Engineering Structures 179 (2019) 556–565

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Accelerated optimization method for low-embodied energy concrete box- T

girder bridge design
⁎
Vicent Penadés-Plàa, Tatiana García-Segurab, , Víctor Yepesa
a
Institute of Concrete Science and Technology (ICITECH), Universitat Politècnica de València, 46022 Valencia, Spain
b
School of Civil Engineering, Universitat Politècnica de València, 46022 Valencia, Spain

A R T I C LE I N FO A B S T R A C T

Keywords: Structural optimization is normally carried out by means of conventional heuristic optimization due to the
Low-embodied energy complexity of the structural problems. However, the conventional heuristic optimization still consumes a large
Post-tensioned concrete amount of time. The use of metamodels helps to reduce the computational cost of the optimization and, along
Box-girder bridge these lines, kriging-based heuristic optimization is presented as an alternative to carry out an accelerated op-
Structural optimization
timization of complex problems. In this work, conventional heuristic optimization and kriging-based heuristic
Metamodel
Kriging
optimization will be applied to reach the optimal solution of a continuous box-girder pedestrian bridge of three
spans with a low embodied energy. For this purpose, different penalizations and different initial sample sizes will
be studied and compared. This work shows that kriging-based heuristic optimization provides results close to
those of conventional heuristic optimization using less time. For the sample size of 50, the best solution differs
about 2.54% compared to the conventional heuristic optimization, and reduces the computational cost by
99.06%. Therefore, the use of a kriging model in structural design problems offers a new means of solving certain
structural problems that require a very high computational cost and reduces the difficulty of other problems.

1. Introduction due to their importance in the area of communications. However, de-

signing a sustainable bridge is not easy, due to the fact that the struc-
The traditional main objective of structural engineering is to reach tural problem is characterized by a large number of design variables
maximum safety with the minimum investment. However nowadays, with multiple combinations. A heuristic optimization process is pre-
due to the increased concern for sustainability, other aspects have also sented as an alternative to achieve a solution within the design space
become important within the field of structural engineering. These as- that reaches the objectives and guarantees the constraints imposed by
pects are usually grouped into the three main objectives (economic, the regulations. This method has been used to optimize many types of
environmental and social) of sustainability [1,2]. In this way, the tra- structures, such as reinforced concrete columns [13,14], reinforced
ditional structural engineering problem becomes a complex problem concrete frames [8], precast concrete floors [15], prestressed concrete
that should be solved by means of a decision-making process [3,4]. precast road bridges [12] and post-tensioned concrete box-girder
Regarding the environmental goal, life-cycle assessment is an accepted bridges [11,16]. However, the structural optimization problem depends
process to obtain the complete environmental profile of a process, on a large number of design variables with several constraints. This
product or service [5–7]. However, a first approximation of the en- results in excessive computational costs [17]. One effective solution to
vironmental assessment can be carried out using a single criterion that carry out the optimization with a lower computational cost is the use of
represents, in a reliable way, the environmental impact. The most re- approximate response surfaces obtained by surrogate models or meta-
presentative criteria are the CO2 emissions and the embodied energy, models. The most common metamodels are polynomial regression,
which also have a direct relationship with the cost [8,9]. This indicates neural networks and kriging [18,19]. The kriging model is one of the
that the optimization of CO2 emissions or embodied energy reduces at most encouraging metamodels used in structural optimization [20] al-
the same time the cost. There are several works that have analyzed though despite this fact, only few works have been carried out using a
structures with a lower CO2 emission [10,11], but the embodied energy kriging-based heuristic optimization to design real structures. This
has been less studied [12]. model provides an optimal interpolation based on regression against
Bridges are one of the most important structures in civil engineering observed values of the surrounding data points, weighted according to

⁎
Corresponding author.
E-mail addresses: vipepl2@cam.upv.es (V. Penadés-Plà), tagarse@upv.es (T. García-Segura), vyepesp@cst.upv.es (V. Yepes).

https://doi.org/10.1016/j.engstruct.2018.11.015
Received 11 June 2018; Received in revised form 3 October 2018; Accepted 7 November 2018
0141-0296/ © 2018 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/BY-NC-ND/4.0/).
V. Penadés-Plà et al. Engineering Structures 179 (2019) 556–565

