local Pareto and
estimate the global Pareto front with reasonable accuracy and spread.
• Ultimately, the approach adopted by this algorithm has notably enhanced its con-
vergence rate. This attribute further bolsters the dependability of the optimal solu-
tions produced by this algorithm in limited iterations. Such capability constitutes
the principal advantage of this algorithm, which served as the primary impetus for
conducting this study.

The main characteristics of this paper can be summarized as follows.

• An archive-based EP method is proposed to increase the convergence speed.

• A new quality indicator is proposed to approximate the algorithm convergence
speed
• The validity of the proposed algorithm is verified for 30 mathematical benchmark
optimization problems, and Chess Rating System for Multi-Objective Evolutionary
Algorithms with Quality Indicator Ensemble (CRS4MOEA/QIE) was used to com-
pare the performance of algorithms correctly.
• The application of the presented algorithm is conducted by examining three real-
world structural design problems.

The rest of the paper is arranged as follows: Sect. 2 summarizes the equilibrium
optimizer algorithm (EO). Afterward, Sect. 3 describes the proposed algorithm for
tackling MOPs. The mathematical and structural optimization problems are studied in
Sects. 4. Finally, in Sect. 5 some concluding remarks are provided.

2 Equilibrium optimizer (EO)

Equilibrium optimizer [35] is a recently introduced single-objective global optimiza-

tion algorithm with exemplary performance in solving engineering optimization prob-
lems. EO is inspired by control volume mass balance models used to estimate both
dynamic and equilibrium states. It uses a mass balance equation to measure the amount
of mass that joins, exits, and is produced in that control volume over a while and seeks
to find the state that achieves the equilibrium of the system.
EO starts with a randomly generated population within the search space. The four
best-so-far particles and their average are used to generate new potential particles in the
EO process. The four candidates help EO have better exploration capability while the

13
 M. Ilchi Ghazaan et al.

average helps in exploitation. These five solutions are nominated as equilibrium candi-
dates and are used to construct a vector called the equilibrium pool:
{ }
⃗ eq,pool = C
C ⃗ eq(1) , C
⃗ eq(2) , C
⃗ eq(3) , C
⃗ eq(4) , C
⃗ eq(ave) (5)

One member of C ⃗ eq,pool is selected randomly as C

⃗ eq and each particle will be
updated according to the following formula:
⃗ (
⃗ + G 1−F
( ) )
⃗ =C
C ⃗ eq + C⃗ −C
⃗ eq .F ⃗ (6)
⃗
𝜆V

⃗ = a1 sign ⃗r − 0.5 (e−𝜆t⃗ − 1) (7)

( )
F

( )
iter
a2 max
iter (8)
( ) iter
t = 1−
max iter
where 𝜆⃗ is a random vector between 0 and 1, and V is considered as unit. F ⃗ assists
EO in having a reasonable balance between exploration and exploitation. iter and
maxiter are the current and the maximum iteration, respectively. a1 and a2 are con-
stant value that controls exploration and exploitation ability.
Another term in Eq. (6) is generation rate. G⃗ used to improve the intensification
operator and is formulated as follows:
⃗ =G
G ⃗ 0F
⃗ (9)

( )
G �������⃗ C
⃗ 0 = GCP ⃗ eq − 𝜆⃗C
⃗ (10)

{
�������⃗ = 0.5r1 r2 ≥ GP
GCP
0 r2 < GP (11)

where r1 and r2 are random numbers between 0 and 1. In this equation, GCP
�������⃗ is the
generation rate control to determine whether it will affect the updating process or
not according to Generation Probability (GP).
Algorithm 1 depicts the steps of the equilibrium optimizer algorithm for single-
objective problems.

13
A fast convergence EO‑based multi‑objective optimization…

13
 M. Ilchi Ghazaan et al.

3 Fast convergence multi‑objective equilibrium optimizer equipped

with archive evolution path mechanism (FC‑MOEO/AEP)

3.1 Overall framework

This section proposes a multi-objective EO-based algorithm with the archive evo-
lution path mechanism. Due to the existence of the unary objective in a single-
objective problem comparisons of solutions can be easily made by regular relational
operators. The matter is different when dealing with the multi-objectives problem as
we have several conflicting objectives. Instead of finding a single solution, we seek
to find a set of solutions that contains the best trade-offs between objectives. For this
purpose, the Maxmin function [36] is used to compare the solutions and rank them.
The maxmin value for the kth solution can be expressed as follows:

Fig. 1  Flowchart for the proposed FC-MOEO/AEP

13
A fast convergence EO‑based multi‑objective optimization…

{ }
(12)
imink
{ ( ) ( )}
fmax = max min fi xk − fi xj
j=1,2,…,n;j≠k i=1,2,…,M

where n is the population size and m is the number of objectives. In this equation,
the min operator is first used to calculate the minimum value over all the objective
functions. Then, the max operator is taken over all agents except for the kth solution.
From this equation, it can be concluded that: (i) any solution for which the maximin
value > 0 is a dominated solution, (ii) any solution with maximin value < 0 is a non-
dominated solution, and (iii) a solution with a smaller maxmin value is located in
sparsely populated areas [37].
The overall framework of the FC-MOEO/AEP is given in Algorithm 2. First, sim-
ilar to other metaheuristic algorithms, the initial population of particles is generated
randomly in the search space and the objective functions are calculated for all popu-
lations. An external archive with a limited capacity is defined and updated in each
iteration. In this update, all the currently nondominated solutions are inserted into
the archive, and dominated solutions are deleted. Afterward, the maxmin value of
archive members is computed and ranked. Archive members with smaller maxmin
values are located in sparsely populated areas; therefore, the greater maxmin value
solutions will be deleted when the archive is full. The ranked archive is also used in
the process of generating new solutions.
In the proposed method, the new solutions are generated by two different mecha-
nisms, EO and AEP. Each of these two mechanisms generates N solutions, i.e., in
each iteration, 2N solutions are generated. In contrast, only N solutions are needed
to form the next population. To select the best N solutions from the generated
solutions, we used the maxmin function again. The N solutions with the smallest
maxmin value are selected as the next-generation population. In this approach, the
generated solutions are examined in terms of dominance and spread simultaneously.
At last, when the terminal condition is fulfilled, the archive members will be
presented as Pareto optimal solutions. Many stopping criteria, such as maximum
iteration number or no improvement of the external archives, can be utilized [38].
To guarantee equitable evaluation of the suggested algorithms, this article utilizes
a consistent stopping criterion of a maximum iteration count, employing the same
population across all algorithms. The flowchart of the FC-MOEO/AEP algorithm
is shown in Fig. 1.

13
 M. Ilchi Ghazaan et al.

3.2 Multi‑objective equilibrium optimizer

To adapt the standard EO to the proposed multi-objective optimization algorithm

framework, the following changes are made:

• The ranked archive is used instead of the equilibrium pool, and all the particles in
the populations are updated using the best non-dominated solutions found so far.
• A roulette-wheel selection is used to choose a solution from the ranked archive to
generate a new solution. So, the selection probability of solutions with a smaller
maximin value will be more. In this way, the algorithm will be guided to search
sparsely populated areas.

