Construction and Building Materials 260 (2020) 120457

Contents lists available at ScienceDirect

Construction and Building Materials

journal homepage: www.elsevier.com/locate/conbuildmat

Intelligent mixture design of steel fibre reinforced concrete using a

support vector regression and firefly algorithm based multi-objective
optimization model
Yimiao Huang a, Junfei Zhang a,⇑, Foo Tze Ann a, Guowei Ma a,b
a
Department of Civil, Environmental and Mining Engineering, The University of Western Australia, Perth 6009, Australia
b
School of Civil and Transportation Engineering, Hebei University of Technology, 5340 Xiping Road, Beichen District, Tianjin 300401, China

h i g h l i g h t s

The strength of SFRC are predicted using FA-SVR;

Multi-objective firefly algorithm is proposed based on the basic firefly algorithm;

The mixture proportion of SFRC is automatically designed using a multi-objective optimization method.

a r t i c l e i n f o a b s t r a c t

Article history: Steel fibre reinforced concrete (SFRC) is widely used in the construction concrete industry as it partakes
Received 16 March 2020 an important role of evolving concrete technology. It consists of steel fibres of various shapes, sizes and
Received in revised form 23 June 2020 geometries that influence the concrete mix composition and mechanical properties. However, compared
Accepted 31 July 2020
to traditional concrete, it is difficult to design the mix proportions because more influencing variables
Available online 3 September 2020
need to be considered to optimise multiple properties including ultimate compressive strength, tensile
or flexural strength and cost. Therefore, the present study proposes an artificial intelligence based
Keywords:
multi-objective optimization model to enable an efficient method of finding the optimum mix design
Steel fibre reinforced concrete
Mixture design
for SFRC. A large dataset including 299 instances for uniaxial compressive strength (UCS) test and 269
Multi-objective optimization instances for flexural strength (FS) test were collected from previous literature. Support vector regression
Uniaxial compressive strength (SVR) model was applied to predict UCS and FS for SFRC. The hyper parameters of SVR models were tuned
Flexural strength using a firefly algorithm (FA) and a sensitivity study was conducted to understand the importance of the
inputs on the output variables for the algorithms. High correlation coefficients (0.91 for UCS and 0.85 for
FS) were achieved on the test dataset. The FA-SVR model was then applied as the objective function for a
developed multi-objective FA to search for the optimal SFRC mixture proportion. Pareto optimal solutions
were obtained and served as a design guide to determine the optimal SFRC mixtures.
Ó 2020 Elsevier Ltd. All rights reserved.

1. Introduction ties [2,3]. Thus, it is challenging for traditional mixture design

Steel Fibre Reinforced Concrete (SFRC) represents a new class
of construction concrete that combines a cementitious matrix
and a discontinuous reinforcement, consisting of steel fibres ran-
domly distributed in the matrix [1]. SFRC has many advantages
including enhanced tensile strength and improvedconcrete dura-
bility compared with traditional concrete. Steel fibres are pro-
duced with various shapes, sizes and geometrics, which
influence the concrete mix compositions and mechanical proper-
⇑ Corresponding author.
E-mail address: junfei.zhang@research.uwa.edu.au (J. Zhang).
methods to find the optimum concrete design due to the com-
plexity of the SFRC mixtures.
Traditional mixture proportion designs are often trial and error
processes which can be time- and cost-consuming [2]. Meanwhile,
traditional methods used for the prediction of mechanical proper-
ties are mainly based on statistical analysis with linear and nonlin-
ear regression equations. Yet, the choice of suitable regression
equations is not a simple task as it requires technique and experi-
ence [4]. Furthermore, the implementation of explicit mathemati-
cal regression equation using empirical-based models could not
establish the highly nonlinear relationship, and thus resulting in
inaccurate outcomes [5,71–73]. Therefore, conventional methods

https://doi.org/10.1016/j.conbuildmat.2020.120457
0950-0618/Ó 2020 Elsevier Ltd. All rights reserved.
2 Y. Huang et al. / Construction and Building Materials 260 (2020) 120457
may be insufficient for modelling complex nonlinear behaviour
and uncertainties.
To tackle the issues behind traditional methods, recent studies
have focused more on using computational optimisation methods
for predicting mechanical properties of SFRC, e.g. the use of Arti-
ficial Neural Network (ANN) to predict concrete compressive or
shear strength [6–10]. One common machine learning (ML)
approach known as the support vector regression (SVR) is another
powerful modelling tool that has been widely used in the data
mining field due to its good generalisation capacity and fast com-
putation ability [11,76]. Many researchers have remarked that the
SVR models can be more efficient and accurate than other AI-
based models in some cases [11–13]. The SVR model is found
to be able to solve both classification and either linear or non-
linear regression problems [14]. Meanwhile, the SVR model is
able to figure a linear regression function for a new higher dimen-
sional feature space while reducing model complexity at the same
time [15,74].
Although the SVR model shows the high effectiveness in
solving prediction problems, one of its major weakness is that
SVR performance is highly dependent on its hyper parameters
which are difficult to tune using traditional optimization meth-
ods [16,75]. Usually, the hyper parameters of SVR are tuned by
the grid search method of which the computation complexity is
very high and hence it can only be applied to solving problems
with very few parameters [17]. To address this problem, nature-
inspired optimization algorithms have been applied in this area,
including particle swarm optimization [18], genetic algorithm
[19] and firefly algorithm (FA) [20]. Among these optimization
algorithms, FA has been widely used in tuning hyper parameter
of ML models due to two advantages over the others: diminish-
ing multimodality and automatic subdivision [21]. Therefore, FA
is used in this study to search for optimal hyper parameters of
SVR.
SFRC mixture optimization is a two-step approach. The first
step is to model the compressive strength of concrete, while the
second step is to search for optimal mixture proportions using
metaheuristic algorithms [77,78]. For the first step, ML models
have been used to model the objectives of concrete [22,23], while
for the second step, current literature mainly focuses on single
objective optimization (usually cost or UCS). For instance, Cheng
et al. optimized high-performance concrete mixture using support
vector machine and genetic algorithm [19]. Behnood and Golaf-
shani used biogeography-based programming to estimate the opti-
mal mix design of silica fume concrete [24]. Naseri optimized cost
of no-slump concrete using genetic algorithm and particle swarm
opitimization [25]. However, in real-world situations, concrete
mixture designs are complicated with multiple objectives and non-
linear constraints. Thus, it is necessary to develop a multi-objective
optimization (MOO) model that can optimise all the objectives
simultaneously.
Therefore, in this study, a MOO model using both FA-SVR and
MOFA is developed to predict mechanical properties and search
optimal mixtures for SFRC. A large dataset including 299
instances for uniaxial compressive strength (UCS) test and 269
instances for flexural strength (FS) test were collected for the
study. Based on the dataset, the mathematical relationships
between input variables and the mechanical properties of SFRC
(UCS and FS) were modelled by the FA-SVR algorithm. Conse-
quently, the multi-objective firefly algorithm (MOFA), was
applied to obtain the Pareto optimised solutions. Furthermore,
a sensitivity study on the variables was also conducted to com-
prehend the relationship and check the dependency between
input and output variables to enhance the understanding of
the variables of SFRC for finding a more accurate SFRC optimum
mix design.
2. Methodology
Fig. 1 shows the whole progress of applying the proposed MOO
model to the SFRC database to predict mechanical properties and
optimise mixture designs for SFRC. Detailed explanation of each
stage is given in the following sections.
2.1. Data description
The database was collected from published journal articles,
including 299 samples for UCS prediction and 269 samples for FS
prediction [26–53]. The input variables include the weight of
cement, coarse aggregate, fine aggregate, superplasticiser, the vol-
ume fraction of steel fibre, and the length and diameter of steel
fibre. The output variables include UCS and FS of SFRC. The statis-
tics of the parameter are summarised in Table 1 and Table 2.
The relationships between the input variables are presented
using a correlation matrix in Fig. 2. The correlations between the
majorities of any two input variables are low (less than 0.5). This
means that those variables are independent of each other, and
thus, they are regarded as important SFRC input variables and no
multicollinearity issue will be produced [54]. It is evident that
water and cement is highly correlated. Therefore, these two
parameters can be replaced by water-to-cement-ratio when pre-
dicting strength properties of SFRC. However, this paper aims to
find optimal mixture proportions of SFRC, i.e.: to find the optimal
amount of each variable to achieve both mechanical and econom-
ical objectives. When the MOFA creates a new mixture proportion,
the volume of each variable should be calculated and the total vol-
ume of the mixture should be 1 m3. Therefore, we must use the
amount instead of the ratio.
2.2. Objective function: FA-SVR based model
2.2.1. Support vector regression (SVR)
Support vector regression (SVR) developed by Vapnik (1995)
has been commonly used in regression. SVR converts nonlinear
problems to linear problems by mapping data from sample space
to higher dimensional characteristic space [55]. This is done by
using kernel parameters to understand the complex relationship
between inputs and outputs. SVR can efficiently withstand noises
[56] and it has brilliant generalisation capability and fast learning
speed [57].
A training dataset of n points is given as follow:
fðx1 ; y1 Þ; ðx2 ; y2 Þ;    ; ðxn ; yn Þg ð1Þ
where each xi is a l-dimensional real director and yi is the scalar
regression value. The regression function can be derived for this
dataset as follow [56]:
f ðxÞ ¼ w  uðxÞ þ b ð2Þ
where uðxÞ is a nonlinear mapping function, w represents the
weight vector, b is the bias; f ðxÞ is required to be as flat as possible.
Flatness in Eq (2) represents the Euclidean norm of w, e.g.
kwk2 needs to be minimised [55]. Function f ðxi Þ could be found if
the deviation between f ðxi Þ and yi is less than e (the largest toler-
ance error) for each instance of xi .
To measure the deviation degree, the e-insensitive factor is used
for the loss function as follows:
(




