applied

sciences
Article
Optimization of Reinforced Concrete Sections under
Compression and Biaxial Bending by Using a Parallel
Firefly Algorithm
Gregorio Sánchez-Olivares * and Antonio Tomás

Department of Mining and Civil Engineering, Universidad Politécnica de Cartagena (UPCT), Paseo Alfonso XIII
52, 30203 Cartagena, Murcia, Spain; [email protected]
* Correspondence: [email protected]; Tel.: +34-968-325-927

Abstract: A simple formulation for the optimal design of reinforced concrete sections under com-
pression and biaxial bending was established in a previous work by the authors. In that work, it was
found that the formulation produced satisfactory results when used together with three algorithms
that belong to the nature-inspired meta-heuristic algorithm group. However, despite the favorable
results obtained, the necessary calculation times were extensive in all the cases. In order to solve
this problem, the authors implemented a parallel calculation strategy in the algorithm that gave
better results in the previous work. It was possible to verify, through two examples, that this strategy
reduces calculation times as more processes are used in parallel, and that the adjustments made in
the algorithm favor reaching designs close to the global optimum independently of the number of
parallel processes adopted.






Keywords: optimization; parallel algorithm; meta-heuristic methods; reinforced concrete sections;
Citation: Sánchez-Olivares, G.;
Tomás, A. Optimization of Reinforced
biaxial loading
Concrete Sections under
Compression and Biaxial Bending by
Using a Parallel Firefly Algorithm.
Appl. Sci. 2021, 11, 2076. https:// 1. Introduction
doi.org/10.3390/app11052076 In recent years, meta-heuristic algorithms have been developed to solve not only
simple engineering problems but also complex ones that present nonlinear objective func-
Academic Editor: Panagiotis tions and numerous nonlinear constraints. Biologically-inspired algorithms belong to the
G. Asteris
nature-inspired meta-heuristic algorithm group. Those algorithms can be divided into two
categories: evolution-based and swarm-based methods [1]. The best-known and most used
Received: 5 February 2021
evolution-based algorithms are Genetic Algorithms (GA) [2,3]. The most popular of the
Accepted: 21 February 2021
Swarm-based methods is Particle Swarm Optimization (PSO), which is inspired by the
Published: 26 February 2021
group behavior of flocks of birds in flight [4].
Physics-based meta-heuristic algorithms, another type of nature-inspired algorithm,
Publisher’s Note: MDPI stays neutral
are based on the physical laws of the universe. The most popular of these is Simulated
with regard to jurisdictional claims in
Annealing (SA) [5]. Besides these algorithms, there are human-based algorithms, the most
published maps and institutional affil-
iations.
popular of which is Harmony Search (HS) [6].
At the same time as these new meta-heuristic algorithms were developed, a number of
“no free lunch” (NFL) theorems were presented [7]. These NFL theorems establish that for
any algorithm, any elevated performance in one type of problem is offset by performance
in another type. Therefore, after a period of proposing new algorithms, some of them
Copyright: © 2021 by the authors.
began to be applied and fitted to specific engineering problems. The optimum design of
Licensee MDPI, Basel, Switzerland.
concrete structures, which has been widely used in building [8], road [9], rail [10], and
This article is an open access article
marine [11] engineering, is one of these specific problems.
distributed under the terms and
conditions of the Creative Commons
In parallel to this period dedicated to the development and application of new algo-
Attribution (CC BY) license (https://
rithms, studies have been oriented toward detecting differences in their performance and
creativecommons.org/licenses/by/ some improvements have been proposed. PSO has been shown to be more efficient than
4.0/). GAs [12,13]. The work by Valvano et al. [14] is oriented toward the development of a new,

Appl. Sci. 2021, 11, 2076. https://doi.org/10.3390/app11052076 https://www.mdpi.com/journal/applsci

Appl. Sci. 2021, 11, 2076 2 of 22

accurate, and efficient Decline Population Swarm Optimization (PD SO) algorithm. This
new method, applied to vibro-acoustic problems, has resulted in increased performance of
the standard PSO. Moreover, the Firefly Algorithm (FA), which is a swarm-based method,
has been shown to be more efficient than GAs and PSO [15]. The work by Gandomi
et al. [16] shows how the use of different chaotic systems to replace the parameters of the
FA improves the efficiency of the algorithm. It is also necessary to mention the work by
Fister et al. [17], which shows the widespread use of the FA in various fields of engineering.
The works by Gandomi et al. [18,19] and Talatahari et al. [20] are particularly interesting.
In the case of complex structural systems, the number of variables and constraints
increases considerably, so the chosen methodology must be based on efficient algorithms
in order to reduce the computational cost [21,22]. However, less effort has been made to
solve structural system problems where elements subject to biaxial bending appear. It is
necessary to keep in mind that under biaxial bending, the computational cost increases
if the equivalent rectangular compressive stress block is not used for calculating stresses
in the compressed zone of the section [23]. Thus, the need to reduce costs is increasing
constantly, and the use of efficient algorithms, such as the FA, in these types of problems, is
essential [24]. Many recent methods are efficient, although not all of them respond well to
real problems with constraints [22].
Meta-heuristic algorithms need a long time to achieve a final design. Parallel com-
puting strategies sharing the computation load over several processors have started to
be applied to reduce this computational time. Thierauf and Cai [25] present a parallel-
evolution method for structural optimization. Firstly, the problem is divided into two
subproblems; one with discrete design variables and the other with continuous ones. Then,
each subproblem is solved using a parallel evolution approach. Leite and Topping [26] re-
view and evaluate different parallel schemes applied to an SA algorithm, where the design
dominium is complex and very constrained, and the evaluation of the objective functions
can result in medium to high computational costs. In the conclusion, the authors warn that
parallelization seems to be the only general strategy able to reduce time and open its appli-
cability to engineering design. Hasanҫebi et al. [27] address an evolutionary meta-heuristic
algorithm using an integrated parallelization technique for the structural optimization
of high-rise steel buildings. The parallelization is based on a master–slave model that
provides optimal solutions with a reduction of time and without a lack of accuracy. Truong
et al. [28] present an approach integrating GAs and OpenMP applied to steel structures
with semi-rigid connections that produce a clear reduction in computational time.
The previous formulation proposed by the authors [29] for the optimal design of
reinforced concrete sections under compression and biaxial bending includes design con-
siderations in accordance with the Eurocode 2 (EC2) [30] or ACI318 [31] standards and can
be implemented in any meta-heuristic algorithm. In this paper, the authors implemented
and checked a parallel FA, where the formulation is considered, with the aim of reducing
computational time in the process. The two problems solved in the previous work by the
authors [29] now show not only that the parallel strategy reduces the time needed to arrive
at a solution but also produces good results regardless of the number of processors used.