13
A fast convergence EO‑based multi‑objective optimization…

• Sensitivity analysis was performed on the parameters a1, a2, and GP to achieve
the best MOEO performance in the FC-MOEO/AEP framework. For this pur-
pose, 60 different combinations of these three parameters were tested on math-
ematical benchmark problems. Based on empirical testing, it was found that the
appropriate values for these parameters in the FC-MOEO/AEP are greater than
the suggested values in the original EO. Generally, this change has increased the
exploration capability of the MOEO.

Algorithm 2 represents the pseudo-code of the MOEO. The ranked archive and
the current population are given to the algorithm as input. A solution is selected
from the archive using the roulette wheel selection mechanism. Then, EO compo-
nents are calculated using Eqs. (7)–(9). Finally, based on the Eq. (6), the particle
will be updated, and N new solutions will be the output of this algorithm.

3.3 Archive evolution path

Song et al. have recently suggested a novel path-based reproduction operator gen-
erate high potential solutions. However, there are two general drawbacks to their
proposed mechanism. The first one is that the evolutionary path (EP) vector is cal-
culated based on the movement of the main population in the decision space. It is
assumed that the population of each iteration is definitely better than the population
of the previous iteration. While in many cases, the population of one generation may
not be better than its previous generation. In these cases, the calculated EP will not

13
 M. Ilchi Ghazaan et al.

Fig. 2  Illustration of how to generate solutions according to the evolution path, ACRep
������������⃗ . the new candi-
date solutions will be located in the hatched areas

help to generate potential solutions and will also mislead the algorithm. The second
drawback is that the EP is calculated using only the current and previous generation
populations. This means that only the information from the evolution between two
successive generations considered, while experimental tests in this study show that
the evolutionary path can be better estimated using several generation populations.
The AEP mechanism is proposed to overcome these drawbacks, which works well
with the equilibrium optimizer engine.
In the AEP mechanism, we use the evolution process of the archive members
that are the best solutions found so far. Therefore, the archive is continuously evolv-
ing, and it can be used to estimate the evolution path correctly. The leading term in
this mechanism is the ACRep
������������⃗ vector that predicts the potential areas in the decision
space. The size and direction of this vector are updated based on the experiences of
the previous steps. The ACRep
������������⃗ defined as follows

������������⃗ = Archive Centerc − Archive Centerj

ACRep (13)

where Archive Centerc and Archive Centerj are the mean point of all archive
members at current iteration and iteration j , respectively. As an example, the
������������⃗ vector calculation process is illustrated in Fig. 2. where Arc_size = 4 and
ACRep
dimension = 3. First, ACRep ������������⃗ is calculated according to Eq. (13). The calculated
vector direction indicates that all three dimensions should increase when generating
potential solutions. Then, the vector size is modified by multiplying a factor 𝛼 , and
the potential area is estimated using 𝛼 × ACRep ������������⃗ . In this case, the solutions will be
located in the hatched areas in Fig. 2 randomly.𝛼 is a self-adaptive parameter decid-
ing how far solutions can go in the direction of ACRep ������������⃗ . The 𝛼 value in each iteration
is modified based on the quality of the generated solutions by the AEP mechanism

13
A fast convergence EO‑based multi‑objective optimization…

in the previous iteration. For this purpose, the success rate parameter, called S_rate,
is introduced. As mentioned before, 2N solutions are generated by MOEO and AEP
mechanisms in each iteration; then, the top N solutions are selected for the next gen-
eration. The success rate is calculated using the number of solutions generated by
AEP that survive into the next generation, divided by the total number of solutions
generated by that ( N ). If the success rate is greater than a specific value, the alpha
will increase by multiplying by 𝜆. Otherwise, the alpha will decrease by dividing
by 𝜆. New generated solutions will be almost based on a random search if the 𝛼 is
too large. On the other hand, if its value is too small, the generated solutions will be
close to the current solutions. So, 𝛼 is bounded with an upper and lower bound to
avoid these situations.
The steps of the AEP are provided in the Algorithm 4. According to this
pseudo-code, first, the centers of the archives are calculated in the current and
previous iterations. Then the evolution vector is calculated and normalized. If
the maximum element of the normalized ACRep ������������⃗ is less than a limit, an ele-
ment of Archive Centercurrent will be re-initialized according to line 5 and line
6. The reason for just selecting one element but not all elements is that the new
Archive Centercurrent with all elements being re-initialized will result in a very con-
fused and less effective evolution path. Generally, the operations between line 4 and
line 8 are to ensure that ACRep
������������⃗ are informative and practical for generating poten-
tial solutions.
After calculating the ACRep
������������⃗ , N solutions must be selected from the archive to
generate new AEP solutions using them. We used the roulette wheel selection to
select these solutions from the archive, so the solutions in sparsely populated areas
will have a better chance of selection. Finally, new solutions are generated using the
following equation
������������⃗ × rand + Cselected
CAEP = 𝛼 × ACRep (14)

Fig. 3  Illustration of archive center in five consecutive iterations and the evolution path vector based on
the evolution experienced in the previous iterations

13
 M. Ilchi Ghazaan et al.

One of the parameters that should be decided is how many evolutionary steps
must be taken into account to calculate the ACRep ������������⃗ . In other words, how many
iteration experiences should be used to estimate the evolutionary path. For a bet-
ter explanation, see Fig. 3, where the archive center is shown in five consecutive
iterations. In this case, if the evolution vector is estimated based on the evolution
experienced in the previous four iterations, next-generation solutions will be gen-
erated using vector B. Else if the evolutionary path is calculated only based on
the experience of one previous iteration, new solutions will be generated using
vector A.
For this purpose, we examined the algorithm’s performance using 30 different
mathematical functions. Experimental studies showed that the algorithm’s best
performance in terms of convergence and diversity of Pareto Front occurs when
the ACRep
������������⃗ is estimated based on the experiences of the previous seven steps. It
also observed that if more than 15 steps are used to estimate the evolution vector,
the exploration capability of the algorithm is significantly reduced, and the algo-
rithm acquires only a part of the Pareto Front in most cases.

13
A fast convergence EO‑based multi‑objective optimization…

3.4 Discussion

The main features of the proposed algorithm and the details of its operation are sum-
marized as follows.

• The two mechanisms, AEP and MOEO, supplement each other. The AEP mech-
anism for generating potential solutions requires an evolving archive. MOEO,
with a structure inspired by the laws of physics, explores the search space and
contributes to the better evolution of the archive by finding nondominated solu-
tions.
  On the other hand, the better the current population, the better solutions gen-
erated by the MOEO mechanism. In each iteration of the proposed algorithm,
the solutions generated by AEP and MOEO are combined, and their bests are
selected as the next generation’s population. So, the population quality in the
next generation will increase and help the MOEO generate high potential solu-
tions.
  Also, using these two mechanisms together makes no helpful information
about the evolution process be ignored. This means that both the experiences of
the evolution process in the objective space and the evolution process in the deci-
sion space are used to achieve optimal solutions.
• There is an intelligent balance between exploration and exploitation ability in
this algorithm. MOEO is designed to have high exploration ability in the initial
iterations and increase exploitation ability in the final iterations. AEP has high
exploitation ability and constantly searches for areas around the archive solu-
tions. So, in the early iterations, MOEO by extensive search and AEP by exploit-
ing around the archive solutions, help to find the optimal solutions.
  On the other hand, when the MOEO exploration ability is reduced in the final
iterations, if the algorithm falls into the local Pareto trap, AEP automatically
increases the algorithm’s exploration ability. Because when the algorithm gets
stuck in the local Pareto, the archive changes will be minimal. In such cases,
as described in Sect. 3, AEP makes random changes in the ACRep ������������⃗ vector and
diversifies the search.
• This algorithm tries to direct the search to sparsely populated areas by using the
roulette wheel selection method (RWS) in the process of generation solutions.
Experience showed that if the search is limited to sparsely populated areas in
each iteration, the diversity of the final Pareto Front will be harmed. Therefore,
we only increased the probability of searching these areas using the RWS to
achieve a good spread of the Pareto front.
• In this paper, the idea of the single-objective EO algorithm is used in the pro-
posed multi-objective algorithm framework because EO has an excellent abil-
ity to optimize various problems, especially in discrete problems. This feature
is suitable for addressing structural problems. However, it can be used the idea
of other single-objective algorithms in this framework. Note that our studies
showed it is better to increase the exploration ability of the algorithm that used
in the proposed framework. Thus, if there is a more robust algorithm than EO