0;
yi  f ðxi Þ
< e
Lðx; y; f Þ ¼
yi  f ðxi Þ
e ¼



ð3Þ

yi  f ðx Þ
 ei ;
yi  f ðx Þ
 e
i i
This loss function shows that only data points situated inside or
outside e-tube will be used as support vectors to construct f ðxÞ
and the training points within the e-tube will not be penalised. In
 Y. Huang et al. / Construction and Building Materials 260 (2020) 120457 3

SVR hyperparameter tuning using FA

Initialize population of firefly

Data set Optimal mixtures of SFRC and

Calculate the fitness value of each Multi-objective model
firefly its objectives

Polynomial model for actual cost Cost of the optimal mixtures

SVR
Training set (70%) Update the light intensity of fireflies
(for UCS)

Rank the fireflies and update the Get FA-SVR model with optimum
UCS of the optimal mixtures
position hyperparameters (for UCS)
10-fold CV in training set
No
Is the stop criterion reached? Find the optimal solution using
Optimal mixtures of SFRC
MOFA
Yes
Test set (30%)
LSSVR hyperparameter tuning using FA
SVR Get FA-SVR model with optimum
FA UCS of the optimal mixtures
(for FS) hyperparameters (for FS)

Fig. 1. Flowchart of the SVR-MOFA system for SFRC mixture4 design.

Table 1
Statistics of input and output variables for the UCS dataset.

Variables Minimum Maximum Mean Std Dev.

Cement, C (kg/m3) 283.00 1174.00 477.48 174.01
Water, W (kg/m3) 122.00 534.50 201.62 77.01
Coarse aggregate, CA (kg/m3) 305 1629.80 1048.33 262.28
Fine aggregate, FA (kg/m3) 233 1355.90 808.94 227.89
Superplasticiser, SP (kg/m3) 0.00 137.50 15.32 30.28
Volume fraction, Vf (%) 0.00 75 3.37 11.59
Length, Lf (mm) 0.00 60.00 38.17 19.07
Diameter, df (mm) 0.00 31.00 1.18 3.94
UCS (MPa) 23.30 190.94 54.35 20.66

Table 2
Statistics of input and output variables for the FS dataset.

Variables Minimum Maximum Mean Std Dev.

Cement, C (kg/m3) 150.00 1031.30 602.84 295.21
Water, W (kg/m3) 82.50 534.50 291.66 162.75
Coarse aggregate, CA (kg/m3) 305 1629.80 1171.26 325.71
Fine aggregate, FA (kg/m3) 150 1355.90 922.53 317.91
Superplasticiser, SP (kg/m3) 0.00 67.50 11.48 19.31
Volume fraction, Vf (%) 0.00 6.00 0.73 0.74
Length, Lf (mm) 0.00 83.00 34.20 18.43
Diameter, df (mm) 0.00 1.60 0.85 0.49
FS (MPa) 0.00 17.30 6.44 2.92

accordance to the structural risk minimisation [56], the problem out the tube. An example of a nonlinear SVR with an e-tube is
can be written as follow: shown in Fig. 3.
Xn To resolve the problems with constraints, Lagrange multipliers
1
RðwÞ ¼ kwk2 þ Lðx; y; f Þ ð4Þ can be used as follows [56]:
2 i¼1

Some errors are allowed, thus, slack variables ni and ni are used to 1 Xn
Lðw; b; n; a; lÞ ¼ kwk2 þ C ðni þ ni Þ
help with the infeasible constraints. Eq (4) can be then changed into 2 i¼1
a convex optimisation function as follow:
X
n X
n

