Journal of Constructional Steel Research 67 (2011) 1096–1105

Contents lists available at ScienceDirect

Journal of Constructional Steel Research

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

A new prediction model for the load capacity of castellated steel beams
Amir Hossein Gandomi a,∗ , Seyed Morteza Tabatabaei b , Mohammad Hossein Moradian c , Ata Radfar d ,
Amir Hossein Alavi e
a
College of Civil Engineering, Tafresh University, Tafresh, Iran
b
Department of Civil Engineering, The Institute of Higher Education of Eqbal Lahoori, Mashhad, Iran
c
Department of Civil Engineering, Islamic Azad University, Science and Research Branch, Tehran, Iran
d
Department of Civil Engineering, Ferdowsi University of Mashhad, Mashhad, Iran
e
School of Civil Engineering, Iran University of Science and Technology, Tehran, Iran

article info abstract

Article history: In this study, a robust variant of genetic programming, namely gene expression programming (GEP), is
Received 29 July 2010 utilized to build a prediction model for the load capacity of castellated steel beams. The proposed model
Accepted 21 January 2011 relates the load capacity to the geometrical and mechanical properties of the castellated beams. The
model is developed based on a reliable database obtained from the literature. To verify the applicability
Keywords: of the derived model, it is employed to estimate the load capacity of parts of the test results that were
Castellated beam
not included in the modeling process. The external validation of the model was further verified using
Failure load
Gene expression programming
several statistical criteria recommended by researchers. A multiple least squares regression analysis is
performed to benchmark the GEP-based model. A sensitivity analysis is further carried out to determine
the contributions of the parameters affecting the load capacity. The results indicate that the proposed
model is effectively capable of evaluating the failure load of the castellated beams. The GEP-based design
equation is remarkably straightforward and useful for pre-design applications.
© 2011 Elsevier Ltd. All rights reserved.

1. Introduction II. Decreasing the weight of structures;

III. Optimum use of existing profiles;
With the development of electrical welding technology in steel IV. By-passing the used plate girders; and
construction, the castellated steel beam (CSB) became available to V. The passage of services through the web openings.
structural engineers. The CSB, initially known as the ‘‘Boyd beam’’,
The seven potential failure modes generally associated with the
was first used in 1910 [1], and then designed and manufactured
castellated beams are as follows [4].
in the early 1930s as roof beams in Czech Republic. The initial
inspiration for the elastic and plastic calculation methods were i. Flexure mechanism formation;
respectively introduced in 1942 and the early 1970s. Castellation ii. Overall beam lateral–torsional buckling;
is the procedure of cutting the web of a rolled section in a zigzag iii. Vierendeel mechanism formation;
pattern [2]. One of the halves is turned round and welded to the iv. Welded joint rupture in the web;
other half. This procedure increases the depth of the original beam v. Web post-shear bucking;
(h) by the depth of the cut (d). As shown in Fig. 1, CSBs are often vi. Web post-compression buckling; and
made from I sections by the castellation process. This shape fits vii. Tee compression buckling.
the dictionary definition of castellated as ‘‘castle-like’’. According
These modes are of two different categories. The first two modes
to Zirakian and Showkati [3], the basic reasons for the fabrication
are similar to the corresponding modes for solid-web beams and
of the castellated beams are as follows.
may be analyzed in almost identical fashions. Modes iii to vii are
I. Augmentation of section height, resulting in the enhancement specific to castellated beams, since they are associated with the
of moment of inertia, section modulus, stiffness, and flexural tees and web posts that bound openings. There is a clear correlation
resistance of sections; between mode iv and shear failure, and between mode vi and
the buckling solid web. However, it is necessary to develop new
analytical models for modes iii to vii. Detailed explanations for
∗
Corresponding author. these failure modes can be found in [5].
E-mail addresses: [email protected] (A.H. Gandomi), Knowles [6] showed that the failure load predicted using
[email protected] (A.H. Alavi). a column in compression approach adopting Blogett’s force
0143-974X/$ – see front matter © 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jcsr.2011.01.014
 A.H. Gandomi et al. / Journal of Constructional Steel Research 67 (2011) 1096–1105 1097

Fig. 1. Castellation process in a CSB.