13
 M. Ilchi Ghazaan et al.

or a more suitable algorithm for use in other fields, it can be replaced within the
framework of the multi-objective algorithm described in this study.
• The aim of this investigation was to design a multi-objective algorithm that can
efficiently converge in problems featuring time-consuming objective functions.
Specifically, the goal was to outperform competing algorithms in terms of find-
ing superior solutions while using fewer evaluations for time-consuming objec-
tive functions. To achieve this objective, several mechanisms were employed,
such as MOEO, AEP, RWS, Alpha, External Archive, Handle full archive, etc.
Not only does this make implementing the algorithm more challenging, it can
also result in an escalation of the algorithm’s computational complexity. In order
to analyze the complexity of the proposed algorithm, we conducted an empirical
analysis, which revealed that the algorithm’s search for a solution necessitates a
reasonable amount of time and space and is comparable to its rivals.

4 Results and discussion

In this section, the performance validation of the proposed FC-MOEO/AEP is

discussed. To prove the validity of the FC-MOEO/AEP, a comprehensive set of
mathematical benchmark functions and three real-world structural problems are
considered.

4.1 Experimental setup

All the experimental studies are carried out on a laptop equipped with Windows 10 Pro
with a 64-bit operating system, Intel(R) Core (TM) i7-7700HQ CPU @ 2.80 GHz, and
16 GB of RAM. All algorithms used in our comparisons are implemented using the
version R2021a of MATLAB. The proposed algorithm is compared with three well-
regarded algorithms to evaluate its efficacy; and they are summarized as follows:

Table 1  Control parameters of Algorithm Parameter Value

the algorithm
MOSMA Archive size 100
NSGA-III Selection Binary tournament
Recombination pc = 0.9
Mutation Polynomialpm = 1∕dim
MOMVO Archive size 100
WEP_max 1
WEP_min 0.2
FC-MOEO/AEP Archive size 100
𝜆 1.7
S_rate 0.5

13
A fast convergence EO‑based multi‑objective optimization…

• Non-dominated Sorting Genetic Algorithm III (NSGA-III) [39].

• Multi-objective Slime Mould Algorithm (MOSMA) [40].
• Multi-Objective Multi-Verse Optimizer (MOMVO) [41].

For a fair comparison, we set the values of the parameters for the other algorithms
as suggested by their authors. The control parameters of all algorithms are listed in
Table 1. For mathematical problems, the number of iterations and the evaluation num-
ber are set to 200 and 20,000, respectively, for all the algorithms. These parameters are
set to 200 and 10,000 for structural problems. Each algorithm is evaluated in all prob-
lems using 30 independent runs.
In general, the performance of all multi-objective algorithms is measured using the
performance indicators. Here, to compare the FC-MOEO/AEP with other algorithms,
Generational Distance (GD) [42] and Inverted Generational Distance (IGD) [43] are
utilized for quantifying the convergence. Also, in order to quantify the distribution of
the obtained Pareto optimal solutions (coverage), Spacing (SP) [44] and Maximum
Spread (MS) [45] metrics are employed. These performance indicators can be com-
puted as follows:
�∑
no
d2
GD =
i=1 i (15)
no
�
∑nt � 2
(di )
IGD =
i=1 (16)
nt
√
√ no ( )2
√ 1 ∑
SP = √ d − di (17)
no − 1 i=1

m
� � � � � �2 ⎞1∕2
⎛ �
1 min fimax , Fimax − max fimin , Fimin
MS = ⎜ ⎟ (18)
⎜ m i=1 Fimax − Fimin ⎟
⎝ ⎠

where no is the number of obtained Pareto optimal solutions, nt is the number of

true Pareto optimal solutions, di and di indicates the Euclidean distance between the
′

i-th Pareto solution obtained and the closest true Pareto solution and d is the average
of all di . In Eq. (18), m is the number of objectives, fimax and fimin are the extreme
values of the ith objective in obtained PF, Fimax and Fimin are the extreme values of
the ith objective in true PF.
In addition to the above performance indicators, we defined a new metric called
CS to measure the convergence speed of algorithms. CS is calculated by Eq. (19),
and its higher values show the higher convergence speed.

13
 M. Ilchi Ghazaan et al.

Fig. 4  Illustration of how to calculate HVI and HVtruePF

Table 2  Mathematical test functions

Prob- Func- No. of No. of Refer- Prob- Func- No. of No. of Refer-
lems tion vari- objec- ences lems tion vari- objec- ences
ables tives ables tives

F1 ZDT1 30 2 [45] F16 MMF14 3 3 [47]

F2 ZDT2 30 2 [45] F17 MMF15 3 3 [47]
F3 ZDT3 30 2 [45] F18 UF1 30 2 [48]
F4 ZDT4 30 2 [45] F19 UF2 30 2 [48]
F5 ZDT6 30 2 [45] F20 UF3 30 2 [48]
F6 MMF1 2 2 [47] F21 UF4 30 2 [48]
F7 MMF2 2 2 [47] F22 UF5 30 2 [48]
F8 MMF4 2 2 [47] F23 UF6 30 2 [48]
F9 MMF5 2 2 [47] F24 UF7 30 2 [48]
F10 MMF7 2 2 [47] F25 DTLZ1 7 3 [49]
F11 MMF8 2 2 [47] F26 DTLZ2 12 3 [49]
F12 MMF10 2 2 [47] F27 DTLZ3 12 3 [49]
F13 MMF11 2 2 [47] F28 DTLZ4 12 3 [49]
F14 MMF12 2 2 [47] F29 DTLZ5 12 3 [49]
F15 MMF13 3 2 [47] F30 DTLZ6 12 3 [49]

MaxIter−1
∑ HVRi+1 − HVRi
CS = (19)
i=1
i+1

13
A fast convergence EO‑based multi‑objective optimization…

where
HVi
HVRi = (20)
HVtruePF

HVi is the hypervolume [46] of the Pareto front obtained in the ith iteration and
HVtruePF is the hypervolume of the true Pareto front. Figure 4 shows the HV value
for a two-objective Pareto front for a better explanation. In this figure, the area dis-
played between the obtained Pareto at the ith iteration and the reference point is
equal to HVi.
HVR value tends to one during the optimization process; if the obtained Pareto is
the same as the true Pareto, its value will be one. In CS, both the convergence value
and the convergence time are considered in estimating the convergence speed.