1 Xn  ai ðe þ ni  yi þ w  uðxi Þ þ bÞ  ai ðe
minw;b;n;n RðwÞ ¼ kwk2 þ C ðni þ ni Þ i¼1 i¼1
2 i¼1 X
n
þ ni þ yi  w  uðxi Þ  bÞ  ðli ni þ li ni Þ ð6Þ
8
> yi  w  uðxÞ  b  e þ ni i¼1
>
>
< w  uðxÞ þ b  y  e þ n
i i
s:t ð5Þ where ai  0; ai  0, li  0 and li  0 are Lagrange multipliers.
>
>
> n i  0
: When the constraint functions heavily oppose one another and
ni  0
objective function is differentiable, Karush-Kuhn-Tucker (KKT) con-
where C is the penalty parameter to determine the differences ditions are to be satisfied for each pair of primal and dual optimal
between the flatness of f ðxÞ and the penalising extent of the sample points [59] in the following manner:
4 Y. Huang et al. / Construction and Building Materials 260 (2020) 120457

8
>
>
ai ðe þ ni  yi þ w  uðxi Þ þ bÞ ¼ 0
>
< a ðe þ n þ y  w  uðx Þ  bÞ ¼ 0
i i i i
ð8Þ
>
>
> ðC  a i Þni ¼ 0
:
ðC  ai Þni ¼ 0

The Langrage dual problem can be derived after solving the above
equations as follows:
!
1X n X n
   Xn
  X n
 
maxi  ai  ai aj  aj xTj xj  e ai  ai þ yi ai  ai
2 i¼1 j¼1 i¼1 i¼1

( Pn  
i¼1 ai  ai ¼ 0
s:t ð9Þ
ai ; ai 2 ½0; C
From Eq. (8), the weight vector can be obtained as
P  
w ¼ ni¼1 ai  ai uðxi Þ, and thus, the regression function can be
derived as:
X
n
 
f ð xÞ ¼ ai  ai uðxi Þx þ b ð10Þ
i¼1

2.2.2. Firefly algorithm

Firefly algorithm (FA) is inspired by the social behaviour of fire-
flies as it simulates the flashing characteristics of fireflies. The fol-
lowing three rules are used for this algorithm [60]:

Fireflies are attracted to each other by the amount of brightness

regardless of gender.
The attractiveness corresponds to the brightness of the firefly,
which both decreases as the distances increases between the
two fireflies. If there are no brighter firefly than oneself, the fire-
fly will fly randomly in the neighbourhood.
The brightness of the firefly is governed by the objective func-
tion of a specific problem.

Fig. 2. Correlation matrix of input variables for UCS dataset (a) and FS dataset (b). A firefly travels towards one with higher brightness based on
the following equation [61]:
 
cr 2ij
xtþ1
i ¼ xti þ b0 e xtj  xti þ aðrand  0:5Þ ð11Þ

where r ij ¼ kxtj  xti k is the Euclidian distance between two fireflies

at positions xtj and xti at the tth iteration; b0 is the attractiveness of
the firefly at r = 0; c is the absorption coefficient which varies
within the range from 0 to 1; a represents the randomisation
parameter; rand 2 ½0; 1 is a vector of random number drawn from
the Gaussian distribution; The adopted pseudocode of FA is shown
in Fig. 4.

2.3. Hyperparameter tuning

2.3.1. Cross fold validation

This study uses FA to tune two essential hyperparameters when
using SVR: the penalty coefficient c and Gaussian kernel parameter
Fig. 3. A support vector regression machine [58]. c. The data points from the UCS database is randomly divided into
70% for training sets and 30% for testing sets, likewise for the FS
8 P dataset [17]. Subsequently, 10-fold cross-validation (CV) is applied
>
>
@L
@w
¼ w  ni¼1 ðai  ai Þuðxi Þ ¼ 0 in the training sets, which splits the datasets into 10 mutually
>
> P
< @L
¼ n ða  a Þ ¼ 0 exclusive subsets (or fold) [15] as shown in Fig. 5. The validation
@b i
i¼1 i
ð7Þ fold is a subset that is used to validate the performance of the
>
>
> C  ai  li ¼ 0
>
: SVR model, which results in the optimal hyper parameters in this
C  ai  li ¼ 0
fold. The training fold consists of the other 9 subsets used to tune
the hyper parameters for 50 iterations. The SVR model with the
According to the KKT conditions, the product between the con- minimum root means square error (RSME) is chosen after the val-
straints and dual variables is 0 at optimal solutions: idation of the different subsets for each fold. The selected hyper
 Y. Huang et al. / Construction and Building Materials 260 (2020) 120457 5

where n is the total number of data samples; yi and yi are the pre-
dicted and actual outputs respectively; y and y are mean values
of the predicted and observed outputs respectively.

2.4. Variable importance measure

A machine learning model (e.g. SVR) can be highly complex, and

as a result, its relationships between inputs and outputs may be
poorly understood. In such cases, the model can be viewed as a
black box, i.e. the output is an ‘‘opaque” function of its inputs. To
increase interpretability from black box machine learning models,
we applies a novel visualization approach based on a Sensitivity
Analysis (SA) that is able to measure the influence on the output
of the proposed SVR when the inputs are varied through their
range of values [62]. SA allows a ranking of the inputs based on
the amount of output changes that are produced due to distur-
bances in a given input. Therefore, SA has been widely used to
explain the model to open the black box [11,12,63]. The sensitivity
is approximated by measuring how the results are affected by the
Fig. 4. Pseudocode for firefly algorithm [60]. change of input variables. SA can be applied locally or globally,
however, local sensitivity cannot locate uncertainty of the model.
parameters are then used to predict UCS and FS of SFRC for the Whereas, global sensitivity analysis (GSA) allows all input vari-
study. ables to be altered concurrently when the whole input and output
spaces are considered. Thus, GSA is used and modelled through a
2.3.2. Performance evaluation bar plot to show the importance of variable (from 0 to 100%) [64].
This study uses the following performance measures to evalu- The procedure is as follows:
ate the prediction accuracy of the SVR model: Root Mean Square
Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percent- 1. Input and output variables are designed in the probability
age Error (MAPE), and correlation coefficient (R). These perfor- space.
mance measures are calculated as follows [15]: 2. By changing one variable at each time and setting the remaining
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi parameters are at their mean values, the equivalent sensitivity
u n
u1 X index for each input variable is calculated.
RSME ¼ t ðy  yi Þ2 ð12Þ
n i¼1 i
A gradient metric was used to measure the changes produced
X
n

by the output based on previous research [64]:
1
y  y

MAE ¼ i ð13Þ

L
c




n i X ya;j  yd
a;j1
i¼1
ga ¼ ð16Þ
L  1
n


j¼2
X  yi


MAPE ¼
1
yi ð14Þ
n
y
where a is the input variable under analysis; yc a;j is the sensitivity
i¼1 i
response for xa;j . Then, the relative importance (Ra ) is derived as
Pn  follow:
i¼1 ðyi  y Þðyi  yÞ

R ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffisffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð15Þ
P n  2 P n  2 X
I
yi  y yi  y Ra ¼ g a = gi ð17Þ
i¼1 i¼1 i¼1

Fig. 5. 10-fold cross validation.

