Applied Mathematical Modelling 48 (2017) 899–910

Contents lists available at ScienceDirect

Applied Mathematical Modelling

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

M5 model tree and Monte Carlo simulation for efficient

structural reliability analysis
Behrooz Keshtegar a,∗, Ozgur Kisi b
a
Department of Civil Engineering, University of Zabol, P.B. 9861335-856 Zabol, Iran
b
Canter for Interdisciplinary Research, International Black Sea University, Tbilisi, Georgia

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

Article history: Practically, the performance of many engineering problems can be defined using a com-
Received 26 August 2016 plex implicit limit state function. Approximation of the accurate failure probability is very
Revised 15 January 2017
time-consuming and inefficient based on Monte Carlo simulation (MCS) for complex per-
Accepted 22 February 2017
formance functions. M5 model tree (M5Tree) model is robust approach for simulation and
Available online 2 March 2017
prediction phenomena, which provides ability to dealing with complex implicit problems
Keywords: by dividing them into smaller problems. By improving the efficiency of reliability method
Reliability analysis using accurate approximated failure probability, an efficient reliability method using the
Monte Carlo simulation MCS and M5Tree is proposed to calibrate the performance function and estimate the fail-
M5 model tree ure probability, respectively. The superiorities including simplicity and accuracy of M5Tree
Failure probability meta-model are investigated to evaluate the actual performance function through five non-
Performance function linear complex mathematical and structural reliability problems. The proposed reliability
method-based MCS and M5Tree improved the computational efforts for evaluating the per-
formance function in reliability analysis. The M5Tree significantly increased the efficiency
of reliability analysis with accurate failure probability.
© 2017 Elsevier Inc. All rights reserved.

1. Introduction

In practice, the actual engineering systems inherently involve various uncertainties in their boundary conditions, mate-
rial properties, dimensions, analytical model and applied loadings that should be considered in the structural design based
on a probabilistic model. The structural reliability analysis approximates the failure probability by a probabilistic model to
estimate the safety degree, which can be evaluated by quantitative measures of these uncertainties. The various structural
reliability analysis methods were implemented and developed to estimate the failure probability such as Monte Carlo simu-
lation (MCS) [1,2], the first order reliability method (FORM) [3–7] and the second order reliability method (SORM) [8,9].
The main efforts in reliability methods are to estimate the failure probability using the limit state function (LSF; i.e.
g(X) = 0). The LSF is separated the design domain into the failure (g(X) < 0) and safety (g(X) > 0) regains. The FORM is de-
veloped to improve the efficiency of reliability analysis. However, these approaches may produce unstable results for some
cases of reliability problems [4,10–12], may provide an unacceptable reliability index for highly nonlinear performance func-
tions [6,12] and may also require more iterations for stabilization of complex or high-dimensional LSFs. The MCS is a general
method to accurate estimation of the failure probability in comparison with analytical methods i.e. (FORM and SORM) for

∗
Corresponding author.
E-mail addresses: [email protected], [email protected] (B. Keshtegar).

http://dx.doi.org/10.1016/j.apm.2017.02.047
0307-904X/© 2017 Elsevier Inc. All rights reserved.
900 B. Keshtegar, O. Kisi / Applied Mathematical Modelling 48 (2017) 899–910

highly nonlinear limit state functions. It is generally required tremendous computational efforts to find the exact solution-
based MCS for real structural engineering problems with implicit performance function and low failure probability due
to large sampling points for simulation and computationally time-consuming for structural analysis [13–16]. Consequently,
there may be excessively computational effort increased for implicit limit state probabilistic models where the finite element
method (FEM) is used to compute the output of performance function and may be obtained unacceptable reliability index
for engineering problems with small failure probability. Therefore, the MCS cannot be simply used to evaluate the failure
probability of the implicit limit state function.
In order to enhance the efficiency of reliability methods and reduce the number of evaluating the LSF-based FEM analy-
sis, the approximation methods using the surrogate models were developed to calibrate the LSF such as the response surface
method (RSM) [17–21], neural networks [1,22,23], Kriging [13–15,24,25] and support vector machines [26–30] in reliability
analysis. These modeling approaches are a collection of statistical and mathematical techniques to calibrate the LSF based
on several input variables (random variables) that are given an explicit approximation of implicit LSF for real structural
problems based on several experimental points of deterministic structural analysis. Usually, the mathematical explicit per-
formance function with accurate predictions may not be easily obtained for nonlinear performance functions or LSFs. The
accurate prediction of these meta-models is essentially important to achieve an accurate failure probability with MCS, more
efficiently. Generally, an efficient meta-model is selected based on the accuracy and cost computation efforts in reliability
analysis.
In this paper, the M5Tree meta-model is applied for reliability analysis using MCS. Different from neural networks, the
advantage of M5Tree is that, it provides simple rules and can be calibrated faster. Furthermore, its rules can be presented
and easily used in practical applications. In contrast to existing data driven methods, M5Tree is an alternative method that
is quite transparent and the optimization of network geometry and internal parameters is not required [31]. The M5Tree
is used to enhance the efficiency of reliability analysis. The accurate predictions of the surrogate model-based M5Tree are
illustrated using five nonlinear mathematical and structural problems with normal and non-normal random variables. The
efficiency and accuracy of the failure probability using the M5Tree model through nonlinear examples are compared with
the FORM and MCS. To the knowledge of the writers, no study has been carried out so far that applies the M5Tree approach
for reliability analysis. The M5Tree method is successfully provided more accurate and robust results than the FORM and is
slightly more efficient than the MCS in structural reliability analysis.