Table 3  Comparison among algorithm on two-objective problems using CRS4MOEA/QIE

Problems FC-MOEO/AEP MOSMA MOMVO NSGA-III
R RD σ R RD σ R RD σ R RD σ

F1 1831 (1) 68 0.059 1696 (2) 64 0.059 1251 (3) 66 0.059 1014 (4) 76 0.059
F2 1782 (1) 65 0.059 1650 (2) 61 0.059 1302 (3) 62 0.059 1125 (4) 72 0.059
F3 1837 (1) 74 0.059 1587 (2) 62 0.059 1278 (3) 60 0.059 1116 (4) 65 0.059
F4 1539 (2) 51 0.060 1765 (1) 57 0.059 1323 (4) 57 0.059 1481 (3) 51 0.059
F5 1867 (1) 77 0.059 1387 (3) 54 0.059 1283 (4) 57 0.059 1436 (2) 54 0.059
F6 1590 (1) 50 0.060 1429 (2) 49 0.060 1414 (3) 49 0.059 1363 (4) 50 0.060
F7 1606 (1) 54 0.059 1575 (2) 51 0.059 1484 (3) 51 0.060 1246 (4) 58 0.060
F8 1679 (1) 56 0.059 1584 (2) 53 0.059 1292 (4) 55 0.059 1366 (3) 54 0.059
F9 1598 (1) 55 0.060 1279 (4) 52 0.060 1481 (2) 50 0.059 1436 (3) 52 0.06
F10 1621 (1) 54 0.060 1526 (2) 50 0.060 1384 (4) 51 0.059 1441 (3) 51 0.059
F11 1581 (2) 51 0.059 1657 (1) 55 0.060 1317 (4) 55 0.059 1492 (3) 51 0.060
F12 1756 (1) 64 0.059 1413 (3) 52 0.059 1289 (4) 53 0.060 1460 (2) 52 0.060
F13 1553 (1) 50 0.059 1482 (2) 49 0.060 1407 (4) 50 0.060 1479 (3) 49 0.060
F14 1636 (1) 54 0.059 1314 (4) 52 0.060 1435 (3) 50 0.060 1549 (2) 50 0.060
F15 1555 (1) 50 0.059 1518 (2) 49 0.059 1354 (4) 51 0.060 1444 (3) 49 0.059
F18 1555 (1) 51 0.059 1363 (4) 58 0.060 1536 (2) 50 0.059 1530 (3) 51 0.059
F19 1631 (1) 51 0.059 1498 (2) 49 0.059 1412 (4) 50 0.059 1421 (3) 49 0.060
F20 1513 (3) 52 0.060 1728 (1) 58 0.060 1317 (4) 60 0.059 1542 (2) 53 0.059
F21 1687 (1) 54 0.059 1427 (2) 49 0.059 1424 (3) 49 0.059 1395 (4) 50 0.059
F22 1699 (1) 58 0.060 1291 (4) 57 0.059 1517 (2) 52 0.059 1401 (3) 52 0.059
F23 1632 (1) 53 0.059 1300 (4) 54 0.059 1538 (2) 51 0.059 1389 (3) 51 0.059
F24 1566 (1) 49 0.059 1330 (4) 51 0.059 1517 (2) 51 0.059 1440 (3) 49 0.059
All F 1626 (1) 49 0.059 1420 (2) 47 0.061 1391 (3) 47 0.060 1389 (4) 47 0.060

13
 M. Ilchi Ghazaan et al.

Table 4  Comparison among algorithm on three-objective problems using CRS4MOEA/QIE

Problems FC-MOEO/AEP MOSMA MOMVO NSGA-III
R RD σ R RD σ R RD σ R RD σ

F16 1548 (2) 50 0.059 1460 (3) 51 0.059 1348 (4) 53 0.059 1733 (1) 56 0.059
F17 1770 (1) 60 0.059 1232 (4) 63 0.059 1453 (3) 54 0.059 1535 (2) 54 0.059
F25 1736 (1) 56 0.059 1204 (4) 67 0.059 1523 (3) 53 0.059 1529 (2) 53 0.059
F26 1704 (1) 58 0.059 1219 (4) 58 0.059 1440 (3) 52 0.059 1490 (2) 52 0.059
F27 1597 (2) 52 0.059 1295 (4) 64 0.059 1531 (3) 52 0.059 1634 (1) 53 0.059
F28 1706 (1) 61 0.059 1328 (3) 54 0.059 1484 (2) 53 0.059 1285 (4) 55 0.059
F29 1550 (1) 49 0.059 1401 (4) 49 0.059 1500 (3) 49 0.060 1531 (2) 49 0.059
F30 1737 (1) 64 0.060 1291 (4) 65 0.060 1374 (3) 56 0.059 1616 (2) 56 0.060
All F 1638 (1) 49 0.059 1288 (4) 51 0.059 1489 (3) 47 0.059 1556 (2) 47 0.060

4.2 Results on mathematical problems

Thirty non-constraint benchmark test functions with known Pareto fronts are
employed for validating the performance of FC-MOEO/AEP. These problems are
described in Table 2 and include ZDT, CEC 2020, CEC 2009 and DTLZ test func-
tions [45, 47–49].
When comparing the performance of multiple multi-objective algorithms over
multiple problems, it is common to calculate a summary statistic, such as the aver-
age or median, of the quality indicator values obtained on each problem for each
algorithm. However, this approach may not fully capture the differences in per-
formance between the algorithms, as it does not take into account the variance or
distribution of the data. Additionally, statistical tests may be required to determine
whether any observed differences in performance are statistically significant.
We used the chess rating system for multi-objective evolutionary algorithms with
quality indicator ensemble (CRS4MOEA/QIE) [50] in this paper. CRS4MOEA/
QIE is based on the well-known Glicko-2 system [51, 52], in which each MOEA
is ranked like a chess player. In the Glicko-2 system, each player receives a rating
R, rating deviation RD, and rating volatility σ. The rating represents the player’s
skill, and it generally increases if players perform better than expected and decreases
if they perform worse than expected. The rating deviation indicates how reliable a
player’s rating is; a small rating deviation means that a player has a reliable rat-
ing, while a high rating deviation means that their rating is unreliable. The vola-
tility measure σ indicates the expected degree of rating fluctuation for a player. A
high volatility measure indicates that a player has an unpredictable performance,
while a low volatility measure means that a player performs consistently. Using this
approach, we can successfully rank MOEAs and produce comparable results while
also providing additional information, allowing us to test the statistical performance
of MOEAs.