6 Y. Huang et al. / Construction and Building Materials 260 (2020) 120457

2.5. Multi-objective optimisation Table 4

Set constraints of SFRC input variables.

Solving a MOO problem is more complicated than solving a Mix parameters Expressions Lower Upper
single-objective optimisation problem [65]. In this study, the bound bound
MOFA is used to solve the SFRC mixture design problem. Cement content C(kg/m3) 300 1200
Fine aggregate content FA(kg/m3) 0 1356
2.5.1. Objective function W/C Ratio C w =C C 0.25 0.6
Steel fibre volume fraction Vf 0 0.06
As mentioned above, the FA-SVR model is used as the objective
Fine aggregate ratio C FA =ðC FA þ C CAÞ 0.3 0.6
function for UCS and FS. The third objective function (cost) is cal- Superplasticiser to cement C SP =ðC SP þ C C ) 0 0.03
culated by polynomial equations as follows: ratio

Costð=m3 Þ ¼ C c Q c þ C w Q w þ C fa Q fa þ C ca Q ca þ C SP Q SP þ C sf Q sf
ð18Þ
haps, the simplest one is to combine all the multiple objectives
where Table 3 identifies the estimated cost and unit weight of each (fk) into a composite single objective (f) using a weighted sum
variable of SFRC used in this study. Qc, Qw, Qfa, Qca and Qsf are the method as follows:
quantity (in kg/m3) for cement, water, fine and coarse aggregate,
X
K X
K
superplasticiser and steel fibre, respectively. f ¼ wk f k ; wk ¼ 1 ð21Þ
K¼1 K¼1
2.5.2. Constraints
To solve the MOO problems, constraints must be set as follows: where weights wk ¼ pKk ; pk are the random numbers generated from
a uniform distribution [0,1] and K represents the uniformly dis-

Range constraints tributed number.
For this study, the multi-objective functions using the weighted
There should be a variance amongst decision variables within a sum method are as follow:
definite range set by the datasets for SFRC mixture optimisation
J 1 ¼ w1  UCS þ w2  cost ð22Þ
which are as follow:
Dimin  Di  Dimax ð19Þ J 2 ¼ w1  FS þ w2  cost ð23Þ
where Di represents the ith decision variable with minimum value
(Dimin) and maximum value (Dimax). X
2
wi ¼ 1 ð24Þ
i¼1

Concrete volume constraints
Since several objectives need to be minimised simultaneously in the
The sum of the concrete must add up to one cubic meter as MOO problems, non-dominated solutions (Pareto optimal fronts)
follow: need to be obtained [66]. The Pareto optimal solutions are known
as ‘‘non-dominated” solutions [67], which means no other objective
Q c Q w Q fa Q ca Q SP Q sf
Vm ¼ þ þ þ þ þ ð20Þ functions can be enhanced without deteriorating other objective
U c U w U fa U ca U SP U sf
functions. The Pareto optimum solution is defined as follows [66]:
where Uc, Uw, Ufa, Uca, USP and Usf are the unit weight of cement, If x 2 S whereby S is the set of feasible solutions,
water, fine aggregate, coarse aggregate, superplasticiser and steel
for all i 2 f1; 2;    ; kg; f i ðxÞ  f i ðx Þ and ð25Þ
fibre, respectively.

Ratio constraints for at least one index j 2 f1; 2;    ; kg; f j ðxÞ < f j ðx Þ ð26Þ

The Pareto optimal solution (x*) is obtained if f(x*) dominates

To solve for SFRC mixture optimisation issues, there is a need to f ðxÞfor every x. The Pareto front is known as the set of Pareto opti-
constrain a series of ratios such as water-to-cement ratio, mal solutions. Fig. 6 illustrates the set of non-dominated solutions
superplasticiser-to-cement ratio and fine aggregate-to-total aggre- such as Point A and B are on the Pareto front. It can be seen for f1,
gate ratio.
Table 4 contains the summarised set constraints of SFRC input
variables for the Pareto front optimisation to predict the optimised
solutions. The set constraints were based on the datasets formed.

2.5.3. Construction of MOFA

There are a variety of ways to dealing with MOO problems using
well-established single-objective optimization algorithms. Per-

Table 3
Unit cost of each variables of SFRC.

Variables Notation Unit price ($/kg) Unit weight (kg/m3)

Cement Cc 0.55 3150
Water Cw 0.001 1000
Fine aggregate Cfa 0.23 1600
Coarse aggregate Cca 0.25 2700
Superplasticiser CSP 1.20 1350
Steel fibre Csf 1 7800 Fig. 6. Model of a Pareto line with feasible points [69].
 Y. Huang et al. / Construction and Building Materials 260 (2020) 120457 7

Fig. 1. The three objective functions are identified as: the trained
model for UCS and FS and the cost of the SFRC mixture. The sum-
marised steps below are used to find the optimal mixture design
for SFRC:

1. Train two SVR models by employing UCS and FS databases

respectively and apply the 10-fold CV and FA to find the optimal
hyperparameters for SVR.
2. Decide on the cost of each SFRC mixture.
3. Develop MOFA using the weighted sum method.
4. Explore the optimum outcomes that can minimise J i with
MOFA.
5. Acquire Pareto fronts of the dual-objective optimum SFRC mix-
ture design.

3. Results and discussions

Fig. 7. Pseudocode for MOFA.
3.1. Results of hyperparameter tuning

B < A, whereas for f2, A < B. The set of feasible solutions such as
Point C is not on the Pareto front as it can be easily dominated
by both A and B.
By this weighted sum method, the FA can be extended to MOFA
to search for Pareto optimal solutions for multi-objective optimiza-
tion problems [68]. The pseudocode is summarized in Fig. 7.
2.5.4. Optimal mixture design of SFRC
The process for searching the optimal mixture design of SFRC
using SVR and MOFA can be stated as follows and displayed in
As discussed above, the hyperparameters of SVR were tuned
using FA and 10-fold cross-validation and the optimal fold obtains
the minimum RMSE. It is observed in Figure 8 that the 7th fold
attains the minimum RMSE in the UCS dataset and the 3rd fold
attains the minimum RMSE in the FS dataset. The normalised RMSE
versus iteration in the lowest fold for UCS and FS datasets are
depicted in Fig. 9. It can be observed that it takes 11 iterations
for the RMSE curve to converge for the UCS dataset and 25 itera-
tions on the FS dataset. This implies that the FA successfully found
the best hyperparameters of SVR. The optimal hyperparameters of

Fig. 8. 10-fold cross validation for hyper parameter tuning on the (a) UCS dataset and (b) FS dataset.

Fig. 9. Normalised RMSE score versus iteration in the best fold for (a) UCS and (b) FS.
8 Y. Huang et al. / Construction and Building Materials 260 (2020) 120457

SVR model for UCS prediction are c is 26.89 and c is 24.23, while for
FS prediction are c is 27.45 and c is 49.51.

