You are on page 1of 19

Engineering with Computers

https://doi.org/10.1007/s00366-021-01534-0

ORIGINAL ARTICLE

Optimal design of truss structures with frequency constraints:


a comparative study of DE, IDE, LSHADE, and CMAES algorithms
H. Moosavian1 · P. Mesbahi2 · N. Moosavian3 · H. Daliri3

Received: 15 June 2020 / Accepted: 20 October 2021


© The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2021, corrected publication 2022

Abstract
The present study examines the performance of three powerful methods including the original differential evolution (DE), the
improved differential evolution (IDE), and the winner of the CEC-2014 competition, LSHADE, in addition to the covariance
matrix adaptation evolution strategy (CMAES) for size optimization of truss structures under natural frequency constraints.
Despite the abundant researches on novel meta-heuristic algorithms in the literature, the application of CMAES, one of the
most powerful and reliable optimization algorithms, on the optimal solution of the truss structures has received scant atten-
tion. For consistent comparison between these algorithms, four stopping criteria are defined and for each of these criteria, all
algorithms are executed 30 times. Statistical analysis of the results for each algorithm is performed, and the mean, standard
deviation, minimum, and maximum for 30 executions of the algorithms are calculated. For the small population size, results
show that the CMAES algorithm not only has the best performance and the least standard deviation values among other
given algorithms in all cases but also finds the best ever optimal solutions for the design of the benchmark truss structures
which have not been reported in other studies. However, by increasing the number of decision variables and the population
size, the CMAES algorithm needs more function evaluations to converge to the global optimal solution with higher accuracy.

Keywords Covariance matrix adaptation evolution strategy (CMAES) · Truss structures · Frequency constraints · Optimal
design · Statistical tests

1 Introduction Over the past decades, many researchers have conducted


several studies to develop an efficient optimization tech-
The dynamic behavior of the structural system is primarily nique for the optimal design of the structures under fre-
influenced by its natural frequencies and the correspond- quency constraints. The primary works in this field deal
ing mode shapes, especially in low-frequency vibration with traditional gradient-based techniques. Kiusalaas and
problems. For instance, when the frequencies of external Shaw [1] proposed a design algorithm for optimizing struc-
dynamic excitations match with one of the natural frequen- tures subject to lower bound frequency constraints based
cies, the resonance phenomenon occurs. Hence, the ability to on the finite element method and Kuhn–Tucker conditions.
control each system’s natural frequencies and consequently Levy and Chai [2] suggested an effective optimality crite-
prevent the destructive outcomes of the resonance phenom- rion technique for the selection of the cross-sectional area
enon is of paramount importance. to control the frequency of structures and linked it to an
algorithm for natural frequency assessment. Another work
for solving such problems was released by Khot [3]. In this
* H. Moosavian method, a basic optimality criterion is extended, and vari-
hashemmoosavian@yahoo.com ous schemes are also proposed to obtain a specified natural
frequency. Sadek [4] utilized an iterative design procedure
1
Department of Civil Engineering, Sharif University based on Kuhn–Tucker conditions to minimize truss struc-
of Technology, Tehran, Iran
tures with frequency constraints. In the latter work, the size
2
Department of Civil Engineering, Iran University of Science of the cross-sections and the coordinates of nodes are con-
and Technology, Tehran, Iran
sidered as the design variables. Grandhi and Vernkayya [5]
3
Department of Civil Engineering, University of British employed a simple resizing algorithm based on optimality
Columbia, Vancouver, BC, Canada

13
Vol.:(0123456789)
Engineering with Computers

criterion (OC) in conjunction with a scaling technique to optimization. Other novel methods such as vibrating particle
minimize the weight of structures under frequency con- systems algorithm [22], cyclical parthenogenesis algorithm
straints. Sedaghati et al. [6] developed a structural optimi- (CPA) [23], and improved version of the spiral optimization
zation algorithm by the integrated force formulation to find algorithm (SPO) [24] have been proved to be reliable and
the optimum weight of the frame-type and truss structures stable with the desirable speed of convergence when applied
with single or multiple natural frequency constraints. Since to the weight minimization of truss structures under natural
the optimal solution in such problems highly depends on the frequency limitations. Miguel and Miguel [25] tested two
selection of initial points, the application of gradient-based meta-heuristic methods of HS and firefly algorithm (FA)
techniques under frequency constraints encounters serious to perform shape and size optimization of various trusses
difficulties. It is worthwhile to mention that this drawback with multiple fundamental frequency constraints. Zuo et al.
is absent in some optimization problems under stress and [26] combined OC and GA and presented a hybrid OC–GA
displacement constraints [7]. In other words, frequency con- algorithm for finding the least weight of truss configuration,
straints are non-convex, highly nonlinear, and implicit with which satisfies multiple dynamic constraints.
respect to the design variables and these features made the Using the modified versions of passing vehicle
traditional techniques impractical. Due to the advancement search (PVS), Tejani et al. [27] obtained the optimal topol-
of the computational capability of computers, metaheuris- ogy of the truss structures under various static and dynamic
tic algorithms have been known as practical and efficient constraints. Vu [28] proposed a new methodology for achiev-
methods to solve intricate optimization problems regard- ing the optimal node coordinates and cross-section areas of
less of starting points and gradient information. Some of truss structures subject to natural frequency constraints. This
the most significant methods in the field are listed as follows: method is constructed using DE and finite element code.
genetic algorithm (GA) [8], differential evolution (DE) [9], Ho-Huu et al. [29] proposed an improved differential evolu-
covariance matrix adaptation evolution strategy (CMAES) tion (IDE) for solving this type of problem. They used an
[10], soccer league competition (SLC) [11–13], and particle adaptive mutation scheme and an elitist selection algorithm
swarm optimization (PSO) [14]. to improve the performance of basic DE. Another modi-
In this regard, Wei et al. [15] employed a hybrid method, fied DE method in this research direction was developed by
called NHGA, to solve shape and size optimization of truss Pham [30]. This method was named ANDE and carried out
structures. They combined a Niche technique and a simplex three improvements in different phases of DE to preserve
search to prevent premature convergence while enhancing an effective balance between global exploration and local
the global searching capacities of basic GA. Subsequently, exploitation. Another research utilizing DE concepts in this
Wei et al. [16] improved the computational cost of NHGA field was reported by Bureerat and Pholdee [31]. They also
by utilizing the advantages of parallel computing and intro- incorporated a constraint handling technique in their main
duced the NHPGA algorithm with better performance com- algorithm to efficiently deal with constraints of the optimi-
pared to the original version. Among various studies on zation problem. Lieu et al. [32] presented a new adaptive
the optimization of truss structures with multiple natural hybrid evolutionary firefly algorithm (AHEFA) to improve
frequency constraints, the contribution of Kaveh and his co- the accuracy of the final solution and to speed up the con-
workers towards the incorporation of novel meta-heuristic vergence rate. Khatibinia and Naseralavi [33] introduced
algorithms is substantial. Kaveh and Zolghadr [17] utilized a new metaheuristic method named OMGSA which com-
the charged system search algorithm and its improved ver- bines a multi-gravitational search algorithm (MGSA) and
sion for the optimization of the aforementioned problem. an orthogonal cross-over to solve truss optimization prob-
Kaveh and Mahdavi [18] adopted Big Bang–Big Crunch lems with dynamic constraints. Another algorithm adopting
(BB–BC) algorithm combined with the quasi-Newton a modified version of symbiotic organism search (SOS) was
method to effectively search for the optimal solution in introduced by Kumar et al. [34] which improves the effi-
both local and global domains. Kaveh and Zolghadr [19] ciency and accuracy of the search procedure. It should be
introduced the democratic PSO algorithm to simultane- noted that an extensive study on the efficiency and suitability
ously benefit from the advantages of PSO and avoid any of nine popular meta-heuristic algorithms was carried out by
premature convergence of the original algorithm. In another Kaveh and Zolghadr [35]. Furthermore, the performances
study, they proposed a new algorithm through hybridization of the improved ray optimization (IRO) algorithm [36], the
of PSO and ray optimization algorithm (PSRO) to maintain a aging mechanism in PSO algorithm (ALC-PSO), and HS
suitable balance between exploitation and exploration of the mechanism combined with ALC-PSO (HALC-PSO) [37]
approach [20]. In order to tackle this highly nonlinear opti- were examined through comparison with other well-known
mization problem, Kaveh and Javadi [21] enhanced the PSO algorithms.
algorithm by equipping the method with exploitation opera- Among the variety of non-gradient-based methods, covar-
tor in harmony search (HS) and motion refinement in ray iance matrix adaptation evolution strategy (CMAES) [10]

