See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/336214481

Critical Evaluation of Metaheuristic Algorithms for Weight Minimization of

Truss Structures

Article  in  Frontiers in Built Environment · October 2019

DOI: 10.3389/fbuil.2019.00113

CITATIONS READS

0 153

2 authors:

Aristotelis E. Charalampakis George C. Tsiatas

University of West Attica University of Patras
48 PUBLICATIONS   432 CITATIONS    79 PUBLICATIONS   965 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Mass Damping Systems View project

Elastic membranes View project

All content following this page was uploaded by George C. Tsiatas on 04 October 2019.

The user has requested enhancement of the downloaded file.

 ORIGINAL RESEARCH
published: 02 October 2019
doi: 10.3389/fbuil.2019.00113

Critical Evaluation of Metaheuristic

Algorithms for Weight Minimization
of Truss Structures
Aristotelis E. Charalampakis 1 and George C. Tsiatas 2*
1
School of Civil Engineering, National Technical University of Athens, Athens, Greece, 2 Department of Mathematics,
University of Patras, Rio, Greece

This study critically compares variants of Genetic Algorithms (GAs), Particle Swarm
Optimization (PSO), Artificial Bee Colony (ABC), Differential Evolution (DE), and Simulated
Annealing (SA) used in truss sizing optimization problems including displacement and
stress constraints. The comparison is based on several benchmark problems of varying
complexity measured by the number of design variables and the degree of static
indeterminacy. Most of these problems have been studied by numerous researchers
using a large variety of methods; this allows for absolute rather than relative comparison.
Rigorous statistical analysis based on large sample size, as well as monitoring of the
success rate throughout the optimization process, reveal and explain the convergence
behavior observed for each method. The results indicate that, for the problem at
Edited by: hand, Differential Evolution is the best algorithm in terms of robustness, performance,
Georgios Eleftherios Stavroulakis, and scalability.
Technical University of Crete, Greece
Keywords: truss weight minimization, genetic algorithm, particle swarm optimization, differential evolution,
Reviewed by:
simulated annealing, artificial bee colony
Makoto Ohsaki,
Kyoto University, Japan
Francesco Tornabene,
University of Salento, Italy INTRODUCTION
*Correspondence:
George C. Tsiatas
Structural optimization has always been a topic of interest for both the research community and the
gtsiatas@upatras.gr practicing engineers. It is commonly acknowledged that many real-life problems, featuring a large
number of design variables, as well as non-linear objective function(s) and constraints, are quite
Specialty section: demanding due to characteristics such as high dimensionality, multimodality, and non-convex
This article was submitted to feasible regions. In order to address this, researchers were led to Mathematical Programming (MP)
Computational Methods in Structural and Optimality Criteria (OC) methods (Feury and Geradin, 1978). In MP, the optimization problem
Engineering, is decomposed into simpler subproblems producing designs of good quality which, however, are not
a section of the journal
necessarily optimal. In OC methods, simple recursion formulas are derived based on assumptions
Frontiers in Built Environment
on the optimum design. These formulas may be simple and practical, but often do not converge for
Received: 02 August 2019 highly redundant structures (Feury and Geradin, 1978). Moreover, they do not address the problem
Accepted: 17 September 2019
properly, in mathematical terms. Interestingly, analytical global sizing optimization of trusses is
Published: 02 October 2019
actually feasible, based on the Cylindrical Algebraic Decomposition algorithm, as shown recently
Citation: in Charalampakis and Chatzigiannelis (2018).
Charalampakis AE and Tsiatas GC
In contemporary optimization research, metaheuristics, i.e., stochastic algorithms inspired
(2019) Critical Evaluation of
Metaheuristic Algorithms for Weight
by principles of evolution, swarm intelligence, or other physical phenomena, enriched with
Minimization of Truss Structures. simple probabilistic and/or statistical methods, have been established as a prominent tool for
Front. Built Environ. 5:113. solving complex problems. These algorithms possess certain important advantages, such as
doi: 10.3389/fbuil.2019.00113 easy implementation and good performance without being dependent on the gradient or other

Frontiers in Built Environment | www.frontiersin.org 1 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

problem-specific information (Eiben and Smith, 2003). As averaged to a single position during the Big Crunch phase.
a result, many applications of metaheuristics in the size Repeating cycles of these phases guide the algorithm toward
optimization of trusses, in particular, have been presented in optimality. Applications of this method for the size optimization
the literature. of trusses can be found in Camp (2007) and Kaveh and Talatahari
Arguably, the most widely used member of the metaheuristic (2009b). Other studies on size optimization of trusses utilize
family is Genetic Algorithms (GAs), which originate from the Colliding Bodies Optimization algorithm (CBO) (Kaveh and
work of Holland (1975). In GAs, a population of individuals Ghazaan, 2015; Kaveh and Mahdavi, 2015a,b), which is based
is evolved using appropriate mutation, crossover, and selection on an analogy with one or two-dimensional collisions between
operators. Based on Darwin’s principles of natural selection bodies, Enhanced Bat Algorithm (EBA) (Kaveh and Zakian,
and survival of the fittest, promising genetic information 2014), Ray optimization (RO) (Kaveh and Khayatazad, 2013),
is propagated into future generations providing solutions to Charged System Search (CSS) (Kaveh and Talatahari, 2010), etc.
the optimization problem. Application of GAs to truss size Notably, Differential Evolution (DE), introduced by Storn
optimization can be found in Koumousis and Georgiou (1994), and Price (1997), is a stochastic optimization method that is
Rajan (1995), and Coello and Christiansen (2000). not based on any natural paradigm. An initial version of the
Another popular method of optimization is Simulated algorithm, termed “Genetic Annealing,” was first published in
Annealing (SA). It is inspired by the annealing process of physical a programmer’s magazine in 1994 (Price et al., 2005). Several
systems which, being at a high-energy state, are gradually cooled applications of DE for the size optimization of trusses can be
down to their minimum energy level. This process was first found in the literature (Wu and Tseng, 2010; Bureerat and
formulated into an optimization algorithm by Kirkpatrick et al. Pholdee, 2016). Other studies on the same topic that are not
(1983). An application of SA for size optimization of trusses can nature-inspired are the Adaptive Dimensional Search (Hasançebi
be found in Lamberti (2008). and Azad, 2015) and the Guided Stochastic Search (Kazemzadeh
More recently, Geem et al. (2001) proposed Harmony Search Azad and Hasançebi, 2015). In the latter, problem-specific
(HS) for solving combinatorial optimization problems. HS was information is utilized as the stochastic optimization process is
inspired by the harmony sought by music players as they actually guided by the principle of virtual work.
improvise the pitches of their instruments. This algorithm has Finally, researchers have attempted to combine the better
also been used for size optimization of trusses (Lee and Geem, features of different algorithms by means of hybridization in
2004; Lamberti and Pappalettere, 2009; Degertekin, 2012). order to achieve lighter truss designs, as in Hybrid Particle
In the category of Swarm Intelligence algorithms, Karaboga Swarm—Swallow Swarm (HPSSO) (Kaveh et al., 2014a),
and Basturk (2008) proposed the Artificial Bee Colony (ABC) Hybrid Particle Swarm—Ant Colony Strategy—Harmony
algorithm, a stochastic algorithm inspired by the foraging Search (HPSACO) (Kaveh and Talatahari, 2009a), and Swarm
behavior of honey bees. An application of ABC for the size Intelligence—Chaos Theory in Chaotic Swarming of Particles
optimization of trusses can be found in Sonmez (2011). (CSP) (Kaveh et al., 2014b).
In the same category, Particle Swarm Optimization (PSO) Metaheuristic algorithms are commonly compared based on
is based on the social sharing of information among members, mathematical functions with clearly defined characteristics, e.g.,
which produces behavioral patterns that offer an evolutionary unimodal/multimodal, basic/expanded/hybrid, separable/non-
advantage (i.e., avoid predators, seek food, and mates). The separable, convex/non-convex. These characteristics cannot be
method, introduced by Kennedy and Eberhart (1995), searches clearly measured when considering truss size optimization. On
the design space by adjusting the velocities of moving “particles.” the other hand, studies on truss size optimization do not focus
As they move, the particles are stochastically attracted toward on the comparison between algorithms. Usually, only the best
both their personal best position and the best position found final design is compared which may have been found using
by the whole “swarm” (Clerc and Kennedy, 2002). Many an excessive computational budget or, in certain cases, with
applications of PSO for the size optimization of trusses can constraint violations. Thus, a need for an overall and unbiased
be found in the literature (Ray and Saini, 2001; Fourie and comparison is evident. There are certain publications in the
Groenwold, 2002; Schutte and Groenwold, 2003; Li et al., literature regarding this topic (e.g., Keane, 1996; Hasançebi et al.,
2007; Perez and Behdinan, 2007; Dimou and Koumousis, 2009; 2009). However, the list of examined algorithms is limited, while
Talatahari et al., 2013). there is also a need for more detailed statistical analysis. As
Recently, Rao et al. (2011) proposed a method called measure of the complexity, the number of design variables D, as
“teaching-learning-based optimization” (TLBO), inspired by the well as the degree of static indeterminacy SI is indicated for each
learning process of students who are influenced by their teachers problem. Some preliminary results on the problem have been
as well as their fellow students. Several applications of TLBO for presented by Charalampakis (2016), which are greatly revised
the size optimization of trusses can be found in the literature and extended herein.
(Degertekin and Hayalioglu, 2013; Camp and Farshchin, 2014; Motivated by the above, in this work, an unbiased framework
Dede and Ayvaz, 2015). for meaningful comparison between different metaheuristic
Further, the literature is abundant with algorithms inspired optimization methods is defined. A strict computational budget
by physical phenomena. In Big Bang—Big Crunch (BB-BC), is predetermined, according to the dimensionality of the
proposed by Erol and Eksin (2006), a random and disordered problem. Performance is always measured with respect to
population, produced at the Big Bang phase, is qualitatively function evaluations, i.e., truss analyses, instead of generations