3.2. Performance of FA-SVR

Fig. 10 presents the correlation between the actual and pre-

dicted values of the UCS and FS, respectively. Although there are
several outliers observed, most of the points lie close to the perfect
fitting curve (R = 1) in both Fig. 10 (a) and (b). This indicates a
robust relationship between the predicted and actual values.
Table 5 summarizes the values of performance metrics (R,
RMSE, MAE and MAPE) on the training and test set for both UCS
and FS predictions. It shows that both datasets obtain strong corre-
lation coefficients between the predicted and actual values, for the
test set of 0.91 and 0.85, respectively. This suggests a high predic-
tion accuracy of the FA-SVR model and indicates that the algorithm
is suitable for the datasets. Furthermore, the RMSE, MAE and MAPE
are low and similar in training and test sets for both UCS and FS,
indicating that the models are well fitted with low error measures.
No over-fitting or under-fitting is occurring. Hence, the FA-SVR
model is proved to be robust and has high accuracy in prediction
of mechinical objectives of SFRC.

3.3. SFRC mixture optimisation

By using MOFA, solutions that minimise the objective functions

are obtained. As shown in Fig. 11, the non-dominated solutions
between cost and UCS/FS can be obtained within the constraints
based on the dataset. It can be observed that an increase in UCS/
FS can only be achieved by increasing the cost of SFRC. The type
of Pareto point chosen for SFRC mixture design is dependent on
the decision-maker. For example, from Fig. 11 (a), if a high UCS is
wanted, solution point A can be considered even though it is more
expensive. If the budget is a priority, solution point C is the cheaper
option with a lower UCS value attained. Solution point B has an
Fig. 10. Actual versus predicted values for UCS (a) and FS (b).
intermediate cost and UCS compared with point A and C.
Fig. 11 (b) reveals that there are gaps in the Pareto front based
on the cost and FS for SFRC. This is due to the limitation of the cur-
Table 5
rent dataset. With the current database, if a decision-maker deci-
Calculated performance measures on the UCS and FS datasets. des to estimate the cost for an FS with no apparent Pareto point,
it could be solved by estimating a suitable curve for the Pareto
Performance measure UCS FS
front to approximate the cost of the FS. Moreover, more Pareto
Training set Test set Training set Test set points could form a more accurate Pareto front if requested as
RMSE (MPa) 4.7 7.6 1.3 1.6 MOFA would take a longer time to attain the optimised solutions
R 0.97 0.91 0.89 0.85 required.
MAE (MPa) 3.1 5.1 0.92 1.3
To find the most cost-effective point in the Pareto front, an effi-
MAPE 0.060 0.11 0.15 0.20
ciency ratio was calculated using the difference in cost divided by
the difference in strength. With the lower the efficiency ratio, it

Fig. 11. Pareto front based on Cost and (a) UCS or (b) FS for SFRC.
 Y. Huang et al. / Construction and Building Materials 260 (2020) 120457 9

indicates that an increase in strength is possible for a slight point 6 has the lowest efficiency ratio, while point 20 has the high-
increase in cost; hence, being more economical. est efficiency ratio. Thus, for UCS of SFRC, the most-cost effective
Table 6 and Table 7 show the efficiency ratio for both UCS and UCS value would be at point 6 when UCS is 41.664 MPa for the cost
FS Pareto front points. For the UCS Pareto front as listed in Table 6, of $33.4/m3. Other cost-effective points include point 16 with UCS
being 77.6 MPa and a cost of $35.10/m3 ,as the efficiency ratio is
Table 6 similar to point 6 but allows a higher UCS value for a slight increase
UCS Pareto front efficiency ratio. in cost.
No. of points UCS (MPa) Cost ($) Efficiency ratio For the FS Pareto front as presented in Table 7, point 3 has the
lowest efficiency ratio, while point 20 has the highest. The effi-
1 25.644 33.163 –
2 34.225 33.197 0.0039 ciency ratios for the FS Pareto points are comparatively high. Thus,
3 36.693 33.233 0.0147 for FS of SFRC, the most cost-effective FS value would be at point 3
4 37.687 33.314 0.0819 when FS is 9.058 MPa for the cost of $27.44/m3. If a higher FS value
5 39.560 33.443 0.0687 is sought after, point 18 when FS is 13.035 MPa for the cost of
6 41.662 33.449 0.0030
7 42.433 33.510 0.0786
$62.499/m3 could be considered.
8 43.064 33.532 0.0355 In general, it is up to the decision-maker to decide their budget
9 47.019 33.608 0.0192 and type of strength and type of mechanical property they would
10 49.555 33.858 0.0986 want to achieve, otherwise, the option of using the most cost-
11 52.604 34.190 0.1090
effective point based on the efficiency ratio is available.
12 61.328 34.426 0.0271
13 62.539 34.437 0.0086
14 69.267 34.797 0.0535
15 70.808 34.883 0.0560 3.4. Variable importance
16 77.577 35.098 0.0318
17 84.799 35.550 0.0626
The results from the sensitivity study show the measurement of
18 85.756 35.685 0.1401
19 86.965 36.294 0.5041
influences of the variables for both UCS and FS as shown in Fig. 12.
20 89.726 41.059 1.7264 The input variables for the UCS dataset are shown in Fig. 12 (a), in
which the most influential variable is the SP with the highest influ-
ence score of 1.9746. This is followed by C (1.6189), while UCS is
Table 7
FS Pareto front efficiency ratio. least sensitive to Vf (0.26). Unexpectedly, SP ranks the highest as
an influential variable compared to other cementitious materials.
No. of points FS (MPa) Cost ($) Efficiency ratio
However, this could be due to the data points inputted into the
1 6.292 27.437 - database from experimental journal-based articles. Many of the
2 6.931 27.440 0.005
journal-based articles do state that SP was added to the experi-
3 9.058 27.442 0.001
4 9.175 27.463 0.180 mental mixture design, yet the proportion amount was not given.
5 9.327 27.466 0.018 Hence, the model could potentially only understand the patterns
6 9.667 27.472 0.019 of dataset variables from the journal articles that provided SP pro-
7 9.668 27.473 0.569 portions. The model could have identified that higher UCS values
8 9.686 27.495 1.239
9 9.793 27.515 0.186
mean SP value is known but that is not necessarily the case.
10 9.863 28.258 10.638 Nonetheless, it is not unexpected that the steel fibre input vari-
11 10.069 28.609 1.713 ables (Lf, df and Vf of 0.4404, 0.3584 and 0.26, respectively) are
12 10.325 29.349 2.882 ranked the lowest as it is known that steel fibres only slightly
13 10.430 29.912 5.355
affects UCS in SFRC [70].
14 10.494 30.249 5.342
15 11.139 38.720 13.128 For FS input variables shown in Fig. 12 (b), the most influential
16 11.370 42.477 16.215 variable is the Vf with the highest influence score of 1.9410. This is
17 12.144 52.760 13.297 followed by W (1.2307), while FS is least sensitive to df (0.5776). It
18 13.035 62.499 10.928 is not unexpected that the highest variable is the Vf which has a
19 14.101 81.326 17.667
20 14.171 87.224 83.679
major influence on SFRC as steel fibre has substantial effects on
the tensile/flexural strength of concrete. However, the geometry