distribution model and an effective length factor would provide a

good agreement with the experimental results. In this approach,
the web post is treated as a column having a width equal to the
narrowest width of the web, a length equal to the clear height of the
castellation, and a thickness equal to the web thickness. The model
recommended in BS 449, Clause 28a, is for 45° dispersion angle.
Applying BS 449 to CSBs requires the consideration of the width to
thickness ratio in order that the section can be correctly classified
for local buckling. The rules apply only to sections castellated to
the profile adopted in the UK. Although many of the rules may well
apply to other types of section, they have not been fully approved.
This is particularly true with unrestrained beams [7].
Developments in computational software and hardware have
enabled several alternative computer-aided data mining ap-
proaches to emerge. As an example, pattern recognition systems Fig. 2. Tree representation of a GP model (X 1 + 3/X2 )2 .
learn adaptively from experience and extract various discrimina-
tors. Artificial neural networks (ANNs) are the most widely used the values of its parameters [22,23]. GP optimizes a population of
pattern recognition procedures. ANNs have been applied to assess computer programs according to a fitness landscape determined
different characteristics of steel beams [8–10]. Despite the accept- by a program’s ability to perform a given computational task. The
able performance of ANNs, they do not commonly give a particular fitness of each program in the population is evaluated using a fit-
function to estimate the outcome using the input values. ness function. Thus, the fitness function is the objective function
Genetic programming (GP) [11] is an alternative approach for GP aims to optimize [23,24].
behavior modeling of structural engineering problems. GP may This classical GP approach is referred to as tree-based GP. A
generally be defined as a supervised machine learning technique population member in tree-based GP is a hierarchically structured
that searches a program space instead of a data space [12,13]. tree comprising functions and terminals. The functions and
Many researchers have employed GP and its variants to discover terminals are selected from a set of functions and a set of terminals.
complex relationships among experimental data [13–16]. Gene For example, the function set F can contain the basic arithmetic
expression programming (GEP) [17] is a recent extension of operations (+, −, ×, /, etc.), Boolean logic functions (AND, OR,
GP. GEP evolves computer programs of different sizes and NOT, etc.), or any other mathematical functions. The terminal set
shapes encoded in linear chromosomes of fixed length. The GEP T contains the arguments for the functions and can consist of
chromosomes are composed of multiple genes, each gene encoding numerical constants, logical constants, variables, etc. The functions
a smaller subprogram. The GEP approach has been shown to and terminals are chosen at random and constructed together
be an efficient alternative to traditional GP [17,18]. There have to form a computer model in a tree-like structure with a root
been some scientific efforts directed at applying GEP to structural point with branches extending from each function and ending in a
engineering tasks [19–21]. terminal. An example of a simple tree representation of a GP model
The main purpose of this paper is to utilize the GEP technique is illustrated in Fig. 2 [23].
to build a predictive model for the failure load of CSBs. A reliable GEP is a linear variant of GP. The linear variants of GP make
database of previously published CSB test results is utilized to a clear distinction between the genotype and the phenotype of
develop the model. The performance of the derived model is an individual. Thus, the individuals are represented as linear
subsequently compared with that of a regression-based model. strings [13,18].

2. Genetic programming 2.1. Gene expression programming

GP is a subset of genetic algorithms (GAs) with a great ability GEP is a natural development of GP first invented by
to automatically evolve computer programs. The evolutionary pro- Ferreira [17]. Most of the genetic operators used in GAs can also
cess followed by GP is inspired by the principle of Darwinian natu- be implemented in GEP with minor changes. GEP consists of five
ral selection. GP was introduced by Koza [11] after experiments on main components: function set, terminal set, fitness function, con-
symbolic regression. The main difference between GP and a GA is trol parameters, and termination condition [13]. Unlike the parse-
the representation of the solution. The GP solutions are computer tree representation in conventional GP, GEP uses a fixed length of
programs that are represented as tree structures and expressed in character strings to represent solutions to the problems, which are
a functional programming language (like LISP) [11]. A GA creates afterwards expressed as parse trees of different sizes and shapes.
a string of numbers that represent the solution. In other words, in These trees are called GEP expression trees (ETs). One advantage
GP, the evolving programs (individuals) are parse trees than can of the GEP technique is that the creation of genetic diversity is
vary in length throughout the run rather than fixed-length binary extremely simplified, as genetic operators work at the chromo-
strings. Traditional optimization techniques, such as the GA, are some level. Another strength of GEP is its unique multigenic nature,
generally used in parameter optimization to evolve the best val- which allows the evolution of more complex programs composed
ues for a given set of model parameters. GP, on the other hand, of several subprograms [13]. Each GEP gene contains a list of sym-
gives the basic structure of the approximation model together with bols with a fixed length that can be any element from a function set
1098 A.H. Gandomi et al. / Journal of Constructional Steel Research 67 (2011) 1096–1105

Fig. 3. Example of expression trees (ETs).