13
Engineering with Computers

has been known as one of the most promising metaheuristic { }



n

methods. This stochastic method employs a derandomized Min 𝜌Li Ai ,


evolution strategy and covariance matrix adaption. Moreo- i=1

ver, to reduce the computational cost, CMAES takes the subjected to


(1)
advantage of parallel computing. The excellent performance 𝜔j ≤ 𝜔U
j , for jth natural frequency,
of this algorithm helps many researchers to apply this algo-
𝜔k ≥ 𝜔Lk , for kth natural frequency,
rithm successfully for solving the most complicated optimi-
zation problems in different engineering fields. For example, ALi ≤ Ai ≤ AU
i , for i = 1, 2, ⋯ , n,
Wang et al. [38] and Su et al. [39] used CMAES for truss
in which n is the total number of truss members and 𝜌 is the
optimization under displacement, stress, and local stability
mass density. Moreover, Li and Ai denote the length and
constraints and compared the results with those obtained
cross-sectional area of the ith member, respectively. ALi and
by other population-based algorithms such as PSO and the
AU represent the lower bound and the upper bound of cross-
evolutionary node shift (ENS). In the work by Ghosh et al. i
sectional area for the ith member of truss. To avoid any unde-
[40], some useful features of classical DE were combined
sired effects such as resonance, the natural frequencies need
with the CMAES algorithm for real parameter optimization.
not fall in the prescribed ranges. For this purpose, 𝜔Lj and 𝜔U
de Melo and Iacca [41] presented a modified CMAES by j

changing the stopping criteria and improving the sampling are introduced in the above equation as the lower and upper
mechanism of the standard CMAES. Then, the algorithm bounds for the jth natural frequency, respectively. The fre-
was investigated for the optimization of several benchmark quency-prohibited bands are determined according to the
functions and various engineering problems. Recently, Yang dominant frequencies of the external agents. Moreover, from
et al. [42] employed CMAES to achieve an optimum design the perspective of control engineering, it is preferable to
of curvilinear stiffener in aircraft panel structures. Moreover, avoid any coincidence of natural frequencies of different
Moosavian and Moosavian [43] compared the performance modes. Thus, manipulation of the natural frequencies ena-
of CMAES with ten popular optimization algorithms for bles a designer to control the final response of the structure
sizing optimization of truss structures considering discrete under different dynamic loadings [5, 44].
decision variables. Statistical analysis of optimal results
showed that this algorithm is efficient and reliable and can
be more investigated for structural design. Despite being 3 Covariance matrix adaptation evolution
employed by the CMAES algorithm in various optimization strategy (CMAES)
problems, the extension of this algorithm for the optimiza-
tion of truss structures under dynamic constraints has not yet In this section, the CMAES algorithm with the proposed
been presented. In this paper, the CMAES algorithm with adaptive penalization for optimization of truss structures is
adaptive penalty function strategy for handling frequency presented. The objective function is denoted f ∶ ℝn → ℝ.
constraints has been implemented and the results are com- CMAES [10] is an iterative stochastic search algorithm for
pared with the DE, IDE, and LSHADE algorithms. In the real-parameter (continuous domain) optimization of non-
next two sections, a brief review of the optimization problem linear, non-convex functions where at each iteration, instead
and then a detailed description of the CMAES algorithm of acting operators on the population, a population of can-
are provided. didate solutions is sampled. In other words, the CMAES
algorithm translates the evolution of the probability distribu-
tion to sample points at each iteration. The algorithm’s main
2 Optimization of truss structures strategy is sampling a population of 𝜆 candidate solutions;
with frequency constraints evaluation of the 𝜆 candidate solutions on f; and adapta-
tion of the sampling distribution using the feedback from
In the truss optimization problem, the total weight of the f obtained. The sampling distribution in CMAES is based
truss structure is minimized such that its natural frequen- on a multivariate normal distribution which is denoted by
cies fulfill some inequality constraints. Throughout the N(𝐦, 𝐂) where 𝐦 is a mean vector of ℝn and 𝐂 is an n × n
optimization, the truss layout is taken fixed, hence, the symmetric positive definite matrix corresponding to the
cross-sectional areas of members are considered as the covariance matrix of the random vector. The random vector
only design variables. In this regard, the design variables is determined based on the set of parameters (𝐦, 𝐂).
are confined to some upper and/or lower bounds. The math- In CMAES, the mean vector 𝐦 represents the best esti-
ematical framework for the optimization problem of the mate of the optimum or favorite solutions, and the covari-
truss structure with the jth natural frequency, 𝜔j can be ance matrix 𝐂 characterizing the geometric shape of the
defined as distribution defines where new solutions are sampled.

13
Engineering with Computers

Moreover, the step-size 𝜎 is also used as a global scaling and cc is the time constant and is defined as a function of
factor for the covariance matrix. Overall, in the first step of number of decision variables n by
the CMAES algorithm, points are sampled according to the 𝜇eff
following formula: 4+
n
cc = . (10)
𝐦 + 𝜎N(𝟎, 𝐂). (2) 2𝜇
n + 4 + eff
n
The adaptation of 𝐦 aims at finding the best estimate of the
The vector 𝐏c moves towards the descent direction
g
optimum, the adaptation of 𝐂 aims at learning the right coor-
dinate system of the problem, and the adaptation of 𝜎 aims (𝐦g+1 − 𝐦g )∕𝜎 g and adds the steps followed by the mean
at achieving fast convergence to an optimum and preventing vector 𝐦 over the iterations implementing some normali-
premature convergence. The basic equation for sampling the zation to dampen previous steps. This operation helps to
search points is given as follows: reduce the effect of the old information and the number of
function evaluations (NFE) [10]. Therefore, the rank-one
g
𝐱i = 𝐦g + 𝜎 g Ni (𝟎, 𝐂g ), g+1 ( g+1 )T
for i = 1, ⋯ , 𝜆. (3) matrix 𝐏c 𝐏c is constructed and added to the covari-
ance matrix 𝐂 to increase the probability of sampling in the
The 𝜆 individuals are sampled and then evaluated and ranked
direction of 𝐏c . The rank-𝜇 matrix is also the stochastic
g+1
according to f:
approximation of the gradient with respect to 𝐂 defined by
∑𝜇
g g g w 𝐲 𝐲T and added to the covariance matrix 𝐂 . By
f (x1∶𝜆 ) ≤ ⋯ ≤ f (x𝜇∶𝜆 ) ≤ ⋯ ≤ f (x𝜆∶𝜆 ), (4) i=1 i i∶𝜆 i∶𝜆
combining the rank-one and rank-𝜇 update, the update of the
where xi∶𝜆 = ith best individual and 𝜇 is the parent popula-
g covariance matrix can be given as
tion size (< 𝜆). Then, the mean 𝐦g is updated by taking the ∑
( g+1 )T 𝜇
weighted mean of the best 𝜇 individuals as 𝐂g+1 = (1 − c1 − c𝜇 )𝐂g + c1 𝐏g+1
c
𝐏c
+ c𝜇
T
wi 𝐲i∶𝜆 𝐲i∶𝜆 ,
i=1

𝜇
g

𝜇
g (11)
(5)
g+1 g g
𝐦 = wi 𝐱i∶𝜆 =𝐦 +𝜎 wi 𝐲i∶𝜆 ,
i=1 i=1 where c1 and c𝜇 are learning rates of the rank-one and rank-𝜇
update, respectively, and defined as
in which
2
g 1 g c1 = , (12a)
𝐲i∶𝜆 = g (𝐱i∶𝜆 − 𝐦g ). (6) (n + 1.3)2 + 𝜇eff
𝜎
In general, 𝜇 = 𝜆∕2 and wi is strictly positive and ⎧ 1 ⎫
2(𝜇eff − 2 + )
⎪ 𝜇eff ⎪

𝜇
c𝜇 = Min⎨1 − c1 , ⎬. (12b)
wi = 1, w1 ≥ w2 ≥ ⋯ ≥ w𝜇 > 0. (7) ⎪ (n + 2)2 + 𝜇eff ⎪
i=1 ⎩ ⎭
By applying Eq. (5), the mean vector moves towards the best Besides the covariance matrix adaptation, the step-size 𝜎 g
∑𝜇
solutions. The term 𝜎 g i=1 wi 𝐲i∶𝜆 is a stochastic approxi-
g
needs to be controlled after every iteration. In this regard, a
mation of the gradient with respect to 𝐦 which depends on conjugate evolution path 𝐏𝜎 ∈ ℝn at iteration g is updated by
g

quantiles of the objective function f. √


The update of the covariance matrix 𝐂g uses two mecha- 1 ( )−1 g+1
𝐏g+1 = (1 − c )𝐏
𝜎 𝜎
g
+ c𝜎 (2 − c𝜎 )𝜇eff 𝐂g (𝐦 − 𝐦g ),
nisms: (a) rank-one update and (b) rank-𝜇 update. The rank-
𝜎 g
𝜎
one update uses the evolution path 𝐏c ∈ ℝn which may be
g (13)
updated by the following equation: where c𝜎 is the learning rate and defined by

1√ 𝜇eff + 2
𝐏g+1
c = (1 − cc )𝐏gc + cc (2 − cc )𝜇eff (𝐦g+1 − 𝐦g ), (8) c𝜎 = . (14)
𝜎g n + 𝜇eff + 5
where 𝜇eff is the variance effective selection mass and is The step-size 𝜎 g is adapted according to
defined by
[ ( g+1 )]
( )−1 c ∥ 𝐏𝜎 ∥
∑ 𝜇 𝜎 g+1 = 𝜎 g exp 𝜎 −1 , (15)
𝜇eff = w2i , (9) d𝜎 E ∥ N(𝟎, 𝐈) ∥
i=1
where d𝜎 is damping parameter and calculated by

13
Engineering with Computers

{ √ } does not involve any user-specified constant. In this study,


𝜇eff − 1
d𝜎 = 1 + 2 Max 0, − 1 + c𝜎 . (16) the adaptive penalty function method presented by Kaveh
N+1
and Zolghadr [17] will be implemented which improves
As the algorithm progresses, the global step-size changes the convergence properties of the CMAES algorithm. It is
dynamically, and the step-size update rule implements to defined as follows:
increase the step-size (exploration) when the length of the { }
conjugate evolution path 𝐏𝜎 is larger than the length of
g 𝛾2
Min f (𝐱) + (1 + 𝛾1 × Λ) f (𝐱) ,
random selection N(𝟎, 𝐈) and decrease it (exploitation) oth- { }

m (19)
erwise. The combination of the step-size adaptation with
Λ= Max (0, hi (𝐱) − di ) ,
covariance matrix adaptation provides linear convergence i=1
on a wide range of functions including ill-conditioned
problems. where Λ denotes the sum of the violated constraints; and the
On the class of convex functions, the covariance matrix parameters 𝛾1 and 𝛾2 are constants representing the explora-
sequence 𝐂g turns out to be proportional to the inverse Hes- tion and the exploitation rates of the search space, respec-
sian of the objective function f, and the algorithm is able to tively. In this paper, 𝛾1 is set to be one, and 𝛾2 is set to linearly
capture second-order information without using any deriva- increase from 1.5 to 6.
tives [45].