Fig. 12. Input variables importance measure for (a) UCS and (b) FS of SFRC mixture design.
10 Y. Huang et al. / Construction and Building Materials 260 (2020) 120457
of the steel fibres (Lf = 0.6825 and df = 0.5776) are not as influential
as the volume of steel fibres in the concrete. Overall, this means
that the amount of steel fibre plays a more significant role in affect-
ing the mechanical properties of SFRC.
4. Conclusions
SFRC mixture design is a fundamental yet difficult task for civil
engineers. To help with the tough process, this study takes the
opportunity to predict the optimised mixture design of SFRC using
the FA-SVR based model. This model is trained using a wide-
ranging database. The following conclusions are drawn from this
study:
1. FA is competent in the tuning of hyper parameters of the SVR
model. The obtained RSME versus iteration curve is able to con-
verge within several iterations (for UCS: 11 iterations, for FS: 25
iterations).
2. The FA-SVR based model can effectively and accurately predict
both UCS and FS based on performance evaluation on the test
set (for UCS, R: 0.91, for FS, R: 0.85).
3. The proposed MOO model can obtain successful Pareto solu-
tions based on the set constraints for this study but more iter-
ations could be done to produce more Pareto solutions.
4. The variable importance measure shows that for the most influ-
encing variables are the volume of steel fibres and SP for FS and
UCS, respectively.
Overall, this is a promising study on AI application towards
prediction of mechanical properties of SFRC and optimisation of
mixture design. It is a possible cost-effective and sustainable
design process that can be employed without the usage of as
much time, resources and labour when compared to traditional
methods. However, it currently only has the basic input variables
and a basic dataset used for the application. In the future work,
more input variables (e.g. types of steel fibres) should be incorpo-
rated in the dataset to enhance the generalizability of the pro-
posed model. Furthermore, the outcomes are based on
experimental results from journal articles. As such, set constraints
could be altered to make it more suitable and reasonable for real
works in the civil engineering industry. Hence, more investiga-
tions must be done to refine the application of the FA-SVR based
model and the dataset.
CRediT authorship contribution statement
Yimiao Huang: Conceptualization, Methodology, Software,
Writing - original draft. Junfei Zhang: Data curation, Supervision,
Writing - review & editing. Foo Tze Ann: Data curation, Writing-
original draft. Guowei Ma: Supervision, Writing - review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared
to influence the work reported in this paper.
Acknowledgements
The work is financially supported by ARC Research Hub for
Nanoscience-Based Construction Material Manufacturing
(IH150100006); ARC Discovery Project (DP160100119): Develop-
ment of nano-reinforced concrete using boron nitride nanosheets.
References
[1] V. Marcos-Meson, A. Michel, A. Solgaard, G. Fischer, C. Edvardsen, T.L. Skovhus,
Corrosion resistance of steel fibre reinforced concrete - A literature review,
Cem. Concr. Res. 103 (2018) 1–20.
[2] T. Ji, T. Lin, X. Lin, A concrete mix proportion design algorithm based on
artificial neural networks, Cem. Concr. Res. 36 (7) (2006) 1399–1408.
[3] L. Soufeiani, S.N. Raman, M.Z.B. Jumaat, U.J. Alengaram, G. Ghadyani, P. Mendis,
Influences of the volume fraction and shape of steel fibers on fiber-reinforced
concrete subjected to dynamic loading – A review, Eng. Struct. 124 (2016)
405–417.
[4] A.T.A. Dantas, M. Batista Leite, K. de Jesus Nagahama, Prediction of
compressive strength of concrete containing construction and demolition
waste using artificial neural networks, Constr. Build. Mater. 38 (2013) 717–
722.
[5] H.-G. Ni, J.-Z. Wang, Prediction of compressive strength of concrete by neural
networks, Cem. Concr. Res. 30 (8) (2000) 1245–1250.
[6] M. Açikgenç, M. Ulasß, K.E. Alyamaç, Using an artificial neural network to
predict mix compositions of steel fiber-reinforced concrete, Arab. J. Sci. Eng. 40
(2) (2015) 407–419.
[7] B.B. Adhikary, H. Mutsuyoshi, Prediction of shear strength of steel fiber RC
beams using neural networks, Constr. Build. Mater. 20 (9) (2006) 801–811.
[8] N. Ahn, H. Jang, D.K. Park, Presumption of shear strength of steel fiber
reinforced concrete beam using artificial neural network model, J. Appl. Polym.
Sci. 103 (4) (2007) 2351–2358.
[9] F. Altun, Ö. Kisßi, K. Aydin, Predicting the compressive strength of steel fiber
added lightweight concrete using neural network, Comput. Mater. Sci. 42 (2)
(2008) 259–265.
[10] A. Behnood, K.P. Verian, M. Modiri Gharehveran, Evaluation of the splitting
tensile strength in plain and steel fiber-reinforced concrete based on the
compressive strength, Constr. Build. Mater. 98 (2015) 519–529.
[11] J. Sun, J. Zhang, Y. Gu, Y. Huang, Y. Sun, G. Ma, Prediction of permeability and
unconfined compressive strength of pervious concrete using evolved support
vector regression, Constr. Build. Mater. 207 (2019) 440–449.
[12] Y. Sun, J. Zhang, G. Li, G. Ma, Y. Huang, J. Sun, Y. Wang, B. Nener, Determination
of Young’s modulus of jet grouted coalcretes using an intelligent model, Eng.
Geol. 252 (2019) 43–53.
[13] K. Yan, C. Shi, Prediction of elastic modulus of normal and high strength
concrete by support vector machine, Constr. Build. Mater. 24 (8) (2010) 1479–
1485.
[14] H. Salehi, R. Burgueño, Emerging artificial intelligence methods in structural
engineering, Eng. Struct. 171 (2018) 170–189.
[15] J.-S. Chou, C.-F. Tsai, Concrete compressive strength analysis using a combined
classification and regression technique, Autom. Constr. 24 (2012) 52–60.