2. Optimization of Concrete Rectangular Cross-Sections

2.1. Optimum Design Problem
The problem of calculating the optimum geometry and reinforcement in a rectangu-
lar concrete cross-section subjected to biaxial bending and axial force may be generally
stated as:
To find the design variable vector

X( X1 , X2 , . . . , Xnv ) (1)

to minimize the objective function

f (X) (2)
 𝑿(𝑋1 , 𝑋2 , … , 𝑋𝑛𝑣 ) (1)
to minimize the objective function

𝑓(𝑿) (2)
Appl. Sci. 2021, 11, 2076 3 of 22
subject to the constraints
𝑔𝑐 (𝑿) ≥ 0; 𝑐 = 1,2, … , 𝑛𝑐 (3)
subject to the constraints
g𝑋c𝑣𝐿(X
≤) 𝑋≥𝑣 0; 𝑈
≤ c𝑋𝑣= ;𝑣
1, =
2, .1,2,
..,…nc, 𝑛𝑣 (4)
(3)
where 𝑿 is the 𝑛𝑣-dimensional XvL ≤design
Xv ≤ X U
variable
v ; v = 1, vector; 𝑓(𝑿) is the objective function;
2, . . . , nv (4)
𝑔𝑐 (𝑿) is the inequality constraint 𝑐; 𝑋𝑣 is the lower bound for the variable 𝑣 and 𝑋𝑣𝑈 is
𝐿
where X is the nv-dimensional design variable vector; f (X) is the objective function; gc (X)
the upper bound for the variable 𝑣; 𝑛𝑐 is the total number of constraints, and 𝑛𝑣 is the
is the inequality constraint c; XvL is the lower bound for the variable v and XvU is the
number of variables.
upper bound for the variable v; nc is the total number of constraints, and nv is the number
The FA reproduces the social behavior of fireflies in a simple and idealized way. Fire-
of variables.
flies communicate with each other, seek prey, and find a partner using different patterns
The FA reproduces the social behavior of fireflies in a simple and idealized way.
of bioluminescent flashes. The characteristics of these flashes are idealized to achieve the
Fireflies communicate with each other, seek prey, and find a partner using different patterns
development of this algorithm. In a similar way to how real fireflies fly through three-
of bioluminescent flashes. The characteristics of these flashes are idealized to achieve the
dimensional space, modifying their spatial coordinates, the fireflies of the algorithm, or
development of this algorithm. In a similar way to how real fireflies fly through three-
individuals, now fly through 𝑛𝑣-dimensional space where the coordinates are now the
dimensional space, modifying their spatial coordinates, the fireflies of the algorithm, or
variables of the
individuals, nowdesign problem
fly through (1). The best space
nv-dimensional designwherecorresponds to an individual
the coordinates are now that
the
flashes more
variables than
of the the others
design problem because it is best
(1). The in a design
positioncorresponds
that minimizesto ana individual
function, called
that
the fitness
flashes more function.
than theThis fitness
others function
because takes
it is in into account
a position the objective
that minimizes function
a function, (2) and
called the
the inequality constraints (3) that are violated
fitness function. This fitness function takes into account (𝑔 𝑐 (𝑿) < 0) together. During the process,
the objective function (2) and the
the algorithm ensures (3)
thatthat
the are
variables are(gwithin the bounds (4) that define the design
inequality constraints violated c (X) < 0) together. During the process, the
space.
algorithm ensures that the variables are within the bounds (4) that define the design space.

2.2. Variables
There are
There aretwelve
twelvedesign
designvariables
variablesthat thatwere
wereconsidered
considered forfor
eacheach individual
individual i (Figure
i (Figure 1):
1): the
the depth
depth of neutral
of neutral axisaxis zi; the
zi ; the angle angle
of the of the neutral
neutral fiberfiber θi; width
θ i ; the the width bi; height
bi ; the the heighthi ; theh i;
the diameter
bar bar diameter of theofright
the right side reinforcement
side reinforcement φ'y,inumber
ϕ’y,i ; the ; the number of barsofon barstheon theside
right rightofsidethe
of the section
section n'bar
n’y,i ; the y,i; the bar diameter
diameter of the leftof the
sideleft side reinforcement
reinforcement ϕy,i ; theφnumber
y,i; the number
of bars of onbars the
left
on thesideleft
of the of the n
sidesection ;
section
y,i the bar
n y,i ; diameter
the bar of the
diameter topof side
the reinforcement
top side reinforcement
ϕ’ x,i ; the number
φ' x,i ; the
of bars onofthe
number topon
bars side
theoftop
the side
section of n’x,i ;section
the the bar n' diameter
x,i; the barof the bottom of
diameter sidethereinforcement
bottom side
ϕx,i ; and the number
reinforcement φx,i; and of the
barsnumber
on the bottom
of bars on sidethe ofbottom
the sectionsidenof x,i .the section nx,i.

Figure 1. Cross-section geometry and design variables for the individual i.

The design variables zi and θ i are continuous, and the rest of them are discrete. The
designer can define limits for these variables. The variables bi and hi adopt values of 5
by 5 mm. The variables n’y,i , ny,i , n’x,i and nx,i take integer values. Finally, the variables
ϕ’y,i , ϕy,i , ϕ’x,i and ϕx,i adopt values corresponding to commercial diameters of reinforcing
steel bars.
Appl. Sci. 2021, 11, 2076 4 of 22

The proposed formulation considers optimization variables instead of design variables.

The design variables θ i , bi , hi , ϕ’y,i , n’y,i , ϕy,i , ny,i , ϕ’x,i , n’x,i , ϕx,i , nx,i were normalized by
dividing them by their upper bounds in order to obtain the optimization variables: ζ i , βi ,
η i , ω’y,i , ν’y,i , ω y,i , νy,i , ω’x,i , ν’x,i , ω x,i , νx,i , respectively. Instead of the design variable zi, the
optimization variable ξ i is considered as follows:

2
ξi = arctan(zi ) (5)
π
Thereby, the optimization variables adopt values within the same order of magnitude.
This ensures that the variables will have the same relative importance when used in the
design algorithm, thus favouring efficiency.

2.3. Fitness Function

The fitness function proposed is as follows:

!
f ( Xi ) mvi
f ( Xi ) f XU
F ( Xi ) = U

− ∑ g j ( Xi )
+ (6)
f X j =1 f XU f ( Xi )

where Xi = (ξ i , ζ i , βi , η i , ω’y,i , ν’y,i , ω y,i , νy,i , ω’x,i , ν’x,i , ω x,i , νx,i ) is the 12-dimensional
optimization variable vector for the individual i; XU is the 12-dimensional optimization
variable upper bound vector; f (Xi ) is the objective function to be minimized; and mvi is
the number of violated constraints (g j < 0) for the individual i.

2.4. Objective Function

The objective function fi of the optimization problem for the individual i is the cost of
RC cross-section per unit length

f (Xi ) = f i = Pc Ac,i + Ps ρs Ast,i + Pf L f ,i (7)

where Pc is the price of concrete per unit volume; Ps is the price of steel per kilogram; ρs
is the density of steel and Pf is the price of formwork per unit area; Ac,i is the area of the
concrete cross-section for the individual i; Ast,i is the total area of steel reinforcement for
the individual i; and L f ,i is the perimeter of form for the individual i.

2.5. Constraints
2.5.1. Reinforcement Constraints
The normalized reinforcement constraints are the following:

As,i
g As,i = −1 ≥ 0 (8)
As,min,i

A0s,i
g A0s,i = −1 ≥ 0 (9)
A0s,min,i
Ast,i
g Ast,i = −1 ≥ 0 (10)
Ast,min,i
where As ,i is the amount of steel in the tension reinforcement for the individual i; Ast, min,i
is the minimum amount of steel in the tension reinforcement allowed for the individual i;
A’s ,i is the amount of steel in the compression reinforcement for the individual i; A’s ,min,i is
the minimum amount of steel in the compression reinforcement allowed for the individual
i; Ast ,i is the total area of reinforcement for the individual i; and Ast ,min,i is the minimum
total area of reinforcement allowed for the individual i.
Appl. Sci. 2021, 11, 2076 5 of 22

2.5.2. Ductility Constraint

The normalized form of the ductility constraint for the individual i is
zi
gduct,i = 1 − ≥0 (11)
zmax,i

where zmax,I is the maximum neutral axis depth for the individual i.