5 The statistical tests


4 Handling variable and frequency
constraints with CMAES algorithm For comparing the performance of the algorithms, we apply
two well-known nonparametric statistical procedures: the
As it was alluded to, there are two types of constraints for Friedman test [46, 47] (nonparametric analysis of variance),
the optimal design of truss structures: (a) design variable and the Wilcoxon signed-rank test [48] which are imple-
constraints that restrict the cross-sectional areas and (b) mented in IBM-SPSS Statistics software.
frequency constraints that restrict the natural frequencies. The Friedman test involves a multiple comparisons test
In the literature, several methods are proposed to handle which may be used to detect significant differences between
these constraints in stochastic optimization algorithms. For the responses of two or more subjects to various treatments.
satisfying the variable constraint, two approaches have been For the results presented herein, the response is the least-
implemented. In the first and most widespread approach, weight values, the subjects are the optimization algorithms,
the violated variables are replaced with the specified upper and the treatments are the multiple executions. The Fried-
and lower bounds, and in the second one, if the variable man tests are conducted separately for each stopping cri-
constraint is violated, the value is set to be the middle of the terion. The test determines the ranks of the responses, or
previous value and the specified upper and lower bounds. In least-weight solutions, for each of the algorithms. For each
this study, the first approach has been utilized to handle the algorithm, the mean of these ranks is evaluated. The Fried-
cross-sectional variable constraints. man test statistic, based on the number of subjects, the num-
For satisfying the frequency constraints, unfeasible solu- ber of treatments, and the mean ranks of the responses for
tions can be rejected, penalized, or repaired. Penalization each treatment, is tested against a Chi-square critical value
is the most popular approach used to handle these types of and indicates whether the various treatments or executions
constraints [45]. This approach transforms the constraint are significantly different. For a significance level of 0.05, if
optimization problem to an unconstrained model as the asymptotic significance or p value is greater than 0.05,
{ } the test indicates the similarity of algorithms executions.
Min f (𝐱) subjected to hi (𝐱) ≤ di ∀i = 1, ⋯ , m, In addition to the Friedman test, the present study
(17) adopts the Wilcoxon signed-rank test to demonstrate the
or alternatively performance of the CMAES with respect to DE, IDE, and
{ } LSHADE algorithms in a pairwise manner. By taking the
∑m
CMAES algorithm as the comparison basis, R+ and R− are
Min f (𝐱) + 𝛾(hi (𝐱) − di ) , (18)
i=1 defined as the sum of the ranks in which the CMAES is
superior and inferior to the other given algorithm, respec-
where m is the number of constraints and 𝛾 is the penalty tively. In other words, if R+ is greater than R− , it can be
function which has a large positive value and needs to be deduced that the CMAES has the better performance than
tuned. A solution to avoid the difficulty of tuning these the other algorithm. According to the given parameters
parameters consists in using an adaptive penalization that of R+ and R−, the associated p values can be computed to

13
Engineering with Computers

determine whether the null hypothesis is rejected or not. For are implemented in the MATLAB programming language
example, if the p value is less than 0.05, two distributions are environment with an Intel(R) Core(TM) i7-4800MQ CPU
not equal at the 5 percent level of significance. P8700 @ 2.7 GHz and 16.00 GB RAM.
The results of the statistical tests will be provided in the
next section for each numerical example separately. 6.1 A ten‑bar planar truss

The first example is devoted to the analysis of a ten-bar


6 Numerical examples planar truss structure as illustrated in Fig. 1. The material
density, 𝜌 and modulus of elasticity, E are considered for all
The performance of the CMAES and three powerful DE- members as 2770 kg/m3 and 6.98 × 1010 N/m2 , respectively.
based algorithms are tested for the determination of the min- Furthermore, the parameter L shown in Fig. 1 is taken as
imum weight of four benchmark truss structures: a 10-bar 9.144 m and all nodes are equipped with 454 kg non-struc-
planar truss, a 72-bar space truss, a 120-bar dome truss, and tural masses. The allowable bounds for cross-section areas
a 200-bar planar truss structure with natural frequency con- A are expressed as follows:
straints. These DE-based algorithms are the original DE, an
improved version of DE (IDE) presented by Ho-Huu et al. 0.645 × 10−4 m2 ≤ A ≤ 50 × 10−4 m2 , (20)
[29], and the winner of the competition on real-parameter and first three natural frequencies 𝜔1, 𝜔2 , and 𝜔3 are sub-
single objective optimization at CEC-2014 (LSHADE) pre- jected to the following constraints:
sented by Tanabe and Fukunaga [49].
Throughout the truss analysis, the material behavior of all 𝜔1 ≥ 7 Hz, 𝜔2 ≥ 15 Hz, 𝜔3 ≥ 20 Hz. (21)
truss members is assumed to behave linearly elastic within
infinitesimal deformation. Furthermore, in the optimiza-
The performance of the algorithms, in terms of the mini-
tion process of truss structures, only the size of the bars is
mum total weight, under the four stopping criteria, is shown
allowed to alter and the truss layouts remain unchanged.
in the Tables 1, 2, and 3 , for the 20, 50, and 100 population
Hence, the cross-section of truss members is taken into
size, respectively. The results show that, for the number of
account as design variables. To compute the natural fre-
population = 20 and the number of function evaluations =
quency of each structure, the weights of truss members are
3000, 6000, 10000, and 20000, the CMAES exhibits the
incorporated in addition to the additional non-structural
best minimum weight of 524.6354, 524.4512, 524.4508,
masses attached to some nodes. The position of non-struc-
and 524.4508, respectively. For the number of population
tural masses will be reported in detail for each case later.
= 50 and the number of function evaluations = 3000, the
For consistent comparison between algorithms, four
LSHADE exhibits the best performance in terms of the best
stopping criteria are defined as: (i) the maximum number
minimum and for the number of function evaluations =
of function evaluations is set to 3000. Given these crite-
6000, 10000, and 20000, the CMAES found the best mini-
ria, the initial speed of convergence and performance of the
mum weight of 524.4663 and 524.4508 and shows the best
algorithms may be evaluated; (ii) the maximum number of
performance. However, the CMAES algorithm has the least
function evaluations is set to 6000; (iii) the maximum num-
standard deviation in all cases which shows that the algo-
ber of function evaluations is set to 10000. Here, the per-
rithm has a consistent behavior in the search process, for the
formance of the algorithms is evaluated for a more mature
evolution, compared with (i); (iv) the maximum number of
function evaluations is set to 20000, for which the evolu-
L L
tion of the algorithms is considered to be fully mature. For
each of these criteria, all algorithms are executed 30 times, 1 2
i.e., 30 sets of initial populations of solutions are generated
based on 30 different random number sequences. Statisti- 7 9
cal analysis of results for all algorithms are performed, and
the mean, standard deviation (STD), minimum (Min), and 5 6
maximum (Max) for the 30 executions of the algorithms are L
calculated. The mean shows the overall performance of the
algorithm for different runs. The standard deviation indicates 8 10
the variation of the final solution with respect to the mean
3 4
value. Small standard deviation values show that an algo-
rithm has a similar behavior regarding its search process for
different initial random populations. All the computations Fig. 1  The ten-bar planar truss

13
Engineering with Computers

Table.1  Statistical results NFE Algorithm Mean Max Min STD Mean rank of the The Wilcoxon signed-
of DE, IDE, LSHADE, and Friedman test rank test (comparison
CMAES corresponding to with CMAES)
the ten-bar planar truss for
3000, 6000, 10000, and 20000 R+ R− p value
function evaluations for 20
populations 3000 DE 539.08 547.30 532.66 3.74 4.00 465 0 0.000
IDE 528.75 533.61 525.59 2.19 2.37 352 113 0.013
LSHADE 528.24 534.29 524.64 3.19 1.97 320 145 0.073
CMAES 526.78 531.70 524.64 2.74 1.67
6000 DE 526.83 531.83 524.70 2.24 3.07 359 106 0.008
IDE 526.21 530.83 524.46 2.72 2.20 371 94 0.003
LSHADE 527.98 533.60 524.64 3.22 3.37 405 60 0.000
CMAES 525.88 530.60 524.45 2.59 1.37
10000 DE 525.27 530.70 524.49 1.82 2.97 330 135 0.045
IDE 526.16 530.82 524.45 2.74 2.73 371 94 0.003
LSHADE 526.64 531.50 524.45 2.97 2.93 416 49 0.000
CMAES 525.88 530.57 524.45 2.59 1.37
20000 DE 525.10 530.66 524.45 1.84 2.43 330 135 0.045
IDE 526.15 530.82 524.45 2.73 2.90 371 94 0.003
LSHADE 526.64 531.49 524.45 2.97 3.30 416 49 0.000
CMAES 525.88 530.57 524.45 2.59 1.37

Table.2  Statistical results NFE Algorithm Mean Max Min STD Mean rank of the The Wilcoxon signed-
of DE, IDE, LSHADE, and Friedman test rank test (comparison
CMAES corresponding to with CMAES)
the ten-bar planar truss for
3000, 6000, 10000, and 20000 R+ R− p value
function evaluations for 50
populations 3000 DE 561.70 578.05 551.56 7.36 4.00 465 0 0.000
IDE 541.74 547.82 531.15 4.14 2.93 464 1 0.000
LSHADE 530.47 534.98 525.98 2.53 1.53 254 211 0.670
CMAES 530.22 535.76 527.21 2.28 1.53
6000 DE 545.55 553.96 536.25 4.04 4.00 465 0 0.000
IDE 530.97 535.31 527.34 2.25 2.93 465 0 0.000
LSHADE 525.56 531.66 524.49 2.24 1.93 406 59 0.000
CMAES 524.52 525.11 524.47 0.12 1.13
10000 DE 537.74 543.45 531.60 3.03 4.00 465 0 0.000
IDE 526.22 528.32 525.14 0.87 2.87 465 0 0.000
LSHADE 525.28 530.86 524.45 2.11 2.07 448 17 0.000
CMAES 524.45 524.45 524.45 0.00 1.07
20000 DE 528.65 533.34 525.95 1.67 3.87 465 0 0.000
IDE 524.61 524.83 524.47 0.08 2.87 465 0 0.000
LSHADE 525.27 530.63 524.45 2.09 2.27 465 0 0.000
CMAES 524.45 524.45 524.45 0.00 1.00

breadth of different initial random populations. For the num- 3 also show the weak performance of the DE algorithm in
ber of population = 100 and the number of function evalua- comparison with IDE, LSHADE, and CMAES in all cases.
tions = 3000, the LSHADE and then IDE gain the first and In the literature, there are abundant attempts to solve
second places. However, CMAES for the 6000, 10000, and this optimization problem utilizing different meta-heuris-
20000 function evaluations has the least mean, maximum, tic algorithms [15–18, 21, 26, 28, 30, 32, 50, 51]. For
minimum, and standard deviation, i.e., in all executions, it example, Gomes [50] used a particle swarm optimiza-
converges to the global minimum solution. Tables 1, 2, and tion (PSO) algorithm to minimize the truss weight and