2. Structural reliability analysis

The main goal of the reliability analysis methods is to approximate the failure probability Pf for evaluating the safety
level of structural components based on the LSF (g(.)). Generally, the failure probability can be defined as follows [10,32]:
 
Pf = ... f X ( x1 , . . . , xn )d x1 . . . d xn , (1)
g(X )≤0

where, g(X) is the limit state function (LSF) in X-space, g(X) ≤ 0 denotes the failure region and fX is the joint probability
density function for n-dimensional independent basic random variables X. The FORM and MCS can be implemented to esti-
mate the failure probability. The FORM is typically provided a good balance between accuracy and efficiency to approximate
the failure probability [33–35] and the MCS is a general reliability method with an accurate estimation of failure probability.

2.1. First-order reliability method

In the FORM, the LSF approximates the linear prediction in normal standard space at the most probable point (MPP
i.e. U∗ ) which is the point of minimum distance to the origin on the limit state surface. The reliability index (i.e. β = U∗ )
is used to estimate the failure probability using searching the MPP as follows [20,33]:
Pf ≈ (−β ) = (U ∗  ), (2)
where,  is the standard normal cumulative distribution function. Generally, the main effort of the FORM determines the
MPP (U∗ ) by the following iterative formula [6,12]:
∇uT g(Uk ) Uk − g(Uk )
Uk+1 = n ( uk ) , (3)
∇uT g(Uk )n(uk )
where, n(uk ) stands for the normalized steepest descent search direction at point Uk which can be determined using the
HL-RF method as follows:
∇u g(Uk )
n ( uk ) = −  , (4)
∇u g(U )
k

in which, ∇ u g(U) = [∂ g/∂ u1 , ∂ g/∂ u2 ,…, ∂ g/∂ un ]T is the gradient vector of the LSF g(U) in normal standard space. U is stan-
dard normal random variables that they are computed by transforming the basic random variables from the original space
(X-space) into the standard normal space (U-space) as U = T(X) i.e. u =  − 1 {FX (x)} [35,36]. The robustness of the FORM is ma-
jor criterion in HL-RF method because this approach is an efficient iterative formula but the HL-RF yields unstable results as
 B. Keshtegar, O. Kisi / Applied Mathematical Modelling 48 (2017) 899–910 901

periodic and chaotic solutions for nonlinear LSFs [11,37,38]. The conjugate search direction methods are successfully applied
for reliability analysis due to their robustness and efficiency [6,7,12,34,35,39]. The results from Refs [6,12] show that the con-
jugate search direction is more efficient than the other improved FORM using steepest descent search direction vector such
as improved HL-RF [40], relaxed HL-RF [32] and stability transformation method [37]. In order to achieve the stabilization of
FORM formula, the conjugate search direction vector for negative unit normal vector is applied by the following relations:
U +d
n ( uk ) =  k k
, (5)
U + d 
k k

where, dk the conjugate search direction vector, which is proposed as follows:

 
∇u g(U )2
dk = −∇u g(Uk ) − 5   d .
k
(6)
∇u g(U ) 2 k−1
k−1

For k = 0, d0 = −∇u g(U0 ).

The proposed conjugate search direction vector in Eq. (6) is major difference between the proposed conjugate FORM and
other conjugate methods such as conjugate HL-RF [34], chaotic conjugate chaos control [12] and conjugate stability trans-
∇u g(U )2
formation method [6]. In this FORM formula based on proposed conjugate FORM (CFORM), the factor ∇ g(U k ) 2 is scaled
u k−1
five-times larger than the other existing conjugate FORM and also does not use line search rules such as Armijo [6,34] or
sufficient descent condition [10,33,41]. Therefore, the proposed conjugate FORM (CFORM) is as simple as the HL-RF but is
simpler than the chaotic conjugate chaos control [12], conjugate HL-RF [34], and chaotic conjugate stability transformation
∇u g(U )2
method [6]. For highly nonlinear LSFs, a large value for factor ∇ g(U k ) 2 (five times larger than other exiting CFORM) is
u k−1
obtained thus a smaller conjugate search direction is obtained in comparison to other existing CFORM thus the difference
rate of the negative unit normal vector in Eq. (6) is less than the other conjugate reliability method. Consequently, the
proposed CFORM may be converged to a fixed point, more simply.

2.2. Monte Carlo simulation

The integration in Eq. (1) can be estimated based on the MCS as follows:
 
Pf = ... Ig(X )≤0 (X ) fX (x1 , . . . , xn )dx1 . . . dxn , (7)
g(X )≤0

where, Ig(X) ≤ 0 (X) is the failure indicator, it equals to one if g(X) ≤ 0 and zero otherwise. Generally, the MCS results can be
used as references due to their accuracy. Using the Eq. (7), The MCS estimator is given as follows [1,13]:
NS
i=1 Ig(X )≤0 (X )
PˆfMC = E[Ig(X )≤0 (X )] = , (8)
NS
NS
where, E[.] is the expectation operator and NS is the number of sample sizes and i=1 Ig(X )≤0 (X ) the number of samples
giving a negative or zero values of LSF (g(X)). The number of sample sizes NS is extremely important to estimate the ac-
curate failure probability. It requires a large random population for a very low failure probability. In order to quantify the
uncertainty on the failure probability, direct MCS (PˆMC ) is commonly measured based on its variance or the coefficient of
f

variation (CoVPMC ). The variance of PˆMC is estimated as follows [13,16]:

f f

2
PˆfMC − (PˆfMC )
V ar (PˆfMC ) ≈ . (9)
NS − 1
The coefficient of variation of PˆMC can be approximated using Eqs. (8) and (9) as follows:
f
 
V ar (PˆfMC ) 1 − PˆfMC
CoVPMC = = . (10)
f
E (PˆfMC ) (NS − 1 ) PˆfMC
If the variation coefficient of PˆMC is low enough (i.e. CoVPMC < 0.05), then the MCS can provide an accurate failure prob-
f f
ability [16]. Consequently, the accuracy of failure probability-based MCS can be measured using CoVPMC . It can be conducted
f
from Eq. (9) that the CoVPMC decreases by increasing the number of sampling size. For example, the required random sample
f
size for MCS should be selected about 4 × 105 samples for a target CoVPMC less than 0.05 and the failure probability up to
f
10−3 . The application of the MCS is easy for complex engineering problems but evaluating the failure probability requires
more calls of the performance function to achieve an accurate estimation. Consequently, the MCS can become impossible
when a complex computer codes with FE method is selected.
902 B. Keshtegar, O. Kisi / Applied Mathematical Modelling 48 (2017) 899–910

Fig. 1. Splitting input space in M5Tree algorithm.