[16] K. Duan, S.S. Keerthi, A.N. Poo, Evaluation of simple performance measures for
tuning SVM hyperparameters, Neurocomputing 51 (2003) 41–59.
[17] C.-W. Hsu, C.-C. Chang, C.-J. Lin, A practical guide to support vector
classification, Taipei, 2003.
[18] Y. Yu, C. Zhang, X. Gu, Y. Cui, Expansion prediction of alkali aggregate
reactivity-affected concrete structures using a hybrid soft computing method,
Neural Comput. Applic. 31 (12) (2019) 8641–8660.
[19] M.-Y. Cheng, D. Prayogo, Y.-W. Wu, Novel genetic algorithm-based
evolutionary support vector machine for optimizing high-performance
concrete mixture, J. Comput. Civ. Eng. 28 (4) (2014) 06014003, https://doi.
org/10.1061/(ASCE)CP.1943-5487.0000347.
[20] J.-S. Chou, N.-T. Ngo, A.-D. Pham, Shear strength prediction in reinforced
concrete deep beams using nature-inspired metaheuristic support vector
regression, J. Comput. Civ. Eng. 30 (1) (2016) 04015002, https://doi.org/
10.1061/(ASCE)CP.1943-5487.0000466.
[21] I. Fister, I. Fister Jr., X.-S. Yang, J. Brest, A comprehensive review of firefly
algorithms, Swarm Evol. Comput. 13 (2013) 34–46.
[22] A. Behnood, E.M. Golafshani, Predicting the compressive strength of silica
fume concrete using hybrid artificial neural network with multi-objective grey
wolves, J. Cleaner Prod. 202 (2018) 54–64.
[23] A. Kandiri, E.M. Golafshani, A. Behnood, Estimation of the compressive
strength of concretes containing ground granulated blast furnace slag using
hybridized multi-objective ANN and salp swarm algorithm, Constr. Build.
Mater. 248 (2020) 118676.
[24] E.M. Golafshani, A. Behnood, Estimating the optimal mix design of silica fume
concrete using biogeography-based programming, Cem. Concr. Compos. 96
(2019) 95–105.
[25] H. Naseri, Cost optimization of no-slump concrete using genetic algorithm and
particle swarm optimization, IJIMT 10 (1) (2019) 33–37.
[26] W. Abbass, M.I. Khan, S. Mourad, Evaluation of mechanical properties of steel
fiber reinforced concrete with different strengths of concrete, Constr. Build.
Mater. 168 (2018) 556–569.
[27] R.V. Balendran, F.P. Zhou, A. Nadeem, A.Y.T. Leung, Influence of steel fibres on
strength and ductility of normal and lightweight high strength concrete, Build.
Environ. 37 (12) (2002) 1361–1367.
[28] A. Bhosale, M.A. Rasheed, S.S. Prakash, G. Raju, A study on the efficiency of steel
vs. synthetic vs. hybrid fibers on fracture behavior of concrete in flexure using
acoustic emission, Constr. Build. Mater. 199 (2019) 256–268.
[29] N. Buratti, C. Mazzotti, M. Savoia, Post-cracking behaviour of steel and macro-
synthetic fibre-reinforced concretes, Constr. Build. Mater. 25 (5) (2011) 2713–
2722.
 Y. Huang et al. / Construction and Building Materials 260 (2020) 120457
[30] J.A. Carneiro, P.R.L. Lima, M.B. Leite, R.D. Toledo Filho, Compressive stress–
strain behavior of steel fiber reinforced-recycled aggregate concrete, Cem.
Concr. Compos. 46 (2014) 65–72.
[31] K.-K. Choi, D.-O. Ku, Flexural behaviour of amorphous metal-fibre-reinforced
concrete, Proc. Inst. Civ. Eng. Struct. Build. 168 (1) (2015) 15–25.
[32] Ö. Eren, K. Marar, Effects of limestone crusher dust and steel fibers on concrete,
Constr. Build. Mater. 23 (2) (2009) 981–988.
[33] C. Fu, H. Ye, K. Wang, K. Zhu, C. He, Evolution of mechanical properties of steel
fiber-reinforced rubberized concrete (FR-RC), Compos. B Eng. 160 (2019) 158–
166.
[34] D. Gao, L. Zhang, M. Nokken, Mechanical behavior of recycled coarse aggregate
concrete reinforced with steel fibers under direct shear, Cem. Concr. Compos.
79 (2017) 1–8.
[35] H. Hu, P. Papastergiou, H. Angelakopoulos, M. Guadagnini, K. Pilakoutas,
Mechanical properties of SFRC using blended manufactured and recycled tyre
steel fibres, Constr. Build. Mater. 163 (2018) 376–389.
[36] S. Iqbal, A. Ali, K. Holschemacher, T.A. Bier, Mechanical properties of steel fiber
reinforced high strength lightweight self-compacting concrete (SHLSCC),
Constr. Build. Mater. 98 (2015) 325–333.
[37] S.J. Jang, G.Y. Jeong, M.H. Lee, K. Rokugo, H.D. Yun, Compressive strength
effects on flexural behavior of steel fiber reinforced concrete, Key Eng. Mater.
709 (2016) 101–104.
[38] R. Lameiras, J.A.O. Barros, M. Azenha, Influence of casting condition on the
anisotropy of the fracture properties of Steel Fibre Reinforced Self-Compacting
Concrete (SFRSCC), Cem. Concr. Compos. 59 (2015) 60–76.
[39] J.-H. Lee, B. Cho, E. Choi, Flexural capacity of fiber reinforced concrete with a
consideration of concrete strength and fiber content, Constr. Build. Mater. 138
(2017) 222–231.
[40] B. Li, L. Xu, Y. Shi, Y. Chi, Q.i. Liu, C. Li, Effects of fiber type, volume fraction and
aspect ratio on the flexural and acoustic emission behaviors of steel fiber
reinforced concrete, Constr. Build. Mater. 181 (2018) 474–486.
[41] M.L. Nehdi, M.F. Najjar, A.M. Soliman, T.M. Azabi, Novel steel fibre-reinforced
preplaced aggregate concrete with superior mechanical performance, Cem.
Concr. Compos. 82 (2017) 242–251.
[42] L. Nguyen-Minh, M. Rovňák, T. Tran-Quoc, K. Nguyenkim, Punching shear
resistance of steel fiber reinforced concrete flat slabs, Procedia Eng. 14 (2011)
1830–1837.
[43] M. Nili, V. Afroughsabet, Combined effect of silica fume and steel fibers on the
impact resistance and mechanical properties of concrete, Int. J. Impact Eng. 37
(8) (2010) 879–886.
[44] R.S. Olivito, F.A. Zuccarello, An experimental study on the tensile strength of
steel fiber reinforced concrete, Compos. B Eng. 41 (3) (2010) 246–255.
[45] M. Paja˛k, T. Ponikiewski, Flexural behavior of self-compacting concrete
reinforced with different types of steel fibers, Constr. Build. Mater. 47 (2013)
397–408.
ß ahin, F. Köksal, The influences of matrix and steel fibre tensile strengths on
[46] Y. S