13
Engineering with Computers

Table.3  Statistical results NFE Algorithm Mean Max Min STD Mean rank of The Wilcoxon
of DE, IDE, LSHADE, and signed-rank test
CMAES corresponding to the Friedman test
(comparison with
the ten-bar planar truss for CMAES)
3000, 6000, 10000, and 20000
function evaluations for 100 R+ R− p value
populations
3000 DE 581.68 595.68 550.61 8.67 3.27 122 343 0.022
IDE 561.68 573.40 543.58 6.86 2.03 4 461 0.000
LSHADE 544.66 553.45 539.75 3.52 1.03 0 465 0.000
CMAES 591.49 654.72 556.90 21.19 3.67
6000 DE 564.68 579.29 549.47 6.69 4.00 465 0 0.000
IDE 544.33 554.22 533.34 3.91 3.00 465 0 0.000
LSHADE 530.35 535.75 526.88 1.90 2.00 465 0 0.000
CMAES 526.90 528.79 525.79 0.73 1.00
10000 DE 552.22 559.53 542.38 3.85 4.00 465 0 0.000
IDE 536.78 542.66 532.38 2.27 3.00 465 0 0.000
LSHADE 524.85 525.29 524.62 0.18 2.00 465 0 0.000
CMAES 524.49 524.53 524.47 0.01 1.00
20000 DE 540.26 546.49 531.19 3.46 4.00 465 0 0.000
IDE 527.93 530.26 525.85 0.89 3.00 465 0 0.000
LSHADE 524.45 524.45 524.45 0.00 1.77 465 0 0.000
CMAES 524.45 524.45 524.45 0.00 1.23

presented the best minimum weight of 537.98 kg without to the minimum weight of 524.4508 kg after 10000 func-
reporting further information about the number of func- tion evaluations, which is even better than the result of
tion evaluations. Kaveh and Mahdavi [18] reported the AHEFA. Table 4 shows the optimum cross-section design
best minimum weight of 524.5499 kg using hybridized obtained by the CMAES algorithm which results in
BB-BC/quasi-Newton algorithm. Kaveh and Javadi [21] 𝜔1 = 7.0000 Hz , 𝜔2 = 16.1913 Hz , and 𝜔3 = 20.0000 Hz.
utilized PSO in combination with HS and ray optimization To achieve more realistic results, in addition to the natu-
algorithm (HRPSO) and obtained the optimal result as ral frequency constraints accounting for the stress, displace-
524.88 kg . Pham [30] applied LSHADE and the enhanced ment, and Euler buckling constraints under various loading
differential evolution (ANDE) algorithms and reported patterns is inevitable. The incorporation of these additional
the best minimum weight of 524.4748 kg and 524.4956 constraints would definitely lead to the reduction in the fea-
kg after 5500 and 6115 function evaluations, respec- sible domain. If the previous optimal solution still remains
tively. Baykasoğlu and Baykasoğlu [52] utilized weighted in the feasible domain under the extra constraints, the opti-
superposition attraction-repulsion (WSAR) approach and mal weight of the structures keeps unchanged. However, in
obtained the minimum weight of 524.4605 kg . Lieu et al. the majority of cases, the introduction of new constraints
[32] used DE, FA, and an adaptive hybrid evolutionary is accompanied by a weight penalty for the structure. This
firefly algorithm (AHEFA) for the optimization of this penalty substantially depends on the magnitude of loads
truss and found the minimum weight of 524.4561 kg, as well as the strength of the material and allowable nodal
524.6413 kg, and 524.4516 kg after 11460, 13060, and displacements. For instance, Canfield et al. [53] studied
5860 function evaluations, respectively. So far, the lat- the introduction of stress constraints on all members of the
ter one which is obtained using AHEFA yields the best above-mentioned ten-bar truss structure (with different fre-
solution of ten-bar planar truss reported in the literature. quency constraints) and reported the weight penalty of 4%
However, the CMAES algorithm in this study converged for this case.

Table.4  Optimum cross-section design of ten-bar truss structure (rounded to five decimal places) obtained from CMAES
Member 1 2 3 4 5 6 7 8 9 10

Cross-section (cm2) 35.13225 14.71431 35.13223 14.71431 0.64500 4.55894 23.70088 23.70088 12.41792 12.41792

13
Engineering with Computers

3
6.2 A 72‑bar space truss
3 1
2 In the second example, the optimization algorithms are
4
4 3 examined for a 72-bar space truss as illustrated in Fig. 2.
2
3 1 The truss members are categorized into 16 groups of design
2 2 variables as demonstrated in Fig. 2. Similar to previous
h 1 7
7 1 examples, the mass density and the modulus of elasticity
2
2 8
are taken as 2770 kg/m3 and 6.98 × 1010 N/m2 for all mem-
7 bers, respectively. The height of each story, h is chosen to
6
7 5
be 1.524 m and an additional mass of 2270 kg is added to
6 6 the four upper nodes of the truss structure. In addition, the
h 5
6 11 5 cross-section areas of all members need to be selected within
6 12 the following range:
11
10
11 9 6.45 × 10−5 m2 ≤ A ≤ 20 × 10−4 m2 . (22)
h 10 10
9 Likewise, there are only two constraints on the first and third
15 9
10 natural frequencies 𝜔1 and 𝜔3 presented as
10 16
15
15 13
𝜔1 = 4 Hz, 𝜔3 ≥ 6 Hz. (23)
h 14 It is worthwhile to mention that the violation tolerance
13
13 associated with the equality constraint is considered as
14
14 14 1 × 10−10 Hz.
14 2h The performance of the algorithms, in terms of the mini-
mum total weight, under the four stopping criteria, is shown
2h
in the Tables 5, 6, and 7 , for the 20, 50, and 100 population
size, respectively. The results show that, for the number of
Fig. 2  The 72-bar space truss population = 20 and the number of function evaluations =
3000 and 6000, the LSHADE algorithm exhibits the best
mean weight and the best minimum weight of 324.3923

Table.5  Statistical results NFE Algorithm Mean Max Min STD Mean rank of the The Wilcoxon signed-
of DE, IDE, LSHADE, and Friedman test rank test (comparison
CMAES corresponding to with CMAES)
the 72-bar space truss for
3000, 6000, 10000, and 20000 R+ R− p value
function evaluations for 20
populations 3000 DE 345.72 366.92 336.29 7.23 4.00 465 0 0.000
IDE 328.58 334.43 326.19 1.63 2.97 460 5 0.000
LSHADE 324.75 325.57 324.39 0.30 1.17 36 429 0.000
CMAES 325.54 327.85 324.45 0.80 1.87
6000 DE 328.17 334.26 325.48 1.73 4.00 465 0 0.000
IDE 324.72 325.57 324.47 0.25 2.67 389 76 0.001
LSHADE 324.31 324.62 324.23 0.10 1.13 33 432 0.000
CMAES 324.51 324.87 324.29 0.16 2.20
10000 DE 324.90 325.97 324.43 0.33 4.00 465 0 0.000
IDE 324.28 324.41 324.23 0.05 2.53 388 77 0.001
LSHADE 324.26 324.52 324.22 0.05 1.77 292 173 0.229
CMAES 324.24 324.35 324.22 0.02 1.70
20000 DE 324.25 324.33 324.22 0.03 3.30 465 0 0.000
IDE 324.25 324.37 324.22 0.04 2.97 465 0 0.000
LSHADE 324.25 324.43 324.22 0.04 2.73 465 0 0.000
CMAES 324.22 324.22 324.22 0.00 1.00

13
Engineering with Computers

Table.6  Statistical results NFE Algorithm Mean Max Min STD Mean rank of the The Wilcoxon signed-
of DE, IDE, LSHADE, and Friedman test rank test (comparison
CMAES corresponding to with CMAES)
the 72-bar space truss for
3000, 6000, 10000, and 20000 R+ R− p value
function evaluations for 50
populations 3000 DE 414.66 441.53 393.83 12.36 4.00 465 0 0.000
IDE 367.49 390.90 349.21 8.56 3.00 465 0 0.000
LSHADE 329.35 332.62 327.03 1.54 1.10 14 451 0.000
CMAES 336.41 351.08 327.41 5.61 1.90
6000 DE 371.37 389.81 357.45 7.52 4.00 465 0 0.000
IDE 337.83 346.09 329.50 3.47 3.00 465 0 0.000
LSHADE 324.43 324.68 324.30 0.10 1.23 65 400 0.000
CMAES 324.59 325.38 324.30 0.24 1.77
10000 DE 349.52 361.93 337.21 5.95 4.00 465 0 0.000
IDE 328.58 332.25 326.65 1.16 3.00 465 0 0.000
LSHADE 324.23 324.25 324.22 0.01 1.00 50 415 0.000
CMAES 324.26 324.49 324.22 0.05 2.00
20000 DE 332.14 335.49 329.20 1.80 4.00 465 0 0.000
IDE 324.88 325.43 324.60 0.19 3.00 465 0 0.000
LSHADE 324.22 324.22 324.22 0.00 1.77 459 6 0.000
CMAES 324.22 324.22 324.22 0.00 1.23