2.5.3. Steel Reinforcement Spacing Constraints

The normalized form of the spacing constraints for the individual i are
si
gsp,min,i = −1 ≥ 0 (12)
smin,i
si
gsp,max,i = 1 − ≥0 (13)
smax,i
where si is the spacing between the steel bars of reinforcement for the individual i; smin,i is
the minimum spacing for the individual i; and smax,i is the maximum spacing closest to the
tension faces for the individual i.

2.5.4. Strength Constraints

Two strength constraints are defined for the individual i to ensure that the cross-section
withstands the design loads. The first constraint assures sufficient strength against the
combined flexure and axial load, the second one ensures that the load eccentricity is the
same as the strength eccentricity of the cross-section. The first constraint in normalized
form for the individual i is:
q
Pu2 + M2x,u + My,u
2
gs1,i = 1− q =
ϕ(Xi ) Pn (Xi )2 + Mx,n (Xi )2 + My,n (Xi )2
q (14)
Pu2 + M2x,u + My,u
2
1− q
2 + M2 2
≥ 0 (ACI 318)
ϕi Pn,i x,n,i + My,n,i

or
q
2 + M2
NEd 2
x,Ed + My,Ed
gs1,i = 1− q =
NRd (Xi )2 + Mx,Rd (Xi )2 + My,Rd (Xi )2
q (15)
2 + M2
NEd 2
x,Ed + My,Ed
1− q
2 2 2
≥ 0 (EC 2)
NRd,i + Mx,Rd,i + My,Rd,i

where Pu is the factored axial force; Mx,u is the factored moment about x-axis at cross-section;
My,u is the factored moment about y-axis at cross-section; ϕI is the strength reduction
factor for the individual i; Pn,i is the nominal axial load normal to the cross-section for the
individual i; Mx,n,i is the nominal moment about x-axis at the cross-section for the individual
i; My,n,i is the nominal moment about y-axis at the cross-section for the individual i; Mx,Ed
is the design value of the applied internal flexural moment about x-axis at the cross-section;
My,Ed is the design value of the applied internal flexural moment about y-axis at the cross-
section; NRd,I is the design axial resistance of the cross-section for the individual i; Mx,Rd,i is
the design moment resistance of the cross-section about x-axis for the individual i; My,Rd,i
is the design moment resistance of the cross-section about y-axis for the individual i.
The second constraint in normalized form for the individual i is
1 cosψi
(1 − cosψi )
   
q cosψi C1 ∈ [1, 2]
gs2,i =− ;q = e C1
− ; (16)
2 C2 C2 ∈ [5, 50]

with
Appl. Sci. 2021, 11, 2076 6 of 22

Pu Pn,i
cosψi = q q
2 + M2 2
Pu2 + M2x,u + My,u
2 Pn,i x,n,i + My,n,i

Mx,u Mx,n,i
+q q
2 + M2 2
(17)
Pu2 + M2x,u + My,u
2 Pn,i x,n,i + My,n,i
My,u My,n,i
+q q
2 + M2 2
(ACI 318)
Pu2 + M2x,u + My,u
2 Pn,i x,n,i + My,n,i

or with
NEd NRd,i
cosψi = q
2 + M2 2
q
2 + M2 2
NEd x,Ed + My,Ed NRd,i x,Rd,i + My,Rd,i

Mx,Ed Mx,Rd,i
+q 2 + M2 2
q
2 + M2 2
(18)
NEd x,Ed + My,Ed NRd,i x,Rd,i + My,Rd,i
My,Ed My,Rd,i
+q 2 q (EC 2)
NEd + M2x,Ed + My,Ed
2 2 + M2
NRd,i 2
x,Rd,i + My,Rd,i

2.6. Optimization Methodology

2.6.1. Firefly Algorithm (FA)
Highly non-linear optimization problems are common in structural design. Most of
them include many variables that can be discrete and/or continuous, and complex, non-
linear constraints [24]. Among the advantages of the FA is that it efficiently overcomes these
problems [1,32]. In the previous formulation, a sample of non-linear functions (objective
function, fitness function, and constraints) was proposed.
The FA is a swarm-based method whose search process depends on two combined
strategies: exploration and exploitation. The former focuses on searching for the global
optimum, whereas the latter focuses on selecting the best design found thus far. During
the process of design, the parameters of the algorithm allocate more weight to one of the
strategies than to the other.
The FA ideally reproduces the behavior of fireflies. Three rules are considered:
(i) All the fireflies in a population have just one gender and any of them can be attracted
to another.
(ii) The attraction between two fireflies in the whole population is directly proportional to
the brightness of their luminescence. This attraction lessens when distances increase.
Observing a pair of fireflies, the less bright one moves toward the brighter one.
(iii) The brightness of a specific firefly is linked to the value of its fitness function.
The attraction β between two fireflies is

1
β (r ) = β 0 (19)
1 + γr m

where γ is the absorption coefficient for a given medium; m is an integer; r is the distance
between these two fireflies; and β0 is the attraction at r = 0.
The movement of a firefly i, which is attracted to another, brighter firefly j, is expressed
by the following expression:
 
k +1 k

k k
 1 1  U 
pi = pi + p j − pi β 0 m + α rand − X − XL (20)
1 + γrij 2

where k is the current iteration; pik is the spatial coordinate vector of the i-th firefly at the
k-th iteration; pkj is the spatial coordinate vector of the j-th firefly at the k-th iteration; α is
the randomization parameter; rand is a random number generator uniformly distributed in
[0, 1]; X L is the optimization variable lower bound vector; and rij is the Cartesian distance

rij = kpik − pkj k (21)

Appl. Sci. 2021, 11, 2076 7 of 22

The potential oscillatory behavior of the algorithm can be avoided by reducing the
randomization parameter α as the process progresses. Research by Gandomi et al. [18] has
established that α ∈ [0.01 , 1] can be taken.
The absorption coefficient γ controls the speed of the process convergence. If γ is
near zero, the attraction is constant, and each firefly can be seen by the whole population.
However, if γ is large, the attraction decreases and the fireflies become almost blind. This
situation is as if they were flying in the fog, with almost random movement.

2.6.2. Modified Version (MPFA) of the Firefly Algorithm

(i) Parallelization and migration.
The authors included a parallel computing strategy based on dividing the whole
population into subpopulations. Each subpopulation searches for the optimum design
of the best solution found in the subpopulation up to that point, taking exploration and
exploitation into account. A code was implemented in MATLAB® (see the Appendix A) to
address the parallel strategy using the parfor MATLAB® command to distribute the work
among processors [33].
Along with the parallel computing strategy, another strategy is needed to enable the
firefly subpopulations to communicate with each other, so that the information that resides
in each subpopulation can be shared by the rest of the subpopulations. In this way, the
authors prevent the parallel computing strategy from merely dividing the population into
subpopulations that seek the optimal design independently.
A way to do this is by migrating fireflies among subpopulations. The migration
process must ensure that information passes among subpopulations without bias, in order
not to force the search in a specific direction. The migration strategy implemented is
based on several steps: (i) the one with the best of all the designs achieved up to that
point is chosen as the origin subpopulation; (ii) individuals from the origin subpopulation
are chosen randomly; (iii) destination subpopulations are chosen randomly, and (iv) the
individuals of the destination subpopulations are substituted for the individuals of the
origin subpopulation.
As the origin subpopulation is the one that has the best design of those found up to that
point in the process of searching for the global optimum, there is a bias in the information
that is transmitted. In fact, if the origin subpopulation is close to a local optimum, then the
movement of the fireflies in that direction is forced by bringing individuals who are close
to that local optimum to other subpopulations. However, this bias does not occur due to
the second modification introduced in the FA, which is as follows.
(ii) Small random displacements.
Two small random displacements on the variables z and θ are developed; one for the
brightest firefly of each subpopulation (local, small, random displacement, LSRD), and an-
other for the brightest firefly of the whole population (global, small, random displacement,
GSRD).  
1  U 
zk+1 = zk + α2 rand − z − zL (22)
2
 