3. M5 model tree (M5Tree) model

In M5Tree algorithm, the parameter space is divided into subspaces and local linear regression models are built for each
of subspaces. M5Tree is able to efficiently learn and tackle tasks with high dimensionality [42]. The M5Tree algorithm is
utilized for creating a model tree, which runs as follows (Fig. 1). Assume that we have a collection T of training examples.
Every case is described by the values of a fixed set of inputs and has a related output value. The point is to build a model
that find relationships between target value of the training cases and corresponding input values [31,43].
The division criterion is satisfied by reducing standard deviation (SD) of the each class values and computing the ex-
pected reduction in SD based error as a consequence of testing every characteristic at that node. The equation for processing
the SD reduction (SDR) is characterized as follows [41,43]:
 |Ti |
SDR = sd (T ) − sd (Ti ), (11)
|T |
which T means an examples set reaches the node; Ti signifies the subset of cases that have the ith results of the potential
set; sd indicates the standard deviation [44]. Because of the splitting procedure, the SD of the data in lower nodes is less
than that at the parent node. All the splits were examined and then the one that maximizes the expected SDR was chosen.
Whereas, the division procedure generally creates a structure of large tree, which may cause over-fitting or poor gener-
alization. To cope with this issue, the large tree is pruned and then pruned sub-trees are replaced with linear regression
functions. This method by generating the model tree generously expands the exactness of estimation [43].
The M5Tree and MCS to estimate the failure probability can be applied in the reliability analysis by the following steps:

Step 1: Generate the input data points. In this stage, the input data points are randomly generated from design space of each
random variable then the performance function is evaluated in each input data point based on the basic random
variables.
Step 2: Computation of the M5Tree model. According to the design point of experiments in step 1, the M5Tree model is
calibrated.
Step 3: Generate the Monte Carlo samples. The random variables are simulated by their statistical properties (e.g. distribution,
mean and standard deviation) with the number of samples NS.
Step 4: Predict the performance function. The sample points of the simulation in step 3 are predicted using the calibrated
M5Tree models in step 2. Therefore, the values of performance function are evaluated at each sample by the M5Tree
which is a predicted model of LSF (gˆ(X )).
Step 5: Estimate the failure probability. In this step, the failure probability is approximated by the predicted LSF at NS data
points in step 4 by the following relation:
NS
i=1 Igˆ(X )≤0 (X )
PˆfM5T ree = . (12)
NS
Step 6: Compute the coefficient of variation for the failure probability. The coefficient of variation is determined to check the
number of sample points. If the coefficient of variation in the below equation is less than the 5%, then it can be said
that an enough samples generated for estimating the failure probability.

1 − PˆfM5T ree
M5T ree
CoV = . (13)
(NS − 1 )PˆfM5T ree
 B. Keshtegar, O. Kisi / Applied Mathematical Modelling 48 (2017) 899–910 903

Table 1
The results of different reliability method for Example 1.

Method Ncall Pf β Abs. error (β )%

∗
MCS 1.2 × 10 6
0.0011 3.0605 –
M5Tree+MC 500 0.0015 2.9693 3.07
HL-RF 10,129 0.2065 0.8187 273.84
CFORM 1798 0.1750 0.9347 227.44
∗
CoV from the result obtained by MCS is 2.74% using 1.2 × 106 samples.

Table 2
The failure probabilities of Example 1 using M5Tree+MC for various number of call functions (Ncall ).

Ncall 100 150 200 250 300 350 400 450

Pf 8.65 × 10−3 3.90 × 10−3 6.77 × 10−3 5.67 × 10−3 6.64 × 10−3 6.53 × 10−4 7.25 × 10−4 1.61 × 10−3
β 2.3803 2.6610 2.4692 2.5322 2.4764 3.2145 3.1845 2.9462
Abs. error (β )% 28.57 15.03 23.95 20.86 23.59 4.79 3.89 3.88

Based on the above relation, CoV M5T ree < 0.05 then this algorithm can stop and the last estimation of failure probability
is considered as the final result of the M5Tree and MCS, else, go to step 3 and generate a new larger samples than the
previous samples and steps 4–6 are employed.

4. Mathematical/structural examples

The performances including efficiency, robustness and accuracy of the M5Tree and MCS (M5Tree+MC) and pro-
posed CFORM are compared to the MCS and FORM-based HL-RF through five nonlinear LSFs. The stopping criterion
(Uk+1 − Uk /Uk  < ε ) in FORM algorithms i.e. CFORM and HL-RF is given as ε = 10−6 and number of evaluating limit
state function (Ncall ), failure probability (Pf ) and reliability index (β ) are applied to illustrate the accuracy and efficiency of
reliability methods. The number of data points to calibrate the M5Tree is used in evaluating LSF and the generated samples
for MCS are implemented for estimating the failure probability in the M5Tree+MC method. The M5Tree models for each LSF
are regressed by the MATLAB toolbox and is then applied based on the MSC approach.