spatial covariance values. This means that kriging considers both global process of conventional heuristic optimization, which is the structural
and local approximations at the same time. Thus, the kriging model analysis and evaluation of the objective function part, is replaced by an
takes into account the local variations of the objective response. In this evaluation of the metamodel. Therefore, the computational cost ne-
context, a methodology that allows optimal designs to be determined cessary for metamodel-based heuristic optimization (Fig. 2) is lower
with adequate accuracy and at reduced time cost is highly desirable. than the computational cost necessary for conventional heuristic opti-
In this work, conventional heuristic optimization and kriging-based mization.
heuristic optimization will be applied to determine an optimized con-
tinuous box-girder pedestrian bridge of three spans with a low embo- 3. Metamodel construction process
died energy. A comparison between both optimization techniques will
be carried out to determine if the kriging-based heuristic optimization The basis of metamodels consists of constructing an approximate
provides reasonable results compared with the conventional heuristic mathematical model of a detailed simulation model, which predicts the
optimization. For this purpose, different coefficients of penalizations output data (objective response) from input data (design variables) in
and sampling sizes will be considered to determine the characteristics the whole design space, more efficiently than the detailed simulation
of the kriging-based heuristic that performs better. After that, a set of models. It could, as such, be called a model of the model. The con-
parameters for the kriging model will be recommended. In Section 2 struction process of a metamodel focuses on three main parts: (a) ob-
both optimization processes will be described. In Section 3, a general taining the initial input dataset points inside the design space, (b)
scheme of the process to construct a metamodel will be shown, focusing choosing the metamodel type to construct the approximate mathema-
on the main methods used in this work, namely latin hypercube sam- tical model and (c) choosing the fitting model. There are a large number
pling and the kriging model. In Section 4, the problem design will be of options for carrying out these steps [25]. Regardless of the choice for
described, and in Section 5 the most important results will be shown. each step, the main objective of constructing a metamodel is to obtain a
Finally, the most important conclusions will be detailed. model with the best accuracy possible to predict the objective response.
The choice of the initial input dataset points or sampling inside the
2. Optimization process design space is defined by the sample size and the position of the points,
because both aspects have an influence on the model construction. On
Optimization is a process that tries to find the best possible solution the one hand, the sample size is fundamentally related to the number of
to a problem that may be defined by one (mono-objective) or several design variables. The sample size must be higher with a larger number
(multi-objective) objective functions, f, that satisfy some constraints, gj. of design variables for the same accuracy of the metamodel, and
therefore the computational cost necessary to construct the model will
f (X ) (1)
be higher. On the other hand, once the sample size has been defined,
gj (X ) ≤ 0 the position of the points must be placed within the design space in
(2)
order to obtain the best possible information. This process is called
where X represents the vector with the design variables chosen for the Design Of Experiments (DoE).
formulation. The DoE can be divided into two different groups. The first group
The optimization process is defined by the algorithm used and es- clusters the classic designs, that include the factorial or fractional fac-
tablishes a set of rules to be followed in solving operational problems. torial designs, central composite designs, Box-Behnken designs, Plackett-
These algorithms can be divided into exact algorithms and heuristic Burman designs, Koshal designs and D-optimal designs [26]. These types of
algorithms. Exact algorithms reach the global optimum by using se- designs tend to spread the sample points around the border of the de-
quential techniques of mathematical programming. Heuristic algo- sign space and only include a few points inside of it. The classic designs
rithms were developed to solve complex and realistic structural opti- are mainly used to construct polynomial metamodels. When the initial
mization problems of discrete variables. These algorithms achieve good input data points were used to construct more advanced metamodels,
solutions without guaranteeing the global optimum, but with a lower other designs, called space-filling designs, were preferred. These types
computational cost. Complex optimization problems, such as structural of designs trend to spread the sample points all over the design space
optimization, are defined for a large number of design variables, and (often with a uniform distribution), so it is possible to take into account
thus the heuristic algorithms have demonstrated the best behavior in the local phenomena in any region of the design space. The most
solving this kind of problem. popular space-filling designs are latin hypercube sampling [27], dis-
Heuristic algorithms try to simulate simple events observed in tance-based designs [28] and low-discrepancy sequences, which group
nature. In general, the traditional heuristic algorithms look for a local Hammersley sequence sampling [29] and the uniform design [30].
optimum, while the metaheuristic algorithms have tools to avoid local In this work, to generate the sample, latin hypercube sampling
optimums in order to find a better solution. Metaheuristic algorithms (LHS) has been considered; its effectiveness in the estimation of the
follow an iterative process in which a complete structural design objective response of the metamodel has been proven in several works
(combination of design variables) is defined to carry out the structural [31,32]. LHS was proposed by McKay et al. in 1979 [27]. This method
analysis and to evaluate aptitude by an objective function (Fig. 1). In determines the N number of non-overlapping intervals for each variable
recent years, some metaheurisitc algorithms have been applied to (in this work these intervals are divided according to a uniform dis-
structural optimization including the variable neighborhood search tribution) from a number of design variables (v) and a number of initial
[10], harmony search [21], threshold function [22], memetic algorithm input dataset points (N). Therefore, the design space is divided into Nv
[23], glowworm swarm algorithm [9] and simulated annealing [11] regions. Each sample point will be located in one region in order that
among others. each point corresponds to a combination of different intervals of each
However, despite the advances in technology, the computational design variable range. In this way, each interval of each design variable
cost of structural heuristic optimization is still very high [24] due to the range will only be associated with one sample point. Consequently, the
finite element structural analysis carried out during all iterations of the LHS guarantees that all of the design variables are represented along
optimization process. This high computational cost can be reduced by their respective ranges. Fig. 3 shows an example with 2 design variables
means of metamodels (also called surrogate models or approximation and 10 initial input dataset points.
models) [17]. These metamodels construct a mathematically approx- Once the sample is defined, the objective response of the initial
imate model of the objective function from a set of points in the design input dataset points is obtained. All of this initial information (inputs
space (initial sampling) to predict the output without the need to carry and outputs) is used to construct the metamodel over all of the design
out a full structural analysis. This means that the slowest part of the space. In this way, the metamodel predicts the objective response

557
V. Penadés-Plà et al. Engineering Structures 179 (2019) 556–565

Fig. 1. General flow chart of conventional heuristic optimization process.

Fig. 2. General flow chart of metamodel-based heuristic optimization.

according to a mathematical function: others as this depends on the problem posed. However, the most
common metamodels are polynomial regression, neural networks and
y = f (x ) = g (x ) + ε (3)
kriging [18,19]. The polynomial-based response surface model is
where x are the input dataset points, f(x) corresponds to the real sometimes difficult to use in complex engineering problems, and the
response (model), g(x) represents the approximate response (meta- neural network-based model requires many sample points and much
model) and ɛ represents the approximation error. There are several computational time for the training of the network [35]. The kriging
mathematical formulations to construct metamodels with different model is a promising metamodel as it is more flexible than polynomial-
characteristics [19,25]. Although these metamodels have been com- based models and less time consuming than neural network-based
pared [32–34], it is not possible to determine if one is better than the techniques [33]. Thus, this work uses the kriging formulation to

558
V. Penadés-Plà et al. Engineering Structures 179 (2019) 556–565

cov [Z (x i ), Z (x j )] = σ 2·R (x i , x j ) (6)

where the process variance σ scales the spatial correlation function R

2

(xi, xj) between two data points. In engineering design, the Gaussian
correlation function (Eq. (7)) is the most commonly used [37] function
that can be defined with only one parameter (θ) that controls the area
of influence of nearby points [35]. A low θ means that all the sample
points have a high correlation, thus the term Z(x) will be similar all over
the design space. As the value θ increases, the points with higher cor-
relation will be closer, thus the term Z(x) will differ depending on the
point in the design space:
m i j2
R (x i , x j ) = e− ∑k = 1 θ | xk − xk | (7)

Finally, each metamodel type has its associated fitting method. In

this case, the kriging formulation uses the search for the Best Linear
Unbiased Predictor (BLUP). Simpson et al. [19] gave a detailed review
of the equations and fitting methods for common metamodel types.