1  U 
θ k+1 = θ k + α2 rand − θ − θL (23)
2
These LSRD and GSRD displacements are essential to improve the performance of the
algorithm as was verified by the authors after running many examples, some of which are
included in this work.
This improvement is due to the fact that LSRD does not only increase the firefly’s
mobility but also that of the whole subpopulation that follows its movements. Additionally,
the brightest firefly of the subpopulation may cease to be so due to this LSRD. Furthermore,
it is possible that the subpopulation considered as the origin is no longer so because the
brightest firefly may now belong to another subpopulation. Finally, GSRD is necessary
when a high number of subpopulations are considered because of the low number of
 Appl. Sci. 2021, 11, 2076 8 of 22

fireflies in each subpopulation, which reduces the effect of LSRD. As a consequence,

021, 11, x FOR PEER REVIEW the combined effect of both LSRD and GSRD helps biases disappear 9 of 22
and exploration to
improve, decreasing the probability of the process ending at a local minimum.
These basic operations were implemented in MATLAB®and are summarized in
Figure 2.

Figure 2. Pseudo-code of modified version (MPFA).

Figure 2. Pseudo-code of modified version (MPFA).
3. Examples
3. Examples 3.1. Cross-Section under Flexure
3.1. Cross-Section underAFlexure
cross-section under a factored flexural moment Mx,u = 400 kNm is studied. The
A cross-section underofamaterials
strength factoredare fc ’ = 30moment
flexural MPa for M concrete
x,u = 400 andkNmfy = 500 MPa forThe
is studied. steel. The clear cover
of reinforcement is 40 mm and the stirrup diameter is 10
strength of materials are fc’ = 30 MPa for concrete and fy = 500 MPa for steel. The clear cover mm. The cost is calculated using a
price of
of reinforcement is 40 mm and P c = 100 €/m 3 for concrete, P = 1.2 €/kg for steel and P = 30 €/m2 for formwork.
the stirrup diameter is 10 s mm. The cost is calculated f using
a price of Pc = 100The€/mlength of formPis
3 for concrete, s =b1.2
+ 2h.€/kgTheforoptimization
steel and Pf =variable
30 €/m2 forξ isformwork.
constrained in the interval
The length of form [−is
0.7,
b +0.7];
2h. the
Theangle of the neutral
optimization variablefiber in [0, π/2] rad;
ξ isθ constrained theinterval
in the width of [- the cross-section
b [0.30, 0.50] m; the height h
0.7, 0.7]; the angle of the neutral fiber θ in [0, π/2] rad; the width of the cross-section b indiameters ϕ’x , ϕx
in in [0.30, 0.90] m; the reinforcement bar
[0.30, 0.50] m; the in [10, 32]
height h inmm;
[0.30,the number
0.90] m; the of reinforcement
reinforcement barbars n’y , ny φ'
diameters inx,[0,
φx 0] [10, n’x , nx in [2, 10].
in and
32] mm; the number of reinforcement bars n'y, ny in [0, 0] and n'x, nx in [2, 10]. The con-this example. The
The constraints (8) to (10), (11), (12) to (15) and (16) are considered in
straints (8) to (10),parameters
(11), (12) toof theand
(15) Equation
(16) are(16) are C1 = in
considered thisCexample.
2 and 2 = 5. The MPFA
The parameters were already
parameters
of the Eq. (16) are C1 = 2 and C2 = 5. The MPFA parameters were already tuned for this
example, as can be seen in [29]. The values adopted in this work are 0 = 1, m = 2.0, γ0 = 10,
α0 = 1, NF = 300, kmax = 1000 and tc = 0.1.
 sidering ACI318 and EC2 standards are shown in Table 1.

Table 1. Optimal design solutions obtained using the MPFA.

Appl. Sci. 2021, 11, 2076 ACI318 9 of 22EC2

z (mm) 152.3 136.3

θ (rad) 0 0
tuned for this example, as can be seen in [29]. The values adopted in this work are β0 = 1,
b (mm)
m = 2.0, γ0 = 10, α0 = 1, NF = 300, kmax = 1000 and tc = 0.1.
300 325
h (mm)obtained from 50 runs using MPFA585
The best solution with eight subpopulations is605
considered toφ' (mm)
bex the optimal design solution. The optimal10design solutions obtained 10
considering ACI318 and EC2 standards are shown in Table 1.
n'x 2 4
φdesign
Table 1. Optimal x (mm)solutions obtained using the MPFA. 32 20
nx 3 6
ACI318 EC2
Cost (€/m) 85.54 86.19
z(mm) 152.3 136.3
θ(rad) 0 0
The average
b(mm)time of 50 runs, considering
300 ACI318 standard325and using MPFA
h(mm) 585 605
different numbers
ϕ’x (mm)
of subpopulations, was
10
obtained. Although there
10
is a difference in
average time depending
n’x on the computer2 used, the average time 4was considered to w
(mm) 32
out the metric x“speed-up”, which refers to how much a parallel FA
ϕ 20 is faster than a seq
nx 3 6
tial FA [34].Cost
The(€/m)speed-up is shown in 85.54
Figure 3. “Efficiency”, which
86.19 is the ratio of sp
up and the number of processors used to solve the example, measures how efficiently
MPFA uses the parallel
The average time ofresources [34]. Efficiency
50 runs, considering ACI318isstandard
shown andin Figure 4.
using MPFA with
The parallelization produces satisfactory results despite its simple implementa
different numbers of subpopulations, was obtained. Although there is a difference in
the average
(Figure time depending
3). However, on the computer
the efficiency quicklyused,
dropsthewith
average time
more was four
than considered to
processors (Fi
work out the metric “speed-up”, which refers to how much a parallel FA is faster than a
4). The computer that was used to run the calculations is a workstation equipped wit
sequential FA [34]. The speed-up is shown in Figure 3. “Efficiency”, which is the ratio of
Intelspeed-up
Xeon CPU E5-1620
and the number v2 3.70 GHzused
of processors with eightthecores.
to solve example, measures how efficiently
the MPFA uses the parallel resources [34]. Efficiency is shown in Figure 4.

Figure
Figure 3. Cross-section under
3. Cross-section underflexure. Speed-up.
flexure. Speed-up.
Appl. Sci. 2021, 11, x FOR PEER REVIEW 11 of 22
Appl. Sci. 2021, 11, 2076 10 of 22

Figure
Figure 4. Cross-section
4. Cross-section under
under flexure.
flexure. Efficiency.
Efficiency.