√
such as {+, −, ×, /, } and a terminal set such as {X1 , X2 , X3 , 3}.
The function set and the terminal set must have the closure prop-
erty: each function must be able to take any value of data type
which can be returned by a function or assumed by a terminal. A
typical GEP gene with the given function and terminal sets can be
as follows [13]:
√
+. × . .X 1 . − . + . + . × .X2 .X1 .X3 .3.X2 .X3 , (1)
where X1 , X2 , and X3 are variables and 3 is a constant; ‘‘.’’ is
element separator for easy reading. The above expression is termed
Karva notation or a K-expression [17,25]. A K-expression can be Fig. 4. A basic representation of the GEP algorithm.
represented by a diagram which is an ET. For example, the above
sample gene can be expressed as in Fig. 3.
The conversion starts from the first position in the K-
expression, which corresponds to the root of the ET, and reads
through the string one by one [13] The above GEP gene can also
be expressed in a mathematical form as
√
X1 ((X1 + 3) − (X2 × X3 )) + (X2 + X1 ). (2)
An ET can inversely be converted into a K-expression by recording
the nodes from left to right in each layer of the ET, from the root
layer down to the deepest one to form the string. As previously
Fig. 5. Example of expression trees (ETs).
mentioned, GEP genes have fixed length, which is predetermined
for a given problem. Thus, what varies in GEP is not the length of
the genes but the size of the corresponding ETs [13]. This means of a CSB plays an important role in the structural performance of
that there exist a certain number of redundant elements, which steel beams [3]. In its current state, behavior modeling of a CSB
are not useful for genome mapping. Hence, the valid length of a K- is more difficult than that of normal I-section beams. In order to
expression may be equal to or less than the length of the GEP gene. provide an accurate assessment of the performance characteristics
To guarantee the validity of a randomly selected genome, GEP em- of CSBs, the effects of all influencing parameters should be
ploys a head–tail method. Each GEP gene is composed of a head and incorporated into the model development. In this study, the GEP
a tail. The head may contain both function and terminal symbols,
approach was utilized to obtain a meaningful relationship between
whereas the tail may contain only terminal symbols [13,17,25].
the load capacity of a CSB (P) and the influencing variables, as
A basic representation of the GEP algorithm is presented in
Fig. 4 [26]. In GEP, the individuals are selected and copied into follows:
the next generation according to the fitness by roulette wheel
P = f (Fyw , hc , B, tw , tf , S , L, LC ), (4)
sampling with elitism. This guarantees the survival and cloning
of the best individual to the next generation. Variation in the where
population is introduced by conducting single or several genetic
operators on selected chromosomes, which include crossover, Fyw (MPa): Minimum web yield stress
mutation, and rotation. The rotation operator is used to rotate two hc (mm): Overall depth
subparts of the element sequence in a genome with respect to a B (mm): Width of flange
randomly chosen point. It can also drastically reshape the ETs. As
tw (mm): Web thickness
an example, the following gene [13]
√ tf (mm): Flange thickness
+ . + . × .X2 .X1 .X3 .3.X2 .X3 .+. × . .X 1 .− (3)
S (mm): Minimum width of the web post
rotates the first five elements of gene (1) to the end. Only the L (m): Span of castellated beam
first seven elements are used to construct the solution function LC : Indicator variable representing different loading conditions
(X2 + X1 ) + (X3 × 3), with the corresponding expression shown in
1: One-point load
Fig. 5 [13].
2: Two-point load
3. Modeling of a CSB 3: Distributed load.

The enhanced performance characteristics of steel beams are The above variables were chosen as the predictor variables on the
generally achieved by many processes such as castellation. The use basis of a literature review [4] and a trial study.
 A.H. Gandomi et al. / Journal of Constructional Steel Research 67 (2011) 1096–1105 1099

a 35 Frequency 100% b 25
Frequency
Cumulative%
100%
Cumulative%
30
80
0% 20 80%
25

Frequency
Frequency
20 60% 15 60%
15 40
0% 10 40%
10
20% 5 20%
5
0 0% 0 0%
One-point Diistributed 230 265 300 334 300
0 403 More
load load Fyw (MPa)

c 25 Frequency
Cumulative% 10
00% d 20
Frequency
Cumulative% 100%
20 80
0% 16 80%
Frequency

Frequency
15 60
0% 12 60%
10 40
0% 8 40%
5 20
0% 4 20%
0 %
0% 0 0%
22
29 306 384 461 539 616 More 67 90 112 135 158
8 180 More
hc (mm) B (mm)

e 25 Frequency
Cumulative%
10
00% f 25 Frequency
Cumulative% 100%
20 80
0% 20 80%
Frequency

Frequency

15 60
0% 15 60%
10 40
0% 10 40%
5 20
0% 5 20%
0 0%
% 0 0%
3.6 4.9 6.3 7.6 9.0 1
10.3 More 4.6 7.3 10.0 12.8 15.5
5 18.2 More
tw (mm) tf (mm)

g 12
Frequency
Cumulative% 1000% h 30
Frequency
Cumulative% 100%
10 80
0% 25 80%
Frequency

8 20
60
0% 60%
6 15
40
0% 40%
4 10
2 20
0% 5 20%
0 0%
% 0 0%
116 145 174 More

L (m)

i 30 100%
25
80%
Frequency
20 Cumulative% 60%
15
40%
10
5 20%

0 0%
74 278 482
2 687 891 1096 More
PExp. (kN)

Fig. 6. Histograms of the variables used for model development.

3.1. Experimental database (Fig. 6). The descriptive statistics of the database used in this study
are given in Table 1. The complete list of the data is given in Table 2.
A reliable database was obtained from the literature to develop Overfitting is one of the major problems in machine learning
the models. The database contains 47 test results for the load generalization. An efficient approach to prevent overfitting is
capacity of CSBs carried out by several researchers [1,27–32] and to test the derived models on a validation set to find a better
presented by Amayreh and Saka [4]. To visualize the distribution generalization [12]. This strategy was considered in this study
of the samples, the data are presented by frequency histograms for improving the generalization of the models. For this aim, the
1100 A.H. Gandomi et al. / Journal of Constructional Steel Research 67 (2011) 1096–1105

Table 1
Descriptive statistics of the variables used for model development.
LC Fyw (MPa) hc (mm) B (mm) tw (mm) tf (mm) S (mm) L (m) PExp (kN)