Frontiers in Built Environment | www.frontiersin.org 2 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

or optimization steps. All function evaluations are carefully of the feasible region, the algorithm will be driven back from the
accounted for. A large sample is collected and analyzed for boundary, inside the feasible region, in a strong manner. Also,
each algorithm and each problem, to ensure valid statistics. it will be difficult for the algorithm to move from one feasible
To quantify the robustness of an algorithm, in this study the region to another if there are more than one, unless they lie very
probability that an arbitrary run will provide a solution of a close to each other. On the contrary, using a small penalty will
certain quality (i.e., the probability of a “successful” design) lead to a fruitless exploration of the infeasible region, and thus
is measured. This is very useful in real-life engineering as it wasted computational effort, because the penalty will be small
provides a meaningful measure of confidence regarding the as compared to the objective value itself. Since, it is expected
optimized design of a structure that has not been studied before that many of the problems examined herein will have optimum
or cannot be studied thoroughly due to resource limitations. solutions lying on the boundary between feasible and infeasible
The lightest solutions which do not violate constraints, produced regions, these issues are of importance. There are many types of
either herein or (to the best of our knowledge) in the literature, penalty functions which cannot be examined here; for a survey
are utilized to screen successful from unsuccessful designs. The on the state-of-the-art, see Coello (2002). Since, it is very difficult
success rate of each algorithm is not only evaluated at the end but to define a criterion to quantify what is a “small” violation of
also monitored throughout the analysis, which reveals interesting the constraints, in this study a static penalty rule is used with a
information regarding the relative convergence rate. small constant term to keep all best solutions strictly within the
Based on this framework, a critical evaluation of a number feasible domain, i.e., no violation at all is accepted. A normalized
of metaheuristic algorithms is presented, including variants of constraint violation function is defined as follows:
GAs, PSO, ABC, SA, and DE. The comparison is based on several
benchmark problems of varying complexity (number of design
j
variables D = 4–200, degree of static indeterminacy SI = 0–50). j vo
Most of these problems have been studied by numerous vn = j
− 1, (2)
va
researchers using a large variety of methods; this allows for
absolute rather than relative comparison. The differences in j
performance between the algorithms are shown and explained. In where, vn is the normalized violation of the jth optimization
j
addition, often a slightly better design is produced, as compared constraint (stress, displacement, buckling stress, etc.), vo is the
j
to the existing literature. corresponding value computed for a candidate solution, and va is
the allowable constraint limit. The penalty function is of the form:
PROBLEM STATEMENT
Nc
The size optimization problem of a truss with D variables can be X 
j

formally stated as follows. P (x) = δ j Avn + B , (3)
Minimize the objective function: j=1

where, A = 106 , B = 103 , Nc = the number of constraints and

f (x) = W (x) + P (x) (1) δ j = the activation key defined as:

which is subject to: (

j
(a) side constraints xL ≤ x ≤ xU (note that vector inequalities are j 1, vn > 0
δ = j . (4)
applied element-by-element); 0, vn ≤ 0
(b) additional constraints (depending on the problem) regarding
element stress, buckling stress and/or node displacements. The constant B is used to create a step-like rule, which adds a
significant penalty for slight violations. This has proved to be
In Equation (1), x = {x1 , x2 , ..., xD } = a vector which contains
particularly effective and ensures that the best solutions found are
the areas of the cross-section of each group of elements, xL and
devoid of any violations.
xU = vectors which define the minimum and maximum areas,
PD
respectively, W (x) = (Li xi ρi ) = structural weight of the METAHEURISTIC ALGORITHMS
i=1
truss, Li and ρi = total length and specific weight of the ith
Basic Features of the Trade Study
group of bars, and P (x) = penalty function, which is a common
The following assumptions/rules have been used equally for
method to transform the constrained optimization problem into
all algorithms:
an unconstrained problem.
The penalty function should follow the so-called “minimum 1) The same configuration was used for each algorithm in all test
penalty rule,” which is not easy to achieve. This rule dictates that problems. For well-established methods, internal parameters
the penalty should be kept as small as possible, yet large enough to were set according to literature. For newly established
keep infeasible solutions from being optimal (Coello, 2002). If the algorithms, the best combination of parameters was found via
penalty is too large and the optimum solution lies at the boundary parametric analysis.

Frontiers in Built Environment | www.frontiersin.org 3 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

2) The number of structural analyses required in the 5. Repetition of steps (2–5) until some termination condition
optimization process was chosen as the best performance is satisfied.
indicator to compare algorithms of different nature
The particular implementation of SGA used herein includes jump
and configuration.
and creep mutation, as well as elitism. Jump mutation randomly
3) Thirty independent runs with different random seeds
flips a bit in the genotype, while creep mutation randomly
were conducted for each test case and each optimization
increases or decreases a gene in the phenotype by a single
algorithm in order to obtain statistically significant results.
step. Elitism directly transfers the better individual(s) from the
The conclusions are supported by t-tests regarding the
previous generation to the next, ensuring that the solution quality
significance of the differences between algorithms.
will not decrease during evolution. The parameters of SGA are set
4) A robust random number generator of L’Ecuyer with Bays-
as follows: gene length Lg = depending on the problem so that the
Durham shuffle and added safeguards (Press et al., 2002)
 Li
i i
step xU − xL / 2 − 1 ≤ 10−3 ∀i ∈ {1, 2, ..., D}; population
g
was used. Henceforth, r = random variable with uniform
distribution in the interval (0,1), sampled anew each time it size P=50; single crossover with probability 0.7; jump mutation
is required, and r = corresponding vector. with probability 1/P ; creep mutation with probability Lc /D /P
5) For each test problem, a value-to-reach (VTR) defines the D
Lig = chromosome length in bits); tournament selection
P
limit between success and failure. The VTR is 1% heavier than (Lc =
i=1
the best feasible design obtained in this study or reported in with 2 individuals; and elitism (with 1 elite individual).
the literature. Hence, “successful” designs lie very close to the
optimum region of design space. The final design is usually Hybrid Genetic Algorithm (HGA)
more fine-tuned when the VTR threshold is reached early in The standard GA formulation is hybridized by combining the
the optimization process. search space reduction method (SSRM) and a local optimization
6) Regarding the computational budget, it obviously needs to algorithm to form the Hybrid Genetic Algorithm (HGA). SSRM
be analogous to the problem difficulty which is generally is a systematic method which gradually reduces the search
acknowledged that increases non-linearly in D. This rapid space. It facilitates the optimization algorithm by focusing
growth in difficulty is commonly referred to as “curse on the promising areas which lead to better solutions. The
of dimensionality” (Bellman, 1957); however, the allocated SSRM embedded in HGA was presented in Charalampakis and
computing resources cannot follow this pattern. For this Koumousis (2008) and later employed in several cases (e.g.,
reason, in this study a linear function is used, i.e., the Marano et al., 2011). It is based on a statistical analysis of a
computational budget is set to 2500D structural analyses and population of P candidate designs, where each one is assigned a
the best design is achieved strictly within this limit. This weight w depending on its quality. For minimization problems:
budget balances two conflicting aspects: (a) it is high enough
to achieve good designs (often better than those reported

in the literature); (b) it is not too excessive to conflict with max f
limitations on computing resources. wk = , k ∈ {1, 2, ..., P} , (5)
fk
7) The unit system utilized is the same as in the original
statement of the problem, in order to avoid rounding errors

where max f = max f1 , f2 , ..., fP is the worst objective value
during comparisons. in the population, which corresponds to a weight of unity. The
Cross-sectional areas of elements are rounded to three decimal weighted mean value of the design variable i is then calculated as:
digits before performing structural analysis in double precision
without any further rounding. Literature designs including more
P
than three decimal digits are exactly reproduced in this paper; P
(wk xik )
however, the total weight is re-evaluated and may be slightly k=1
different from the source value. The inversion of the stiffness mi = , i = {1, 2, ..., D} , (6)
P
P
matrix is performed using Gauss-Jordan elimination with full (wk )
pivoting (Press et al., 2002). k=1

where xik is the value of the ith variable of the kth candidate
Algorithm Description design. The descriptive weighted standard deviation of the design
Standard Genetic Algorithm (SGA) variable i is given as:
The pseudocode of the so-called standard genetic algorithm
(SGA) is as follows: v
u P
uP
w (x − mi )2