The parallelization
“Accuracy”, which is theproduces satisfactory
mean value results
of several despite itsruns
independent simple
whenimplementation
compared
(Figure 3). However, the efficiency quickly drops with more than
to a reference value, measures how efficiently the MPFA uses the parallel resources four processors (Figure
[34]. 4).
The computer that was used to run the calculations is a workstation equipped
“Robustness” is linked to the scatter of obtained results [34] (for example, standard devi- with an Intel
Xeon CPU E5-1620 v2 3.70 GHz
ation) for several independent runs of MPFA. with eight cores.
The“Accuracy”,
mean cost, which
accuracyis the mean value
(reference valueofofseveral
Cost =independent runs when
85.54, from Table compared
1), and robust- to
a reference value, measures how efficiently the MPFA uses the
ness (standard deviation) from 50 runs considering ACI318 standard and using MPFA parallel resources [34]. “Ro-
bustness”
with is linked
a different numberto the scatter of obtained
of subpopulations wereresults [34] (for
obtained. example,
These standard
metrics are showndeviation)
in
for several independent runs of MPFA.
Tables 2–4, considering only LSRD, only GSRD, and both LSRD and GSRD, respectively.
The meanpurposes,
For comparison cost, accuracy (reference
the average valueofofthe
values Cost = 85.54, frommetrics
performance Table 1), and also
were robustness
ob-
(standard deviation) from 50 runs considering
tained, and they are shown in the last row of Tables 2–4. ACI318 standard and using MPFA with a
different number of subpopulations were obtained. These metrics are shown in Tables 2–4,
considering
Table only LSRD,
2. Cross-section only GSRD,
under flexure. Local,and both
small, LSRDdisplacement
random and GSRD,(LSRD).
respectively. For compari-
Performance
son
metrics. purposes, the average values of the performance metrics were also obtained, and they
are shown in the last row of Tables 2–4.
Subpopulations Mean Cost Accuracy Robustness
Table 2. Cross-section
1 under flexure. Local,
85.81 small, random displacement
0.9969 (LSRD). Performance
0.51 metrics.
2 85.83 0.9966 0.52
Subpopulations Mean Cost Accuracy Robustness
3 85.78 0.9972 0.76
1 85.81 0.9969 0.51
4 85.68 0.9984 0.35
2 85.83 0.9966 0.52
53 85.86
85.78 0.9963
0.9972 0.81
0.76
64 85.68
85.68 0.9983
0.9984 0.36
0.35
75 85.86
85.76 0.9963
0.9974 0.81
0.47
6 85.68 0.9983 0.36
8 85.98 0.9949 0.95
7 85.76 0.9974 0.47
98 86.26
85.98 0.9917
0.9949 1.32
0.95
109 85.94
86.26 0.9953
0.9917 0.99
1.32
1110 85.94
86.52 0.9953
0.9887 0.99
1.49
12 11 86.52
86.35 0.9887
0.9906 1.49
1.72
12 86.35 0.9906 1.72
1313 86.84
86.84
0.9851
0.9851
1.90
1.90
1414 87.05
87.05 0.9827
0.9827 1.94
1.94
1515 88.02
88.02 0.9719
0.9719 2.40
2.40
1616 87.98
87.98 0.9723
0.9723 2.68
2.68
Averagedvalue
Averaged value 86.33
86.33 0.9909
0.9909 1.20
1.20
Appl. Sci. 2021, 11, 2076 11 of 22

Table 3. Cross-section under flexure. Global, small, random displacement (GSRD). Performance metrics.

Subpopulations Mean Cost Accuracy Robustness

1 85.81 0.9969 0.51
2 86.08 0.9937 0.66
3 85.93 0.9955 0.62
4 85.73 0.9978 0.43
5 85.74 0.9976 0.45
6 85.74 0.9976 0.46
7 85.90 0.9958 1.04
8 85.76 0.9974 0.47
9 86.18 0.9926 1.96
10 85.79 0.9971 0.75
11 86.04 0.9941 1.08
12 85.80 0.9969 0.51
13 86.22 0.9921 1.22
14 85.89 0.9959 0.82
15 86.66 0.9871 2.22
16 86.33 0.9908 1.34
Averaged value 85.98 0.9949 0.91

Table 4. Cross-section under flexure. LSRD and GSRD. Performance metrics.

Subpopulations Mean Cost Accuracy Robustness

1 85.81 0.9969 0.51
2 85.73 0.9977 0.43
3 85.71 0.9980 0.39
4 85.73 0.9978 0.43
5 85.73 0.9978 0.41
6 85.65 0.9988 0.27
7 85.71 0.9980 0.37
8 85.81 0.9968 0.51
9 85.99 0.9947 1.05
10 86.24 0.9918 1.43
11 85.85 0.9964 0.62
12 85.96 0.9951 0.99
13 86.06 0.9939 0.94
14 86.20 0.9923 1.20
15 86.32 0.9910 1.51
16 86.29 0.9914 1.38
Averaged value 85.93 0.9955 0.78

The method is able to obtain designs close to the global optimum as shown in
Tables 2–4. When the method considers both LSRD and GSRD (Table 4), the minimum
average value of the mean costs, the maximum average value of the accuracies, and the
minimum average value of the standard deviations are obtained.
The performance metrics are also shown in Figures 5–7. It can be observed that the
metrics tend to worsen in all cases when the number of subpopulations increases. This
tendency is greater when LSRD is used alone and above eight subpopulations. The other
cases have a lower tendency. The sensitivity of the metrics with respect to the number of
subpopulations is reduced when LSRD and GSRD are taken into account at the same time.
1, x FOR PEER REVIEW

1, x FOR PEER REVIEW

Appl. Sci. 2021, 11, 2076 12 of 22

Figure 5. Cross-section under flexure. Mean cost.

Figure 5. Cross-section under flexure. Mean cost.
Figure 5. Cross-section under flexure. Mean cost.

Figure 6. Cross-section under flexure. Robustness.

Figure 6. Cross-section under flexure. Robustness.

Figure 6. Cross-section under flexure. Robustness.

x FOR PEER REVIEW

Appl. Sci. 2021, 11, 2076 13 of 22

Figure 7. Cross-section under flexure. Accuracy.

Figure 7. Cross-section under flexure. Accuracy.
3.2. Cross-Section under Biaxial Bending
3.2. Cross-Section under Biaxial
A rectangular Bending
cross-section under biaxial bending [35] using MPFA is studied. The
design loadings considered are Pu = 200 kN, Mx,u = 300 kNm and My,u = 250 kNm. The
A rectangular cross-section
strength of materials under
are fc ’ = 25.5 MPa for biaxial
concrete bending [35]forusing
and fy = 500 MPa MPFA
steel. The clear is st
cover of reinforcement is 40 mm and the stirrup diameter
design loadings considered are3 Pu = 200 kN, Mx,u = 300 kNm and My,u2 = 250 is 10 mm. The cost is calculated
using a price of Pc = 100 €/m for concrete, of Ps = 1.2 €/kg for steel and of Pf = 15 €/m
strength of for
materials
formwork. are fc’ = of
The length 25.5
formMPa
is 2b + for concrete
2h. The optimizationandvariable
fy = 500 MPa for steel
ξ is constrained
in the interval [−is
cover of reinforcement 0.7,40 mm
0.7]; the and the
angle of thestirrup diameter
neutral fiber is 10
θ in [0, π/2] rad;mm. Theofcost is
the width
the cross-section b in [0.30, 0.50] m; the height h in [0.30, 0.90] m; the reinforcement bar
using a price of Pc ϕ’=y ,100
diameters ϕy , ϕ’€/m
3 for concrete, of Ps = 1.2 €/kg for steel and of Pf =
x , ϕx in [10, 32] mm; the number of reinforcement bars n’y , ny in [0, 10]
formwork. andThethelength
number of form is 2bbars
of reinforcement + 2h.n’x , The optimization
nx in [2, variable
10]. The constraints (8) to (10),ξ(12)
is to
constra
(15) and (16) are considered in this example. The parameters of Equation (16) are C1 = 2
interval [-0.7,
and 0.7];
C2 = 5.the angleparameters
The MPFA of the neutral
are number fiber θ inNF[0,
of fireflies π/2]
= 300; rad; the
maximum width of
number
section b inof[0.30, 0.50]
iterations kmax =m; theattractiveness
1000; height h atinr =[0.30, 0.90]
0 β0 = 1; m; the parameter
randomization reinforcement α0 = 1; bar
absorption coefficient γ0 = 10; m = 2.0 and tc = 0.1.
φ'y, φy, φ'x, φx The
in [10, 32] mm;
best solution thefrom
obtained number
50 runs of reinforcement
using bars n'y, nyisin [0, 1
MPFA with eight subpopulations
number of reinforcement
considered to be thebars optimal n'design
x, nx in [2, 10].
solution. TheThe constraints
optimal (8) obtained
design solutions to (10), (12)
considering ACI318 and EC2 standards are shown in Tables 5–8.
(16) are considered in this example. The parameters of Eq. (16) are C1 = 2 and
MPFA parameters are number of fireflies NF = 300; maximum number of itera
1000; attractiveness at r = 0 β0 = 1; randomization parameter α0 = 1; absorption
γ0 = 10; m = 2.0 and tc = 0.1.
The best solution obtained from 50 runs using MPFA with eight subpop
considered to be the optimal design solution. The optimal design solutions obt
sidering ACI318 and EC2 standards are shown in Tables 5–8.