Mean – 314.894 419.862 112.481 6.741 9.497 110.604 2.095 250.417

Standard error – 4.959 15.500 4.048 0.246 0.488 8.317 0.171 25.658
Standard deviation – 34.000 106.260 27.750 1.685 3.345 57.017 1.171 175.905
Sample variance – 1156.014 11 291.089 770.065 2.840 11.187 3250.943 1.370 30 942.431
Kurtosis – 3.177 0.185 1.197 1.728 3.278 −1.497 4.496 28.351
Skewness – 1.043 0.339 0.567 0.690 1.242 0.095 2.095 4.675
Range 2 208 464.7 136.3 8.14 16.37 174.42 4.8 1226.5
Minimum 1 230 229 66.9 3.56 4.59 28.58 1 73.5
Maximum 3 438 693.7 203.2 11.7 20.96 203 5.8 1300
Sum – 14800 19 733.5 5286.6 316.81 446.35 5198.41 98.46 11 769.6

Table 2
The experimental database used for the model construction.
No. LC Fyw (MPa) hc (mm) B (mm) tw (mm) tf (mm) S (mm) L (m) PExp (kN) PGEP (kN) PLSR (kN)

1 One-point 320 380 150 7.1 10.7 40 5.8 176.5 148.3 74.6
2 One-point 335 500 150 7.1 10.7 100 5.8 73.5 132.0 221.8
3 One-point 290.3 451.8 123.4 7.55 10.72 150 2.14 285 269.2 221.0
4 One-point 293.2 606.2 145.6 7.31 11.4 200 1.5 280 297.0 337.7
5 One-point 335 440 150 7.1 10.7 70 5.8 145 134.0 177.7
6 One-point 294.7 524.8 124.4 7.03 10.79 180 2.6 280 239.1 212.3
7 Two-point 355 693.7 153.5 11.7 11.7 114.3 2.5 1300 1226.8 1149.0
8 Distributed 335 381 101.6 5.84 6.83 89 1.3 295 226.4 330.8
9 Distributed 320 229 76.2 5.84 9.58 38.1 1.32 147 141.6 105.9
10 One-point 290.8 451.4 123.9 7.51 10.7 150 1.15 275 282.5 266.5
11 One-point 352.9 380.5 66.9 3.56 4.59 66.55 1.83 94.8 158.9 150.2
12 Distributed 292.8 381 101.6 5.84 6.83 127 1.8 186 189.7 125.5
13 Distributed 292.8 381 127 9.14 14.02 63.5 1.8 249.1 312.6 365.5
14 One-point 352.9 380.5 66.9 3.56 4.59 66.55 1.22 100.9 156.2 179.8
15 Two-point 292.8 381 114.3 7.62 12.83 63.5 1.6 299 200.0 215.2
16 One-point 295.5 526.7 124.7 7.08 10.68 176 2 232 261.9 255.7
17 One-point 277 605.5 145.4 7.3 11.34 203 1.6 288 269.0 267.5
18 One-point 277 605.5 143.3 7.27 11.25 199 1.5 253 268.8 273.4
19 One-point 293.2 603.6 143.7 7.28 11.27 196 1 226 301.5 361.3
20 One-point 320 381 101.6 5.84 6.83 44.45 3 265.5 218.6 168.0
21 One-point 290 524.3 124.6 7.04 10.7 179 2.6 275 234.3 196.6
22 Distributed 335 229 76.2 5.84 9.58 38.1 1.32 137.2 157.4 164.8
23 One-point 320 381 101.6 5.84 6.83 34.93 3 269 237.7 171.9
24 One-point 304 460.4 103.1 7.21 10.73 143 1.1 267 274.0 307.3
25 Distributed 292.8 342.9 117.8 10.16 20.96 57.15 1.3 279.6 234.9 337.0
26 One-point 395 381 101.6 5.84 6.83 28.58 3 352 365.0 468.9
27 Distributed 292.8 266.7 101.6 6.35 9.83 44.5 1.5 113.7 131.4 63.6
28 Distributed 320 229 76.2 5.84 9.58 38.1 1.15 196 137.7 114.1
29 One-point 290.8 451.6 124.1 7.62 10.66 146 2 280 280.2 239.9
30 Distributed 230 280 100 10.8 5.7 40 3.33 194.1 314.4 322.5
31 One-point 335 381 101.6 5.84 6.83 127 1.6 290 237.9 261.8
32 Distributed 292.8 381 101.6 5.84 6.83 101.6 2.4 89 174.1 106.5
33 One-point 335 381 101.6 5.84 6.83 127 2.4 303.4 227.8 222.9
34 One-point 352.9 380.5 66.9 3.56 4.59 66.55 2.44 84.4 158.9 120.5
35 One-point 438 381 101.6 5.84 6.83 165.1 3.5 277.2 298.2 558.4
36 Distributed 292.8 381 101.6 5.84 6.83 88.9 2.5 89 170.5 106.8
37 Distributed 335 381 101.6 5.84 6.83 101.6 1.37 300 229.5 322.3
38 Distributed 335 381 101.6 5.84 6.83 165.1 1.6 358.6 249.0 285.6
39 One-point 304 459.6 103 7.15 10.7 157 2.3 273 254.8 236.5
40 One-point 290 524.3 124.4 7.02 10.73 176 1.35 280 257.8 255.9
41 One-point 295.5 526.1 124.7 7.08 10.77 176 1.35 260 271.0 284.9
42 One-point 335 229 76.2 5.84 9.58 38.1 1.32 117.6 148.6 125.6
43 Distributed 335 381 101.6 5.84 6.83 165.1 1.75 310.3 247.3 278.3
44 One-point 294.7 524.1 124.3 7.07 10.73 177 1.35 240 269.5 279.4
45 Two-point 352.9 380.5 66.9 3.56 4.59 66.55 1.22 92.7 163.9 199.4
46 Distributed 292.8 381 203.2 10.16 19.89 63.5 1.3 310.5 282.1 352.5
47 One-point 297.3 450.3 123.4 7.54 10.68 150 1.15 279 295.5 294.9