4. Problem design

In this section, a comparison of conventional heuristic optimization

and kriging-based heuristic optimization will be discussed. First of all,
Fig. 3. Latin hypercube sampling (v = 2 and n = 10). the structure considered (a continuous concrete box-girder pedestrian
bridge) and all of the characteristics involved will be described. After
construct the metamodel. that, the optimization problem associated with the bridge will be de-
Kriging is a metamodel that has its origins in geostatic applications fined. Finally, both optimization processes will be explained. This final
involving spatially and temporally correlated data and was developed point includes the design variables considered in each case, as well as
by the South African mining engineer called Danie Gerhardus Kirge. how each approach deals with the constraints.
Later, many researches contributed to the problem of optimal spatial
prediction, but the approach was formalized by Matheron in 1963 [36] 4.1. Box-girder pedestrian bridge description
who used the term kriging in honor of the contribution of Danie Ger-
hardus Kirge [20]. The idea behind kriging is that the deterministic The bridge is a continuous concrete box-girder pedestrian bridge
response y(x) can be described as: deck with three continuous spans of 40–50–40 m length (the relation-
ship between the external span and the central span follows the op-
y (x ) = f (x ) + Z (x ) (4)
timum of 80%). This type of bridge is commonly used due to its
structural performance, low dead load and construction conditions. The
where f(x) is the known approximation function, and Z(x) is a reali-
pedestrian bridge deck has a constant width of 3 m, and the remaining
zation of a stochastic process with mean zero, variance σ2 and non-zero
geometrical dimensions of the cross-section are defined by the seven
covariance. The first term of the equation, f(x), is similar to a regression
variables of (Fig. 4): depth (h), bottom slab width (b), web inclination
model that provides a global approximation of the design space (Eq.
width (d), top slab thickness (es), external cantilever section thickness
(5)). The second term, Z(x), creates local deviations so that the kriging
(ev), bottom slab thickness (ei) and webs slab thickness (ea). The value
model interpolates the initial sample points (Eq. (6)). In many cases, f
of these variables is limited for a range. The depth range is 1.25–2.5 m,
(x) is simply a constant term and the method is then called ordinary
the bottom slab width range is 1.2–1.8 m, the width of the web in-
kriging. If f(x) is set to 0, implying that the response y(x) has a mean of
clination range is 0–0.4 m, the web slab thickens is 0.3–0.6 m and the
zero, the method is called simple kriging [37].
other slab thickness ranges are 0.15–0.4 m. The haunch (t), is calculated
n from the values of other variables (Eq. (8)) according to Schlaich and
f (x ) = ∑ βi·fi (x ) Scheff’s [38] recommendation. In addition, the haunch must provide
i=1 (5)
the space to contain the ducts in the high and low points.

Fig. 4. Box-girder cross-section.

559
V. Penadés-Plà et al. Engineering Structures 179 (2019) 556–565

Fig. 5. Pedestrian bridge and duct layout.

Fig. 6. Longitudinal and transversal reinforcing steel disposition.

t = max { b − 2·ea
5 }
, ei .
(8)
Table 1
Main parameters of the analysis.
Material parameters
The strength of the concrete is defined by the variable fck that can
Maximum aggregate size 20 mm
take a value inside of the range 35–100 MPa. The post-tensioned steel Reinforcing steel B-500-S
formed by 0.6 in. strands is prestressed to 195.52 kN. The ducts are Post-tensioned steel Y1860-S7
symmetrically distributed through the webs with a parabolic layout. Strand diameter Φs = 0.6″
The maximum eccentricity is present where the bending moment is the Tensioning time 7 days