Example 1. The nonlinear high-dimensional problem [45]


D −1
2
[100(xi+1 − x2i ) + (xi − 1 ) ] − 150.
2
g= (14)
i=1

In which, D is the dimensions of the Rosenbrock problem in Eq. (14) that is given as D = 10, and xi (i = 1, 2, 3, …, 10) are
standard normal random variables. The results including the number of evaluating LSF (Ncall ), failure probability (Pf ), relia-
bility index (β ) and absolute error for reliability index (Abs. error (β )%) as (β − β MC )/β × 100) using the MCS, M5Tree+MCS,
HL-RF and CFORM are tabulated in Table 1. It is demonstrated that the MCS provides an accurate failure probability 0.0011
(β = 3.0605) with enough sample points (Cov = 2.74 < 5%). The HL-RF and CFORM are converged to stable results but inaccu-
rate reliability indices are provided as 0.8187 and 0.9347, respectively. However, the proposed M5Tree+MC is more accurate
(β = 2.9693) than the FORM-based steepest descent search direction and conjugate search direction for this nonlinear exam-
ple. The M5Tree+MC is slightly more efficient than the MCS and can be produced a suitable failure probability with accurate
estimation in comparison with FORM. The M5Tree+MC method is more efficient and accurate than the FORM algorithms
(i.e. HL-RF and CFORM) for this example.
The failure probability, reliability index and absolute errors of reliability index for various random train data points to
calibrate the M5Tree models are listed in Table 2. As seen, the M5Tree models, which are calibrated with the random
input data more than 300 experiments, can be evaluated the failure probability for Example 1, more accurately. The failure
probabilities with 10 0–20 0 input experiments are more accurate and efficient than the FORM methods (see Tables 1 and 2).
The M5Tree+MC method is produced an appropriate failure probability with a small data evaluations of LSF, while it is more
efficient than the MCS and FORM for this highly nonlinear problem. The accuracy of the failure probabilities is depended on
the train data set points for highly nonlinear performance functions, thus an appropriate selection of data point can improve
the predictions of LSF using the M5Tree model.

Example 2. Roof truss

A roof truss subject to the uniform load as shown in Fig. 2, chords and compression bars of the truss are made by
steel reinforced concrete, and the bottom chords and tension bars are made by steel. According to structural mechanics, the
perpendicular deflection of truss peak node C (c) can be given for performance function as follows [16]:


ql 2 3.81 1.13
g = 0.03 − + , (15)
2 AcEc AsEs
904 B. Keshtegar, O. Kisi / Applied Mathematical Modelling 48 (2017) 899–910

Fig. 2. Schematic view of roof truss example.

Table 3
Basic random variables for roof truss.

Random variable q(N/m) l(m) As(m2 ) Ac(m2 ) Es(Pa) Ec(Pa)

−4
Mean 20,0 0 0 12 9.82 × 10 0.04 1 ×1011
2 × 1010
Standard deviation 1400 0.12 5.9852 × 10−5 0.0048 6 × 109 1.2 × 109

Table 4
The results of different reliability methods for roof truss example.

Method Ncall Pf β Abs. error (β )%

MCS 9.5 ×105 0.00949∗ 2.3460 –

HL-RF – – Chaos Not converged
CFORM 806 0.00770 2.4227 3.17
M5Tree+MC 100 0.00883 2.3727 1.12
200 0.00930 2.3536 0.32
300 0.00957 2.3426 −0.15
400 0.00951 2.3452 −0.04
500 0.00953 2.3445 −0.07
∗
CoV from the result obtained by MCS is 1.05% using 9.5 ×105 samples.

where Ac, As are the cross sectional areas of reinforced concrete and steel bars, respectively, Ec, Es are their corresponding
elastic modulus, l is the length of the truss. This example involves six normal random variables, whose statistical properties
are presented in Table 3.
The results of the failure probabilities for MCS, M5Tree+MC, HL-RF and CFORM methods are listed in Table 4. The
coefficient of variation for the result obtained by MCS method is 1.048% using 9.5 × 105 samples. It can be seen from
Table 4 that the results of M5Tree+MC are all close to the result of MCS while the HL-RF yields unstable results as chaotic
solutions. The MCS, M5Tree+MC and CFORM give a reliability index after 9.5 × 105 , 10 0–50 0 and 806 number of call func-
tions, respectively. The M5Tree+MC need 100 calls of the actual function to obtain an accurate and robust result compared
to the FORM-based HL-RF and conjugate gradient methods. Therefore, the proposed M5Tree+MC method needs a few sam-
ples to estimate the actual LSF and to approximate the accurate failure probability, more efficiently.

Example 3. A dynamic system with two degree of freedom primary-secondary

The force capacity of the secondary spring, which its schematic view is shown in Fig. 3, is considered for performance
function as follows [35,39]:

g = F s − Ks × P (E[x2s ] )1/2 , (16)

where, P is the peak factor which is considered as 3 for simplicity and E[x2s ] is mean-square relative displacement response
which can be given as follows:

π S0 ξa ξs (ξ p ω3p + ξs ωs3 )ω p
E[x2s ] = × , (17)
4ξs ωs ξ p ξs (4ξa2 + θ 2 ) + γ ξa2 4ξa ωa4
 B. Keshtegar, O. Kisi / Applied Mathematical Modelling 48 (2017) 899–910 905

Fig. 3. Schematic view of a dynamic system with two-degree of freedom.

Table 5
Statistical properties of random variables for two degree dynamic system.