available data sets were randomly divided into learning, validation, were further employed to measure the performance of the optimal
and testing subsets. The learning data were taken for training model obtained by GEP on data that played no role in building the
(genetic evolution). The validation data were used to specify the models. To obtain consistent data division, several combinations
generalization capability of the models on data they did not train of the training and testing sets were considered. The selection
on (model selection). Thus, both the learning and the validation was in a way that the maximum, minimum, mean, and standard
data were involved in the modeling process and were categorized deviation of parameters were consistent in the training and testing
into one group referred to as ‘‘training data’’. The model with the data sets. Of the 47 data sets, 38 data vectors were taken for the
best performance on both the learning and the validation data sets training process (30 sets for learning and 8 sets for validation). The
was finally selected as the outcome of the runs. The testing data remaining 9 sets were used for the testing of the derived model.
 A.H. Gandomi et al. / Journal of Constructional Steel Research 67 (2011) 1096–1105 1101

3.2. Performance measures Table 3

Parameter settings for the GEP algorithm.

The best model was chosen on the basis of a multi-objective Parameter settings
strategy as follows. General
i. The simplicity of the model, although this was not a predomi- Chromosomes 50, 150, 300
nant factor. Genes 1, 2, 3
ii. Providing the best fitness value on the learning set of data. Head size 3, 5, 8
Tail size 9
iii. Providing the best fitness value on a validation set of data. Dc size 9
The first objective was controlled by the user through the Gene size 26
Linking function Addition
parameter settings (e.g., number of genes or head size). For
the other objectives, the following objective function (OBJ) was Complexity increase
constructed as a measure of how well the model predicted output Generations without change 2000
agrees with the experimental results. The best GEP model was then Number of tries 3
deduced by the minimization of the following function: Maximum complexity 5
Genetic operators
No.Learning − No.Validation
 
MAELearning
OBJ = Mutation rate 0.044
No.Training R2Learning Inversion rate 0.1
IS transposition rate 0.1
2No.Validation MAEValidation
+ , (5) RIS transposition rate 0.1
No.Training R2Validation One-point recombination rate 0.3
Two-point recombination rate 0.3
where No.Train , No.Learning , and No.Validation are respectively the Gene recombination rate 0.1
number of training, learning, and validation data. R and MAE are Gene transposition rate 0.1
respectively the correlation coefficient and mean absolute error Numerical constants
given in the form of the following relationships: Constants per gene 2
n Data type Integer
(hi − hi )(ti − ti ) Lower bound −10
∑
i =1
Upper bound 10
R=  (6)
n n
(hi − hi ) (ti − ti )2
∑ 2
∑
were considered for the head size and the number of genes. For
i=1 i =1 the number of genes greater than 1, the addition linking function
n
∑ was used to link the mathematical terms encoded in each gene.
|hi − ti | One level was considered for the other parameters based on some
i=1
MAE = , (7) previously suggested values [13,19–21] and also after making
n several preliminary runs and observing the performance behavior.
in which hi and ti are respectively the actual and calculated outputs There are 3 × 3 × 3 = 27 different combinations of the parameters.
for the ith output; hi and ti are the average of the actual and All of these combinations were tested, and 10 replications for
calculated outputs, and n is the number of samples. It is well each combination were carried out. Therefore, the overall number
known that R alone is not a good indicator for predicting the of runs was equal to 27 × 10 = 270. The period of time
accuracy of a model. This is because, on equal shifting of the output acceptable for evolution to occur without improvement in best
values of a model, the R value does not change. The constructed fitness is set via the generations without change parameter. After
objective function takes into account the changes of R and MAE 2000 generations considered herein, a mass extinction or a neutral
simultaneously. Higher R and lower MAE values result in lowering gene was automatically added to the model. In this study, basic
OBJ, and hence indicate a more precise model. In addition, the arithmetic operators and mathematical functions were utilized to
above function takes the effects of different data divisions into get the optimum GEP model. The mean absolute error function
account for the learning and validation data. was used to calculate the overall fitness of the evolved programs.
The program was run until there was no longer any significant
3.3. Development of an empirical model using GEP improvement in the performance of the models. The GEP algorithm
was implemented using GeneXproTools [33].
Eight input parameters, Fyw , hc , B, tw , tf , S , L, and LC , were used
to create the GEP model. Various parameters involved in the GEP 3.3.1. GEP-based formulation for the load capacity of a CSB
predictive algorithm are shown in Table 3. The parameter selection The GEP-based formulation of the failure load (P) in terms of
will affect the model generalization capability of GEP. Several Fyw , hc , B, tw , tf , S , L, and LC is as given below:
runs were conducted to come up with a parameterization of GEP   
that provided enough robustness and generalization to solve the Fyw − B
PGEP (kN ) = tw L − L + (tw − 4)(tw − LC )
problem. The number of programs in the population that GEP will S
evolve is set by the population size (number of chromosomes). 
A run will take longer with a larger population size. The proper
(hc + Fyw )(Fyw − tw − 216)

3
population size depends on the number of possible solutions +
and the complexity of the problem. Three levels were set for
the population size. The chromosome architectures of the models
 
S
− (tf − LC ) + ( tf + hc − 10) tw − .
2

3 2
evolved by GEP include the head size and the number of genes. (8)
6
The head size determines the complexity of each term in the
evolved model. The number of terms in the model is determined The expression tree of the above formulation is given in Fig. 7.
by the number of genes per chromosome. Each gene codes for A comparison of the predicted against experimental failure load
a different sub-expression tree or sub-ET. Three optimal levels values is shown in Table 2 and Fig. 8.
1102 A.H. Gandomi et al. / Journal of Constructional Steel Research 67 (2011) 1096–1105

Fig. 9. Experimental versus predicted failure load values using the LSR model.