1. Random initialization of the population;
u k=1 k ik
u
2. Fitness calculation; si = u . (7)
P
3. Selection of individuals to form the new population;
u P
t
(wk )
4. Crossover and mutation; k=1

Frontiers in Built Environment | www.frontiersin.org 4 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

The new trial lower and upper bounds of the design variable i are where vmax
0 = γ (xU − xL ) and γ = a scalar parameter. The
formed symmetrically around mi : position vector of a particle m at the next time instant k + 1 is
given by:
i =m −qs
x̄L,j i i
i , (8)
x̄U,j = mi + q si xm m m
k+1 = xk + vk+1 . (12)
where j is an index for SSRM steps and q is a scalar parameter. Note that the time step 1t between the distinct time
Finally, the new range of values of the design variable i is instants is taken equal to unity. The velocity vector vm
k+1 is
the intersection (common part) of the previous range and the given as:
trial range:
g
h
i i
i h
i i
i h
i i
i vm m m m m
k+1 = wk vk + c1 r1 ◦ (pk − xk ) + c2 r2 ◦ (pk − xk ), (13)
xL,j+1 , xU,j+1 = xL,j , xU,j ∩ x̄L,j , x̄U,j , (9)
where, wk = inertia factor at time instant k; c1 , c2 = cognitive and
Based on Equation (9), the search space is not allowed to expand. social parameters, respectively; pmk = best-ever position vector
Instead, it is gradually reduced on a variable-by-variable basis, g
of particle m up to and including time instant k; pk = best-
if justified by the statistical analysis. The scalar q controls the ever position vector amongst all particles up to and including
aggressiveness of SSRM. Small values may stochastically lead to time instant k; and the ◦ operator indicates element-by-element
exclusion of a promising area in subsequent SSRM steps; large multiplication. Note that these formulas are compatible with the
values cancel the beneficial effect of SSRM. In this study, q = 3 diverse “classical” version of PSO (Wilke et al., 2007). The side
is set which is a rather conservative choice (Charalampakis and constraints of Equations (10) and (11) are enforced after each
Koumousis, 2008). The chromosomes of the existing population time step.
are substituted by their closest counterpart in the new mapping. It is known that standard PSO suffers from convergence rate
If a variable falls out of a new bound, it is set equal to it. problems, due to the delicate balance between exploration and
Regarding variable discretization, the gene length is set to exploitation that is required. This was clear in the problems
10 bits irrespective of the problem at hand. This leads to a considered in this study, so an enhanced variant is used
relatively small chromosome length, which greatly facilitates the instead. This variant (EPSO) is based on the work by Fourie
GA. Refining of the solutions is achieved through SSRM, which and Groenwold (2002) and has been successfully applied in a
is triggered automatically. For the test problems considered in parameter identification problem (Charalampakis and Dimou,
this study, this happens every 5,000, 10,000, 20,000, and 50,000 2010, 2015). The differences with respect to the basic PSO
function evaluations when D = {4}, D = {8, 10, 17}, D = {29} algorithm are the following:
and D = {200}, respectively.
A local optimization algorithm, namely the Greedy Descend ◦ If a period of h consecutive steps passes without the best
Hill Climber (GDHC) (Eiben and Smith, 2003), is also embedded solution found by the whole swarm being improved,
in the HGA. The final 5,000 function evaluations of each run then it is postulated that the velocities are large and
are dedicated to GDHC. The best solution found so far becomes “overshooting” prevents the algorithm from locating
seed; this solution is continuously improved by alternating the better solutions. Therefore, a reduction is applied to
bits of the chromosome from left to right, keeping the best result both the inertia factor and the maximum velocity,
as a reference. When a full cycle has been concluded without as follows:
improvement of the best result, the local optimum has been
found and the process is terminated. If the limit of 5,000 function  
evaluations is reached prior to finding the local optimum, the g g wk+1 = a wk wk+1 = wk
If f (pk ) = f (pk−h ) then max else max .
reference solution becomes final. vmax
k+1 = β vk vmax
k+1 = vk
(14)
Enhanced Particle Swarm Optimization (EPSO)
In the basic PSO algorithm, the population consists of P ◦ The craziness operator imposes a random velocity vector to
particles. Each particle is assigned a position and velocity vector, a particle which hence strays from the swarm and examines
determined at time instant k by xk and vk , respectively. The other regions of the search space. A probability Pcr is used to
particles are initialized randomly in the box-constrained design activate the operator, as follows:
space, so that:
If r < Pcr then randomly assign vm max m max
k+1 with -vk+1 ≤ vk+1 ≤ vk+1 .
xL ≤ xm
0 ≤ xU ∀m ∈ {1, 2, ..., P} . (10) (15)
The initial velocities of the particles are also chosen randomly:

◦ The worst individual is moved to the best ever position found

− vmax
0 ≤ vm max
0 ≤ v0 ∀m ∈ {1, 2, ..., P} , (11) by the swarm.

Frontiers in Built Environment | www.frontiersin.org 5 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

g
◦ If the velocity vm
k leads to an improvement of pk then, instead
x = xL + r ◦ (xU − xL ) . (20)
of Equation (12), the so-called “elite velocity” rule is applied
for particle m: In this study, the parameters of the ABC algorithm are set as
follows: SN = 50 and LIMIT = SN/2 × D.

Differential Evolution (DE)

g
xm m
k+1 = pk + c3 r3 ◦ vk , (16) In the classic DE algorithm, a population of P individuals is
initialized randomly within the design space, as follows:
where c3 = a scalar parameter. Note the difference between the
current implementation and the work by Fourie and Groenwold
(2002), where a single random variable r3 is multiplied by the xL ≤ xi,0 ≤
xU ∀i ∈ {1, 2, ..., P} 
velocity vector vm
k.
Px,g = xi,g
, i ∈ {1, 2, ..., P} , g ∈ 0, 1, ..., gmax , (21)
In this study, the parameters of EPSO are set as follows: P = xi,g = xj,i,g , j ∈ {1, 2, ..., D} .
20, c1 = 0.5, c2 = 1.6, γ = 0.4, w0 = 1.40, h = 3, a = 0.99,
β = 0.95, Pcr = 0.22, and c3 = 1.30. where Px,g = array of P vectors (= candidate solutions); xi,g = D-
dimensional vector (= candidate solution); gmax = the maximum
Artificial Bee Colony (ABC) number of generations; i, g, j = indices for vectors, generations,
According to the ABC method, the colony of artificial bees is and design variables; and the parentheses

indicate an array.
divided into two equal groups, i.e., the employed bees and the A mutated population Pv,g = vi,g is created from the current
onlookers. Also, the number of food sources (possible solutions) population Px,g in each generation g, as given by:
around the hive is equal to the number of employed bees.
Initially, the ABC initializes randomly SN/2 food source

positions, where SN = total colony size. Next, the food vi,g = xr0,g + F xr1,g − xr2,g , (22)
source positions are subjected to repeated cycles of improvement
as follows. where, r0 , r1 and r2 are random integers in {1, 2, ..., P}, mutually
The employed bee i modifies the position vector of the different and also different from the index i; xr0,g = base vector;
associated food source i: and F = a scalar parameter. Note that after the formation of the
mutated population using Equation (22), all design variables are

relocated within their respective boundaries, if necessary.