Table 5. Optimal design solutions with invariant section. EC2.

Appl. Sci. 2021, 11, 2076 14 of 22

Table 5. Optimal design solutions with invariant section. EC2.

Case 1 Case 2 Case 3

ϕ’y = ϕy
ϕ’y = ϕy = ϕ’x = ϕx
ϕ’x = ϕx
n’y = ny
n’y = ny
n’x = nx
n’x = nx
Gil-Martín et al. [35] MPFA MPFA MPFA
z (mm) - 251.8 216.3 263.2
θ (rad) - 1.0819 1.2200 1.2041
b (mm) 400 400 400 400
h (mm) 700 700 700 700
ϕ’y (mm) 14.4 20 10 -
n’y 6 5 5 -
ϕy (mm) 14.4 20 10 14
ny 6 5 5 10
ϕ’x (mm) 14.4 20 32 10
n’x 8 2 2 2
ϕx (mm) 14.4 20 32 14
nx 8 2 2 8
Ast (mm2 ) 4560.1 4398.2 4002.4 2928.0
Relative Ast (%) 103.7 100.0 91.0 66.6
Cost (€/m) 103.5 102.0 98.3 88.3
Relative cost (%) 101.5 100.0 96.4 86.6

Table 6. Optimal design solutions with variable section. EC2.

Case 1 Case 2 Case 3

ϕ’y = ϕy
ϕ’y = ϕy = ϕ’x = ϕx
ϕ’x = ϕx
n’y = ny
n’y = ny
n’x = nx
n’x = nx
Gil-Martín et al. [35] MPFA MPFA MPFA
z(mm) - 269.9 250.4 280.3
θ(rad) - 0.8660 1.0035 1.1029
b(mm) - 495 460 465
h(mm) - 575 620 615
ϕ’y (mm) - 12 12 -
n’y - 8 2 -
ϕy (mm) - 12 12 10
ny - 8 2 10
ϕ’x (mm) - 12 32 10
n’x - 10 2 2
ϕx (mm) - 12 32 16
nx - 10 2 9
Ast (mm2 ) - 4070.2 3667.0 2752.0
Relative Ast (%) - 100.0 90.1 67.6
Cost (€/m) - 98.5 95.1 86.6
Relative cost (%) - 100.0 96.5 87.9
Appl. Sci. 2021, 11, 2076 15 of 22

Table 7. Optimal design solutions with invariant section. ACI318.

Case 1 Case 2 Case 3

ϕ’y = ϕy
ϕ’y = ϕy = ϕ’x = ϕx
ϕ’x = ϕx
n’y = ny
n’y = ny
n’x = nx
n’x = nx
Gil-Martín et al. [35] MPFA MPFA MPFA
z(mm) - 222.7 203.2 242.1
θ(rad) - 1.1585 1.2390 1.1597
b(mm) - 400 400 400
h(mm) - 700 700 700
ϕ’y (mm) - 16 12 10
n’y - 9 9 4
ϕy (mm) - 16 12 16
ny - 9 9 10
ϕ’x (mm) - 16 25 12
n’x - 4 3 3
ϕx (mm) - 16 25 16
nx - 4 3 5
Ast (mm2 ) - 5217.6 4978.7 3669.4
Relative Ast (%) - 100.0 95.4 70.3
Cost (€/m) - 109.6 107.4 95.2
Relative cost (%) - 100.0 98.0 86.9

Table 8. Optimal design solutions with variable section. ACI318.

Case 1 Case 2 Case 3

ϕ’y = ϕy
ϕ’y = ϕy = ϕ’x = ϕx
ϕ’x = ϕx
n’y = ny
n’y = ny
n’x = nx
n’x = nx
Gil-Martín et al. [35] MPFA MPFA MPFA
z(mm) - 251.8 243.7 264.6
θ(rad) - 0.8095 0.9426 0.8786
b(mm) - 500 485 485
h(mm) - 565 605 615
ϕ’y (mm) - 14 12 10
n’y - 7 6 4
ϕy (mm) - 14 12 14
ny - 7 6 9
ϕ’x (mm) - 14 25 10
n’x - 8 3 3
ϕx (mm) - 14 25 14
nx - 8 3 7
Ast (mm2 ) - 4618.1 4300.6 3012.8
Relative Ast (%) - 100.0 93.1 65.2
Cost (€/m) - 103.2 102.1 90.9
Relative cost (%) - 100.0 98.9 88.1

The average time of 50 runs, considering ACI318 standard and using MPFA with
different numbers of subpopulations, was obtained. The speed-up and efficiency are shown
in Figures 8 and 9.
Sci. 2021, 11, x FOR PEER REVIEW

ci. 2021, 11, x FOR PEER REVIEW

Appl. Sci. 2021, 11, 2076 16 of 22

Figure 8. Cross-section under biaxial bending. Speed-up.

Figure 8. Cross-section under biaxial bending. Speed-up.
Figure 8. Cross-section under biaxial bending. Speed-up.

Figure 9. Cross-section under biaxial bending. Efficiency.

Figure 9. Cross-section under biaxial bending. Efficiency.
As in the previous example, the parallelization produces satisfactory results (Figure 8),
Figure 9. Cross-section
although
The mean value,under
the efficiency quickly biaxial
drops
accuracy withbending. Efficiency.
more than four
(reference processors (Figure 9).
value of Cost = 90.9, from Tabl
The mean value, accuracy (reference value of Cost = 90.9, from Table 8), and robustness
ness (standard
(standard deviation)
deviation) from 50 runsfrom 50 runs
considering considering
ACI318 ACI318
standard and using MPFA standard
with a a
The mean value, accuracy (reference value of Cost = 90.9, from Table
with a different number of subpopulations were obtained and shown in T
ness (standard deviation) from 50 runs considering ACI318 standard an
sidering only LSRD, only GSRD, and both LSRD and GSRD, respectively.
with a different number of subpopulations were obtained and shown in T
purposes, the average values of the performance metrics were also obtain
Appl. Sci. 2021, 11, 2076 17 of 22

different number of subpopulations were obtained and shown in Tables 9–11, considering
only LSRD, only GSRD, and both LSRD and GSRD, respectively. For comparison purposes,
the average values of the performance metrics were also obtained, and they are shown in
the last row of Tables 9–11.

Table 9. Cross-section under biaxial bending. LSRD. Performance metrics.

Subpopulations Mean Cost Accuracy Robustness

1 91.92 0.9889 0.71
2 92.02 0.9878 0.62
3 91.91 0.9890 0.58
4 91.75 0.9907 0.50
5 92.10 0.9869 0.91
6 92.24 0.9855 0.69
7 92.28 0.9850 0.77
8 92.39 0.9839 0.92
9 92.68 0.9808 1.06
10 92.64 0.9812 1.05
11 92.71 0.9805 1.29
12 92.69 0.9807 1.17
13 92.68 0.9808 0.99
14 92.92 0.9783 1.07
15 92.64 0.9812 0.93
16 93.05 0.9769 1.21
Averaged value 92.41 0.9836 0.91

Table 10. Cross-section under biaxial bending. GSRD. Performance metrics.