Fig. 10. Contributions of the predictor variables in the GEP model.

GEP technique in comparison with a classic statistical approach.

The LSR method is extensively used in primary regression analysis
due to its interesting nature. LSR minimizes the sum-of-squared
residuals for each equation, accounting for any cross-equation
restrictions on the parameters of the system. If no restrictions
exist, this technique becomes identical to estimating each equation
by single-equation ordinary least squares. The EViews software
package [35] was used to perform the regression analysis. The
formulation of the failure load (P), for the best results using the
LSR analysis, is as given below:
Fig. 7. Expression tree for the load capacity of a CSB; where d1 = LC , d2 = PLSR. (kN ) = 19.564LC + 3.926Fyw + 0.937hc − 0.455B
Fyw , d3 = hc , d4 = B, d5 = tw , d6 = tf , d7 = S , d8 = L.
+ 111.794tw − 19.869tf − 0.403S − 48.641L − 1772.193. (9)
A comparison of the experimental against predicted failure load
values is shown in Table 2 and Fig. 9.

4. Sensitivity analysis

Sensitivity analysis is of utmost concern for selecting the

important input variables. The contribution of each predictor
variable in the GEP model was evaluated through a sensitivity
analysis. For this aim, frequency values of the input variables were
obtained. A frequency value equal to 1.00 for an input indicates that
this variable has appeared in 100% of the best 30 programs evolved
by GEP. This methodology is a common approach in GP-based
analyses [13–15]. The frequency values of the predictor variables
are presented in Fig. 10. According to these figures, the failure load
Fig. 8. Experimental versus predicted failure load values using the GEP model. is more sensitive to LC , Fyw , and tw . This can be regarded as an
expected case from a structural engineering point of view.
3.4. Development of an empirical model using regression analysis
5. Performance analysis
In the conventional material modeling process, regression
analysis is an important tool in building models. In the present A new model was developed for the estimation of the failure
study, a multivariable least squares regression (LSR) analysis [34] load upon a reliable database. Based on a logical hypothesis [36], if
was initially performed to get an idea of the predictive power of the a model gives R > 0.8, and the MAE values are at the minimum,
 A.H. Gandomi et al. / Journal of Constructional Steel Research 67 (2011) 1096–1105 1103

a b

c d

e f

g h

Fig. 11. The ratio between the experimental and predicted failure load values with respect to the design parameters (average = 0.988).

there is a strong correlation between the predicted and measured Table 4

Statistical parameters of the GEP model for external validation.
values [23]. The model can therefore be judged as very good. Based
on the results, the proposed GEP model with high R and low MAE Item Formula Condition GEP
values is able to predict the target values to an acceptable degree 1 R ∑n
R > 0.8 0.904
of accuracy. The performance of the model on the training and i=1 (hi ×ti )
2 k= 2 0.85 < K < 1.15 1.003
∑n hi
testing data suggests that it has both good predictive ability and i=1 (hi ×ti )
3 k′ = 0.85 < K ′ < 1.15 0.974
generalization performance. ti2
R2 −Ro2
4 m= m < 0.1 −0.225
The models derived using soft computing techniques (e.g., neu- R2
R2 −Ro2
ral networks or GP-based approaches) mostly have a predictive 5 n= R′2
, n < 0.1 −0.203
∑n
(t −ho )2
Ro2 = 1 − ∑in=1 (t −t̄i )2 , hoi = k × ti
i
capability within the data range used for their development. The where
i=1 i i
1.000
amount of data used for the training process of these techniques ∑n
i=1 (hi −ti )
o 2

is an important issue, as it bears heavily on the reliability of the fi- Ro′2 = 1− ∑n , tio = k′ ×hi 0.983
i=1 ( h i −h̄ i)
2

nal models. To cope with this limitation, Frank and Todeschini [37]
argue that the minimum ratio of the number of objects over the
the predicted and experimental values (Ro2 ) or the coefficient be-
number of selected variables for model acceptability is 3. They also
tween the experimental and predicted values (Ro′2 ) should be close
suggest that considering a ratio equal to 5 is more reasonable. In to 1 [23]. The considered validation criteria and the relevant re-
the present study, this ratio is higher, and is equal to 47/8 = 5.9. sults obtained by the models are presented in Table 4. The results of
The above facts ensure that the final GEP model has prediction the model validity indicate that the derived GEP model is strongly
power and is not a chance correlation. Furthermore, new criteria valid.
recommended by Golbraikh and Tropsha [38] were checked for In addition, Fig. 11 shows the ratio of the experimental to
the external validation of the models on the validation data sets. the GEP predicted failure against different parameters. As the
It is suggested that at least one of the slopes of the regression scattering increases in these figures, the accuracy of the model
lines (k or k′ ) through the origin should be close to 1. The perfor- decreases. It can be observed that, with the exception of LC , the
mance indices of m and n should be lower than 0.1. Also, either scattering slightly decreases with increasing different parameters.
the squared correlation coefficient (through the origin) between In the case of LC , the results do not exhibit any noticeable
1104 A.H. Gandomi et al. / Journal of Constructional Steel Research 67 (2011) 1096–1105