x̄ij = xij + (2r − 1) xij − xkj , (17) In the next step, a trial population Pu,g = ui,g is created,
consisting of members taken from both the parent and mutated
where i, k ∈ {1, 2, ..., SN/2} = random indices with i 6= k populations, as follows:
and j ∈ {1, 2, ..., D} = a randomly chosen dimension of the
D-dimensional vector xi to be modified. A greedy selection is 

performed between the current and modified vector; the best
vj,i,g , if r ≤ Cr or j = jrand
ui,g = uj,i,g = , (23)
either becomes or remains the food source i. xj,i,g , otherwise
Next, the employed bees share the information with the
onlooker bees. An onlooker bee chooses each food source with where jrand = a random index in {1, 2, ..., P} which makes certain
a probability given by: that a minimum of one design variable will be taken from the
mutant vector vi,g ; and Cr = a parameter with values in the
range [0,1].
fiti The last step of DE is a greedy selection criterion. In the case
pi = , (18)
SN/2
P of minimization problems, it is given by:
fitk
k=1



ui,g , if f ui,g ≤ f xi,g
where fiti = the fitness of food source i. For minimization xi,g+1 = . (24)
xi,g , otherwise
problems, this can be evaluated by:
The aforementioned classic DE implementation is denoted as
1 rand/1/bin (Price et al., 2005), or DE1 herein for short. In
fiti = , (19) general, it is acknowledged that DE1 exhibits intense exploration
1 + fi
capability and thus is more suitable for multimodal problems
where fi = the objective value of food source i. The onlooker bee (Qin et al., 2009). Following recommendations in Price et al.
also modifies the selected food source according to Equation (17) (2005) and Rönkkönen et al. (2005), F = 0.5 and a high value
and applies a greedy selection criterion between the current and of Cr = 0.9 is selected which is anticipated to lead to good
modified vector. performance with non-separable functions. The population size
Finally, if a food source has not been improved over a period of is taken equal to 50 for all problems.
LIMIT cycles, then the associated employed bee becomes a scout. Another popular DE variant is denoted as best/1/bin (Price
This means that the food source is re-initialized as: et al., 2005) in which the currently best vector of the population

Frontiers in Built Environment | www.frontiersin.org 6 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

is used as a base vector. In addition, jitter is introduced to F and, avoid zero success rate. Parameter F, which is closely related to
thus, Equation (22) becomes: the convergence speed, is sampled from a normal distribution
with mean value 0.5 and standard deviation 0.3, denoted by

N (0.5, 0.3). Regarding parameter Cr , it has been observed that
vi,g = xbest,g + Fj xr1,g − xr2,g good settings generally fall into a small range of values for a given
, (25)
Fj = F + d (r − 0.5) problem (Qin et al., 2009). In SaDE, Cr for the kth strategy is
sampled from a normal distribution N (CRmk , 0.1), where CRmk
where d = 0.001 = the magnitude of jitter. This variant is a parameter initialized by the value 0.5. A memory named
was tested in this study and found to be too greedy, as it CRMemoryk stores the values of Cr that produced trial vectors
showed great initial performance which was quickly followed which successfully competed with existing vectors and entered
by stagnation in the more difficult problems. Instead, another the next population using the kth strategy. When the memory
DE variant is included which is denoted as DE3 or “rand- overflows, the earliest entries are replaced. At generation G > LP,
best/1/bin” (Charalampakis and Dimou, 2015). Let us define rb as CRmk is overwritten by the median value of CRMemoryk . For
the expected ratio of evaluations with a random base vector to the more implementation details, see Qin et al. (2009).
total number of evaluations. If r < rb , then evolution proceeds
according to rand/1/bin with jitter. Conversely, best/1/bin is Simulated Annealing (SA)
used. Based on the information presented in a separate section, The implementation of the SA algorithm used herein is based
a high value rb = 0.90 is chosen to promote exploration. on the work by Balling (1991). The algorithm begins with the
From the very beginning, DE has proved to be a very efficient creation of D random designs in the box-constrained design
and robust optimization algorithm. Of course, its performance space xL ≤ x ≤ xU . The best of these designs becomes the current
is dependent to a certain degree on both the strategy for the design xc . This preliminary loop is done in order to avoid the
generation of trial vectors and the values of control parameters. possibility that the starting point for the optimization is too poor.
Although there exist suggested values for parameters, there is Following the selection of a suitable cooling schedule, the current
no specific setting that is equally suitable for all problems or design xc is subjected to small perturbations to create a candidate
even at different optimization stages of a single problem. For design xa . Whenever, xa is better than xc , it replaces it with a
this purpose, several adaptive algorithms have been proposed, probability of 1. For a minimization problem, this is expressed
such as SaDE (Qin et al., 2009), JADE (Zhang and Sanderson, mathematically as:
2009), and SaNSDE (Yang et al., 2008). In this study, SaDE is
considered as the final DE variant, in which four competing
strategies are employed simultaneously, namely rand/1/bin, 1f = f (xa ) − f (xc ) ≤ 0 ⇒ P (xa → xc ) = 1. (28)
rand-to-best/2/bin, rand/2/bin, and current-to-rand/1, with K =
4 = the total number of strategies. The probability of selecting If xa is worse than xc , it still replaces it with a probability given by:
each strategy is initialized to 1/K and the stochastic universal
selection method is used to determine which strategy is going 1f = f (xa ) − f (xc ) > 0 ⇒ P (xa → xc ) = e−1f /(Ktc ) ≤ 1, (29)
to be adapted for each vector in the population. The successes
and failures of each strategy are recorded to the corresponding where, K = Boltzman parameter and tc = the current system
memory of the last LP = 50 generations, where LP = learning temperature, corresponding to the cooling cycle c. The Boltzman
period. Once the memory overflows, new results replace the parameter is not kept constant during optimization. Instead, it is
earliest ones. At generation G > LP, the probability of choosing updated prior to using Equation (29) as follows:
the kth strategy is given by:
 
KNa Na + 1f 
Sk,G KNa +1 = , (30)
pk,G = , (26) Na + 1
K
P
Sk,G where, Na = number of accepted designs so far, and KNa =
k=1
previous Boltzman parameter. Initially, Na = 0 and K0 = 1. The
where starting and final system temperatures are given by:

G−1 1 1
P
nsk,g ts = − , tf = −
, (31)
ln (Ps ) ln Pf
g=G−LP
Sk,G = + ε, (27)
G−1 G−1
P
nsk,g +
P
nfk,g where, Ps and Pf = the starting and the final probability of
g=G−LP g=G−LP acceptance, respectively. The temperature is reduced gradually in
Nc cooling cycles, as follows:
where Sk,G = the success rate of the kth strategy; nsk,g , nfk,g =
the number of successes and failures, respectively, of strategy
k at generation g; and ε = 0.01 a small constant value to tc+1 = cf × tc , (32)

Frontiers in Built Environment | www.frontiersin.org 7 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

where, cf ∈ (0, 1) = cooling factor, given by: cycle encourage acceptance of poorer designs (ii) poorer designs
with small 1f have higher probability of acceptance than poorer
  1 designs with large 1f (iii) the temperature is gradually decreased
tf Nc −1 during the cooling process and, thus, the probability of accepting
cf = . (33)
ts a poorer design is reduced accordingly. This gradually shifts the
focus from the exploration of the search space to the exploitation
At each cooling cycle, a number of inner loops of of the most promising solution.
perturbation/improvement of the current design xc is executed. In this study, the parameters of the SA algorithm are set as
For each loop, an array of integers representing the design follows: Ps = 0.5, Pf = 1E − 07, Nc = 300, Is = 1, If = 3, d =
variables (1 to D) is shuffled. According to the sequence 0.01 (xU − xL ).
indicated in the shuffled array, each design variable i is in turn
perturbed to form the candidate design xa according to:
TEST PROBLEMS
xia = xic + (2r − 1) di , (34) 18-Bar Truss (D = 4, SI = 0)
The planar 11-node, 18-bar truss shown in Figure 1 is subjected
where di = the magnitude of perturbation of variable i. Note that to concentrated loads of P = 20 kips acting on the free nodes of
the side constraints xL ≤ xa ≤ xU are re-enforced after using the upper chord. The same material is used for all members (E =
Equation (34). It has been observed that inner loops are more 10,000 ksi and ρ = 0.1 lb/in3 ). A symmetrical stress limit σmax =
important at low temperatures. For this reason, the repetitions −σmin = 20 ksi is imposed for both tension and compression. In
I of the inner loop are not kept constant but instead, they are addition, the Euler buckling limit of the ith bar is calculated as:
determined as follows:
βEAi
! σiE = − , (36)

tc − ts L2i
I = round Is + If − Is , (35)
tf − ts
where, Li , Ai = length and area of the ith bar, respectively, and
where, Is and If = the starting and the final number of repetitions β = 4 = a constant determined from geometry. The problem
of inner loops, respectively. Note that in SA the current design features four design variables (D = 4) which correspond to the
is occasionally replaced by a poorer design due to Equation (29). following bar groups: (i) 1, 4, 8, 12, and 16; (ii) 2, 6, 10, 14, and
This means that even if a better design replaces the current design 18; (iii) 3, 7, 11, and 15; (iv) 5, 9, 13, and 17. The areas of each
due to Equation (28), it may not be the best ever design. For our group of elements can vary between 0.10 and 50 in2 .
purposes, it is desirable to keep track of the best design found This problem has been solved in the literature either
so far during the optimization process. This best design xb is considering only sizing variables (Lee and Geem, 2004; Sonmez,
updated after each function evaluation. 2011), or combining layout and sizing variables (Lamberti,
The driving force of the algorithm is based on three pillars: 2008). Note that the truss is statically determinate, which
(i) high temperatures observed at the beginning of the cooling means that the force acting in each bar is independent of the

FIGURE 1 | Planar 18-bar truss.