13
A fast convergence EO‑based multi‑objective optimization…

Fig. 5  Results of Pareto optimal front realized by FC-MOEO/AEP algorithm on a MMF10, b MMF11, c
MMF12, d MMF13, e MMF14, and f MMF15

Tables 3 and 4 show the results of the two-objective and three-objective func-
tions, respectively. Each row in the tables represents one tournament containing
a single problem, with the best results marked in boldface and the rank indi-
cated in parentheses. Since CRS4MOEA/QIE tournaments can include multiple

13
 M. Ilchi Ghazaan et al.

Fig. 6  Evolutionary trajectories of mean IGD value obtained by FC-MOEO/AEP and MOSMA on a
ZDT1, b ZDT2, and c UF6

problems, a new tournament was conducted to evaluate all problems together,

and its result is presented in the last row.
According to Tables 3 and 4, it is evident that FC-MOEO/AEP achieved bet-
ter results than its competitors on almost all test functions. Based on the results
in the last row, we can say with 95% of confidence that FC-MOEO/AEP over-
all strength for two-objective and three-objective problems lies in the interval
(1528, 1724) and (1540, 1739), respectively. Therefore, the proposed algorithm
generally outperforms other algorithms considering all aspects of algorithm
capability, including convergence, coverage, and speed.
Figure 5 shows the Pareto optimal fronts attained by the proposed FC-MOEO/
AEP on the test functions F13–F17 to observe the outcomes qualitatively.
F12–F17 are challenging test functions with the local PF or multi-model PS
[23]. It can be seen that the algorithm on these problems did not fall into the trap
of local Pareto and was able to approximate the global Pareto as well.
Figure 6 plots the evolutionary trajectories of mean IGD value obtained on
three problems (ZDT1, ZDT2, UF6) to analyze the convergence speed of algo-
rithms. As reported in Ref., MOSMA has promising performance in terms of
convergence speed and is faster than several multi-objective algorithms. This
study also observes that MOSMA exhibits higher convergence speeds than
NSGA-III and MOMVO. Nevertheless, Fig. 6a and b show that FC-MOEO/AEP
can obtain nearly the same MOSMA results with less than a quarter of computa-
tional resources (evaluations) in ZDT1 and ZDT2. This proves that the proposed
algorithm has excellent performance in terms of convergence speed. Addition-
ally, Fig. 6 shows that the algorithm with a larger CS value has a faster evolu-
tionary trajectory of mean IGD value. This indicates that the proposed perfor-
mance indicator for comparing the convergence speed of algorithms is reliable
and useful.

4.3 Results on structural design problems

The proposed FC-MOEO/AEP and other selected algorithms are applied to three
structural design optimization problems, including:

13
 Table 5  Details of structural design problems
Design problem No. of vari- No. of elements Type Objective functions Selectable sections Design guide
ables

Tower truss 32 582 Discrete (1) Total structural volume 140 W-shaped standard steel sections AISC [53]
(2) Displacement at nodes
A fast convergence EO‑based multi‑objective optimization…

Double-layer grid 31 1520 Discrete (1) Total structural weight 37 Pipe sections standard steel sections AISC [53]
(2) Displacement at nodes
Frame 11 105 Discrete (1) Total structural weight 267 W-shaped standard steel sections AISC [53]
(2) Inter-story drift

13
 M. Ilchi Ghazaan et al.

Fig. 7  Schematic of the 582-bar tower truss

(1) A tower truss problem

(2) A double-layer grid problem
(3) A frame problem

The details of these three problems are summarized in Table 5.

The design provisions are considered as constraints of the optimization problems.
The penalty function method is utilized as a constraint-handling approach, so the con-
strained objective function is expressed as follows:
)𝜀
(21)
(
P({X}) = 1 + 𝜀1 .𝜐 2 × Obj({X})

where P({X}) is the penalized objective functions to be minimized, Obj({X}) pre-

sents the value of objective functions based on Table 5 for each problem, and 𝜐
denotes the sum of the violations of the design constraints. Here, 𝜀1 is set to unity,
and 𝜀2 is calculated by
iter
𝜀2 = 1.5 + 1.5 × (22)
iter_max
where iter is the current iteration number, and iter_max is the total number of itera-
tions for the optimization process.

13
A fast convergence EO‑based multi‑objective optimization…

Fig. 8  Location of the thirty Pareto front produced by a FC-MOEO/AEP, b NSGA-III, c MOSMA, and d
MOMVO algorithm on 582-bar tower truss

4.3.1 A spatial 582‑bar tower truss problem

The schematic of a spatial 582-bar tower truss shown in Fig. 7 is divided into 32
groups, considering the structural symmetry. A single load case is considered: lateral
loads of 1.12 kips applied in both x- and y-directions and a vertical load of − 6.74 kips
applied in the z-direction at all nodes of the tower. Based on the AISC code, the allow-
able tensile stress (𝜎i+) for tension members and the allowable stress limit for compres-
sion members (𝜎i−) are calculated by
𝜎i+ = 0.6Fy (23)

⎧ �� 𝜆2i 𝜆3i
� �� � �
5 3𝜆i
⎪ 1− 2C2
Fy 3
+ 8Cc
− 8Cc3
for 𝜆i < Cc
𝜎i− = ⎨ 12𝜋 2 E c
(24)
⎪ 23𝜆2 for 𝜆i ≥ Cc
⎩ i

where Fy and E are respectively the yield stress and modulus of elasticity of steel, 𝜆i
is the slenderness ratio (𝜆i = kLi ∕ri ), k is the effective length factor of members, Li is
the length of the ith member, ri is the radius of gyration of the ith member and Cc is

13
 M. Ilchi Ghazaan et al.

Fig. 9  Compare the best Pareto front produced by FC-MOEO/AEP and MOSMA

Fig. 10  3D view of the 1520-bar double-layer grid problem

the slenderness
√ ratio dividing the elastic and inelastic buckling regions
(Cc = 2𝜋 2 E∕Fy ).
According to the provisions of ASD-AISC, the maximum slenderness ratios are
limited to 300 and 200 for tension and compression members, respectively. Nodal
displacements in all coordinate directions must be less than ± 3.15 inches. In this
design problem, the total structural volume and the displacement at nodes in all
directions have to be minimized simultaneously.
As mentioned before, 30 independent runs were carried out for optimization
problems. To make the results more comprehensible, in Fig. 8, the location of the
thirty Pareto obtained by each algorithm is shown as a heat map. The denser areas

13
A fast convergence EO‑based multi‑objective optimization…

Fig. 11  Top view of the 1520-bar double-layer grid problem and member groups: a all members with
simple supports, b bottom layer members, c top layer members, and d web members

are marked as warmer colors. It can be seen that MOMVO and NSGAIII algorithms
have weaknesses in searching different areas of the objective space. Almost all of the
Pareto proposed by these two algorithms are limited to specific regions. Whereas,
based on Fig. 8a and c, the FC-MOEO/AEP and MOSMA algorithms have provided
various optimal solutions.
Since the MOSMA performed better than the NSGAIII and MOMVO, only the
optimized non-dominated solutions obtained by MOSMA and FC-MOEO/AEP are
compared in Fig. 9. The best Pareto fronts from these two multi-objective optimiza-
tion methods are presented in Fig. 9. It can be seen that the PF set obtained using
MOSMA is dominated by the one obtained using FC-MOEO/AEP.