Subpopulations Mean Cost Accuracy Robustness

1 91.92 0.9889 0.71
2 91.68 0.9915 0.55
3 91.87 0.9894 0.69
4 91.87 0.9895 0.60
5 91.96 0.9885 0.64
6 91.97 0.9884 0.61
7 91.98 0.9882 0.85
8 92.26 0.9853 0.76
9 92.17 0.9862 0.74
10 92.39 0.9839 0.95
11 92.11 0.9869 0.73
12 92.45 0.9832 0.93
13 92.25 0.9854 0.88
14 92.51 0.9826 0.76
15 92.45 0.9832 0.81
16 92.59 0.9817 0.89
Averaged value 92.15 0.9864 0.76
 Appl. Sci. 2021, 11, 2076 18 of 22

Table 11. Cross-section under biaxial bending. LSRD and GSRD. Performance metrics.

Subpopulations Mean Cost Accuracy Robustness

1 91.92 0.9889 0.71
2 91.87 0.9895 0.60
3 91.87 0.9895 0.62
4 91.97 0.9884 0.71
5 92.09 0.9871 0.77
6 92.02 0.9878 0.77
7 92.20 0.9859 0.80
8 92.25 0.9854 0.82
9 92.17 0.9862 0.79
Appl. Sci. 2021, 11, x FOR PEER REVIEW 19 of 22
10 92.20 0.9859 0.83
11 92.27 0.9852 0.80
12 92.25 0.9853 0.79
13 92.19 0.9860 0.82
16
14 92.49
92.47 0.9828
0.9831 0.79 0.86
15
Averaged value 92.28
92.16 0.9851
0.9864 0.89 0.77
16 92.49 0.9828 0.86
Averaged value 92.16 0.9864 0.77
The method is able to obtain designs close to the global optimum as shown in Tables
9 to 11. When the method considers only GSRD (Table 10), the minimum average value of
the meanThecosts,
method theismaximum
able to obtain designs
average valueclose to accuracies,
of the the global optimum as shown in
and the minimum average
Tables 9–11. When the method considers only GSRD (Table 10), the minimum
value of the standard deviations are obtained. Very similar values, but somewhat higher, average
value of the mean costs, the maximum average value of the accuracies, and the minimum
are obtained when LSRD and GSRD (Table 11) are taken into account together.
average value of the standard deviations are obtained. Very similar values, but somewhat
The performance metrics are also shown in Figures 10–12. It can be observed that the
higher, are obtained when LSRD and GSRD (Table 11) are taken into account together.
metricsThetend to worsen
performance in allare
metrics cases
alsowhen
shownthe number
in Figures of subpopulations
10–12. It can be observedincreases.
that the This
tendency is greater
metrics tend when
to worsen LSRD
in all casesiswhen
usedthe
alone and of
number above six subpopulations.
subpopulations The other
increases. This
cases have is
tendency a lower
greatertendency.
when LSRD Theis sensitivity of the
used alone and metrics
above with respect toThe
six subpopulations. theother
number of
cases have a lower
subpopulations tendency.
is reduced whenThe sensitivity
LSRD andofGSRDthe metrics withinto
are taken respect to theat
account number of time.
the same
subpopulations is reduced when LSRD and GSRD are taken into account at the same time.

Figure
Figure Cross-section under
10.Cross-section
10. underbiaxial
biaxialbending.
bending.Mean
Meancost.
cost.
 Appl. Sci. 2021, 11, 2076 Figure 10. Cross-section under biaxial bending. Mean cost. 19 of 22

ppl. Sci. 2021, 11, x FOR PEER REVIEW

Figure
Figure 11. Cross-section under
11.Cross-section biaxial
under bending.
biaxial Robustness.
bending. Robustness.

12.Cross-section
Figure 12.
Figure Cross-sectionunderunder
biaxial bending.
biaxialAccuracy.
bending. Accuracy.
4. Conclusions and Final Remarks
4. Conclusions and Final
A simple formulation Remarks
for the optimal design of reinforced concrete sections under
compression and biaxial bending was established in a previous work by the authors. In
A simple formulation for the optimal design of reinforced concrete se
that work, to verify that the formulation could work well in different types of metaheuristic
compression andalgorithm
algorithms, a genetic biaxial and
bending was
two firefly established
algorithms intoaobtain
were used previous work by th
the solutions
to two
that problems
work, under the
to verify conditions
that of two different
the formulation design
could codeswell
work (EC2 in
anddifferent
ACI318). types
ristic algorithms, a genetic algorithm and two firefly algorithms were used
solutions to two problems under the conditions of two different design cod
ACI318). As meta-heuristic algorithms have the common disadvantage of nee
Appl. Sci. 2021, 11, 2076 20 of 22

As meta-heuristic algorithms have the common disadvantage of needing to evaluate the

objective function many times during the exploration phase, the time required to obtain
the solution was extensive.
For this reason, the authors studied and implemented a simple parallel computing
strategy and the corresponding modifications and adjustments in the FA, which is explained
in detail in this work, with the aim of: (i) reducing the necessary calculation time and (ii)
avoiding any lack of precision regardless of the number of parallel processes adopted.
In accordance with these two objectives, the two examples show that:
(i) The speed-up increases as more parallel processes are considered. The trend is almost
linear up to six parallel processes. From six onwards, there is no significant increase.
It can also be seen that the efficiency quickly drops with more than four parallel
processes. These results depend on the computer used to run the calculations but not
on the proposed design method. To achieve better results, using another computer
whose architecture allows more processes in parallel would be sufficient.
(ii) The method achieves designs close to the global optimum despite the number of
parallel processes considered. Small random displacements (LSRD and GSRD) have
proven to be essential to avoiding the bias produced by the migration between sub-
populations. LSRD shows its effect when there are few subpopulations of many
individuals. In contrast, GSRD shows its effect when there are many subpopulations
with few individuals. The combined use of LSRD and GSRD slightly improves the
results and reduces their sensitivity in relation to the number of parallel processes
considered. It should be noted that the proposed method was tested considering only
up to 16 parallel processes (or 16 subpopulations). More research should be done con-
sidering more parallel processes. To facilitate this task for other potential researchers
or for other uses, the authors included the MATLAB®code for the complete method
in the Appendix A.
This cross-sectional design method can be integrated into a more general design
scheme for three-dimensional reinforced concrete structures by making the necessary
changes and adjustments. To facilitate this integration, it was previously designed to be
able to solve different load situations, as shown in the two examples included in this work.

Supplementary Materials: The following are available online at https://www.mdpi.com/2076-341

7/11/5/2076/s1.
Author Contributions: Conceptualization, A.T. and G.S.-O.; methodology and algorithm implemen-
tation, G.S.-O.; checking the results, A.T.; writing-original draft preparation, G.S.-O.; writing-review
and editing, A.T. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data presented in this study are available on request from the
corresponding author.
Conflicts of Interest: The authors declare that they have no known competing financial interests or
personal relationships that could have appeared to influence the work reported in this paper.