trend. It is worth noting that the observed deviation between the • They derived equation is very simple and can readily be used
experimental and predicted failure load is not only due to the for practical pre-planning and pre-design purposes via hand
deficiency of the proposed model. It can partly be attributed to calculations.
uncertainties, errors, and inconsistencies in the data used for the • An observation from the results of the sensitivity analysis is that
training and testing of the model. the most important parameters governing the behavior of the
The results presented in Figs. 8 and 9 indicate that the GEP- load capacity of CSBs are the loading condition, minimum web
based formula significantly outperforms the LSR model on both yield stress, and web thickness.
the training and the testing sets. However, no rational model has • The constitutive equation derived using GEP is basically
been developed for the prediction of failure load including the different from the conventional constitutive models based on
variables considered in this study. Thus, it is not possible to conduct first principles (e.g., elasticity and plasticity theories) [23]. One
a comparative study between the results of this research and those of the distinctive features of the GEP-based model is it is
in hand. based on the experimental data rather than on assumptions
Note that one of the major advantages of the GEP approach made in developing the conventional models. Consequently, as
over traditional regression analysis is its ability to derive explicit more data become available, this model can be retrained and
relationships for failure load without assuming prior forms of the improved without repeating the development procedures from
existing relationships. The best solution (equation) evolved by this the beginning.
technique is determined after controlling numerous preliminary
models, even millions of linear and nonlinear models. However, Acknowledgement
one of the goals of introducing expert systems, such as the GP-
based approaches, into the design processes is better handling of The authors are thankful to Professor Mohammad Mehdi Alinia
the information in the pre-design phase. In the initial steps of (Amirkabir University of Technology, Tehran, Iran) for his support
design, information about the features and properties of targeted and stimulating discussions.
output or process are often imprecise and incomplete [39].
Nevertheless, it is idealistic to have some initial estimates of References
the outcome before performing any extensive laboratory work.
The GEP approach employed in this research is based on the [1] Aglan AA, Redwood RG. Web buckling in castellated beams. Proceedings of the
data alone to determine the structure and parameters of the Institution of Civil Engineers, Part 2: Research and Theory 1974;57:307–20.
[2] Dougherty BK. Castellated beams: a state of the art report. Technical report.
models. Thus, the derived constitutive models can be particularly Journal of the South African Institution of Civil Engineering. 35 (2). 1993.
valuable in the preliminary design stages. For more reliability, p. 12–20.
the results of the GEP-based analyses are suggested to be treated [3] Zirakian T, Showkati H. Distortional buckling of castellated beams. Journal of
Constructional Steel Research 2006;62:863–71.
as a complement to conventional computing techniques (such [4] Amayreh L, Saka MP. Failure load prediction of castellated beams using
as the finite element method). In any case, the importance of artificial neural networks. Asian Journal of Civil Engineering (Building and
engineering judgment in the interpretation of the results obtained Housing) 2005;6(1–2):35–54.
[5] Kerdal D, Nethercot DA. Failure modes for castellated beams. Journal of
should not be underestimated. In order to develop a sophisticated Constructional Steel Research 1984;4:295–315.
prediction tool, GEP can be combined with advanced deterministic [6] Knowles PR. Castellated beams. Proceedings of the Institution of Civil
models. Assuming that the deterministic model captures the key Engineers, Part 1: Design and Construction 1991;90:521–36.
[7] Knowles PR. Design of castellated beams. (for use with BS 5950 and BS 449).
physical mechanisms, it needs appropriate initial conditions and Croydon (UK): Constrado; 1985.
carefully calibrated parameters to make accurate predictions. An [8] Pala M, Caglar N. A parametric study for distortional buckling stress on cold-
idea could be to calibrate the parameters by the use of GEP which formed steel using neural networks. Journal of Constructional Steel Research
2007;63:686–91.
takes into account historic data sets as well as the laboratory [9] Caglar N, Pala M, Elmas M, Eryilmaz DM. A new approach to determine the
test results. This allows integrating the uncertainties related to base shear of steel frame structures. Journal of Constructional Steel Research
testing conditions which the conventional constitutive models do 2009;65:188–95.
[10] Guzelbey IH, Cevik A, Erklig A. Prediction of web crippling strength of cold-
not explicitly account for [23]. formed steel sheetings using neural networks. Journal of Constructional Steel
Research 2006;62(10):962–73.
[11] Koza JR. Genetic programming: on the programming of computers by means
6. Conclusion of natural selection. Cambridge (MA): MIT Press; 1992.
[12] Banzhaf W, Nordin P, Keller R, Francone F. Genetic programming—an
A robust variant of GP, namely GEP, was utilized to formulate introduction. In: On the automatic evolution of computer programs and its
application. Heidelberg, San Francisco: Dpunkt, Morgan Kaufmann; 1998.
the load capacity of castellated steel beams. An accurate empirical [13] Gandomi AH, Alavi AH, Mirzahosseini MR, Moghadas Nejad F. Non-
model was derived for the prediction of the failure load. A reliable linear genetic-based models for prediction of flow number of asphalt
database from previously published failure load test results was mixtures. Journal of Materials in Civil Engineering, ASCE 2011;23(3):
1–18.
used to develop the model. The following conclusions are drawn [14] Alavi AH, Gandomi AH, Sahab MG, Gandomi M. Multi expression program-
based on the results presented. ming: a new approach to formulation of soil classification. Engineering with
Computers 2010;26(2):111–8.
• The proposed GEP-based model is capable of predicting the [15] Gandomi AH, Alavi AH, Sahab MG. New formulation for compressive strength
failure loads of CSBs with high accuracy. The validity of the of CFRP confined concrete cylinders using linear genetic programming.
model was tested for a part of test results beyond the training Materials and Structures 2010;43(7):963–83.
[16] Gandomi AH, Alavi AH, Kazemi S, Alinia MM. Behavior appraisal of steel semi-
data domain. Furthermore, the GEP prediction model efficiently rigid joints using linear genetic programming. Journal of Constructional Steel
satisfies the conditions of different criteria considered for its Research 2009;65(8–9):1738–50.
external validation. The validation phases confirm the efficiency [17] Ferreira C. Gene expression programming: a new adaptive algorithm for
solving problems. Complex Systems 2001;13(2):87–129.
of the model for its general application to the load capacity [18] Oltean M, Grosan C. A comparison of several linear genetic programming
estimation of CSBs. techniques. Complex Systems 2003;14(4):1–29.
• Due to the nonlinearity in collapse behavior of CSBs, the [19] Pala M. Genetic programming-based formulation for distortional buckling
stress of cold-formed steel members. Journal of Constructional Steel Research
GEP model produces considerably better outcomes than the 2008;64:1495–504.
multivariable linear regression-based model. [20] Cevik A. A new formulation for web crippling strength of cold-formed steel
• The proposed model simultaneously takes into account the role sheeting using genetic programming. Journal of Constructional Steel Research
2007;63:867–83.
of several important factors representing the failure load of CSB [21] Cevik A. Genetic programming based formulation of rotation capacity of wide
behavior. flange beams. Journal of Constructional Steel Research 2007;63:884–93.
 A.H. Gandomi et al. / Journal of Constructional Steel Research 67 (2011) 1096–1105 1105