maximum or minimum (Fig. 5). At these points, the distance considered Geometrical parameters
between the duct and the surface is 0.2 m. The distance from the piers Pedestrian bridge width B=3m
Number of spans 3
to the point of inflection is defined by 5% of the length of each span. In
Central span length L1 = 50 m
addition, the position of the reinforced steel is defined according the External span length L2 = 40 m
Fig. 6. Longitudinal reinforcement is defined by the number of bars per Clearance 5m
meter and their diameter, placed at the top slab (LRn1, LRØ1), the flange Diaphragm thickness 1.2 m
(LRn2, LRØ2, LRn3, LRØ3), the web (LRn4, LRØ4), the bottom slab (LRn5, Exposure related parameters
LRØ5) and the core (LRn6, LRØ6). In addition, extra bending reinforce- External ambient conditions IIb
ment is divided into two systems. One covers the top slab at the support Regulation related parameters
zone (L/5 on both sides of the piers), with a diameter defined by LRØ7 Regulations EHE-08/IAP-11/Eurocodes
and the same number of bars per meter as LRn1. The other is placed at Service working life 100 years
the bottom slab throughout the rest of the external span (LRØ8) and the Loading related parameters
central span (LRØ9). The number of bars per meter is, for both loca- Reinforced concrete self-weight 25 kN/m3
tions, equal to LRn5. The diameter can change according to 0, 10, 12, Asphalt layer self-weight 24 kN/m3
Mean asphalt thickness 47.5 mm
16, 20, 25 and 32 mm. Regarding transverse reinforcement, the dia- Bridge railing self-weight 1 kN/m
meter of the standard reinforcement (TRØ1, TRØ2, TRØ3, TRØ4, TRØ5, Live load 5 kN/m2
TRØ6, TRØ7) is set with the same spacing (TRS) for construction re- Differential settling 5 mm
quirements.
Traditional scaffolding is used in the construction stage with a
clearance of 5 m. The formwork is disposed over the scaffolding to give gj (x1, x2 , x3, ⋯, x n ) ≤ 0 (10)
the shape of the cross section of the bridge. In addition, lighting is used
to lighten the self-weight of the bridge. Table 1 defines the other con- where x1, x2, x3, .., xn are the design variables.
ditions employed in this study such as the materials, the load actions on The objective function evaluates the embodied energy for the total
the structure, the exposure class and the regulations used. number of construction units considering the material used and the
placement embodied energy defined in Eq. (9). The embodied energy of
each unit (ei), shown in Table 2, were obtained from the BEDEC ITEC
4.2. Optimization problem description
database [39]. The embodied energy of concrete is determined for each
compressive strength grade according to the mix design, including the
In this work, the problem of continuous concrete box-girder pe-
embodied energy of raw materials extraction, manufacture and trans-
destrian bridge deck optimization involves a single-objective optimi-
portation. The measurements (mi) concerning the construction units are
zation of the embodied energy of the structure. Hence, this optimization
evaluated from the design defined using the design variables.
aims to minimize the embodied energy (Eq. (9)) and satisfy the con-
The structural constraints represented by Eq. (10) check the servi-
straints (Eq. (10)).
ceability and ultimate limit states (SLS and ULS) of Vertical shear,
Embodied energy = ∑ ei × mi (x1,x2,⋯.,x n ) Longitudinal shear, Punching shear, Bending, Torsion, Torsion com-
i = 1, n (9) bined with bending and shear, Cracking, compression and tension

560
V. Penadés-Plà et al. Engineering Structures 179 (2019) 556–565

Table 2 verify all of the constraint defined by the regulations, kriging-based

Unit energy. heuristic optimization only defines the design variables that the en-
Unit measurements Energy (kWh) gineers would take into account in their design (geometrical variables
and the concrete strength) to later calculate the amount of the post-
3
m of scaffolding 20.4 tensioned steel and the reinforced steel required according to the
m2 of formwork 8.7
standards. Therefore, the post-tensioned and reinforced steel are not
m3 of lighting 1137.5
kg of steel (B-500-S) 10.44
variables, and consequently, the design space is greatly reduced.
kg of post-tensioned steel (Y1860-S7) 12.99 First of all, a specific sample size (N) over all the design space is
m3 of concrete HP-35 612.22 obtained according to LHS, then, the embodied energy is calculated for
m3 of concrete HP-40 646.61 each of these points. Due to the complexity of structural problems, there
m3 of concrete HP-45 681
are regions of the design space for which certain combinations of the
m3 of concrete HP-50 715.39
m3 of concrete HP-55 749.77 geometrical design variables are not possible (for example h <
m3 of concrete HP-60 784.16 es + ev + 2 * t). This is because the embodied energy of the bridge
m3 of concrete HP-70 852.94 cannot be obtained in some points of the LHS. To solve this constraint
m3 of concrete HP-80 921.72
and attempt to conduct the optimization for feasible designs, two re-
m3 of concrete HP-90 990.49
m3 of concrete HP-100 1059.27
sponse surfaces will be constructed. The first one is determined by the
feasible solutions of the LHS. With this response surface, the objective
response of the unfeasible solutions of the LHS will be estimated and a
stress, vibration. Note that the code [40] provides different equations penalization is applied to those solutions. To prevent too much skewing
for conventional and high-strength concrete (concrete with a char- of the response surface, the penalization is imposed depending on the
acteristic compressive strength greater than 50 MPa). In addition, the case: if the total embodied energy is higher than the minimum embo-
geometrical and constructability requirements are verified, following died energy of the set of feasible solutions, the embodied energy is not
the Spanish regulations for this type of structure [40,41] as well as the modified (case 1). Otherwise, if the total embodied energy is lower than
Eurocodes [42,43]. It is worth mentioning that the analysis and the the minimum embodied energy of the set of feasible solutions, a pe-
verification of the limit states are coded in Matlab. nalization is imposed to avoid reaching unfeasible optimum solutions
The algorithm used to carry out the optimization problem is simu- (case 2). A study of the penalization imposed will be carry out in the
lated annealing (SA) [44] due to its versatile acceptance criterion. Many next section. Finally, on grouping all of the feasible and non-feasible
works use SA to carry out conventional heuristic optimization [8,45]. In solutions, a second response surface will be determined. This response
this work, the initial temperature is calibrated following Medina’s [46] surface is constructed considering all of the LHS points, and thus all of
method, which proposes that the initial temperature is halved when the the design space will be represented. In addition, the penalization
percentage of acceptances is greater than 40%, and doubled when it is avoids the optimization tending towards to unfeasible solutions.
less than 20%. After that, the temperature decreases according to a Once the final response surface is obtained, a validation process that
coefficient of cooling k following the equation T = k * T, when a compares the real embodied energy and the estimated embodied energy
Markov chain ends. In this work, the calibration revealed that a coef- of nine random data points is carried out in order to determine the
ficient of cooling of 0.8 and a length of the Markov chain of 1000 are accuracy of the model. Then, heuristic optimization by means of the SA
appropriate. The algorithm finishes after three Markov chains show no algorithm is carried out to determine the final energy-optimized con-
improvement. tinuous concrete box-girder pedestrian bridge. Finally, the estimated
optimized solution will be checked. In the case that this solution is
feasible, the process finishes, but if the solution is unfeasible, a new
4.3. Optimization process
initial population by LHS will be generated and the entire process is
repeated. This procedure aims to study the influence of the initial po-
As stated above, in this work, a comparison between conventional
pulation (N) on the accuracy of the model, the optimization and the
heuristic optimization and kriging-based optimization will be carried
computational cost. Fig. 7 shows the scheme followed in this kriging-
out. The main difference between these processes is that, while in
based heuristic optimization.
conventional heuristic optimization, before obtaining the objective re-
sponse, all of the constraints of the bridge are checked at each step of
5. Results
the optimization, in kriging-based heuristic optimization, the objective
response is estimate throughout a mathematical approximation.
In this section, the results of the comparison between conventional
heuristic optimization and kriging-based heuristic optimization are
4.3.1. Conventional heuristic optimization shown. For this purpose two main objectives are proposed in this study:
In conventional optimization, in addition to the seven geometrical (1) to obtain the characteristics of the kriging model that provides good
variables and the concrete strength, the reinforced steel and the pre- results of the optimization process, and (2) to study if the kriging-based
stressed steel are also variables. Reinforced steel is defined by 23 heuristic optimization reaches acceptable results compared with the
variables, 15 for the longitudinal reinforcement and 8 for the transverse conventional heuristic optimization. The comparative study is carried
reinforcement (see Fig. 6). Once the initial box-girder pedestrian bridge out based on the mean results and best result of nine optimized solu-
is completely defined, the SA algorithm carries out movements of the tions.
design variables to compare the objective response obtained after each Before comparing the optimization methodologies, a sensitivity
movement until the energy-optimized box-girder pedestrian bridge is study is carried out to study the coefficient of penalization (p) applied
reached according to the process defined in Section 2. Each movement to the unfeasible solutions of case 2. Different p values have been
requires the complete verification of the SLS and ULS, entailing a high considered including 1, 1.25 and 1.5, and applied to the highest po-
computational cost. Fig. 2 shows a scheme of the conventional heuristic pulation considered (N = 500) in order to determine the influence of
optimization considered. this parameter on the kriging model. Table 3 shows the results of a
group of nine solutions for each different coefficient of penalization.
4.3.2. Kriging-based heuristic optimization The first three columns refer to the mean results and the last one to the
In contrast to the conventional heuristic optimization in which the best result of each group. The first column shows the accuracy of the
bridge is defined completely at the beginning of each iteration to later kriging model, evaluated as the mean difference of the real embodied