Frontiers in Built Environment | www.frontiersin.org 8 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

cross-sectional areas of the bars. The forces can be evaluated has been evaluated in Charalampakis and Chatzigiannelis (2018).
analytically using simple equilibrium equations (Charalampakis This result is used as a target to evaluate the relative performance
and Chatzigiannelis, 2018). of metaheuristic methods by setting VTR = 9568.715 × 1.01
The optimization becomes more challenging if there are = 9664.402 lb.
displacement constraints. The exact global optimum for a 6 in. Table S1 shows that all DE variants, EPSO, and SA practically
tip displacement constraint and four-element groups (D = 4), converged to the analytical solution of the optimization problem,
obtained with the Cylindrical Algebraic Decomposition method, while HGA and ABC exhibited larger statistical dispersion. The

FIGURE 2 | Size optimization of the planar 18-bar truss (A) average progress of the best solution (B) evolution of success rate.

FIGURE 3 | Spatial 25-bar tower.

Frontiers in Built Environment | www.frontiersin.org 9 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

average progress of the best solution, as well as the evolution 25-Bar Tower (D = 8, SI = 7)
of the success rate, are shown in Figure 2. It can be seen The spatial 25-bar tower with 10 nodes shown in Figure 3 has
that EPSO and DE variants find the optimum region very been studied extensively in the literature. The same material is
quickly; 100% success rate is reached after only ∼350D analyses. used for all members (E = 10,000 ksi, ρ = 0.1 lb/in3 ). The cross-
SA ranks next, requiring ∼1250D analyses for 100% success sectional area of each bar may vary between 0.01 and 35 in2 . Bars
whereas HGA reaches ∼95% success at the end. Although it are grouped in eight groups (D = 8) with different compressive
appears that ABC is not trapped into local optima, its progress stress limit but the same tensile stress limit (see Table 1). In
is slow, and many structural analyses are required to obtain addition, displacements of top nodes 1 and 2 must be <0.35 in. in
good results. all directions. Two independent loading conditions are applied,
as shown in Table 2.
Table S2 shows that DE1 and DE3 obtained the same best
design, which is slightly better than the one found in literature
TABLE 1 | Member grouping and stress limits for the spatial 25-bar tower.
in terms of weight [including MSPSO (Talatahari et al., 2013) by
Member Members Compressive Tensile stress a very small margin]. This design was hence used as reference
group stress limit [ksi] limit [ksi] (VTR = 545.172 × 1.01 = 550.623 lb). The statistical data shown
in Table S3 indicate that EPSO ranked right after DE in terms
1 1 35.092 40
of best weight but was less robust than ABC and HGA. The
2 2, 3, 4, 5 11.590 40
average progress of the best solution and the variation of the
3 6, 7, 8, 9 17.305 40
success rate of each algorithm measured against the number
4 10, 11 35.092 40
of structural analyses are compared in Figure 4. It can be seen
5 12, 13 35.092 40
that DE1 and DE3 could reach 100% success after ∼750D
6 14, 15, 16, 17 6.759 40 analyses, while SaDE follows at ∼1250D analyses. SA, being
7 18, 19, 20, 21 6.959 40 the only method not employing a population of solutions but
8 22, 23, 24, 25 11.082 40 rather perturbing/improving a single solution, was the slowest
to converge.
TABLE 2 | Load cases for the spatial 25-bar tower.

Node Px [kips] Py [kips] Pz [kips] 10-Bar Truss (D = 10, SI = 2)

Figure 5 shows the geometry and forces acting on a cantilever
Load case I
truss consisting of 10 bars and 6 nodes which has been analyzed
1 0 20 −5
by many researchers. Two loading cases are considered: in case
2 0 −20 −5 I, P1 = 100 kips and P2 = 0; in case II, P1 = 150 kips and P2 =
Load case II 50 kips. The same material is used for all members (E = 10,000
1 1 10 −5 ksi and ρ = 0.1 lb/in3 ). The cross-sectional area of each bar may
2 0 10 −5 vary between 0.10 and 35 in2 . The stress limit is set to ±25 ksi,
3 0.5 0 0 while the displacement of the free nodes must be limited to 2 in.
6 0.5 0 0 in all directions.

FIGURE 4 | Size optimization of the spatial 25-bar tower (A) average progress of the best solution (B) evolution of success rate.

Frontiers in Built Environment | www.frontiersin.org 10 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

FIGURE 5 | Planar 10-bar truss.

FIGURE 6 | Size optimization of the planar 10-bar truss (load case I) (A) average progress of the best solution (B) evolution of success rate.

Regarding load case I, Table S4 shows that the best result was final results are comparable to EPSO. HGA follows next, reaching
discovered by DE3 and used as reference (VTR = 5060.855 × ∼60% at the end. The small difference between SGA and HGA
1.01 = 5111.464 lb). It is clear from Figure 6 that EPSO, DE1 during the first 10,000 analyses is due to the coarser discretization
and DE3 variants find the optimum region very quickly, as 100% of the search space (16 and 10 bits per variable for SGA and HGA,
success rate is reached after only ∼500D analyses. SaDE follows respectively). The SSRM is triggered at 10,000 analyses while
next, producing very good results but with a slower convergence GDHC is introduced at 20,000 analyses, as shown in Figure 6A.
rate. Interestingly, SA may be the slowest to converge, but its Table S5 reveals the very small standard deviation of the results

Frontiers in Built Environment | www.frontiersin.org 11 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

FIGURE 7 | Size optimization of the planar 10-bar truss (load case II) (A) average progress of the best solution (B) evolution of success rate.

FIGURE 8 | Planar 17-bar truss.

obtained by DE1 and DE3, as well as the small final ranges for all (E = 30000 ksi and ρ = 0.268 lb/in3 ). The structure is subject to
design variables. a vertical force of 100 kips at node 9. The cross-sectional area of
Regarding load case II, Table S6 shows that the best solution each bar may vary between 0.1 and 50 in2 . The stress limit is 50
was discovered by DE3 and used as reference (VTR = 4676.932 ksi for both tension and compression, while the displacement of
× 1.01 = 4723.701 lb). The average progress of the best solution free nodes must be < ±2 in. Since no design variable linking was
and the evolution of the success rate are presented in Figure 7, used, the problem dimensionality is D = 17.
whereas a statistical analysis of the results is given in Table S7. It can be seen from Table S8 that the best design was obtained
The same conclusions can be drawn as in load case I. Once again, by DE3. The corresponding structural weight was taken as
the final variable ranges for DE variants are very narrow. target to compute VTR = 2581.895 × 1.01 = 2607.714 lb. The
statistical data are given in Table S9 confirm the robustness of
17-Bar Truss (D = 17, SI = 3) DE1 and DE3 that achieved a standard deviation on optimized
The planar 17-bar with 9 nodes shown in Figure 8 has been weight equal to 0.047 and 0.086 lb, respectively, and converged
studied in Berke and Khot (1987), Adeli and Kumar (1995), and practically to the same optimized weight. The other algorithms
Lee and Geem (2004). The same material is used for all members ranked in the following order: SaDE, EPSO, SA, HGA, and ABC.

Frontiers in Built Environment | www.frontiersin.org 12 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

FIGURE 9 | Size optimization of the planar 17-bar truss (A) average progress of the best solution (B) evolution of success rate.