[22] Javadi AA, Rezania M. Applications of artificial intelligence and data [30] Van Oostrom J, Sherbourne AN. Plastic analysis of castellated beams—part II:
mining techniques in soil modeling. Geomechanics and Engineering 2009;1: analysis and tests. Computers and Structures 1972;2:111–40.
53–74. [31] Maalek S, Burdekin FM. Weld quality requirements for castellated beams.
[23] Alavi AH, Ameri M, Gandomi AH, Mirzahosseini MR. Formulation of flow Structural Engineer 1991;13(69):243–54.
number of asphalt mixes using a hybrid computational method. Construction [32] Sherbourne AN, Van Oostrom J. Plastic analysis of castellated beams—part I:
and Building Materials 2011;25:1338–55. interaction of moment, shear and axial force. Computers and Structures 1972;
[24] Torres RS, Falcão AX, Gonçalves MA, Papa JP, Zhang B, Fan W, et al. A 2:79–109.
genetic programming framework for content-based image retrieval. Pattern [33] GEPSOFT. GeneXpro Tools. Version 4.0. 2006. Available at:
Recognition 2009;42:283–92. http://www.gepsoft.com.
[25] Ferreira C. Gene expression programming: mathematical modeling by an
[34] Ryan TP. Modern regression methods. New York: Wiley; 1997.
artificial intelligence. 2nd ed. Germany: Springer-Verlag; 2006.
[35] Maravall A, Gomez V. EViews software, ver. 5. Irvine (CA): Quantitative Micro
[26] Sadat Hosseini SS, Gandomi AH. Short-term load forecasting of power systems
Software, LLC; 2004.
by gene expression programming. Neural Computing and Applications, (in
press) doi:10.1007/s00521-010-0444-y. [36] Smith GN. Probability and statistics in civil engineering. London: Collins; 1986.
[37] Frank IE, Todeschini R. The data analysis handbook. Amsterdam: Elsevier;
[27] Redwood RG, Demirdjian S. Castellated beam web buckling in shear. Journal
1994.
of Structural Engineering 1998;124(10):1202–7.
[28] Hosain MU, Speirs WG. Experiments on castellated steel beams. Journal [38] Golbraikh A, Tropsha A. Beware of q2 . Journal of Molecular Graphics and
of American Welding Society, Welding Research Supplement 1973;8(52): Modelling 2002;20:269–76.
329S–42. [39] Kraslawski A, Pedrycz W, Nyström L. Fuzzy neural network as instance
[29] Okubo T, Nethercot DA. Web post strength in castellated beams. Proceedings generator for case-based reasoning system: an example of selection of heat
of the Institution of Civil Engineers, Part 2: Research and Theory 1985;79: exchange equipment in mixing. Neural Computing and Applications 1999;8:
533–57. 106–13.