Table.7  Statistical results NFE Algorithm Mean Max Min STD Mean rank of the The Wilcoxon signed-
of DE, IDE, LSHADE, and Friedman test rank test (comparison
CMAES corresponding to with CMAES)
the 72-bar space truss for
3000, 6000, 10000, and 20000 R+ R− p value
function evaluations for 100
populations 3000 DE 452.65 483.60 422.84 15.94 3.10 12 453 0.000
IDE 418.79 436.32 394.00 11.74 2.03 1 464 0.000
LSHADE 352.73 364.22 336.76 6.02 1.00 0 465 0.000
CMAES 513.18 642.25 430.32 49.37 3.87
6000 DE 418.06 433.68 389.06 9.99 4.00 465 0 0.000
IDE 371.92 381.70 361.89 5.26 3.00 465 0 0.000
LSHADE 328.92 331.53 327.15 1.20 1.27 69 396 0.000
CMAES 331.21 337.89 325.89 2.43 1.73
10000 DE 390.54 411.58 371.78 10.19 4.00 465 0 0.000
IDE 347.64 355.00 341.50 3.90 3.00 465 0 0.000
LSHADE 324.60 325.20 324.31 0.19 1.73 380 85 0.002
CMAES 324.46 324.83 324.28 0.13 1.27
20000 DE 357.24 370.44 346.43 6.14 4.00 465 0 0.000
IDE 331.01 333.74 328.48 1.24 3.00 465 0 0.000
LSHADE 324.22 324.22 324.22 0.00 1.07 463 2 0.000
CMAES 324.22 324.22 324.22 0.00 1.93

kg, and 324.2278 kg, respectively, and the CMAES algo- 50 and the number of function evaluations = 3000, 6000,
rithm has the second-best mean and best minimum weight and 10000, the LSHADE algorithm exhibits the best perfor-
of 324.4521 kg, and 324.2888, respectively. For the number mance and for the number of function evaluations = 20000,
of function evaluations = 10000 and 20000, the CMAES the CMAES algorithm finds the best minimum weight of
algorithm finds the best minimum weight of 324.2222 kg, 324.2209 kg and shows the least standard deviation. How-
and 324.2209 kg and shows the best performance among ever, the LSHADE algorithm has also a standard deviation
all algorithms. In Table 6, for the number of population = close to zero. The IDE and DE algorithms are taken the third

13
Engineering with Computers

and fourth rank, respectively. In Table 7, for the number all bars in the truss are assumed to be 7971.810 kg/m3 and
of population = 100 and the number of function evalua- 2.1 × 1011 N/m2 , respectively. To the nodes with the height
tions = 3000 and 6000 the LSHADE algorithm exhibits the of 7.00, 5.85, and 3.00 m the non-structural masses 3000,
fastest convergence, and the CMAES algorithm is placed 500, and 100 kg are attached, respectively. The geometrical
in the second rank. However, both LSHADE and CMAES properties of the 120-bar dome truss structure are depicted
for the 20000 function evaluations have almost similar in Fig. 3. The allowable range of cross-sectional areas in this
mean, maximum, minimum, and standard deviation values. structure is given as
These results also show the weak performance of DE and
IDE algorithms in comparison with LSHADE and CMAES 0.0001 m2 ≤ A ≤ 0.01293 m2 . (24)
algorithms in most cases.
Furthermore, the first two natural frequencies 𝜔1 and 𝜔2
This truss structure with mentioned properties has been
must hold the following inequality conditions:
optimized by several researchers in the following studies [17,
21, 28–30, 32, 51, 54]. For instance, Kaveh and Javadi [21] 𝜔1 ≥ 9 Hz, 𝜔2 ≥ 11 Hz. (25)
obtained the minimum weight of this truss as 324.497 kg
with the aid of HRPSO. Tejani et al. [51] implemented mod-
ified symbiotic organisms search and found the best mini- The performance of the algorithms, in terms of the mini-
mum weight of 324.6897 kg after 4000 function evaluations. mum total weight, under the four stopping criteria, is shown
Ho-Huu et al. [29] used the IDE algorithm and found the best in the Tables 9, 10, and 11 for the 20, 50, and 100 popula-
minimum weight of 324.2441 kg after 11620 function evalu- tion size, respectively. The results show that, for the num-
ations. Pham [30] applied LSHADE and ANDE algorithms ber of population = 20 and the number of function evalu-
and obtained the best minimum weight of 324.2402 kg and ations = 6000, 10000, and 20000, the CMAES algorithm
324.2262 kg after 7500 and 7430 function evaluations, exhibits the best mean weight and the best minimum weight
respectively. Recently, Baykasoğlu and Baykasoğlu [52] by of 8707.2432 kg, and the LSHADE algorithm found the
employing the WSAR method obtained the minimum weight second-best minimum weight of 8707.2511 kg. Moreover,
of 324.2221 kg which is the best ever reported in the litera- for the number of population = 50 and the number of func-
ture. They selected the number of populations as 10 in the tion evaluations = 3000, the LSHADE algorithm exhibits
algorithm and found the optimum result after 2150 function the best performance in terms of the best mean and least
evaluations with a 0.0009 standard deviation. However, in standard deviation. For the number of function evaluations
the present study, the CMAES and LSHADE algorithms = 6000, 10000, and 20000, the CMAES finds the best mini-
converged to the best minimum weight of 324.2209 kg after mum weight of 8707.2456, 8707.2432, and 8707.2432 kg
20000 function evaluations, which are even better than and shows the best performance. However, both the CMAES
the result of LSHADE given by Pham [30] and WSAR by and LSHADE algorithms have the least standard deviation
Baykasoğlu and Baykasoğlu [52]. Further details about the in these cases. A similar manner can be observed for the
optimum cross-section design obtained by the latter scenario number of population = 100 and the number of function
are demonstrated in Table 8. The mentioned results give evaluations = 3000, in which the LSHADE algorithm exhib-
𝜔1 = 𝜔2 = 4.0000 Hz and 𝜔3 = 6.0000 Hz. its the fastest convergence, and for the number of function
evaluations = 6000, 10000, and 20000, the CMAES gains
6.3 A 120‑bar dome truss structure the first rank. In other words, the CMAES algorithm has the
least mean, maximum, minimum, and its standard deviation
In this example, the size optimization problem of a 120-bar is zero, i.e., in all executions it converges to the global mini-
dome truss structure with frequency constraints is investi- mum solution. These tables also show the weak performance
gated. The mass density and the modulus of elasticity of of DE and IDE algorithms in comparison with LSHADE and
CMAES in most cases.

Table.8  Design area of the 72-bar truss structure (rounded to five decimal places) obtained from CMAES
Group member 1 2 3 4 5 6 7 8

Cross-section (cm2) 3.46173 7.84905 0.64500 0.64500 7.97092 7.92595 0.64500 0.64500
Group member 9 10 11 12 13 14 15 16

Cross-section (cm2) 12.67536 7.97434 0.64500 0.64500 17.09355 8.00203 0.64500 0.64500

13
Engineering with Computers

Fig. 3  The 120-bar dome


structure
7 6 7
7 7
6 6
7 5 5 7
5 3 5
5 4 4 3 5
7 3 4 4 7

6 5 4 4 5 6
4 2 2 4
7 3 3 7
5 2 2 5
4 1 4
7 1 1 7
5 4 2 1 2 4 5
1
6 3 1 1 3 6
5 4 1 1 2 4 5
2
7 1 1 7
1
4 4
5 2 2 5
7 3 3 7
4 2 2 4
6 5 4 5 6
4
7 3 4 4 3 7
5 4 5
4 3
7 5 5
5 5 7
6 6
7 7
7 6 7

6.94 m
12.04 m
15.89 m

7.00 m
5.85 m
3.00 m

In the literature, several algorithms have been used to after 20000 function evaluations. Lieu et al. [32] employed
optimize this problem [17–20, 22, 29, 32, 35–37, 51, 54]. the DE, FA, and AHEFA algorithms for the optimization
For example, the minimum weight of this dome truss struc- of this truss and found the minimum weight of 8707.2715
ture has been determined by hybridized BB-BC/quasi-New- kg, 8708.2350 kg, and 8707.2559 kg after 6000, 13840, and
ton algorithm as 8789.50 kg [18] and by PSRO as 8892.33 kg 3560 function evaluations, respectively. Among all attempts,
[20]. Kaveh and Zolghadr [19] used a modified PSO for the WSAR optimal solution with 8707.2537 kg [52] and 1580
truss optimization and presented the best minimum weight function evaluations is the best result obtained in the litera-
of 8890.48 kg after 6000 function evaluations. Kaveh and ture so far. The CMAES and LSHADE algorithms in this
Ilchi Ghazaan also solved this problem by employing IRO study converged to the best minimum weight of 8707.2432
[36], ALC-PSO [37], HALC-PSO [37], and vibrating parti- kg, respectively, after 6000 and 10000 function evaluations,
cle system algorithm [22] and obtained the optimal weight which are better than the result of other outputs reported in
as 8895.42 kg , 8890.70 kg , 8889.96 kg , and 8888.74 kg , the literature. The standard deviation of the CMAES method
respectively. Ho-Huu et al. [29] used DE and IDE algo- is given as 0.0004 which is less than 0.342 proposed by [52].
rithms and found the best minimum weight of 8707.2808 For more details, Table 12 shows the optimum cross-section
kg and 8707.2898 after 8620 and 4060 function evaluations. design obtained by the proposed algorithms. The data results
In this study, the IDE algorithm converged to 8707.243 kg in 𝜔1 = 9.0000 Hz and 𝜔2 = 11.0000 Hz.