Appendix A
To execute the following MATLAB® functions, it is necessary to select an appropriate
number of workers in the “Parallel preferences” menu, “Parallel tool” option before execu-
tion. There, the user must set the “Preferred number of workers in a parallel survey” to a
value greater than the number of subpopulations considered. This value is related to the
number of processor cores.
Every function included here must be located in an independent .m file.
Appl. Sci. 2021, 11, 2076 21 of 22

References
1. Yang, X.-S. Nature-Inspired Metaheuristic Algorithms, 2nd ed.; Luniver Press: Frome, UK, 2010.
2. Holland, J.H. Adaptation in Natural and Artificial Systems; University of Michigan Press: Ann Anbor, MI, USA, 1975.
3. Holland, J.H. Genetic algorithms. Sci. Am. 1992, 267, 66–72. [CrossRef]
4. Kennedy, J.; Eberhart, R. Particle swarm optimization. In Proceedings of the 1995 IEEE International Conference on Neural
Networks, Perth, WA, Australia, 27 November–1 December 1995; Volume 4, pp. 1942–1948.
5. Kirkpatrick, S.; Gelatt, C.D.; Vecchi, M.P. Optimization by simulated annealing. Science 1983, 220, 671–680. [CrossRef]
6. Geem, Z.W.; Kim, J.H.; Loganathan, G. A new heuristic optimization algorithm: Harmony search. Simulation 2001, 76,
60–68. [CrossRef]
7. Wolpert, D.H.; Macready, W.G. No Free Lunch Theorems for Optimization. IEEE Trans. Evol. Comput. 1997, 1, 67–82. [CrossRef]
8. Kunnath, S.K.; Reinhorn, A.M.; Abel, J.F. A computational tool for evaluation of seismic performance of reinforced concrete
buildings. Comput. Struct. 1991, 41, 157–173. [CrossRef]
9. Carbonell, A.; González-Vidosa, F.; Yepes, V. Design of reinforced concrete road vaults by heuristic optimization. Adv. Eng. Softw.
2011, 42, 151–159. [CrossRef]
10. Song, Y.; Wang, Z.; Liu, Z.; Wang, R. A spatial coupling model to study dynamic performance of pantograph-catenary with
vehicle-track excitation. Mech. Syst. Signal Process. 2021, 151, 107336. [CrossRef]
11. Kondratova, I.L.; Montes, P.; Bremner, T.W. Natural marine exposure results for reinforced concrete slabs with corrosion inhibitors.
Cem. Concr. Compos. 2003, 25, 483–490. [CrossRef]
12. Panda, S.; Padhy, N.P. Comparison of particle swarm optimization and genetic algorithm for FACTS-based controller design.
Appl. Soft Comput. 2008, 8, 1418–1427. [CrossRef]
13. Artale, V.; Milazzo, C.L.R.; Orlando, C.; Ricciardello, A. Comparison of GA and PSO approaches for the direct and LQR tuning of
a multirotor PD controller. Am. Inst. Math. Sci. 2017, 13, 2067–2091. [CrossRef]
14. Valvano, S.; Orlando, C.; Alaimo, A. Design of a noise reduction passive control system based on viscoelastic multilayered plate
using PD SO. Mech. Syst. Signal Process. 2019, 123, 153–173. [CrossRef]
15. Yang, X.-S. Firefly Algorithms for Multimodal Optimization. In Stochastic Algorithms: Foundations and Applications, SAGA 2009,
Lecture Notes in Computer Sciences; Springer: Berlin/Heidelberg, Germany, 2009; Volume 5792, pp. 169–178.
16. Gandomi, A.H.; Yang, X.-S.; Talatahari, S.; Alavi, A.H. Firefly algorithm with chaos. Commun. Nonlinear Sci. Numer. Simul. 2013,
18, 89–98. [CrossRef]
17. Fister, I., Jr.; Yang, X.-S.; Brest, J. A comprehensive review of firefly algorithms. Swarm Evol. Comput. 2013, 13, 34–46. [CrossRef]
18. Gandomi, A.H.; Yang, X.-S.; Alavi, A.H. Mixed variable structural optimization using Firefly Algorithm. Comput. Struct. 2011, 89,
2325–2336. [CrossRef]
19. Gandomi, A.H.; Kashani, A.R.; Roke, D.A.; Mousavi, M. Optimization of retaining wall design using recent swarm intelligence
techniques. Eng. Struct. 2015, 103, 72–84. [CrossRef]
20. Talatahari, S.; Gandomi, A.H.; Yun, G.J. Optimum design of tower structures using Firefly Algorithm. Struct. Des. Tall Spéc. Build.
2012, 23, 350–361. [CrossRef]
21. Akin, A.; Saka, M.P. Optimum detailing design of reinforced concrete plane frames to ACI318-05 using the harmony search
algorithm. In Proceedings of the Eleventh International Conference on Computational Structures Technology, Paper 72, Dubrovnik,
Croatia, 4–7 September 2012; Topping, B.H.V., Ed.; Civil Comp Press: Stirlingshire, UK, 2012. [CrossRef]
22. Akin, A.; Saka, M.P. Harmony search algorithm based optimum detailing design of reinforced concrete plane frames subject to
ACI318-05 provisions. Comput. Struct. 2015, 147, 79–95. [CrossRef]
23. Nilson, A.H.; Darwin, D.; Dolan, C.W. Design of Concrete Structures, 14th ed.; McGraw-Hill: New York, NY, USA, 2010.
24. Sahab, M.G.; Toropov, V.V.; Gandomi, A.H. A review on traditional and modern structural optimization: Problems and techniques.
In Metaheuristic Applications in Structures and Infrastructures; Elsevier Inc.: London, UK, 2013; pp. 25–47.
25. Thierauf, G.; Cai, J. Parallel evolution strategy for solving structural optimization. Eng. Struct. 1997, 19, 318–324. [CrossRef]
26. Leite, J.P.B.; Topping, B.H.V. Parallel simulated annealing for structural optimization. Comput. Struct. 1999, 73, 545–564. [CrossRef]
27. Hasanҫebi, O.; Bahҫecioǧlu, T.; Kurҫ, Ö.; Saka, M.P. Optimum design of high-rise steel buildings using an evolution strategy
integrated parallel algorithm. Comput. Struct. 2011, 89, 2037–2051. [CrossRef]
28. Truong, V.H.; Nguyen, P.C.; Kim, S.E. An efficient method for optimizing space steel frames with semi-rigid joints using practical
advanced analysis and the micro-genetic algorithm. J. Constr. Steel Res. 2017, 128, 416–427. [CrossRef]
29. Sánchez-Olivares, G.; Tomás, A. Improvements in meta-heuristic algorithms for minimum cost design of reinforced concrete
rectangular sections under compression and biaxial bending. Eng. Struct. 2017, 130, 162–179. [CrossRef]
30. Technical Committee CEN/TC250. Eurocode 2: Design of Concrete Structures—Part 1-1: General Rules and Rules for Buildings; (EN
1992-1-1:2004/A1:2014); European Committee for Standardization: Brussels, Belgium, 2014.
31. ACI Committee 318. Building Code Requirements for Structural Concrete and Commentary; (ACI CODE-318-19); American Concrete
Institute: Farmington Hills, MI, USA, 2019.
32. Talbi, E. Metaheuristics: From Design to Implementation; John Wiley & Sons: Hoboken, NJ, USA, 2009.
33. MATLAB® Primer (R2014a); The MathWorks, Inc.: Natick, MA, USA, 2014.
Appl. Sci. 2021, 11, 2076 22 of 22

34. Gomes, G.M.; Esposito, A. A Parallelised Firefly Algorithm for Structural Size and Shape Optimization with Multimodal
Constraints. In Cuckoo Search and Firefly Algorithm; Theory and Applications, Studies in Computational Intelligence, 516;
Springer: London, UK, 2014; pp. 291–314.
35. Gil-Martín, L.M.; Hernández-Montes, E.; Aschheim, M. Optimal reinforcement of RC columns for biaxial bending. Mater. Struct.
2010, 43, 1245–1256. [CrossRef]