561
V. Penadés-Plà et al. Engineering Structures 179 (2019) 556–565

Fig. 7. Kriging-based heuristic optimization.

Table 3 the mean embodied energy decreases from 771,853 kWh to

Study of coefficient of penalization. 744,156 kWh when the coefficient of penalization decreases from 1.5 to
Mean results Best result 1. It shows that considering a coefficient of penalization of one im-
proves the following optimization. This demonstrates that the kriging
Surface Embodied Optimized- Embodied surface has a better behavior when there are smaller variations in its
accuracy (%) energy (kWh) solutions energy (kWh) objective response. Thus, the coefficient of penalization considered to
accuracy (%)
carry out this study will be p = 1.
p=1 3.88% 744,156 4.01% 701,910 Once the coefficient of penalization is determined, nine kriging
p = 1.25 3.52% 750,254 4.83% 704,500 surfaces are obtained for each initial sample size (N = 10, N = 20,
p = 1.5 3.99% 771,853 4.58% 731,210 N = 50, N = 100, N = 200, N = 500) to evaluate the influence of the
sample size on the accuracy of the results. The accuracy of this kriging
surface is evaluated through the mean of the difference between the
energy and the estimated embodied energy of nine random points. The
real embodied energy and the predicted embodied energy of a random
second and third columns show the mean embodied energy of nine
sample of the design space. Fig. 8 shows that the accuracy of the kriging
optimized solutions in kWh and the accuracy of these nine optimized
surface increases with the number of initial samples but a horizontal
solutions. Finally, the last column shows the best optimized solution in
convergence is observed from N = 50, in which the accuracy of the
kWh. Both the mean embodied energy and the best embodied energy
surface is 4.04%. From N = 10 to N = 50 the accuracy of the kriging
improve when the coefficient of penalization decreases. For example,

562
V. Penadés-Plà et al. Engineering Structures 179 (2019) 556–565

12% 18%
16%
Validation of the model (%)

10%

Coefficient of variation (%)

14%
8% 12%
10%
6%
8%
4% 6%
4%
2%
2%
0% 0%
0 100 200 300 400 500 0 100 200 300 400 500
Initial sampling size (N) Initial sampling size (N)
Fig. 8. Validation of kriging response surface. Fig. 10. Comparison of the coefficient of validation.

1200000
the best solution obtained with N = 50 differs only 2.54% with respect
to the best solution of the conventional heuristic optimization. Besides,
Mean embodied energy (kWh)