Random variable Mean Standard deviation Random variable Mean Standard deviation

Mp 1 0.1 ξp 0.05 0.02

Ms 0.01 0.001 ξs 0.02 0.01
Kp 1 0.2 Fs 15 1.5
Ks 0.01 0.002 S0 100 10

Table 6
The results of different reliability methods for two degree dynamic system.

Method Ncall Pf β CPU-time (sec) Abs. error (β )%

MCS 7.5 ×105 0.00424∗ 2.6321 51.2 –

HL-RF – Periodic − 2 as – Not converged
{4.9808, 4.2172}
M5Tree+MC 300 0.00572 2.5292 0.2 4.07
400 0.00365 2.6830 0.4 1.90
500 0.00473 2.5950 0.7 1.43
∗
CoV from the result obtained by MCS is 1.77% using 7.5 × 105 samples.

ω p +ωs ξ p + ξs
in which, mass ratio γ = Ms
Mp , average frequency ωa = 2 and average damping ratio ξa = 2 , tuning parameter
ω −ω
θ = pωa s and intensity of the white noise S0 . The means and standard deviations of basic random variables with Log-
normal distribution are given in Table 5.
The results of the failure probability (Pf ), reliability index (β ), number of call function to evaluate the failure probability
(Ncall ) and the reliability index errors for different reliability methods such as FORM-based HL-RF, CFORM, MCS and proposed
M5Tree+MC are tabulated in Table 6. As seen, the HL-RF method yields the unstable results as periodic-2 i.e. β = {4.9808,
4.2172}, but the proposed CFORM is converged to stable results after 2954 number of call functions as β = 2.0245. The
MCS provides a failure probability of 0.00424 (β = 2.6321) with CoV about 1.77% using 7.5 × 105 samples. The CFORM is
slightly more efficient than the MCS but it produces a reliability index with an about 30% error compared to MCS method.
The proposed M5Tree+MC method can provide a suitable failure probability, more efficiently. The M5Tree+MC with 300
samples points to calibrate the M5Tree model is about 10-times faster than the proposed CFORM and is more accurate than
the CFORM for this example.

Example 4. A steel frame

A steel frame under a lateral load P, and external distributed load Was shown in Fig. 4 is given using an implicit LFS
based on the horizontal displacement of roof () as follows [6]:

g = 0.05 − , (18)
where,  is evaluated using the finite element model based on the three frame elements. The schematic view of steel frame
with span L and height H and three elements are plotted in Fig. 4. The Young’s modulus (E), cross-sectional (A), moment of
initial (I) and dimensions are used to compute the stiffness matrix of beam (E2 , A2 , I2 , L) and columns (E2 , A1 , I1 , A2 , I2 , H)
elements. The horizontal displacement of steel frame roof is computed based on a twelve normal and non-normal random
variables, whose statistical properties are tabulated in Table 7.
The obtained results using different reliability methods such as MCS, HL-RF, proposed CFORM and proposed M5Tree+MC
are shown in Table 8. It can be seen that the traditional FORM formula using HL-RF yields unstable results as chaotic
solutions while the proposed CFORM formula converges to stable result of reliability index 3.2103 after 2215 evaluating call
function to achieve stabilization. The CFORM is more robust than the HL-RF and produces an accurate reliability index. The
proposed M5Tree+MC obtains a very close reliability index with the MCS compared to the CFORM, more efficiently. The
M5Tree+MC method is about 7-times faster than the CFORM and is more accurate than the CFORM for this example.
906 B. Keshtegar, O. Kisi / Applied Mathematical Modelling 48 (2017) 899–910

Fig. 4. The steel frame structure.

Table 7
Statistical properties of the random variables for steel frame example.

Random variable Mean Standard deviation Distribution Random variable Mean Standard deviation Distribution

A1 (m2 ) 91 × 10−4 9.1 × 10−4 Lognormal E1 (kg/m2 ) 2.38 ×1010 1.9 × 109 Normal
I1 (m4 ) 80.9 × 10−6 8.09 × 10−6 Lognormal E2 (kg/m2 ) 2.17 × 1010 1.74 × 109 Normal
A2 (m2 ) 53.8 × 10−4 5.38 × 10−4 Lognormal L(m) 6 0.48 Normal
I2 (m4 ) 83.6 × 10−6 8.36 × 10−6 Lognormal H(m) 4.5 0.36 Normal
A3 (m2 ) 106 × 10−4 10.6 × 10−4 Lognormal W(kg/m) 4500 630 Gumbel
I3 (m4 ) 112.6 × 10−6 11.26 × 10−6 Lognormal P(kg) 20 0 0 350 Gumbel

Table 8
The results of different reliability methods for steel frame.

Method Ncall Pf β CPU-time (sec) Abs. error (β )%

−4 ∗
MCS 1.2 × 10 6
10.93 × 10 3.0639 2967 –
HL-RF – Chaos – Not converged
CFORM 2215 6.63 × 10−4 3.2103 59.4 4.56
M5Tree+MC 300 9.23 × 10−4 3.1141 6.3 1.61
∗
CoV from the result obtained by MCS is 2.76% using 1.2 × 106 samples.