The average progress of the best solution and the variation of heavier than the target best design of Coster and Stander (1996),
the success rate of each algorithm with respect to the number while SaDE produced designs that are, on average, 3.21% heavier.
of structural analyses are compared in Figure 9. It can be seen Remarkably, all DE variants exhibited at least one order lower
that DE1, DE3 could reach 100% success after ∼900D analyses. standard deviations than the other algorithms. HGA ranked
SaDE also reached 100 success but much later. EPSO was the next in terms of average weight and standard deviation. ABC
best algorithm in the early stages of the optimization process but was comparable to HGA as far as best weight is concerned,
about 20% of the runs failed to succeed at the end. SA achieved a but its standard deviation was more than three times larger.
65% success rate when measured at the end of the analyses. The average progress of the best design and the evolution of
success rate with respect to the number of structural analyses
200-Bar Truss (D = 29, SI = 50) are plotted in Figure 11 for all algorithms. SA shows an almost
Figure 10 shows a planar 200-bar truss consisting of 200 bars and linear convergence rate, ranking last. It can be seen that only
77 nodes. The same material is used for all members (E = 30,000 DE1 and DE3 reached a significant success rate, ranging between
ksi and ρ = 0.268 lb/in3 ). The structure is subjected to three 85 and 90%, demonstrating excellent fine-tuning capabilities.
independent loading conditions: (i) 1.0 kips acting in the positive SaDE barely missed producing successful runs, with the best
x-direction at nodes 1, 6, 15, 20, 29, 34, 43, 48, 57, 62, and 71; (ii) design of 25794.837 lb achieved within 2500D analyses. Finally,
10 kips acting in the negative y-direction at nodes 1, 2, 3, 4, 5, 6, Table S12 indicates that the superiority of DE3 is statistically very
8, 10, 12, 14, 15, 16, 17, 18, 19, 20, 22, 24, 26, 28, 29, 30, 31, 32, 33, significant against all algorithms with the exception of DE1, with
34, 36, 38, 40, 42, 43, 44, 45, 46, 47, 48, 50, 52, 54, 56, 57, 58, 59, 60, which no safe conclusion can be drawn.
61, 62, 64, 66, 68, 70, 71, 72, 73, 74, 75; (iii) the first two load cases
acting together. Bars are linked in 29 groups (D = 29), as shown 200-Bar Truss (D = 200, SI = 50)
in Table 3, and their cross-sectional area may vary between 0.1 In order to compare the algorithms in a large-scale problem,
and 35 in2 . The structure is optimized with respect to limitations the previously described planar 200-bar truss is re-examined
on member stresses that cannot exceed ±10 ksi; no displacement without bar grouping; thus D = 200. Note that the best designs
constraints are imposed. of the previous example are actually valid in this case as well,
This optimization problem has been solved with both although not optimal anymore. Table S13 presents the statistical
gradient-based and metaheuristic algorithms as an example of data gathered from optimization runs. It can be seen that the
average-scale structure. Some optimum designs for this problem best design was found by DE1 (VTR = 21999.715 × 1.01 =
are presented in Table S10. Overall, the two best solutions 22219.712 lb); this design is quoted in Table S14. DE1 and
were obtained by Coster and Stander (1996), although these DE3 were overall the best algorithms, followed by SaDE. Next
designs slightly violate a few constraints. In order to correct this, comes ABC although the standard deviation of this algorithm
cross-sectional areas of some element groups were increased by was about 25% larger than that of HGA. The average progress
0.001 in2 . The modified designs turned feasible and the VTR is of the best design measured against the number of structural
computed as 25452.332 × 1.01 = 25706.855 lb. analyses is shown in Figure 12A. It is confirmed that DE1 and
The statistical data given in Table S11 indicate that DE DE3 are definitely superior over the other algorithms, exhibiting
variants are definitely superior over the other metaheuristic remarkable scalability with respect to problem dimensionality.
algorithms considered in this study. In particular, DE1 and SA again shows an almost linear convergence rate, which is
DE3 converged to designs that are, on average, only 0.58–0.67% not enough to surpass other algorithms within the selected

Frontiers in Built Environment | www.frontiersin.org 13 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

TABLE 3 | Bar group members for the planar 200-bar truss (D = 29).

Group Members

1 1, 2, 3, 4
2 5, 8, 11, 14, 17
3 19, 20, 21, 22, 23, 24
4 18, 25, 56, 63, 94, 101, 132, 139, 170, 177
5 26, 29, 32, 35, 38
6 6, 7, 9, 10, 12, 13, 15, 16, 27, 28, 30, 31, 33, 34,
36, 37
7 39, 40, 41, 42
8 43, 46, 49, 52, 55
9 57, 58, 59, 60, 61, 62
10 64, 67, 70, 73, 76
11 44, 45, 47, 48, 50, 51, 53, 54, 65, 66, 68, 69, 71,
72, 74, 75
12 77, 78, 79, 80
13 81, 84, 87, 90, 93
14 95, 96, 97, 98, 99, 100
15 102, 105, 108, 111, 114
16 82, 83, 85, 86, 88, 89, 91, 92, 103, 104, 106, 107,
109, 110, 112, 113
17 115, 116, 117, 118
18 119, 122, 125, 128, 131
19 133, 134, 135, 136, 137, 138
20 140, 143, 146, 149, 152
21 120, 121, 123, 124, 126, 127, 129, 130, 141, 142,
144, 145, 147, 148, 150, 151
22 153, 154, 155, 156
23 157, 160, 163, 166, 169
24 171, 172, 173, 174, 175, 176
25 178, 181, 184, 187, 190
26 158, 159, 161, 162, 164, 165, 167, 168, 179, 180,
182, 183, 185, 186, 188, 189
27 191, 192, 193, 194
28 195, 197, 198, 200
29 196, 199

FIGURE 10 | Planar 200-bar truss.

computational budget depended on the problem dimensionality,
careful accounting for all function evaluations and monitoring of
the success rate of each algorithm during the analyses.
computational budget. Figure 12B shows that success rate never Regarding the algorithms, SGA provides a baseline
exceeded 10%, because, for such a difficult problem, the VTR is performance while the HGA, which combines a SSRM with a
too strict for a computational budget that increases linearly in D. local optimizer, showed improved performance with respect
to SGA. The difference became more clear for larger problems
CONCLUDING REMARKS because of the relatively small chromosome length, as the
coarse discretization appears to be very beneficial to the HGA.
In this study the relative performance of variants of Genetic Refinement of the solutions is achieved through SSRM. The local
Algorithms (SGA, HGA), Particle Swarm Optimization (EPSO), search initiated 5,000 function evaluations before the end of
ABC, Differential Evolution (DE1, DE3, and SaDE), and SA in each run. Except for the largest problem (200 bar truss with 200
sizing optimization problems of truss structures including up to design variables), the local optimum was usually found before
200 design variables was investigated. The comparison was based the end of each run. The local search was invariably beneficial
on a framework that is both unbiased and meaningful. In order since the normal progress of the GA at this stage of optimization
to clearly assess the differences between algorithms, rigorous is usually very limited. Intermediate applications of the local
statistical analysis with large samples was used, based on common search may improve performance; however, it is known that with

Frontiers in Built Environment | www.frontiersin.org 14 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

FIGURE 11 | Size optimization of the planar 200-bar truss (D = 29) (A) average progress of the best solution (B) evolution of success rate.

FIGURE 12 | Size optimization of the planar 200-bar truss (D = 200) (A) average progress of the best solution (B) evolution of success rate.

GAs there is always the possibility of premature convergence ABC is not easily trapped into local optima but has slow
when a single individual is much better than the other members convergence because a single design variable is changed per
of the population. mutant vector. The present results indicate that ABC could not
Simple PSO was tested in the preliminary versions of this compete with EPSO and DE variants in most test cases.
study and produced very poor results; it was clear that the As far as it concerns DE variants, DE1 is based on rand/1/bin
algorithm could not maintain a stable exploration/exploitation and DE3 is a stochastic mixture of rand/1/bin and best/1/bin.
balance. Thus, an enhanced PSO algorithm (EPSO) was These variants combined particularly good convergence rate
examined instead. It resulted very competitive in small to with robustness, stability, and scalability. DE3 showed improved
medium problems but failed in the larger problems. This convergence rate in the small-scale problems. A very greedy
is due to the scheme of Equation (14) employed to reduce DE variant, based solely on best/1/bin, was also tested in
maximum velocity and inertia factor. If h = 3 steps the preliminary versions of this study and found to be very
have passed without improving the global best position competitive only in small problems. Regarding SaDE, it appears
of the swarm, it is assumed that large velocities prevent that the advantage of its adaptive nature is traded-off by a
progress due to overshooting. However, in large-scale problems, small hysteresis (delay) in the convergence rate, as compared
unjustified reduction of maximum velocity and inertia factor to DE1 and DE3. Note that the latter algorithms utilize fixed
leads to search stagnation, as the particles literally freeze settings that were expected to work well based on the specific
in place. problem characteristics. This information may not be available