1100000
Fig. 10 shows that the increment in the initial sample size reduces the
1000000 coefficient of variance of the nine solutions, reaching a lower value than
the coefficient of variance of the nine solutions obtained in the con-
900000 ventional heuristic optimization. While the coefficient of variance of
the conventional heuristic optimization is 3.79%, the coefficient of
800000
variance of the kriging-based heuristic optimization is 3.67% when the
700000 sample size is N = 500. These results show that a satisfactory solution
can be obtained with an initial sample size of N = 50, but a higher
600000 initial sample size improves the accuracy of the model and the mean
0 100 200 300 400 500 embodied energy. Thus, it can be said that the kriging model is robust
Initial sampling size (N) for optimization problems.
Fig. 9. Comparison of the mean of embodied energy of bridges. It must not be forgotten that the main advantage of the kriging-
based heuristic optimization is the computational cost saving as the
objective response of each iteration is directly obtained. The kriging-
model improves from 11.11% to 4.04% (upgrading of 7.07%). How-
based heuristic optimization required 1804.11 s for an initial sample
ever, the accuracy of the kriging model from N = 50 to N = 500 im-
size of N = 500, while the conventional heuristic optimization required
proves from 4.04% to 3.88% (upgrading of only 0.16%).
19617.14 s. This is a reduction of 90.80% in the computational cost.
Once the kriging surface is obtained, the optimization is carried out.
Note that the greater part of the computing time in the kriging-based
For each initial sample size, different characteristics of the two opti-
heuristic optimization is due to the generation of the initial population.
mizations have been compared. Fig. 9 shows the mean embodied en-
Regarding the conventional heuristic optimization, more than 80% of
ergy of nine optimized box-girder pedestrian bridges. The horizontal
the computing time is spent in the analysis and the verification of the
dashed line represents the mean embodied energy obtained by con-
ultimate and serviceability limit states, as well as the geometrical and
ventional heuristic optimization, while the solid line represents the
constructability requirements. Table 4 shows in more detail the time
mean embodied energy obtained by the kriging-based heuristic opti-
savings achieved for the other initial sample sizes.
mization according to the sample size. The mean embodied energy of
Table 4 summarizes the most important results in comparing the
the nine optimized bridges obtained by the conventional heuristic op-
optimization approaches. The first six rows represent the different in-
timization is 713,504 kWh. This result improves by 4.30% the best
itial sample sizes of the kriging-based heuristic optimizations, and the
mean embodied energy of the nine optimized bridges obtained by the
last row represents the conventional heuristic optimization. The col-
kriging-based heuristic optimization (corresponding to N = 500). Fur-
umns represent the results of the different characteristics studied. The
thermore, as can be seen in Table 4, the best solutions of each group of
first six columns show the main results of the nine optimized bridges,
nine obtained by kriging-based heuristic optimization are close to the
and the last two columns show the best optimized bridges for each case.
best solution of the conventional heuristic optimization. For example,
The first column shows the accuracy of the kriging model, evaluated as

Table 4
Overview of results obtained.
Method N Mean results Best result

Surface Time (s) Time comparison Embodied Energy Coefficient of Embodied Comparison with
accuracy (%) with CH (%) energy (kWh) comparison with variance (%) energy (kWh) CH (%)
CH (%)

Kriging-based heuristic 10 11.11% 26.73 99.86% 1,130,127 58.39% 15.81% 814,840 19.49%
optimization (KH) 20 5.83% 236.71 98.79% 844,816 18.2% 13.67% 721,400 5.79%
50 4.04% 185.08 99.06% 783,726 9.84% 6.65% 699,240 2.54%
100 4.16% 510.10 97.40% 762,350 6.85% 5.20% 700,800 2.77%
200 3.88% 1497.33 92.37% 767,034 7.50% 3.94% 701,910 2.93%
500 3.88% 1804.11 90.80% 744,157 4.30% 3.67% 701,910 2.93%

Conventional heuristic 19617.14 713,505 3.79% 681,917

optimization (CH)