13
Engineering with Computers

Table.9  Statistical results NFE Algorithm Mean Max Min STD Mean rank of the The Wilcoxon
of DE, IDE, LSHADE, and Friedman test signed-rank test
CMAES corresponding to (comparison with
the 120-bar dome truss for CMAES)
3000, 6000, 10000, and 20000
function evaluations for 20 R+ R− p value
populations
3000 DE 8752.82 8802.22 8717.33 22.30 3.97 465 0 0.000
IDE 8710.82 8717.95 8708.40 2.06 2.60 444 21 0.000
LSHADE 8712.67 8741.60 8707.25 8.98 2.03 346 119 0.019
CMAES 8707.72 8712.95 8707.29 1.01 1.40
6000 DE 8708.75 8712.06 8707.53 1.14 3.50 465 0 0.000
IDE 8707.56 8708.20 8707.25 0.30 2.67 465 0 0.000
LSHADE 8712.48 8741.60 8707.25 9.07 2.77 429 36 0.000
CMAES 8707.24 8707.25 8707.24 0.00 1.07
10000 DE 8707.32 8707.79 8707.25 0.11 2.73 465 0 0.000
IDE 8707.52 8708.15 8707.25 0.28 3.30 465 0 0.000
LSHADE 8712.13 8741.60 8707.24 9.13 2.90 426 39 0.000
CMAES 8707.24 8707.24 8707.24 0.00 1.07
20000 DE 8707.26 8707.38 8707.24 0.03 2.33 465 0 0.000
IDE 8707.52 8708.15 8707.25 0.27 3.50 465 0 0.000
LSHADE 8712.07 8741.60 8707.24 9.15 3.10 465 0 0.000
CMAES 8707.24 8707.24 8707.24 0.00 1.07

Table.10  Statistical results NFE Algorithm Mean Max Min STD Mean rank of The Wilcoxon
of DE, IDE, LSHADE, and the Friedman signed-rank test
CMAES corresponding to test (comparison with
the 120-bar dome truss for CMAES)
3000, 6000, 10000, and 20000
function evaluations for 50 R+ R− p value
populations
3000 DE 9864.17 11160.72 9171.96 414.84 4.00 465 0 0.000
IDE 8960.27 9172.83 8804.23 79.64 3.00 465 0 0.000
LSHADE 8743.64 8772.42 8719.93 14.28 1.33 115 350 0.15
CMAES 8756.99 8793.77 8716.52 19.23 1.67
6000 DE 8936.92 9146.87 8750.72 86.49 4.00 465 0 0.000
IDE 8726.09 8740.40 8713.57 6.44 3.00 465 0 0.000
LSHADE 8707.53 8708.03 8707.29 0.18 2.00 465 0 0.000
CMAES 8707.26 8707.29 8707.25 0.01 1.00
10000 DE 8750.41 8777.87 8725.73 13.84 4.00 465 0 0.000
IDE 8708.90 8710.75 8707.67 0.70 3.00 465 0 0.000
LSHADE 8707.24 8707.24 8707.24 0.00 1.92 465 0 0.000
CMAES 8707.24 8707.24 8707.24 0.00 1.08
20000 DE 8709.63 8711.36 8707.90 0.99 4.00 465 0 0.000
IDE 8707.24 8707.25 8707.24 0.00 2.90 465 0 0.000
LSHADE 8707.24 8707.24 8707.24 0.00 1.55 465 0 0.000
CMAES 8707.24 8707.24 8707.24 0.00 1.55 465 0 0.000

6.4 A 200‑bar planar truss structure is obtained based on these algorithms. In this example, all
members of the structure are assumed to have Young’s mod-
To examine the effects of large-scale problems on the per- ulus of 2.1 × 1011 N/m2 and mass density of 7860 kg/m3. The
formance of the DE, IDE, LSHADE, and CMAES algo- additional masses of 100 kg are added to the nodes 1–5 as
rithms, the optimal design of a 200-bar planar truss structure illustrated by blue solid circles in Fig. 4. Furthermore, all

13
Engineering with Computers

Table.11  Statistical results NFE Algorithm Mean Max Min STD Mean rank of The Wilcoxon
of DE, IDE, LSHADE, and the Friedman signed-rank test
CMAES corresponding to test (comparison with
the 120-bar dome truss for CMAES)
3000, 6000, 10000, and 20000
function evaluations for 100 R+ R− p value
populations
3000 DE 10947.22 12953.72 9376.79 796.49 3.73 432 33 0.000
IDE 10060.98 10946.97 9492.05 368.55 2.60 252 214 0.700
LSHADE 9125.52 9360.02 8880.54 110.47 1.00 0 465 0.000
CMAES 9999.26 10853.63 9336.30 408.42 2.67
6000 DE 9791.62 10487.24 9200.15 295.08 4.00 465 0 0.000
IDE 8967.46 9169.94 8867.47 74.37 3.00 465 0 0.000
LSHADE 8754.80 8809.39 8717.72 19.18 2.00 465 0 0.000
CMAES 8710.47 8715.49 8707.81 1.72 1.00
10000 DE 9152.13 9365.20 8933.53 121.55 4.00 465 0 0.000
IDE 8756.88 8805.22 8734.49 15.41 3.00 465 0 0.000
LSHADE 8708.95 8710.87 8707.88 0.71 2.00 465 0 0.000
CMAES 8707.25 8707.25 8707.24 0.00 1.00
20000 DE 8772.58 8855.80 8742.26 22.64 4.00 465 0 0.000
IDE 8711.16 8713.71 8709.43 1.11 3.00 465 0 0.000
LSHADE 8707.24 8707.24 8707.24 0.00 1.50 465 0 0.000
CMAES 8707.24 8707.24 8707.24 0.00 1.50

Table.12  Design area of the 120-bar dome truss structure (rounded to five decimal places) obtained from CMAES
Group member 1 2 3 4 5 6 7

Cross-section (cm2) 19.49937 40.38900 10.60733 21.11262 9.84202 11.77150 14.83839

truss members are grouped into 29 independent variables as maximum, mean, and standard deviation. However, increas-
depicted in Fig. 4. The first three fundamental frequencies of ing the number of function evaluations to 20000 improves
the structure need to satisfy the following relations: the performance of the CMAES algorithm such that it finds
the best result with a significantly lower standard deviation.
𝜔1 ≥ 5 Hz, 𝜔2 ≥ 10 Hz, 𝜔3 ≥ 15 Hz. (26) Unlike the previous cases, for the population size = 100,
In addition, the cross-section areas of all members must be LSHADE exhibits the best performance in terms of mean,
selected from the interval of 0.1 cm2 and 30 cm2. maximum, minimum, and standard deviation for all function
The performances of the DE, IDE, LSHADE and evaluations = 3000, 6000, 10000, and 20000. The powerful
CMAES algorithms are given in Tables 13, 14, and 15 for mechanism of the LSHADE in the higher population sizes
the population sizes of 20, 50, and 100, respectively. For the is originated from its ability in the updating population size.
population size = 20 and the function evaluations = 3000 Nevertheless, it can be expected that the CMAES will out-
and 6000, the LSHADE algorithm obtains the least mini- perform the LSHADE algorithm for the larger number of
mum and mean values. On the other hand, for the function function evaluations.
evaluation of 10000, the LSHADE algorithm still gives the This example has been already optimized and reported
best minimum solution, while the IDE algorithm gives the in the former studies. Kaveh and Zolghadr [17] utilized
least value for maximum, mean, and standard deviation. By the charged system search algorithm and its modified ver-
increasing the number of function evaluations to 20000, sion in combination with BB–BC algorithm to optimize
the CMAES excels at DE, IDE, and LSHADE algorithm in this structure and obtained the optimal weight of 2259.86
terms of the minimum and the standard deviation. For popu- kg and 2298.61 kg, respectively. Tejani et al. [51] applied
lation size = 50 and function evaluations = 3000, 6000, and SOS algorithm together with its improved version using
10000, the performance of the LSHADE algorithm remains adaptive benefit factors and determined the minimum
superior to other given algorithms with the least minimum, weight of this structure as 2180.3210 kg and 2164.8840 kg,