the fracture energy of high-strength concrete, Constr. Build. Mater. 25 (4)
(2011) 1801–1806.
[47] A. Sivakumar, M. Santhanam, Mechanical properties of high strength concrete
reinforced with metallic and non-metallic fibres, Cem. Concr. Compos. 29 (8)
(2007) 603–608.
[48] D.V. Soulioti, N.M. Barkoula, A. Paipetis, T.E. Matikas, Effects of fibre geometry
and volume fraction on the flexural behaviour of steel-fibre reinforced
concrete, Strain 47 (s1) (2011) e535–e541.
[49] P.R. Tadepalli, Y.L. Mo, T.T.C. Hsu, Mechanical properties of steel fibre concrete,
Mag. Concr. Res. 65 (8) (2013) 462–474.
[50] M. Taylor, F.D. Lydon, B.I.G. Barr, Toughness measurements on steel fibre-
reinforced high strength concrete, Cem. Concr. Compos. 19 (4) (1997) 329–
340.
[51] D.-Y. Yoo, N. Banthia, J.-M. Yang, Y.-S. Yoon, Size effect in normal- and high-
strength amorphous metallic and steel fiber reinforced concrete beams,
Constr. Build. Mater. 121 (2016) 676–685.
[52] D.-Y. Yoo, Y.-S. Yoon, N. Banthia, Flexural response of steel-fiber-reinforced
concrete beams: Effects of strength, fiber content, and strain-rate, Cem. Concr.
Compos. 64 (2015) 84–92.
[53] X. Zheng, T. Ji, S.M. Easa, B. Zhang, Z. Jiang, Tensile basic creep behavior of
lightweight aggregate concrete reinforced with steel fiber, Constr. Build.
Mater. 200 (2019) 356–367.
11
[54] D.E. Farrar, R.R. Glauber, Multicollinearity in regression analysis: The problem
revisited, Rev. Econ. Stat. 49 (1) (1967) 92–107.
[55] A.J. Smola, B. Schölkopf, A tutorial on support vector regression, Stat. Comput.
14 (3) (2004) 199–222.
[56] D. Basak, S. Pal, D.C. Patranabis, Support vector regression, Neural Inform.
Process. Lett. Rev. 11 (10) (2007) 203–224.
[57] C.J. Burges, A tutorial on support vector machines for pattern recognition, Data
Min. Knowl. Disc. 2 (2) (1998) 121–167.
[58] S.K. Lahiri, K.C. Ghanta, The support vector regression with the parameter
tuning assisted by a differential evolution technique: Study of the critical
velocity of a slurry flow in a pipeline, Chem. Ind. Chem. Eng. 14 (3) (2008) 191–
203.
[59] S. Boyd, L. Vandenberghe (Eds.), Convex Optimization, Cambridge University
Press, 2004.
[60] Y. Xin-She, Firefly Algorithms for Multimodal Optimization, arXiv.org (2010).
[61] X.-S. Yang, M. Karamanoglu, 1 - Swarm Intelligence and Bio-Inspired
Computation: An Overview, in: X.-.-S. Yang, Z. Cui, R. Xiao, A.H. Gandomi, M.
Karamanoglu (Eds.), Swarm Intelligence and Bio-Inspired Computation,
Elsevier, Oxford, 2013, pp. 3–23.
[62] P. Cortez, M.J. Embrechts, Opening black box data mining models using
sensitivity analysis, 2011 IEEE Symposium on Computational Intelligence and
Data Mining (CIDM), IEEE, 2011, pp. 341-348.
[63] Yuantian Sun, Junfei Zhang, Guichen Li, Yuhang Wang, Junbo Sun, Chao Jiang,
Optimized neural network using beetle antennae search for predicting the
unconfined compressive strength of jet grouting coalcretes, Int. J. Numer. Anal.
Methods Geomech. 43 (4) (2019) 801–813.
[64] Paulo Cortez, Mark J. Embrechts, Using sensitivity analysis and visualization
techniques to open black box data mining models, Inf. Sci. 225 (2013) 1–17.
[65] M.K.A. Ariyaratne, T.G.I. Fernando, S. Weerakoon, Solving systems of nonlinear
equations using a modified firefly algorithm (MODFA), Swarm Evol. Comput.
48 (2019) 72–92.
[66] Adil Baykasoğlu, Ahmet Öztasß, Erdoğan Özbay, Prediction and multi-objective
optimization of high-strength concrete parameters via soft computing
approaches, Expert Syst. Appl. 36 (3) (2009) 6145–6155.
[67] Dobroslav Grygar, René Fabricius, An efficient adjustment of genetic algorithm
for pareto front determination, Transp. Res. Procedia 40 (2019) 1335–1342.
[68] Xin-She Yang, Multiobjective firefly algorithm for continuous optimization,
Eng. Comput. 29 (2) (2013) 175–184.
[69] M. Klammer, J.N. Dybowski, D. Hoffmann, C. Schaab, Pareto optimization
identifies diverse set of phosphorylation signatures predicting response to
treatment with dasatinib, PLoS One 10 (6) (2015) e0128542.
[70] D.J. Hannant, 4.11 - Cement-based Composites, in: A. Kelly, C. Zweben (Eds.),
Comprehensive Composite Materials, Pergamon, Oxford, 2000, pp. 323–362.
[71] Junfei Zhang, Guowei Ma, Yimiao Huang, Farhad Aslani, Brett Nener, Modelling
uniaxial compressive strength of lightweight self-compacting concrete using
random forest regression, Construction and Building Materials 210 (2019)
713–719.
[72] Junfei Zhang, Dong Li, Yuhang Wang, Predicting uniaxial compressive strength
of oil palm shell concrete using a hybrid artificial intelligence model, Journal of
Building Engineering 30 (2020).
[73] Junfei Zhang, Dong Li, Yuhang Wang, Toward intelligent construction:
prediction of mechanical properties of manufactured-sand concrete using
tree-based models, Journal of Cleaner Production 258 (2020).
[74] Junfei Zhang, Yimiao Huang, Guowei Ma, Junbo Sun, Brett Nener, A
metaheuristic-optimized multi-output model for predicting multiple
properties of pervious concrete, Construction and Building Materials 249
(2020).
[75] Junfei Zhang, Yimiao Huang, Guowei Ma, Brett Nener, Multi-objective beetle
antennae search algorithm, arXiv preprint arXiv:2002.10090 (2020).
[76] Junfei Zhang, Dong Li, Yuhang Wang, Predicting tunnel squeezing using a
hybrid classifier ensemble with incomplete data, Bulletin of Engineering
Geology and the Environment (2020).
[77] Junfei Zhang, Yimiao Huang, Farhad Aslani, Guowei Ma, Brett Nener, A hybrid
intelligent system for designing optimal proportions of recycled aggregate
concrete, Journal of Cleaner Production (2020).
[78] Junfei Zhang, Yimiao Huang, Yuhang Wang, Guowei Ma, Multi-objective
optimization of concrete mixture proportions using machine learning and
metaheuristic algorithms, Construction and Building Materials 253 (2020).