563
V. Penadés-Plà et al. Engineering Structures 179 (2019) 556–565

the mean difference of the real embodied energy and the estimated means of commercial software. In this way, structural engineers can
embodied energy of nine random points. The second and third columns obtain the response objective of a small sample size through commer-
show the mean computational time of nine optimized solutions in cial software without the necessity of writing code, and after that,
seconds and the percentage with respect to the conventional heuristic achieve an optimized structure in simple terms. Thus, the use of the
optimization. The fourth and fifth columns show the mean embodied kriging model in structural design has a high potential in the research
energy of nine optimized solutions in kWh and the percentage with field as well as practical engineering.
respect to the conventional heuristic optimization. The sixth column
shows the coefficient of variance of the nine optimized solutions. Fi- Acknowledgments
nally, the seventh and eighth columns show the best optimized solution
of each group of nine in kWh and the percentage with respect to the The authors acknowledge the financial support of the Spanish
conventional heuristic optimization. This table can be used as a re- Ministry of Economy and Competitiveness, along with FEDER funding
ference for defining the initial sample size. Note that the design space of (Project: BIA2017-85098-R).
this work is formed by 8 variables. Depending on the preferred char-
acteristics, one sample size will be adjusted more than the others. References
However, taking into account all of the characteristics, the initial
sample size that shows the best behavior is N = 50. This initial sample [1] Gervásio H, Simões Da Silva L. A probabilistic decision-making approach for the
size provides a satisfying mean embodied energy (783,726 kWh) with a sustainable assessment of infrastructures. Expert Syst Appl 2012;39:7121–31.
https://doi.org/10.1016/j.eswa.2012.01.032.
low coefficient of variance (6.65%) and gives the best solution [2] Aghdaie MH, Zolfani SH, Zavadskas EK. Prioritizing constructing projects of mu-
(699,240 kWh) whose cross section variables are b = 1.2 m, nicipalities based on AHP and COPRAS-G: a case study about footbridges in Iran.
h = 1.35 m, d = 0 m, ev = 0.15 m, es = 0.15 m, ea = 0.35 m, Balt J Road Bridg Eng 2012;7:145–53. https://doi.org/10.3846/bjrbe.2012.20.
[3] Zavadskas E, Antucheviciene J, Vilutiene T, Adeli H. Sustainable decision-making in
ei = 0.15 m, and fck = 60 MPa. These results have been obtained with a civil engineering, construct building technology. Sustainability 2017;10:14. https://
99.06% reduction in time spent with respect to the conventional doi.org/10.3390/su10010014.
heuristic optimization, whose cross section variables are b = 1.35 m, [4] Penadés-Plà V, García-Segura T, Martí J, Yepes V. A review of multi-criteria deci-
sion-making methods applied to the sustainable bridge design. Sustainability
h = 1.3 m, d = 0 m, ev = 0.15 m, es = 0.2 m, ea = 0.4 m, ei = 0.2 m,
2016;8:1295. https://doi.org/10.3390/su8121295.
and fck = 50 MPa. In addition, Figs. 8–10 show that the initial sample [5] Ferreiro-Cabello J, Fraile-Garcia E, Martinez-Camara E, Perez-de-la-Parte M.
size of N = 50 is close to the results of N = 500, but saving 89.74% of Sensitivity analysis of life cycle assessment to select reinforced concrete structures
with one-way slabs. Eng Struct 2017;132:586–96. https://doi.org/10.1016/J.
the computational cost. However, as mentioned previously, the sample
ENGSTRUCT.2016.11.059.
size of N = 500 improves the coefficient of variance. [6] Penadés-Plà V, Martí JV, García-Segura T, Yepes V. Life-cycle assessment: a com-
parison between two optimal post-tensioned concrete box-girder road bridges.
6. Conclusions Sustainability 2017;9:1864. https://doi.org/10.3390/su9101864.
[7] Gervasio H, Simoes da Silva L. Comparative life-cycle analysis of steel-concrete
composite bridges. Struct Infrastruct Eng 2008;4:251–69. https://doi.org/10.1080/
In this work, a conventional heuristic optimization and a kriging- 15732470600627325.
based heuristic optimization have been compared. The results show [8] Camp CV, Huq F. CO2 and cost optimization of reinforced concrete frames using a
big bang-big crunch algorithm. Eng Struct 2013;48:363–72. https://doi.org/10.
that the use of the kriging model provides a response surface with a 1016/j.engstruct.2012.09.004.
goof accuracy that improves with an increase in the initial sample size. [9] Yepes V, Martí JV, García-Segura T. Cost and CO2 emission optimization of precast-
Therefore, the objective response of a problem can be obtained without prestressed concrete U-beam road bridges by a hybrid glowworm swarm algorithm.
Autom Constr 2015;49:123–34. https://doi.org/10.1016/j.autcon.2014.10.013.
any structural analysis and with a high accuracy. The results of kriging- [10] Molina-Moreno F, Martí JV, Yepes V. Carbon embodied optimization for buttressed
based heuristic optimization are close to the solutions reached in the earth-retaining walls: implications for low-carbon conceptual designs. J Clean Prod
conventional heuristic optimization cases with a significantly high re- 2017;164:872–84. https://doi.org/10.1016/j.jclepro.2017.06.246.
[11] García-Segura T, Yepes V. Multiobjective optimization of post-tensioned concrete
duction of computational cost. box-girder road bridges considering cost, CO2 emissions, and safety. Eng Struct
The sensitivity analysis of the penalization imposed on the un- 2016;125:325–36. https://doi.org/10.1016/j.engstruct.2016.07.012.
feasible designs shows that the kriging model has a better behavior with [12] Martí JV, García-Segura T, Yepes V. Structural design of precast-prestressed con-
crete U-beam road bridges based on embodied energy. J Clean Prod
the lowest penalization. In addition, the study of the optimization ob-
2016;120:231–40. https://doi.org/10.1016/j.jclepro.2016.02.024.
tained according to the initial sample size shows that the best solutions [13] Park H, Kwon B, Shin Y, Kim Y, Hong T, Choi S. Cost and CO2 emission optimization
obtained are similar for the different sample sizes, but that the mean of steel reinforced concrete columns in high-rise buildings. Energies
and the coefficient of variance improve with the initial sample size. We 2013;6:5609–24. https://doi.org/10.3390/en6115609.
[14] Nigdeli SM, Bekdas G, Kim S, Geem ZW. A novel harmony search based optimi-
can conclude that the initial sample size that performs best is N = 50. zation of reinforced concrete biaxially loaded columns. Struct Eng Mech
For this case, the accuracy of the response surface is within 4.04% and 2015;54:1097–109. https://doi.org/10.12989/sem.2015.54.6.1097.
the mean energy of the optimum solutions differ by 9.84% compared to [15] de Albuquerque AT, El Debs MK, Melo AMC. A cost optimization-based design of
precast concrete floors using genetic algorithms. Autom Constr 2012;22:348–56.
the conventional heuristic optimization, but with a reduction in the https://doi.org/10.1016/J.AUTCON.2011.09.013.
computational cost of the 99.06%. Regarding the best solution, the [16] García-Segura T, Yepes V, Alcalá J, Pérez-López E. Hybrid harmony search for
comparison shows that the use of kriging increases the optimum energy sustainable design of post-tensioned concrete box-girder pedestrian bridges. Eng
Struct 2015;92:112–22. https://doi.org/10.1016/j.engstruct.2015.03.015.
by 2.54%. However, if the main objective is to reduce the coefficient of [17] Simpson TW, Booker AJ, Ghosh D, Giunta AA, Koch PN, Yang R-J. Approximation
variance, the initial size that performs better is N = 500. For this case, methods in multidisciplinary analysis and optimization: a panel discussion. Struct
the solutions obtained have a coefficient of variance of 3.67%, even Multidiscip Optim 2004;27:302–13. https://doi.org/10.1007/s00158-004-0389-9.
[18] Bäckryd RD, Ryberg A-B, Nilsson L. Multidisciplinary design optimisation methods
lower than the 3.79% that corresponds to the conventional heuristic for automotive structures. Int J Automot Mech Eng 2017;14:2229–8649. https://
optimization. Thus, structural engineers must consider an appropriate doi.org/10.15282/ijame.14.1.2017.17.0327.
initial sample size depending on the characteristics of the problem. [19] Simpson TW, Poplinski JD, Koch PN, Allen JK. Metamodels for computer-based
engineering design: survey and recommendations. Eng Comput 2001;17:129–50.
In conclusion, the use of the kriging model in structural design of-
https://doi.org/10.1007/PL00007198.
fers a new way to solve a number of structural problems that require a [20] Cressie N. The origins of kriging. Math Geol 1990;22:239–52. https://doi.org/10.
very high computational cost and reduces the difficulty of other pro- 1007/BF00889887.
blems. On the one hand, due to the lower computational cost, kriging- [21] García-Segura T, Yepes V, Frangopol DM, Yang DY. Lifetime reliability-based op-
timization of post-tensioned box-girder bridges. Eng Struct 2017;145:381–91.
based heuristic optimization can be used to obtain the best solution for [22] Kutylowski R, Rasiak B. Application of topology optimization to bridge girder de-
problems involving several criteria and yields robust designs. On the sign. Struct Eng Mech 2014;51:39–66. https://doi.org/10.12989/sem.2014.51.1.
other hand, kriging-based heuristic optimization can be used to opti- 039.
[23] Martí JV, Yepes V, González-Vidosa F. Memetic algorithm approach to designing
mize structural problems with a lower number of design variables by