13
Engineering with Computers

6.1 m 6.1 m 6.1 m 6.1 m 7 Results and discussion


1 1 1 1
As shown in Tables 1, 2, 3, 5, 6, 7, 9, 10, 11, 13, 14, and
6 6 6 6 6 6 6 6
2 2 2 2 2 3.66 m 15, it is evident that the data obtained by the Friedman and
4 3 3 3 3 3 3 4 the Wilcoxon signed-rank test provide a reasonable frame-
5 6 6 5
6 6 5
6 6 5
6 6 5 3.66 m work for comparison between the performances of all four
7 7 7 7
algorithms.
The Friedman mean rank of the algorithms denotes the
11 11 11 11 11 11 11 11
8 8 8 8 8 3.66 m mean of the ranks pertinent to the least-weight solutions
4 9 9 9 9 9 9 4 over each of the algorithms and consequently the less rank
10 11 11 10 11 11 10 11 11 10 11 11 10 3.66 m indicates more reliable results. Accordingly, the Friedman
12 12 12 12
test results conclude that the DE algorithm is inferior to the
others due to the largest mean rank in the majority of cases.
16 16 16 16
13
16
13
16
13
16
13
16
13 3.66 m This weak performance of DE is amplified for larger deci-
4 14 14 14 14 14 14 4 sion variables. Moreover, in most cases especially for 6000,
16 16 15 16 16 16 16 16 16 10000, and 20000 function evaluations, the IDE gains the
15 15 15 15 3.66 m
17 17 17 17
second largest mean rank after the DE algorithm. Unlike
the two mentioned algorithms, the ranks of the CMAES and
18
21 21
18
21 21
18
21 21
18
21 21
18 3.66 m LSHADE algorithm highly depend on the number of deci-
4 19 19 19 19 19 19 4 sion variables, function evaluations, and the population size.
21 21 21
In the proposed examples, increasing the population size
20 21 20 21 20 21 20 21 21 20 3.66 m
from 20 to 100 has an adverse effect on both the optimal
22 22 22 22
result and the standard deviation of all algorithms. Never-
23
26 26
23
26 26
23
26 26
23
26 26
23 3.66 m theless, the performance of the LSHADE algorithm remains
4 24 24 24 24 24 24 4 stable compared to others. This is due to the fact that the
26 26 26 LSHADE algorithm updates the population size during the
25 26 25 26 25 26 25 26 26 25 3.66 m
optimization procedure and hence, increasing the population
27 27 27 27
size yields a lower mean rank for the LSHADE algorithm.
29 29
On the other hand, the higher number of function evalua-
tions enhances the associated performance of all algorithms.
28 28 28 28 9.144 m However, this process causes the CMAES algorithm distin-
guishes itself from others due to the remarkable improve-
ment. This distinctive feature of the CMAES gives rise to the
least mean rank of the algorithm as the number of function
evaluations rises. It should be noted that, the rejection of the
Fig. 4  The 200-bar planar truss structure null hypothesis for all Friedman tests (i.e., Asymp. Sig. or p
value< 0.05), in all cases, indicates that the performance of
the algorithms executions differs significantly.
respectively. Lieu et al. [32] optimized the weight of this The results of the Wilcoxon signed-rank test fairly agree
problem via DE (2160.7747 kg), FA (2441.3773 kg), and with those of the Friedman test, but they further clarify the
AHEFA (2160.7445 kg) algorithms. In the literature, the performance of the CMAES compared to other algorithms.
best result has been reported by the WSAR method [52] with In all the examples with fixed population size, it is observed
10 populations and 6110 function evaluations which yields that R+ grows significantly as the number of function evalu-
the optimal weight of 2159.590 kg and standard deviation ations increases. This behavior confirms the marked posi-
of 0.04438. In the present study, however, the CMAES algo- tive impacts of the number of function evaluations on the
rithm finds the best optimal weight of 2159.5474 kg with a CMAES performance.
standard deviation of 0.26943 for 20 populations and 20000 Based on the statistical analysis and results obtained in
function evaluations. For the latter case, the cross-section this work, the CMAES algorithm shows the most reliable
areas associated with each group member are reported in performance with the least standard deviation values in
Table 16. It is worthwhile to note that these data lead to most cases. However, raising the population size and the
𝜔1 = 5.0000 Hz, 𝜔2 = 12.2098 Hz, and 𝜔3 = 15.0203 Hz for number of independent variables have adverse effects on
the first three fundamental frequencies. the performance of the CMAES algorithm. To compensate

13
Engineering with Computers

Table.13  Statistical results NFE Algorithm Mean Max Min STD Mean rank of The Wilcoxon
of DE, IDE, LSHADE, and the Friedman signed-rank test
CMAES corresponding to test (comparison with
the 200-bar planar truss for CMAES)
3000, 6000, 10000, and 20000
function evaluations for 20 R+ R− p value
populations
3000 DE 4057.51 5266.33 3234.49 438.88.29 4.00 465 0 0.000
IDE 2759.96 3174.66 2561.18 126.51 2.23 67 398 0.000
LSHADE 2327.94 2771.37 2184.38 138.75 1.03 0 465 0.000
CMAES 3036.95 3657.53 2545.79 297.88 2.73
6000 DE 2800.46 3321.67 2414.02 183.25 4.00 465 0 0.000
IDE 2200.79 2298.24 2177.54 22.46 1.93 60 405 0.000
LSHADE 2193.98 2343.55 2164.45 43.30 1.47 61 404 0.000
CMAES 2255.36 2501.89 2166.90 71.11 2.60
10000 DE 2321.46 2541.06 2231.33 66.17 4.00 465 0 0.000
IDE 2163.51 2168.29 2161.34 1.87 1.43 22 443 0.000
LSHADE 2169.45 2234.23 2160.64 13.97 2.07 145 320 0.073
CMAES 2173.91 2226.91 2160.50 12.93 2.50
20000 DE 2167.80 2173.82 2164.23 2.35 3.90 465 0 0.000
IDE 2160.04 2160.97 2159.66 0.39 1.57 311 154 0.109
LSHADE 2162.99 2180.47 2160.01 3.75 3.07 462 3 0.000
CMAES 2159.91 2160.77 2159.55 0.27 1.47

Table.14  Statistical results NFE Algorithm Mean Max Min STD Mean rank of the The Wilcoxon
of DE, IDE, LSHADE, and Friedman test signed-rank test
CMAES corresponding to (comparison with
the 200-bar planar truss for CMAES)
3000, 6000, 10000, and 20000
function evaluations for 50 R+ R− p value
populations
3000 DE 7008.27 8219.26 5860.67 547.08 3.80 408 57 0.000
IDE 5252.20 6070.84 4635.62 333.06 2.10 11 454 0.000
LSHADE 2796.56 3085.78 2484.11 144.23 1.00 0 465 0.000
CMAES 6253.90 8684.88 4954.69 749.11 3.10
6000 DE 5359.99 6173.51 4452.36 429.46 4.00 465 0 0.000
IDE 3613.27 4060.21 3162.14 212.30 3.00 465 0 0.000
LSHADE 2231.31 2348.24 2180.40 50.45 1.00 0 465 0.000
CMAES 2635.35 3108.82 2360.45 156.41 2.00
10000 DE 4341.10 5076.67 3848.74 244.26 4.00 465 0 0.000
IDE 2800.74 3072.42 2584.91 102.29 3.00 465 0 0.000
LSHADE 2176.72 2263.97 2162.50 23.93 1.10 35 430 0.000
CMAES 2227.06 2324.12 2166.97 38.44 1.90
20000 DE 3321.26 3622.97 2962.23 173.86 4.00 465 0 0.000
IDE 2246.75 2298.26 2209.60 23.26 3.00 465 0 0.000
LSHADE 2162.63 2174.36 2160.35 3.36 1.93 440 25 0.000
CMAES 2160.34 2162.67 2159.67 0.55 1.07

for this drawback, a higher number of function evaluations 8 Conclusions


is required. This algorithm also discovered the best optimal
results for all four truss structures which were not presented In this study, the CMAES algorithm, one of the most robust
in other studies. The adaptive penalty function proposed and reliable self-adaptation mechanisms, was developed to
in this paper, also makes the CMAES algorithm free from solve the size optimization of truss structures under mul-
parameter tuning in different problems. tiple frequency constraints. For this purpose, an adaptive

13
Engineering with Computers

Table.15  Statistical results NFE Algorithm Mean Max Min STD Mean rank of The Wilcoxon
of DE, IDE, LSHADE, and the Friedman signed-rank test
CMAES corresponding to test (comparison with
the 200-bar planar truss for CMAES)
3000, 6000, 10000, and 20000
function evaluations for 100 R+ R− p value
populations
3000 DE 9137.41 10942.18 6372.60 943.35 2.90 0 465 0.000
IDE 7557.75 8731.50 6559.78 533.42 2.10 0 465 0.000
LSHADE 3848.56 4177.81 3550.58 167.80 1.00 0 465 0.000
CMAES 222507.91 36642.59 14484.71 4867.11 4.00
6000 DE 7521.35 9005.59 5579.96 660.83 4.00 0 465 0.000
IDE 5550.44 6441.31 5017.87 309.39 2.93 5 460 0.000
LSHADE 2733.00 3025.48 2508.02 110.36 1.00 0 465 0.000
CMAES 4918.12 5563.04 4325.03 337.77 2.07
10000 DE 6021.02 7092.24 5347.13 434.10 4.00 465 0 0.000
IDE 4265.00 4564.06 3844.55 180.65 3.00 465 0 0.000
LSHADE 2228.92 2326.33 2182.02 35.27 1.00 0 465 0.000
CMAES 2675.67 3198.34 2381.88 150.75 2.00
20000 DE 4714.17 5150.23 4155.77 248.80 4.00 465 0 0.000
IDE 3049.38 3221.75 2815.49 112.33 3.00 465 0 0.000
LSHADE 2163.54 2175.83 2160.62 3.23 1.17 49 416 0.000
CMAES 2177.38 2240.73 2163.10 18.05 1.83

Table.16  Design area of the 200-bar planar truss structure (rounded to five decimal places) obtained from CMAES
Group member 1 2 3 4 5 6 7 8 9 10

Cross-section (cm2) 0.30349 0.45043 0.10000 0.10000 0.50880 0.82103 0.10000 1.42386 0.10000 1.58807
Size of grouped members 11 12 13 14 15 16 17 18 19 20

Cross-section (cm2) 1.15781 0.12438 2.97083 0.10000 3.25442 1.58406 0.25604 5.09849 0.10000 5.44473
Group member 21 22 23 24 25 26 27 28 29

Cross-section (cm2) 2.09734 0.69048 7.66664 0.10000 7.95770 2.79246 10.52297 21.32787 10.66537

penalty function was added to the algorithm which elimi- separately. The results reveal that the LSHADE algorithm
nated its parameter selection process in different problems. has more reliable performance compared to others for a
Subsequently, the performance of the proposed algorithm smaller number of function evaluations. However, in the
was compared with three powerful DE-based optimiza- cases with the larger number of function evaluations, the
tion algorithms, the original DE, an improved version of CMAES algorithm provides the best results among other
DE (IDE), and the winner of the CEC-2014 competition given algorithms as well as the least standard deviation
(LSHADE), through size optimization of four benchmark values and mean ranks. Moreover, the CMAES algorithm
truss structures with multiple natural frequency constraints: found the least weight design in all trusses which had not
a 10-bar planar truss, a 72-bar space truss, a 120-bar dome been reported in other previous studies. In other words, the
truss, and a 200-bar planar truss. To compare the perfor- present work suggests that the CMAES algorithm com-
mance of each algorithm consistently, four stopping crite- bined with the proposed penalty function can be a robust
ria were introduced and all algorithms were executed 30 and reliable technique for a wide range of populations and
times associated with each criterion. Then, the statistical decision variables. By increasing the population size and/
analysis of the results of each algorithm was performed or the decision variables, the algorithm requires a larger
and the Friedman as well as the Wilcoxon signed-rank test number of function evaluations to obtain the global best
for the 30 executions of the algorithms were carried out solution.