13
 M. Ilchi Ghazaan et al.

4.3.2 A 1520‑bar double‑layer grid problem

Figure 10 shows the 3D view of a diagonal on a diagonal grid. This structure has
1520 members and 401 nodes, and the bottom layer is simply supported at the nodes
shown in Fig. 11. A span of 40 × 40 m is considered for this structure and the height
is equal to 3 m. All connections are assumed to be ball jointed. Each top layer joint
is subjected to a concentrated vertical load of 16kN. The cross-sectional areas of the
members are categorized into 31 groups as depicted in Fig. 11. Different methods
investigated this case as a single objective (weight optimization) problem. However,
this study considers two conflicting objectives for this large-scale design problem:
minimization simultaneously (1) the total structural weight and (2) maximum dis-
placement of all nodes in all coordinate directions.
According to AISC, the constraint conditions for grid structures are briefly
explained below. The displacement constraint is represented by Eq. (25), where 𝛿i
and 𝛿imax are the displacements and allowable displacement for the ith node, respec-
tively, and m is the number of nodes. Displacement limitations of span/600 were
imposed on all nodes in the vertical direction.
𝛿i ≤ 𝛿imax , i = 1, 2, … , m (25)

Fig. 12  Location of the thirty Pareto front produced by a FC-MOEO/AEP, b NSGA-III, c MOSMA, and
d MOMVO algorithm on 1520-bar double-layer grid

13
A fast convergence EO‑based multi‑objective optimization…

Fig. 13  Compare the best Pareto front produced by FC-MOEO/AEP and MOMVO

Tension and compression member constraint are calculated using Eqs. (26) and
(27), respectively.
{
𝜙t Fy Ag ; 𝜙t = 0.90
pu ≤ pr ; pr = min
𝜙t Fu Ae ; 𝜙t = 0.75 (26)

pu ≤ pr ; pr = 𝜙c Fcr Ag ; 𝜙c = 0.85 (27)

� Fy
� �
⎧ 0.658 Fe F ; KL E
y ≤ 4.71
r Fy 𝜋2E
(28)
⎪
Fcr = ⎨ � ; Fe = � �2
KL E KL
⎪ 0.877Fe ;
r
> 4.71 Fy
⎩ r

where pu and pr are the required strength and the nominal axial strength respec-
tively, Fy and Fu are the yield stress and tensile strength, respectively, E is modulus
of elasticity, Ag and Ae are the gross cross-sectional area and the effective net cross-
sectional area of a member, r is the radius of gyration, k is the effective length factor,
and L is the length of member. Also, the maximum allowable Slenderness ratios
( kL∕r ) are 200 and 300 for compression and tension members.
The location of the Pareto fronts obtained during the 30 runs by each algo-
rithm is shown in the heat map of Fig. 12. This figure indicates that the FC-
MOEO/AEP algorithm outperforms other algorithms in terms of spacing/spread.
The MOMVO has the best convergence among the three competitor algorithms
because its Pareto is dominated almost all optimal solutions represented by
MOSMA and NSGAIII. So, in Fig. 13, the best Pareto provided by the MOMVO
is compared with the best Pareto of the FC-MOEO/AEP to demonstrate the con-
vergence quality of the FC-MOEO/AEP.

13
 M. Ilchi Ghazaan et al.

Fig. 14  Schematic of the 3-bay 15-story frame

13
A fast convergence EO‑based multi‑objective optimization…

As seen in this figure, the MOMVO algorithm is dominated FC-MOEO/AEP’s

solutions in a limited interval [(7.5e6, 0.065), (1.3e5, 0.04)]. Nevertheless, con-
sidering the full Pareto, we can say that the performance of the proposed algo-
rithm has been better than MOMVO.

4.3.3 A 3‑bay 15‑story frame problem

Figure 14 represents the schematic of a 3-bay 15-story frame. The applied loads and
the numbering of member groups are also shown in this figure. The effective length
factors of the members are calculated as kx ≥ 0 for a sway-permitted frame, and the
out-of-plane effective length factor is specified as ky = 1. Each column is considered
non-braced along its length, and the non-braced length for each beam member is speci-
fied as one-fifth of the span length.
Limitations on displacement and strength are imposed according to the provisions of
the AISC. Lateral displacement and inter-story drift constraints is defined as Eqs. (29)
and (30)
ΔT
≤R (29)
H

di
≤ R, i = 1, 2, … , ns (30)
hi

where ΔT is the maximum lateral displacement, H is the height of the frame struc-
ture, di is the inter-story drift, hi is the story height of the ith floor, ns is the total
number of stories, and R is the maximum drift index which is equal to 1/300.
Strength constraints are imposed as follows:
{ Pu M P
2𝜙c Pn
+ 𝜙 Mu ≤ 1, for 𝜙 Pu < 0.2
Pu
b n
8Mu
c n
Pu (31)
𝜙P
+ 9𝜙 M
≤ 1, for 𝜙P
≥ 0.2
c n b n c n

where Pu is the required strength, Pn is the nominal axial strength, 𝜙c is the resist-
ance factor (𝜙c = 0.9 for tension, 𝜙c = 0.85 for compression), Mu is the required
flexural strengths, Mn is the nominal flexural strengths, and 𝜙b denotes the flexural
resistance reduction factor (𝜙b = 0.9).
The nominal tensile strength for yielding and the nominal compressive strength
of a member are calculated by Eq. (32, 33)
Pn = Ag ⋅ Fy (32)

Pn = Ag ⋅ Fcr (33)
where

13
 M. Ilchi Ghazaan et al.

Fig. 15  Location of the thirty Pareto front produced by a FC-MOEO/AEP, b NSGA-III, c MOSMA, and
d MOMVO algorithm on 15-story frame

Fig. 16  Compare the best Pareto front produced by FC-MOEO/AEP and NSGA-III

13
A fast convergence EO‑based multi‑objective optimization…

� �
⎧ 𝜆2
⎪ Fcr = �0.658� c Fy , for 𝜆c ≤ 1.5
⎨ 0.877 (34)
⎪ Fcr = 𝜆2 Fy , for 𝜆c > 1.5
⎩ c

√
kl Fy
𝜆c = (35)
r𝜋 E

where Ag is the cross-sectional area of a member, Fy is the yield stress, E is modulus

of elasticity, r is the radius of gyration, l is the length of member, and k is the effec-
tive length factor that is compute by
√ ( )
1.6GA GB + 4 GA + GB + 7.5
k= (36)
GA + GB + 7.5

where GA and GB are stiffness ratios of columns and girders at the two end joints, A
and B, of the column section, respectively.
It is desired to minimize the total weight of the structure and the maximum of
inter-story drift, simultaneously. The distributions of the resulting Pareto fronts from
four optimization methods are presented in Fig. 15. Also, the algorithm that has
the best convergence between competitors, namely NSGAII, is compared with FC-
MOEO/AEP in Fig. 16. Based on the presented results, it can be concluded that the
FC-MOEO/AEP algorithm has an acceptable performance in this design problem as
well.

4.3.4 Brief discussions

Since the true Pareto of these multi-objective design problems is unknown, we

evaluated the performance of the proposed algorithm by comparing it with three
of the most powerful algorithms introduced in recent years. Studying the results
obtained from the algorithms showed that the FC-MOEO/AEP has better perfor-
mance than its three competing algorithms in terms of convergence and cover-
age. It was also observed that the three competing algorithms performed non-
uniformly in solving various problems. For example, the MOSMA algorithm
has given excellent results in the truss problem, but it has the worst performance
compared to other algorithms in solving the grid problem. In return, FC-MOEO/
AEP showed a reliable and stable performance in solving structural problems.
Therefore, the FC-MOEO/AEP method can be used with high confidence to opti-
mize various problems in structural engineering.