Frontiers in Built Environment | www.frontiersin.org 15 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

for other problems, which makes SaDE and similar self-adapting DATA AVAILABILITY STATEMENT
algorithms a very good choice.
SA is the only method examined in this study which All datasets generated for this study are included in the
perturbs/improves a single solution. The lack of synergistic manuscript/Supplementary Files.
information within a population had the obvious effect that
SA did not show explosive initial performance. Nevertheless, AUTHOR CONTRIBUTIONS
its progress was clear and it is evident that the Metropolis test
(Equation 29) is a powerful method to escape local optima. AC had the research idea, drafted the article, and contributed to
For most of the small problems, this was enough to allow the numerical analysis of the examples and the statistical analysis
SA to rank among the best. For the planar 200-bar truss, the of the results. GT contributed to the conception and design of the
large dimensionality did not allow SA to compete within the work, and interpretation of the results.
strict computational budget. A more sophisticated control of
the Boltzmann parameter than Equation (30) would improve SUPPLEMENTARY MATERIAL
its performance.
Overall, DE was superior over the rest of the algorithms and The Supplementary Material for this article can be found
found very competitive designs, which in some cases were even online at: https://www.frontiersin.org/articles/10.3389/fbuil.
better than those reported in the literature. 2019.00113/full#supplementary-material

REFERENCES Coello, C. A., and Christiansen, A. D. (2000). Multiobjective optimization

of trusses using Genetic Algorithms. Comput. Struct. 75, 647–660.
Adeli, H., and Kumar, S. (1995). Distributed Genetic Algorithm doi: 10.1016/S0045-7949(99)00110-8
for structural optimization. J. Aerosp. Eng. 8, 156–163. Coster, J. E., and Stander, N. (1996). Structural optimization using augmented
doi: 10.1061/(ASCE)0893-1321(1995)8:3(156) Lagrangian methods with secant Hessian updating. Struct. Optim. 12, 113–119.
Balling, R. J. (1991). Optimal steel frame design by Simulated Annealing. J. Struct. doi: 10.1007/BF01196943
Eng. 117, 1780–1795. doi: 10.1061/(ASCE)0733-9445(1991)117:6(1780) Dede, T., and Ayvaz, Y. (2015). Combined size and shape optimization of
Bellman, R. (1957). Dynamic Programming. Princeton, NJ: Princeton structures with a new meta-heuristic algorithm. Appl. Soft Comput. 28,
University Press. 250–258. doi: 10.1016/j.asoc.2014.12.007
Berke, L., and Khot, N. S. (1987). “Structural optimization using optimality Degertekin, S. O. (2012). Improved harmony search algorithms for sizing
criteria,” in Computer Aided Optimal Design: Structural and Mechanical optimization of truss structures. Comput. Struct. 92–93, 229–241.
Systems, eds E. Atrek, R. H. Gallagher, K. M. Ragsdell, and O. C. Zienkiewicz doi: 10.1016/j.compstruc.2011.10.022
(Berlin, Heidelberg: Springer), 271–311. doi: 10.1007/978-3-642-83051-8_7 Degertekin, S. O., and Hayalioglu, M. S. (2013). Sizing truss structures
Bureerat, S., and Pholdee, N. (2016). Optimal truss sizing using an adaptive using teaching-learning-based optimization. Comput. Struct. 119, 177–188.
Differential Evolution algorithm. J. Comput. Civ. Eng. 30: 04015019. doi: 10.1016/j.compstruc.2012.12.011
doi: 10.1061/(ASCE)CP.1943-5487.0000487 Dimou, C. K., and Koumousis, V. K. (2009). Reliability-based optimal design of
Camp, C. V. (2007). Design of space trusses using big bang– truss structures using Particle Swarm Optimization. J. Comput. Civ. Eng. 23,
big crunch optimization. J. Struct. Eng. 133, 999–1008. 100–109. doi: 10.1061/(ASCE)0887-3801(2009)23:2(100)
doi: 10.1061/(ASCE)0733-9445(2007)133:7(999) Eiben, A. E., and Smith, J. E. (2003). Introduction to Evolutionary Computing. New
Camp, C. V., and Farshchin, M. (2014). Design of space trusses using York: Springer. doi: 10.1007/978-3-662-05094-1
modified teaching–learning based optimization. Eng. Struct. 62–63, 87–97. Erol, O. K., and Eksin, I. (2006). A new optimization method: Big Bang–
doi: 10.1016/j.engstruct.2014.01.020 Big crunch. Adv. Eng. Softw. 37, 106–111. doi: 10.1016/j.advengsoft.2005.
Charalampakis, A. E. (2016). “Comparison of metaheuristic algorithms for size 04.005
optimization of trusses,” in 11th HSTAM International Congress on Mechanics Feury, C., and Geradin, M. (1978). Optimality criteria and mathematical
(Athens), 27–30. programming in structural weight optimization. Comput. Struct. 8, 7–17.
Charalampakis, A. E., and Chatzigiannelis, I. (2018). Analytical solutions for the doi: 10.1016/0045-7949(78)90155-4
minimum weight design of trusses by cylindrical algebraic decomposition. Fourie, P. C., and Groenwold, A. A. (2002). The Particle Swarm Optimization
Arch. Appl. Mech. 88, 39–49. doi: 10.1007/s00419-017-1271-8 algorithm in size and shape optimization. Struct. Multidiscip. Optim. 23,
Charalampakis, A. E., and Dimou, C. K. (2010). Identification of Bouc–Wen 259–267. doi: 10.1007/s00158-002-0188-0
hysteretic systems using Particle Swarm Optimization. Comput. Struct. 88, Geem, Z. W., Kim, J. H., and Loganathan, G. V. (2001). A new heuristic
1197–1205. doi: 10.1016/j.compstruc.2010.06.009 optimization algorithm: harmony search. Simulation 76, 60–68.
Charalampakis, A. E., and Dimou, C. K. (2015). Comparison of evolutionary doi: 10.1177/003754970107600201
algorithms for the identification of Bouc-Wen hysteretic systems. J. Comput. Hasançebi, O., and Azad, S. K. (2015). Adaptive dimensional search: a new
Civ. Eng. 29:04014053. doi: 10.1061/(ASCE)CP.1943-5487.0000348 metaheuristic algorithm for discrete truss sizing optimization. Comput. Struct.
Charalampakis, A. E., and Koumousis, V. K. (2008). Identification of Bouc– 154, 1–16. doi: 10.1016/j.compstruc.2015.03.014
Wen hysteretic systems by a hybrid evolutionary algorithm. J. Sound Vib. 314, Hasançebi, O., Çarbaş, S., Dogan, E., Erdal, F., and Saka, M. P. (2009).
571–585. doi: 10.1016/j.jsv.2008.01.018 Performance evaluation of metaheuristic search techniques in the optimum
Clerc, M., and Kennedy, J. (2002). The particle swarm - explosion, stability, and design of real size pin jointed structures. Comput. Struct. 87, 284–302.
convergence in a multidimensional complex space. IEEE Trans. Evol. Comput. doi: 10.1016/j.compstruc.2009.01.002
6, 58–73. doi: 10.1109/4235.985692 Holland, J. H. (1975). Adaptation in Natural and Artificial Systems. Ann Arbor, MI:
Coello, C. A. (2002). Theoretical and numerical constraint-handling University of Michigan Press.
techniques used with evolutionary algorithms: a survey of the state Karaboga, D., and Basturk, B. (2008). On the performance of Artificial
of the art. Comput. Methods Appl. Mech. Eng. 191, 1245–1287. Bee Colony (ABC) algorithm. Appl. Soft Comput. 8, 687–697.
doi: 10.1016/S0045-7825(01)00323-1 doi: 10.1016/j.asoc.2007.05.007

Frontiers in Built Environment | www.frontiersin.org 16 October 2019 | Volume 5 | Article 113