Example 5. A planar ten-bar structure

The schematic diagram of a planar ten-bar, linear-elastic, truss structure is shown in Fig. 5 which is given for better illus-
trating the engineering application with an implicit LSF. The section area, length and Young’s modulus of all the horizontal
and vertical bars are A, L and E, respectively. According to the analysis of the finite element model, the limited vertical
displacement of node 3 (V3 ) for this structure is given for implicit limit state function as follows [39]:
 
g = 0.004 − V3 . (19)
Young’s modulus of the material for all bars (E), the length of horizontal and vertical bars (L), the section area of each bar
(Ai , i = 1,2,3,…10) and the point loads (P1 , P2 , P3 ) are fifteen random variables for this example, whose statistical properties
with normal and non-normal distribution are listed in Table 9. The maximum displacement of the 10-bar truss at the mean
point is obtained based on a finite element model as 0.00291501, which is almost same with the extracted results from Ref
[15] i.e. 0.002915 m.
The reliability results for different reliability methods are listed in Table 10. It can be clearly seen that the proposed
CFORM is converged to stable reliability index but the HL-RF is not converged. The reliability indexes are obtained using
MCS, CFORM and M5Tree+MC as 1.5226, 1.7827 and 1.5809 after 5 × 105 , 1258 and 300 samples, respectively. The proposed
M5Tree+MC shows high efficiency and accuracy compared to the MCS and CFORM, respectively. The proposed M5Tree+MC
produces stable results for this example, more efficiently.
The results of the complicated engineering problems from Tables 8 and 10 illustrate that the proposed CFORM is more ro-
bust than the HL-RF and can estimate an appropriate reliability index for moderately performance function but the proposed
M5Tree+MC is more accurate and efficient than the CFORM. The proposed M5Tree+MC method has the computational ef-
ficiency and accuracy for approximating the failure probabilities. Therefore, the proposed M5Tree+MC can be implemented
for reliability analysis of these kinds of real engineering problems with accurate results.
 B. Keshtegar, O. Kisi / Applied Mathematical Modelling 48 (2017) 899–910 907

Fig. 5. The schematic view of planar ten-bar structure.

Table 9
Statistical properties of random variables for ten-bar structure.

Random variable Ai (m2 ) L (m) E(Gpa) P1 (kN) P2 (kN) P3 (kN)

Mean 0.001 1 100 80 10 10

Coefficient of variation 0.15 0.05 0.05 0.1 0.1 0.1
Distribution Normal Log-normal Normal Gumbel Gumbel Gumbel

Table 10
The results of different reliability methods for 10-bar truss.

Method Ncall Pf β CPU-time (sec) Abs. error (β )%

∗
MCS 5 × 10 5
0.063936 1.5226 6463 –
HL-RF – Chaos – Not converged
CFORM 1258 0.037319 1.7827 31.2 14.59
M5Tree+MC 300 0.056952 1.5809 8.3 3.69
∗
CoV from the result obtained by MCS is 0.54% using 5 ×105 samples.

Example 6. A five-bay ten-storey steel frame

The analytical configuration, dimensions and applied external loads including wind (P), dead (D) and live (L) loads of the
five-bay ten-storey are semantically shown in Fig. 6. This example is a real steel frame with rigid joints that it has 6 fixed
boundary conditions, 60 rigid beam-column connection points, 110 members which consist 50 beams and 60 columns with
180° of freedom (DOF). The maximum deflection of top storey is applied for implicit limit stat function as follows:

g = 0.075 − . (20)

The finite element model is applied to evaluate the maximum deflection of top storey. The stiffness matrix of each
element is computed using the cross-sectional (A) and moment of initial (I) of beams and columns which are listed in
Table 11. Two categories C1 and C2 are used for columns with three HEB profiles in storey 1–3, 4–6 and 7–10 for each
column category. The IPE330 profile is applied for all beams.
This example involves fifteen basic variables including ten different lateral loads at each story P1 –P10 , floor dead load
(Ds ); floor live load (Ls ); roof dead load (Dr ); roof live load (Lr ); and Young’s modulus (E) whose statistical properties are
listed in Table 12.
The reliability results including failure probability (Pf ), reliability index (β ), number of call performance function in
Eq. (20) (Ncall ), total times for evaluating the performance function (CPU-time) and relative errors of the reliability indices
of the different reliability methods including CFORM and M5Tree+MC with MCS are tabulated in Table 13. It can be con-
ducted that the HL-RF method yields the unstable results as chaotic solutions while the proposed CFORM is converged,
more robustly. The proposed M5Tree+MC provides an accurate reliability index (i.e. abs. error (β ) = 1.27%) compared to
the CFORM (i.e. abs. error (β ) = 8.48%) result. The MCS computationally obtains a reliability index of 2.4573 with 2 ×105
samples that total CPU times to evaluate the reliability index is 53,908 s (898.45 min). However, the total CPU times for pro-
posed M5Tree+MC are captured 33 s (0.55 min) that this approach is about 36-times faster than the CFORM. The surrogate
model-based M5tree can improves the computational efforts and the accuracy of the reliability analysis, more affectively.
908 B. Keshtegar, O. Kisi / Applied Mathematical Modelling 48 (2017) 899–910

Fig. 6. Schematic view of five-bay ten-storey frame.

Table 11
The column and beam profiles for five-bay ten-storey frame.

Group C1 C2 Beam

Storey 1–3 4–6 7–10 1–3 4–6 7–10 All

Profile HEB280 HEB240 HEB200 HEB340 HEB300 HEB240 IPE330
A (cm2 ) 131 106 78.1 171 149 106 62.8
I (cm4 ) 19,270 11,260 5700 36,660 25,170 11,260 11,770

5. Conclusions

This paper proposes two reliability methods, first order reliability method (FORM) using conjugate search direction
(CFORM) and M5 tree model with Monte Carlo simulation (M5Tree+MC). The CFORM is developed to improve the robust-
ness of FORM formula. The M5Tree meta-model and Monte Carlo simulation (MCS) are combined to enhance the efficiency
and accuracy of the reliability analyses.
The proposed CFORM method exhibits stable results compared to traditional FORM-based steepest descent search di-
rection (i.e. HL-RF) but the failure probabilities are inaccurately obtained using the proposed CFORM for highly nonlinear
performance functions. The MCS provides an accurate failure probabilities but this approach is computationally inefficient
for the actual time-consuming performance function. The M5Tree meta-model with MCS is used to predict the performance
function to improve the efficiency and accuracy of reliability method. Only a few sample points were evaluated for the
actual performance function for accurately prediction using M5Tree model. Thus, the MCS can accurately approximate the
failure probability using the M5Tree meta-model to predict the actual limit state functions of complex engineering problems.
Consequently, the proposed M5Tree+MC may improve the accuracy of reliability analysis, more efficiently.
 B. Keshtegar, O. Kisi / Applied Mathematical Modelling 48 (2017) 899–910 909

Table 12
Statistical properties of random variables for five-bay ten-storey frame.