13
 M. Ilchi Ghazaan et al.

5 Conclusion

Many engineering problems require the handling of several conflicting objective

functions. Evolutionary algorithms have been considered as an efficient approach
to handle these complex optimization problems. These approaches search for
optimal solutions by evaluating objective functions several times. However,
evaluating objective functions is time-consuming and costly in many engineer-
ing applications such as structural optimal design problems. This paper proposes
a new evolutionary algorithm called FC-MOEO/AEP, which has a high conver-
gence speed and is suitable for solving such problems. In addition to high con-
vergence speed, this algorithm created an intelligent balance between exploration
and exploitation. This ability allows the algorithm to save itself easily from the
local Pareto and estimate the global Pareto front with reasonable accuracy and
spread.
FC-MOEO/AEP is a multi-objective version of EO equipped with an AEP mech-
anism. AEP considers the evolution path in the decision space and predicts the
potential regions for optimal solutions. In AEP, the potential regions are estimated
using a vector called the evolution vector, which its magnitude and direction are
updated in each iteration.
The FC-MOEO/AEP was tested on 30 mathematical benchmark problems,
including ZDT, DTLZ, CEC 2020, and CEC 2009. The four performance indica-
tors are utilized for quantifying coverage and convergence of Pareto. In addition to
these performance indicators, we defined a new metric called CS to measure the
convergence speed of algorithms. CS uses the hypervolume value obtained in each
iteration and finally presents a number representing the algorithm’s convergence
speed value. The higher the value of this number for an algorithm, the higher the
convergence speed of that algorithm. Ultimately, the CRS4MOEA/QIE mechanism
was used to compare the results comprehensively and fairly. It was observed that the
proposed algorithm could achieve superior results than its competitors and perform
better in 83% of problems. Also, three real-size structural problems with various
constraints and discrete variables were investigated. The performance of the FC-
MOEO/AEP in these three problems demonstrated that this algorithm is applicable
in practical problems and has reliable performance. Therefore, the proposed algo-
rithm can be used with high confidence to optimize various engineering problems,
especially structural design problems.
It is advisable to explore the utilization of an alternative optimization methodol-
ogy, rather than EO, within the FC-MOEO/AEP framework. Furthermore, studying
many objectives optimization problems using FC-MOEO/AEP is recommended for
future works.

Author contributions MIG: Conceptualization, Supervision, Writing—Review and EditingPG: Concep-

tualization, Supervision, Writing—Review and EditingAR: Software, Formal analysis, Validation, Writ-
ing—Original Draft.

Funding No funds, grants, or other support was received.

13
A fast convergence EO‑based multi‑objective optimization…

Declarations
Conflict of interest The authors have no relevant financial or non-financial interests to disclose.

References
1. Stewart T, Bandte O, Chakraborti N et al (2008) Real-world applications of multiobjective optimi-
zation. In: Branke J et al (eds) Multiobjective optimization: interactive and evolutionary approaches.
Springer, Berlin, pp 285–327. https://​doi.​org/​10.​1007/​978-3-​540-​88908-3
2. Ivanov SY, Ray AK (2016) Application of multi-objective optimization in the design and operation
of industrial catalytic reactors and processes. Phys Sci Rev. https://​doi.​org/​10.​1515/​psr-​2015-​0017
3. Abitha R, Vennila SM, Zaheer IM (2022) Evolutionary multi-objective optimization of artificial
neural network for classification of autism spectrum disorder screening. J Supercomput 78:11640–
11656. https://​doi.​org/​10.​1007/​s11227-​021-​04268-4
4. Cao L, Liu Y (2022) Optimization design and research of simulation system for urban green eco-
logical rainwater by genetic algorithm. J Supercomput 78:11318–11344. https://​doi.​org/​10.​1007/​
s11227-​021-​04192-7
5. Li S, Chen H, Wang M, Heidari A, Mirjalili S (2020) Slime mould algorithm: a new method for
stochastic optimization. Future Gener Comput Syst 111:300–323. https://​doi.​org/​10.​1016/j.​future.​
2020.​03.​055
6. Yang Y, Chen H, Heidari A, Gandomi A (2021) Hunger games search: visions, conception,
implementation, deep analysis, perspectives, and towards performance shifts. Expert Syst Appl
177:114864. https://​doi.​org/​10.​1016/j.​eswa.​2021.​114864
7. Ahmadianfar I, Heidari A, Noshadian S, Chen H, Gandomi A (2022) INFO: an efficient optimiza-
tion algorithm based on weighted mean of vectors. Expert Syst Appl 195:116516. https://​doi.​org/​10.​
1016/j.​eswa.​2022.​116516
8. Mirjalili S, Dong JS (2020) Multi-objective optimization using artificial intelligence techniques.
Springer, Cham. https://​doi.​org/​10.​1007/​978-3-​030-​24835-2
9. Das I, Dennis JE (1998) Normal-boundary intersection: a new method for generating the pareto sur-
face in nonlinear multicriteria optimization problems. SIOPT 8:631–657. https://​doi.​org/​10.​1137/​
S1052​62349​63075​10
10. Kim IY, de Weck OL (2005) Adaptive weighted-sum method for bi-objective optimization: Pareto
front generation. Struct Multidiscipl Optim 29:149–158. https://​doi.​org/​10.​1007/​s00158-​004-​0465-1
11. Messac A, Mattson CA (2002) Generating well-distributed sets of Pareto points for engineering
design using physical programming. Optim Eng 3:431–450. https://​doi.​org/​10.​1023/A:​10211​79727​
569
12. Meignan D, Knust S, Frayret JM et al (2015) A review and taxonomy of interactive optimization
methods in operations research. ACM Trans Interact Intell Syst 5:1–43. https://​doi.​org/​10.​1145/​
28082​34
13. Deb K, Pratap A, Agarwal S et al (2002) A fast and elitist multiobjective genetic algorithm: NSGA-
II. IEEE Trans Evol Comput 6:182–197. https://​doi.​org/​10.​1109/​4235.​996017
14. Mirjalili S, Gandomi AH, Mirjalili SZ et al (2017) Salp swarm algorithm: a bio-inspired optimizer
for engineering design problems. Adv Eng Softw 114:163–191. https://​doi.​org/​10.​1016/j.​adven​
gsoft.​2017.​07.​002
15. Wei Q, Huang D, Zhang Y (2021) Artificial chicken swarm algorithm for multi-objective
optimization with deep learning. J Supercomput 77:13069–13089. https://​doi.​org/​10.​1007/​
s11227-​021-​03770-z
16. Akbari R, Hedayatzadeh R, Ziarati K, Hassanizadeh B (2012) A multi-objective artificial bee col-
ony algorithm. Swarm Evol Comput. https://​doi.​org/​10.​1016/j.​swevo.​2011.​08.​001
17. Zhang Q, Li H (2007) MOEA/D: a multiobjective evolutionary algorithm based on decomposition.
IEEE Trans Evol Comput. https://​doi.​org/​10.​1109/​TEVC.​2007.​892759
18. Zitzler E, Kunzli E (2004) Indicator-based selection in multiobjective search. In: Yao X et al (eds)
Parallel problem solving from nature. Springer, Berlin, pp 832–842. https://​doi.​org/​10.​1007/​978-3-​
540-​30217-9_​84