13
Engineering with Computers

Declarations optimizer for enhancing the particle swarm optimization algo-


rithm. Acta Mech 225(6):1595–1605
22. Kaveh A, Ghazaan MI (2017) Vibrating particles system algo-
Conflict of interest The authors declare that they have no conflict of
rithm for truss optimization with multiple natural frequency con-
interest.
straints. Acta Mech 1:307–322
23. Kaveh A, Zolghadr A (2017) Cyclical parthenogenesis algorithm
for layout optimization of truss structures with frequency con-
References straints. Eng Optim 49(8):1317–1334
24. Kaveh A, Mahjoubi S (2019) Hypotrochoid spiral optimization
approach for sizing and layout optimization of truss structures
1. Kiusalaas J, Shaw RCJ (1978) An algorithm for optimal struc-
with multiple frequency constraints. Eng Comput 35:1443–1462
tural design with frequency constraints. Int J Numer Methods Eng
25. Miguel LFF, Miguel LFF (2012) Shape and size optimization of
13(2):283–295
truss structures considering dynamic constraints through modern
2. Levy R, Chai K (1979) Implementation of natural frequency anal-
metaheuristic algorithms. Expert Syst Appl 39(10):9458–9467
ysis and optimality criterion design. Comput Struct 10(1):277–282
26. Zuo W, Bai J, Li B (2014) A hybrid OC-GA approach for fast and
3. Khot N (1985) Optimization of structures with multiple frequency
global truss optimization with frequency constraints. Appl Soft
constraints. Comput Struct 20(5):869–876
Comput 14:528–535
4. Sadek EA (1986) Dynamic optimization of framed structures with
27. Tejani GG, Savsani VJ, Bureerat S, Patel VK, Savsani P (2019)
variable layout. Int J Numer Methods Eng 23(7):1273–1294
Topology optimization of truss subjected to static and dynamic
5. Grandhi RV, Venkayya VB (1988) Structural optimization with
constraints by integrating simulated annealing into passing vehicle
frequency constraints. AIAA J 26(7):858–866
search algorithms. Eng Comput 535:499–517
6. Sedaghati R, Suleman A, Tabarrok B (2002) Structural optimi-
28. Vu TV (2015) Weight minimization of trusses with natural
zation with frequency constraints using the finite element force
freqency constraints. In: Conference: WCSMO-11
method. AIAA J 40(2):382–388
29. Ho-Huu V, Vo-Duy T, Luu-Van T, Le-Anh L, Nguyen-Thoi T
7. Sarcheshmehpour M, Estekanchi HE, Moosavian H (2020) Opti-
(2016) Optimal design of truss structures with frequency con-
mum seismic design of steel framed-tube and tube-in-tube tall
straints using improved differential evolution algorithm based on
buildings. Struct Des Tall Spec Build 29(14):e1782
an adaptive mutation scheme. Autom Constr 68:81–94
8. Goldberg DE, Holland JH (1988) Genetic algorithms and machine
30. Pham HA (2016) Truss optimization with frequency constraints
learning. Mach Learn 3:95–99
using enhanced differential evolution based on adaptive direc-
9. Storn R, Price K (1997) Differential evolution—a simple and effi-
tional mutation and nearest neighbor comparison. Adv Eng Softw
cient heuristic for global optimization over continuous spaces. J
102:142–154
Glob Optim 11:341–359
31. Bureerat S, Pholdee N (2015) Optimal truss sizing using an
10. Hansen N (2006) The CMA evolution strategy: a comparing
adaptive differential evolution algorithm. J Comput Civ Eng
review. Springer, Berlin, pp 75–102
30:04015019
11. Moosavian N, Roodsari BK (2014) Soccer league competition
32. Lieu QX, Do DT, Lee J (2018) An adaptive hybrid evolutionary
algorithm: a novel meta-heuristic algorithm for optimal design of
firefly algorithm for shape and size optimization of truss structures
water distribution networks. Swarm Evol Comput 17:14–24
with frequency constraints. Comput Struct 195:99–112
12. Moosavian N (2015) Soccer league competition algorithm for
33. Khatibinia M, Naseralavi SS (2014) Truss optimization on shape
solving knapsack problems. Swarm Evol Comput 20:14–22
and sizing with frequency constraints based on orthogonal multi-
13. Moosavian N, Roodsari B (2014) Soccer league competition algo-
gravitational search algorithm. J Sound Vib 333(24):6349–6369
rithm, a new method for solving systems of nonlinear equations.
34. Kumar S, Tejani GG, Mirjalili S (2019) Modified symbiotic
Int J Intell Sci 4:7–16
organisms search for structural optimization. Eng Comput
14. Kennedy J, Eberhart R (1995) Particle swarm optimization, in
35:1269–1296
Proceedings of ICNN’95—international conference on neural
35. Kaveh A, Zolghadr A (2014) Comparison of nine meta-heuristic
networks 4:1942–1948
algorithms for optimal design of truss structures with frequency
15. Wei L, Mei Z, Guangming W, Guang M (2005) Truss optimization
constraints. Adv Eng Softw 76:9–30
on shape and sizing with frequency constraints based on genetic
36. Kaveh A, Ilchi Ghazaan M (2015) Layout and size optimization
algorithm. Comput Mech 35:361–368
of trusses with natural frequency constraints using improved
16. Wei L, Tang T, Xie X, Shen W (2011) Truss optimization on
ray optimization algorithm. Iran J Sci Technol Trans Civ Eng
shape and sizing with frequency constraints based on parallel
39(C2+):395–408
genetic algorithm. Struct Multidiscip Optim 43:665–682
37. Kaveh A, Ilchi Ghazaan M (2015) Hybridized optimization algo-
17. Kaveh A, Zolghadr A (2012) Truss optimization with natural fre-
rithms for design of trusses with multiple natural frequency con-
quency constraints using a hybridized CSS-BBBC algorithm with
straints. Adv Eng Softw 79:137–147
trap recognition capability. Comput Struct 102–103:14–27
38. Wang D, Zhang W, Jiang J (2002) Truss shape optimization with
18. Kaveh A, Mahdavi VR (2013) Optimal design of structures with
multiple displacement constraints. Comput Methods Appl Mech
multiple natural frequency constraints using a hybridized BB-BC/
Eng 191(33):3597–3612
quasi-newton algorithm. Period Polytech Civ Eng 57(1):27–38
39. Su GS, Zhang Y, Wu ZX, Yan LB (2012) Optimization design
19. Kaveh A, Zolghadr A (2014) Democratic PSO for truss layout
of trusses based on covariance matrix adaptation evolution strat-
and size optimization with frequency constraints. Comput Struct
egy algorithm, advances in design technology. Appl Mech Mater
130:10–21
215:133–137 (Trans Tech Publications Ltd, 11)
20. Kaveh A, Zolghadr A (2014) A new PSRO algorithm for fre-
40. Ghosh S, Das S, Roy S, Islam SM, Suganthan P (2012) A dif-
quency constraint truss shape and size optimization. Struct Eng
ferential covariance matrix adaptation evolutionary algorithm for
Mech 52(3):445–468
real parameter optimization. Inf Sci 182(1):199–219 (Nature-
21. Kaveh A, Javadi SM (2014) Shape and size optimization of trusses
Inspired Collective Intelligence in Theory and Practice)
with multiple frequency constraints using harmony search and ray

13
Engineering with Computers

41. de Melo VV, Iacca G (2014) A modified covariance matrix adap- 49. Tanabe R, Fukunaga AS (2014) Improving the search performance
tation evolution strategy with adaptive penalty function and restart of SHADE using linear population size reduction. In: 2014 IEEE
for constrained optimization. Expert Syst Appl 41(16):7077–7094 congress on evolutionary computation (CEC), pp. 1658–1665
42. Yang W, Yue Z, Li L, Yang F, Wang P (2017) Optimization design 50. Gomes HM (2011) Truss optimization with dynamic con-
of unitized panels with stiffeners in different formats using the straints using a particle swarm algorithm. Expert Syst Appl
evolutionary strategy with covariance matrix adaptation. Proc Inst 38(1):957–968
Mech Eng Part G J Aerosp Eng 231(9):1563–1573 51. Tejani GG, Savsani VJ, Patel VK (2016) Adaptive symbiotic
43. Moosavian N, Moosavian H (2017) Testing soccer league com- organisms search (SOS) algorithm for structural design optimi-
petition algorithm in comparison with ten popular meta-heuristic zation. J Comput Des Eng 3(3):226–249
algorithms for sizing optimization of truss structures. Int J Eng 52. Baykasoğlu A, Baykasoğlu C (2021) Weighted superposition
30:926–936 attraction-repulsion (WSAR) algorithm for truss optimization
44. Grandhi R (1993) Structural optimization with frequency con- with multiple frequency constraints. Structures 30:253–264
straints—a review. AIAA J 31(12):2296–2303 53. Canfield RA, Venkayya VB, Grandhi RV (1989) Structural opti-
45. Bouzarkouna Z, Ding DY, Auger A (2012) Well placement opti- mization with stiffness and frequency constraints. Mech Struct
mization with the covariance matrix adaptation evolution strategy Mach 17(1):95–110
and meta-models. Comput Geosci 16:75–92 54. Ho-Huu V, Nguyen-Thoi T, Truong-Khac T, Le-Anh L, Vo-Duy T
46. Friedman M (1937) The use of ranks to avoid the assumption of (2018) An improved differential evolution based on roulette wheel
normality implicit in the analysis of variance. J Am Stat Assoc selection for shape and size optimization of truss structures with
32(200):675–701 frequency constraints. Neural Comput Appl 29:167–185
47. Friedman M (1940) A comparison of alternative tests of signifi-
cance for the problem of m rankings. Ann Math Stat 11(1):86–92 Publisher's Note Springer Nature remains neutral with regard to
48. Derrac J, García S, Molina D, Herrera F (2011) A practical tuto- jurisdictional claims in published maps and institutional affiliations.
rial on the use of nonparametric statistical tests as a methodology
for comparing evolutionary and swarm intelligence algorithms.
Swarm Evol Comput 1(1):3–18

13

You might also like