Random variable Unit Mean Standard deviation COV Distribution

P1 kN 17.37 2.0844 0.12 Gumbel

P2 kN 17.83 2.1396 0.12 Gumbel
P3 kN 18.46 2.2152 0.12 Gumbel
P4 kN 20.89 2.5068 0.12 Gumbel
P5 kN 21.54 2.5848 0.12 Gumbel
P6 kN 22.63 2.7156 0.12 Gumbel
P7 kN 23.54 2.8248 0.12 Gumbel
P8 kN 23.86 2.8632 0.12 Gumbel
P9 kN 24.45 2.934 0.12 Gumbel
P10 kN 13.35 1.602 0.12 Gumbel
Ds kN/m 20.85 2.085 0.1 Weibull
Ls kN/m 8.76 1.314 0.15 Weibull
Dr kN/m 30.68 3.068 0.1 Weibull
Lr kN/m 35.04 5.256 0.15 Weibull
E GPa 210 21 0.1 Normal

Table 13
The results of different reliability methods for five-bay ten-storey frame.

Method Ncall Pf β CPU-time (sec) Abs. error (β )%

MCS 2 × 105 0.007∗ 2.4573 53,908 –

HL-RF – Chaos – Not converged
CFORM 25,668 0.00363 2.6848 1214 8.48
M5Tree+MC 500 0.00763 2.4264 33 1.27
∗
CoV from the result obtained by MCS is 2.66% using 2 × 105 samples.

Results of validating examples demonstrate that the proposed M5Tree+MC method performs very well in both efficiency
and accuracy compared to the CFORM and is slightly more efficient than MCS in explicit and implicit engineering prob-
lems. Furthermore, the proposed M5Tree+MC method can also be applied for reliability analysis of highly nonlinear real
engineering problems with high dimensional and non-normal random variables.

References

[1] M. Papadrakakis, N.D. Lagaros, Reliability-based structural optimization using neural networks and Monte Carlo simulation, Comput. Methods Appl.
Mech. Eng. 191 (2002) 3491–3507.
[2] V. Papadopoulos, D.G. Giovanis, N.D. Lagaros, M. Papadrakakis, Accelerated subset simulation with neural networks for reliability analysis, Comput.
Methods Appl. Mech. Eng. 223 (2012) 70–80.
[3] A.M. Hasofer, N.C. Lind, Exact and invariant second-moment code format(for reliability analysis in multivariate problems), American Society of Civil
Engineers, J. Eng. Mech. Div. 100 (1974) 111–121.
[4] B. Keshtegar, M.O. Sadegh, Instabilities control of reliability analysis using hybrid HL-RF and conjugate optimization algorithms, Fuzzy and Intelligent
Systems (CFIS), in: 2015 4th Iranian Joint Congress on, IEEE, 2015, pp. 1–4.
[5] G. Periçaro, S. Santos, A. Ribeiro, L. Matioli, HLRF–BFGS optimization algorithm for structural reliability, Appl. Math. Model. 39 (2015) 2025–2035.
[6] B. Keshtegar, Chaotic conjugate stability transformation method for structural reliability analysis, Comput. Methods Appl. Mech. Eng. 310 (2016)
866–885.
[7] B. Keshtegar, Limited conjugate gradient method for structural reliability analysis, Eng. Comput. (2016) 1–9, doi:10.10 07/s0 0366-0 0 016-0 0493-0 0367.
[8] J. Lim, B. Lee, I. Lee, Second-order reliability method-based inverse reliability analysis using Hessian update for accurate and efficient reliability-based
design optimization, Int. J. Numer. Methods Eng. 100 (2014) 773–792.
[9] D. Yoo, I. Lee, H. Cho, Probabilistic Sensitivity Analysis for Novel Second-Order Reliability Method using Generalized Chi-Squared Distribution, Struct.
Multidiscip. Optim. 50 (2014) 787–797.
[10] B. Keshtegar, I. Lee, Relaxed performance measure approach for reliability-based design optimization, Struct. Multidiscip. Optim. 54 (2016) 1439–1454.
[11] Z. Meng, G. Li, B.P. Wang, P. Hao, A hybrid chaos control approach of the performance measure functions for reliability-based design optimization,
Comput. Struct. 146 (2015) 32–43.
[12] B. Keshtegar, Stability iterative method for structural reliability analysis using a chaotic conjugate map, Nonlinear Dyn. 84 (2016) 2161–2174.
[13] B. Echard, N. Gayton, M. Lemaire, AK-MCS: an active learning reliability method combining Kriging and Monte Carlo simulation, Struct. Saf. 33 (2011)
145–154.
[14] B. Echard, N. Gayton, M. Lemaire, N. Relun, A combined importance sampling and Kriging reliability method for small failure probabilities with
time-demanding numerical models, Reliab. Eng. Syst. Saf. 111 (2013) 232–240.
[15] L. Zhang, Z. Lu, P. Wang, Efficient structural reliability analysis method based on advanced Kriging model, Appl. Math. Model. 39 (2015) 781–793.
[16] H. Zhao, Z. Yue, Y. Liu, Z. Gao, Y. Zhang, An efficient reliability method combining adaptive importance sampling and Kriging metamodel, Appl. Math.
Model. 39 (2015) 1853–1866.
[17] S. Chakraborty, A. Sen, Adaptive response surface based efficient finite element model updating, Finite Elem. Anal. Des. 80 (2014) 33–40.
[18] X.-J. Meng, S.-K. Jing, L.-X. Zhang, J.-H. Liu, H.-C. Yang, A new sampling approach for response surface method based reliability analysis and its
application, Adv Mech. Eng. 7 (2015) 305473.
[19] N. Roussouly, F. Petitjean, M. Salaun, A new adaptive response surface method for reliability analysis, Probab. Eng. Mech. 32 (2013) 103–115.
[20] S. Zenab, R. Chowdhury, A semi-analytical framework for structural reliability analysis, Computer Methods in Applied Mechanics and Engineering 289
(2015) 475–497.
[21] S. Chakraborty, R. Chowdhury, Multivariate function approximations using the d-MORPH algorithm, Appl. Math. Model. 39 (2015) 7155–7180.
[22] J. Cheng, Q. Li, Reliability analysis of structures using artificial neural network based genetic algorithms, Comput. Methods Appl. Mech. Eng. 197 (2008)
3742–3750.
910 B. Keshtegar, O. Kisi / Applied Mathematical Modelling 48 (2017) 899–910