564
V. Penadés-Plà et al. Engineering Structures 179 (2019) 556–565

precast-prestressed concrete road bridges with steel fiber reinforcement. J Struct [34] Kim B-S, Lee Y-B, Choi D-H. Comparison study on the accuracy of metamodeling
Eng 2015;141:4014114. https://doi.org/10.1061/(ASCE)ST.1943-541X.0001058. technique for non-convex functions. J Mech Sci Technol 2009;23:1175–81. https://
[24] García-Segura T, Yepes V, Frangopol DM. Multi-objective design of post-tensioned doi.org/10.1007/s12206-008-1201-3.
concrete road bridges using artificial neural networks. Struct Multidiscip Optim [35] Forrester AIJ, Keane AJ. Recent advances in surrogate-based optimization. Prog
2017;56:139–50. https://doi.org/10.1007/s00158-017-1653-0. Aerosp Sci 2009;45:50–79. https://doi.org/10.1016/J.PAEROSCI.2008.11.001.
[25] Barton RR, Meckesheimer M. Metamodel-based simulation optimization. vol. 13; n. [36] Matheron G. Principles of geostatistics. Econ Geol 1963;58:1246–66.
d. https://doi.org/10.1016/S0927-0507(06)13018-2. [37] Simpson TW, Mauery TM, Korte J, Mistree F. Kriging models for global approx-
[26] Myers RH, Montgomery DC, Anderson-Cook CM. Response surface methodology: imation in simulation-based multidisciplinary design optimization. AIAA J
process and product optimization using designed experiments. Toronto (Canada): 2001;39:2233–41. https://doi.org/10.2514/3.15017.
Wiley; 1995. [38] Schlaich J, Scheef H. Concrete box-girder bridges. Zürich (Switzerland): Int. Assoc.
[27] McKay MD, Beckman RJ, Conover WJ. Comparison of three methods for selecting Bridg. Struct. Eng; 1982.
values of input variables in the analysis of output from a computer code. [39] Catalonia Institute of Construction Technology. BEDEC PR/PCT ITEC material da-
Technometrics 1979;21:239–45. https://doi.org/10.1080/00401706.1979. tabase; 2016.
10489755. [40] Ministerio de Fomento. EHE-08: Code on structural concrete. Madrid (Spain); 2008.
[28] Johnson ME, Moore LM, Ylvisaker D. Minimax and maximin distance designs. J Stat [41] Ministerio de Fomento. IAP-11: Code on the actions for the design of road bridges.
Plan Inference 1990;26:131–48. https://doi.org/10.1016/0378-3758(90)90122-B. Madrid (Spain); 2011.
[29] Kalagnanam JR, Diwekar UM. An efficient sampling technique for off-line quality [42] European Committee for Standardization. EN 1001-2:2003. Eurocode 1: Actions on
control. Technometrics 1997;39:308. https://doi.org/10.2307/1271135. structures – Part 2: Traffic loads bridges. Brussels (Belgium); 2003.
[30] Fang K-T, Lin DKJ, Winker P, Zhang Y. Uniform design: Theory and application. [43] European Committee for Standardisation. EN1992-2:2005. Eurocode 2: Design of
Technometrics 2000;42:237. https://doi.org/10.2307/1271079. concrete structures – Part 2: Concrete Bridge-Design and detailing rules. Brussels;
[31] Chuang CH, Yang RJ, Li G, Mallela K, Pothuraju P. Multidisciplinary design opti- 2005.
mization on vehicle tailor rolled blank design. Struct Multidiscip Optim [44] Kirkpatrick S, Gelatt CD, Vecchi MP. Optimization by simulated annealing. Science
2008;35:551–60. https://doi.org/10.1007/s00158-007-0152-0. 1983;220:671–80. https://doi.org/10.1126/science.220.4598.671.
[32] Jin R, Chen W, Simpson TW. Comparative studies of metamodelling techniques [45] Martí JV, González-Vidosa F, Yepes V, Alcalá J. Design of prestressed concrete
under multiple modelling criteria. Struct Multidiscip Optim 2001;23:1–13. https:// precast road bridges with hybrid simulated annealing. Eng Struct 2013;48:342–52.
doi.org/10.1007/s00158-001-0160-4. https://doi.org/10.1016/j.engstruct.2012.09.014.
[33] Li YF, Ng SH, Xie M, Goh TN. A systematic comparison of metamodeling techniques [46] Medina JR. Estimation of incident and reflected waves using simulated annealing. J
for simulation optimization in decision support systems. Appl Soft Comput Watery Port Coast Ocean Eng 2001;127:213–21.
2010;10:1257–73. https://doi.org/10.1016/J.ASOC.2009.11.034.

565