Charalampakis and Tsiatas Critical Evaluation of Metaheuristic Algorithms

Kaveh, A., Bakhshpoori, T., and Afshari, E. (2014a). An efficient hybrid particle Perez, R. E., and Behdinan, K. (2007). Particle swarm approach for
swarm and swallow swarm optimization algorithm. Comput. Struct. 143, 40–59. structural design optimization. Comput. Struct. 85, 1579–1588.
doi: 10.1016/j.compstruc.2014.07.012 doi: 10.1016/j.compstruc.2006.10.013
Kaveh, A., and Ghazaan, M. I. (2015). A comparative study of CBO and Press, W. H., Teukolsky, S. A., Vetterling, W. T., and Flannery, B. P. (2002).
ECBO for optimal design of skeletal structures. Comput. Struct. 153, 137–147. Numerical Recipes in C++: the Art of Scientific Computing. Cambridge:
doi: 10.1016/j.compstruc.2015.02.028 Cambridge University Press.
Kaveh, A., and Khayatazad, M. (2013). Ray optimization for size and Price, K. V., Storn, R. M., and Lampinen, J. A. (2005). Differential Evolution : a
shape optimization of truss structures. Comput. Struct. 117, 82–94. Practical Approach to Global Optimization. Berlin; Heidelberg: Springer-Verlag.
doi: 10.1016/j.compstruc.2012.12.010 Qin, A. K., Huang, V. L., and Suganthan, P. N. (2009). Differential Evolution
Kaveh, A., and Mahdavi, V. R. (2015a). Colliding bodies optimization for size algorithm with strategy adaptation for global numerical optimization. IEEE
and topology optimization of truss structures. Struct. Eng. Mech. 53, 847–865. Trans. Evol. Comput. 13, 398–417. doi: 10.1109/TEVC.2008.927706
doi: 10.12989/sem.2015.53.5.847 Rajan, S. D. (1995). Sizing, shape, and topology design optimization
Kaveh, A., and Mahdavi, V. R. (2015b). Two-dimensional colliding bodies of trusses using Genetic Algorithm. J. Struct. Eng. 121, 1480–1487.
algorithm for optimal design of truss structures. Adv. Eng. Softw. 83, 70–79. doi: 10.1061/(ASCE)0733-9445(1995)121:10(1480)
doi: 10.1016/j.advengsoft.2015.01.007 Rao, R. V., Savsani, V. J., and Vakharia, D. P. (2011). Teaching–learning-based
Kaveh, A., Sheikholeslami, R., Talatahari, S., and Keshvari-Ilkhichi, M. (2014b). optimization: a novel method for constrained mechanical design optimization
Chaotic swarming of particles: a new method for size optimization of truss problems. Comput. Des. 43, 303–315. doi: 10.1016/j.cad.2010.12.015
structures. Adv. Eng. Softw. 67, 136–147. doi: 10.1016/j.advengsoft.2013.09.006 Ray, T., and Saini, P. (2001). Engineering design optimization using a swarm with
Kaveh, A., and Talatahari, S. (2009a). Particle swarm optimizer, ant colony strategy an intelligent information sharing among individuals. Eng. Optim. 33, 735–748.
and harmony search scheme hybridized for optimization of truss structures. doi: 10.1080/03052150108940941
Comput. Struct. 87, 267–283. doi: 10.1016/j.compstruc.2009.01.003 Rönkkönen, J., Kukkonen, S., and Price, K. V. (2005). “Real-
Kaveh, A., and Talatahari, S. (2009b). Size optimization of space trusses parameter optimization with Differential Evolution,” in 2005 IEEE
using Big Bang–Big Crunch algorithm. Comput. Struct. 87, 1129–1140. Congress on Evolutionary Computation (Edinburgh: IEEE), 506–513.
doi: 10.1016/j.compstruc.2009.04.011 doi: 10.1109/CEC.2005.1554725
Kaveh, A., and Talatahari, S. (2010). Optimal design of skeletal structures via Schutte, J. F., and Groenwold, A. A. (2003). Sizing design of truss
the charged system search algorithm. Struct. Multidiscip. Optim. 41, 893–911. structures using particle swarms. Struct. Multidiscip. Optim. 25, 261–269.
doi: 10.1007/s00158-009-0462-5 doi: 10.1007/s00158-003-0316-5
Kaveh, A., and Zakian, P. (2014). Enhanced bat algorithm for optimal design Sonmez, M. (2011). Artificial Bee Colony algorithm for optimization of truss
of skeletal structures. Asian J. Civ. Eng. 15, 179–212. Available online at: structures. Appl. Soft Comput. 11, 2406–2418. doi: 10.1016/j.asoc.2010.09.003
https://ajce.bhrc.ac.ir/Portals/25/PropertyAgent/2905/Files/5878/179.pdf Storn, R. M., and Price, K. V. (1997). Differential Evolution – a simple and efficient
Kazemzadeh Azad, S., and Hasançebi, O. (2015). Computationally efficient discrete heuristic for global optimization over continuous spaces. J. Glob. Optim. 11,
sizing of steel frames via guided stochastic search heuristic. Comput. Struct. 156, 341–359. doi: 10.1023/A:1008202821328
12–28. doi: 10.1016/j.compstruc.2015.04.009 Talatahari, S., Kheirollahi, M., Farahmandpour, C., and Gandomi, A. H.
Keane, A. J. (1996). “A brief comparison of some evolutionary optimization (2013). A multi-stage particle swarm for optimum design of truss
methods,” in Modern Heuristic Search Methods, eds. V. J. Rayward-Smith, I. structures. Neural Comput. Appl. 23, 1297–1309. doi: 10.1007/s00521-012-
H. Osman, C. R. Reeves, and G. D. Smith (Chichester: John Wiley), 255–272. 1072-5
Kennedy, J., and Eberhart, R. C. (1995). “Particle swarm optimization,” in IEEE Wilke, D. N., Kok, S., and Groenwold, A. A. (2007). Comparison of linear
International Conference on Neural Networks IV (Piscataway, NJ: IEEE Press), and classical velocity update rules in Particle Swarm Optimization: notes
1942–1948. doi: 10.1109/ICNN.1995.488968 on diversity. Int. J. Numer. Methods Eng. 70, 962–984. doi: 10.1002/nm
Kirkpatrick, S., Gelatt, C. D., and Vecchi, M. P. (1983). Optimization by Simulated e.1867
Annealing. Sci. New Ser. 220, 671–680. doi: 10.1126/science.220.4598.671 Wu, C.-Y., and Tseng, K.-Y. (2010). Truss structure optimization using
Koumousis, V. K., and Georgiou, P. G. (1994). Genetic Algorithms in adaptive multi-population Differential Evolution. Struct. Multidiscip. Optim.
discrete optimization of steel truss roofs. J. Comput. Civ. Eng. 8, 309–325. 42, 575–590. doi: 10.1007/s00158-010-0507-9
doi: 10.1061/(ASCE)0887-3801(1994)8:3(309) Yang, Z., Tang, K., and Yao, X. (2008). “Self-adaptive Differential Evolution with
Lamberti, L. (2008). An efficient Simulated Annealing algorithm for neighborhood search,” in 2008 IEEE Congress on Evolutionary Computation
design optimization of truss structures. Comput. Struct. 86, 1936–1953. (IEEE World Congress on Computational Intelligence) (Hong Kong: IEEE),
doi: 10.1016/j.compstruc.2008.02.004 1110–1116. doi: 10.1109/CEC.2008.4630935
Lamberti, L., and Pappalettere, C. (2009). “An improved harmony-search Zhang, J., and Sanderson, A. C. (2009). JADE: adaptive Differential Evolution
algorithm for truss structure optimization,” in 12th International Conference with optional external archive. IEEE Trans. Evol. Comput. 13, 945–958.
Civil Structural and Environmental Engineering Computing, ed. B. H. V. doi: 10.1109/TEVC.2009.2014613
Topping (Stirlingshire: Civil-Comp Press).
Lee, K. S., and Geem, Z. W. (2004). A new structural optimization method Conflict of Interest: The authors declare that the research was conducted in the
based on the harmony search algorithm. Comput. Struct. 82, 781–798. absence of any commercial or financial relationships that could be construed as a
doi: 10.1016/j.compstruc.2004.01.002 potential conflict of interest.
Li, L. J., Huang, Z. B., Liu, F., and Wu, Q. H. (2007). A heuristic
particle swarm optimizer for optimization of pin connected Copyright © 2019 Charalampakis and Tsiatas. This is an open-access article
structures. Comput. Struct. 85, 340–349. doi: 10.1016/j.compstruc.2006. distributed under the terms of the Creative Commons Attribution License (CC BY).
11.020 The use, distribution or reproduction in other forums is permitted, provided the
Marano, G. C., Quaranta, G., and Monti, G. (2011). Modified Genetic original author(s) and the copyright owner(s) are credited and that the original
Algorithm for the dynamic identification of structural systems using publication in this journal is cited, in accordance with accepted academic practice.
incomplete measurements. Comput. Civ. Infrastruct. Eng. 26, 92–110. No use, distribution or reproduction is permitted which does not comply with these
doi: 10.1111/j.1467-8667.2010.00659.x terms.

Frontiers in Built Environment | www.frontiersin.org 17 October 2019 | Volume 5 | Article 113

View publication stats