[23] A. Chojaczyk, A. Teixeira, L. Neves, J. Cardoso, C.G. Soares, Review and application of artificial neural networks models in reliability analysis of steel
structures, Struct. Saf. 52 (2015) 78–89.
[24] B. Gaspar, A. Teixeira, C.G. Soares, Assessment of the efficiency of Kriging surrogate models for structural reliability analysis, Probab. Eng. Mech. 37
(2014) 24–34.
[25] X. Huang, J. Chen, H. Zhu, Assessing small failure probabilities by AK–SS: an active learning method combining Kriging and subset simulation, Struct.
Saf. 59 (2016) 86–95.
[26] H. Dai, B. Zhang, W. Wang, A multiwavelet support vector regression method for efficient reliability assessment, Reliab. Eng. Syst. Saf. 136 (2015)
132–139.
[27] F. Kang, Q. Xu, J. Li, Slope reliability analysis using surrogate models via new support vector machines with swarm intelligence, Appl. Math. Model 40
(2016) 6105–6120.
[28] H.-S. Li, Z.-Z. Lü, Z.-F. Yue, Support vector machine for structural reliability analysis, Appl. Math. Mech. 27 (2006) 1295–1303.
[29] C.M. Rocco, J.A. Moreno, Fast Monte Carlo reliability evaluation using support vector machine, Reliab. Eng Syst. Saf. 76 (2002) 237–243.
[30] X.-H. Tan, W.-H. Bi, X.-L. Hou, W. Wang, Reliability analysis using radial basis function networks and support vector machines, Comput. Geotech. 38
(2011) 178–186.
[31] A. Etemad-Shahidi, J. Mahjoobi, Comparison between M5 model tree and neural networks for prediction of significant wave height in Lake Superior,
Ocean Eng. 36 (2009) 1175–1181.
[32] B. Keshtegar, An enhanced HL-RF Method for the computation of structural failure probability based on relaxed approach, J. Civ. Eng. Infrast. 46 (2013)
69–80.
[33] B. Keshtegar, P. Hao, A hybrid loop approach using the sufficient descent condition for accurate, robust and efficient reliability-based design optimiza-
tion, J Mech Des. (2016) doi, 10.
[34] B. Keshtegar, M. Miri, Reliability analysis of corroded pipes using conjugate HL–RF algorithm based on average shear stress yield criterion, Eng. Fail.
Anal. 46 (2014) 104–117.
[35] B. Keshtegar, M. Miri, Introducing Conjugate gradient optimization for modified HL-RF method, Eng. Comput. 31 (2014) 775–790.
[36] B. Keshtegar, P. Hao, Z. Meng, A self-adaptive modified chaos control method for reliability-based design optimization, Struct. Multidiscip. Optim. 55
(2017) 63–75.
[37] D. Yang, Chaos control for numerical instability of first order reliability method, Commun. Nonlinear Sci. Numer. Simul. 15 (2010) 3131–3141.
[38] G. Li, Z. Meng, H. Hu, An adaptive hybrid approach for reliability-based design optimization, Struct. Multidiscip. Optim. 51 (2015) 1051–1065.
[39] B. Keshtegar, A hybrid conjugate finite-step length method for robust and efficient reliability analysis, Appl. Math. Model. 45 (2017) 226–237.
[40] P.-L. Liu, A. Der Kiureghian, Optimization algorithms for structural reliability, Struct. Saf. 9 (1991) 161–177.
[41] B. Keshtegar, J. Piri, O. Kisi, A nonlinear mathematical modeling of daily pan evaporation based on conjugate gradient method, Comput. Electron Agric.
127 (2016) 120–130.
[42] B. Bhattacharya, D.P. Solomatine, Neural networks and M5 model trees in modelling water level–discharge relationship, Neurocomputing 63 (2005)
381–396.
[43] A. Rahimikhoob, M. Asadi, M. Mashal, A comparison between conventional and M5 model tree methods for converting pan evaporation to reference
evapotranspiration for semi-arid region, Water Res. Manag. 27 (2013) 4815–4826.
[44] Y. Wang, I.H. Witten, Induction of model trees for predicting continuous classes, (1996).
[45] B. Keshtegar, M.O. Sadeq, Gaussian global-best harmony search algorithm for optimization problems, Soft Comput. (2016) 1–13, doi:10.1007/
s0 050 0- 016- 2274- z.