13
 M. Ilchi Ghazaan et al.

19. Tian Y, Cheng R, Zhang X et al (2018) An indicator-based multiobjective evolutionary algorithm

with reference point adaptation for better versatility. IEEE Trans Evol Comput. https://​doi.​org/​10.​
1109/​TEVC.​2017.​27496​19
20. Bader J, Zitzler E (2011) HypE: an algorithm for fast hypervolume-based many-objective optimiza-
tion. Evol Comput. https://​doi.​org/​10.​1162/​EVCOa​00009
21. Das S, Suganthan PN (2010) Differential evolution: a survey of the state-of-the-art. IEEE Trans
Evol Comput. https://​doi.​org/​10.​1109/​TEVC.​2010.​20590​31
22. Chiang TC, Lai YP (2011) MOEA/D-AMS: improving MOEA/D by an adaptive mating selection
mechanism. IEEE Trans Evol Comput (CEC). https://​doi.​org/​10.​1109/​CEC.​2011.​59497​89
23. Abdel-Basset M, Mohamed R, Abouhawwash M (2021) Balanced multi-objective optimization
algorithm using improvement based reference points approach. Swarm Evol Comput. https://​doi.​
org/​10.​1016/j.​swevo.​2020.​100791
24. Kaveh A, Mahdavi VR (2018) Multi-objective colliding bodies optimization algorithm for design of
trusses. J Comput Des Eng. https://​doi.​org/​10.​1016/j.​jcde.​2018.​04.​001
25. Kaveh A, Ilchi Ghazaan M (2020) A new VPS-based algorithm for multi-objective optimization
problems. Eng Comput. https://​doi.​org/​10.​1007/​s00366-​019-​00747-8
26. Liu L, Huo J (2018) Apple image recognition multi-objective method based on the adaptive har-
mony search algorithm with simulation and creation. Information. https://​doi.​org/​10.​3390/​info9​
070180
27. Kaveh A, Ghazaan MI (2018) Meta-heuristic algorithms for optimal design of real-size structures.
Springer, Cham
28. Amini F, Ghaderi P (2013) Hybridization of harmony search and ant colony optimization for opti-
mal locating of structural dampers. Appl Soft Comput. https://​doi.​org/​10.​1016/j.​asoc.​2013.​02.​001
29. Zheng Y, Zhang L, Pan Y, He Z (2020) Multi-objective structural optimization of a wind turbine
tower. J Shanghai Jiaotong Univ Sci. https://​doi.​org/​10.​1007/​s12204-​020-​2190-3
30. Alkayem NF, Cao M, Zhang Y et al (2018) Structural damage detection using finite element model
updating with evolutionary algorithms: a survey. Neural Comput Appl. https://​doi.​org/​10.​1007/​
s00521-​017-​3284-1
31. Gentils T, Wang L, Kolios A (2017) Integrated structural optimisation of offshore wind turbine sup-
port structures based on finite element analysis and genetic algorithm. Appl Energy. https://​doi.​org/​
10.​1016/j.​apene​rgy.​2017.​05.​009
32. Igel C, Hansen N, Roth S (2007) Covariance matrix adaptation for multi-objective optimization.
Evol Comput. https://​doi.​org/​10.​1162/​evco.​2007.​15.1.1
33. He X, Zhou Y, Chen Z (2019) An evolution path-based reproduction operator for many-objective
optimization. IEEE Trans Evol Comput. https://​doi.​org/​10.​1109/​TEVC.​2017.​27852​24
34. Song W, Du W, Fan C et al (2020) A novel path-based reproduction operator for multi-objective
optimization. Swarm Evol Comput. https://​doi.​org/​10.​1016/j.​swevo.​2020.​100741
35. Faramarzi A, Heidarinejad M, Stephens B et al (2020) Equilibrium optimizer: a novel optimization
algorithm. Knowl Based Syst. https://​doi.​org/​10.​1016/j.​knosys.​2019.​105190
36. Menchaca-Mendez A, Coello CAC (2016) Selection mechanisms based on the maximin fitness
function to solve multi-objective optimization problems. Inf Sci. https://​doi.​org/​10.​1016/j.​ins.​2015.​
11.​008
37. Li X (2004) Better spread and convergence: particle swarm multiobjective optimization using the
maximin fitness function. In: Deb K (ed) Genetic and evolutionary computation. Springer, Berlin,
pp 117–128
38. Martí L, Garcia J, Berlanga A et al (2016) A stopping criterion for multi-objective optimization evo-
lutionary algorithms. Inf Sci. https://​doi.​org/​10.​1016/j.​ins.​2016.​07.​025
39. Deb K, Jain H (2013) An evolutionary many-objective optimization algorithm using reference-
point-based nondominated sorting approach, part I: solving problems with box constraints. IEEE
Trans Evol Comput. https://​doi.​org/​10.​1109/​TEVC.​2013.​22815​35
40. Premkumar M, Jangir P, Sowmya R et al (2021) MOSMA: multi-objective slime mould algorithm
based on elitist non-dominated sorting. IEEE Access. https://​doi.​org/​10.​1109/​ACCESS.​2020.​30479​
36
41. Mirjalili S, Jangir P, Mirjalili SZ et al (2017) Optimization of problems with multiple objectives
using the multi-verse optimization algorithm. Knowl Based Syst. https://​doi.​org/​10.​1016/j.​knosys.​
2017.​07.​018
42. Van Veldhuizen DA, Lamont GB (1998) Multiobjective evolutionary algorithm research: a history
and analysis. University Park, Citeseer

13
A fast convergence EO‑based multi‑objective optimization…

43. Sierra MR, Coello Coello CA (2005) Improving PSO-based multi-objective optimization using
crowding, mutation and ∈-dominance. In: Coello Coello CA et al (eds) Evolutionary multi-criterion
optimization. Springer, Berlin, pp 505–519
44. Coello CAC, Pulido GT, Lechuga MS (2004) Handling multiple objectives with particle swarm
optimization. IEEE Trans Evol Comput. https://​doi.​org/​10.​1109/​TEVC.​2004.​826067
45. Zitzler E, Deb K, Thiele L (2000) Comparison of multiobjective evolutionary algorithms: empirical
results. Evol Comput. https://​doi.​org/​10.​1162/​10636​56005​68202
46. Zitzler E, Thiele L (1999) Multiobjective evolutionary algorithms: a comparative case study and the
strength Pareto approach. IEEE Trans Evol Comput. https://​doi.​org/​10.​1109/​4235.​797969
47. Liang JJ, Qu BY, Gong DW et al (2019) Problem definitions and evaluation criteria for the CEC
2019 special session on multimodal multiobjective optimization. In: 2019 IEEE congress on evolu-
tionary computation
48. Zhang Q, Zhou A, Zhao S et al (2008) Multiobjective optimization test instances for the CEC 2009
special session and competition. University of Essex, Nanyang Technological University, Colchester
49. Deb K, Thiele L, Laumanns et al (2002) Scalable multi-objective optimization test problems. 2002
IEEE congress on evolutionary computation
50. Ravber M, Mernik M, Črepinšek M (2017) Ranking multi-objective evolutionary algorithms
using a chess rating system with quality indicator ensemble. 2017 IEEE congress on evolutionary
computation
51. Glickman ME (2013) Example of the Glicko-2 system. BU
52. Veček N, Mernik M, Črepinšek M (2014) A chess rating system for evolutionary algorithms: a
new method for the comparison and ranking of evolutionary algorithms. Inf Sci. https://​doi.​org/​10.​
1016/j.​ins.​2014.​02.​154
53. AISC (2001) Specifcation for structural steel buildings. American Institute of Steel Construction,
Chicago

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under
a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such publishing agreement and
applicable law.

13