Computers and Geotechnics 87 (2017) 212–228

Contents lists available at ScienceDirect

Computers and Geotechnics

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

Research Paper

System reliability analysis of soil slopes with general slip surfaces using
multivariate adaptive regression splines
Subhadeep Metya a,b,⇑, Tanmoy Mukhopadhyay a, Sondipon Adhikari a, Gautam Bhattacharya b
a
College of Engineering, Swansea University, Swansea, United Kingdom
b
Civil Engineering Department, Indian Institute of Engineering Science and Technology (IIEST), Shibpur, India

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

Article history: A data driven multivariate adaptive regression splines (MARS) based algorithm for system reliability
Received 27 September 2016 analysis of earth slopes having random soil properties under the framework of limit equilibrium method
Received in revised form 18 January 2017 of slices is considered. The theoretical formulation is developed based on Spencer method (valid for gen-
Accepted 21 February 2017
eral slip surfaces) satisfying all conditions of static equilibrium coupled with a nonlinear programming
technique of optimization. Simulated noise is used to take account of inevitable modeling inaccuracies
and epistemic uncertainties. The proposed MARS based algorithm is capable of achieving high level of
Keywords:
computational efficiency in the system reliability analysis without significantly compromising the accu-
Slope stability
General slip surface
racy of results.
System reliability analysis Ó 2017 Elsevier Ltd. All rights reserved.
Multivariate adaptive regression splines
Monte Carlo simulation
Noise

1. Introduction evaluated directly by generating a large number of potential slip

surfaces and performing Monte Carlo simulation (MCS) based on
It is now widely recognized that the soil parameters are uncer- calculating the minimum factor of safety among them for each
tain, and thereby the conventional factor of safety based determin- set of sampled values of soil properties (realisation) [12,20]. In
istic slope stability analyses are increasingly being replaced by Method 2, on the other hand, a few representative slip surfaces
slope reliability analyses under a probabilistic framework. In are first identified from amongst the large number of potential slip
recent years, numerous slope reliability studies have been reported surfaces, and the system failure probability is then evaluated by
in the literature [1,2,3,4,6,5] assuming the single mode of failure, performing Monte Carlo simulation based on calculating the min-
i.e., failure by sliding along a potential slip surface. There is also imum factor of safety among these representative slip surfaces for
a growing appreciation that a slope can have many potential slip each realisation [21,16,22]. The results obtained from these two
surfaces constituting a series system; and the probability of failure methods are found to be practically the same. Further, because
of the slope is greater than that associated with any one of these the Monte-Carlo simulation requires high computational effort
slip surfaces [7,8,9,10,11,12,13,14,15,16,17]. It has been commonly and time, several response surface methods (RSMs) have also been
opined that the total or overall probability of failure of a slope used for system reliability evaluations, e.g., polynomial-based RSM
(under the framework of system reliability) is the ultimate goal [23], Kriging methodology [24,25], the artificial neural network
to achieve. (ANN) [26], the support vector machine (SVM) [27], the high
In the early studies [8,9,18], system failure probability of a slope dimensional model representation (HDMR) [28], and others. An
is reported in terms of two bounds [19] which are sometimes excellent review on the application of various RSMs for slope reli-
widely separated. Most of the other studies fall under the two cat- ability analysis is available in the literature [29].
egories of methods suggested by Zhang et al. [13], namely, Method Alternatively, in order to evaluate the system reliability of a soil
1 and Method 2. In Method 1, the system failure probability is slope, for each set of sampled values of soil properties in a Monte
Carlo analysis, one could find the critical slip surface and its corre-
sponding minimum factor of safety [30,31] instead of finding the
⇑ Corresponding author at: College of Engineering, Swansea University, Swansea
SA1 8EN, United Kingdom.
minimum factor of safety among a fixed set of slip surfaces gener-
E-mail addresses: [email protected], [email protected] ated before the simulation starts (as in Method 1 and Method 2).
(S. Metya). The overall probability of failure (or system failure probability),

http://dx.doi.org/10.1016/j.compgeo.2017.02.017
0266-352X/Ó 2017 Elsevier Ltd. All rights reserved.
 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228 213

pF, s, will then be the ratio of the number of times the minimum parts, namely, computation of factor of safety of a given or trial slip
factor of safety is less than 1.0 to the total number of simulation surface, and then search for the critical slip surface having the min-
runs. While the former approach is more logical of the two imum factor of safety FSmin (known as the deterministic critical slip
approaches, the latter approach (as in Method 1 and Method 2) surface) using an optimization technique. The Spencer method of
is computationally simpler and thus more commonly used in the slices valid for general slip surfaces [39] is regarded as one of the
literature. It is therefore necessary to investigate to what extent rigorous methods as it does not make any a priori assumption
the results obtained based on the two approaches differ, as well regarding the shape of the slip surface and satisfies all conditions
as which approach leads to a more conservative estimate of the of equilibrium [40]. The computation of factor of safety (FS) in
safety of a slope as a system. Further, whichever approach is used, Spencer method involves solution of a pair of nonlinear stability
it requires considerable computational effort and time. To enhance equations. An efficient method of solution first formulates the
the computational efficiency, the relationship between the mini- problem as an optimization problem and then solves it using a
mum factor of safety and the uncertain parameters is approxi- powerful optimization technique [41]. Therefore, the deterministic
mated by the multivariate adaptive regression splines (MARS) slope stability analysis in this study leads to a 2-tier analysis and
based surrogate model in this study. Although the application of the optimization problem in each tier of analysis has been solved
MARS to geotechnical engineering field is not very common, only using the sequential quadratic programming (SQP) [42] technique
found recently in slope stability analysis by Liu and Cheng [32], which has been rated as a powerful optimization technique [31]
its application in other fields of engineering is very much appreci- and can be easily implemented in the MATLAB platform with its
ated recently [33,34,35,36,37]. Moreover, previous studies have optimization toolbox.
focused mainly on the application of various RSMs for approximat- Under the framework of single mode of failure, similar to the
ing the relationship between the factor of safety and uncertain deterministic analysis, the probabilistic slope stability analysis
parameters. In this study, however, the MARS has been made use can be viewed as the problem of locating the slip surface corre-
of not only to approximate the relationship between the minimum sponding to the lowest value of reliability index bmin (or the highest
factor of safety and uncertain parameters, the location of the crit- value of the probability of failure), called the probabilistic critical
ical slip surfaces are also predicted by MARS. Thus the objective of slip surface of the slope. The first order reliability method, (FORM)
the present study is also to explore the potential of MARS as an which is widely accepted as the most versatile among the approx-
efficient mapping route in slope reliability analysis. imate methods of reliability analysis [43], has been adopted in this
Furthermore, in almost all the previous studies on system reli- study. The computational procedure for the determination of the
ability analysis of earth slopes, the shape of the slip surface is probabilistic critical slip surface based on FORM involves a 3-tier
assumed as circular and the Bishop’s simplified method (BSM) is optimization: (i) evaluation of performance function requiring
used as a slope stability model. As the BSM does not satisfy hori- the evaluation of Spencer’s factor of safety involves the first tier
zontal force equilibrium, it is commonly regarded as an approxi- of optimization; (ii) evaluation of the reliability index, b based on
mate method. The shape of actual slip surface is also, in general, FORM involves the second tier of optimization, and (iii) the search
not circular except in a homogeneous slope without discontinuities for the surface of minimum reliability index (bmin) involves the
of any kind [38]. Keeping the above in view, in this study, while no third tier of optimization. The SQP technique has been employed
assumptions have been made regarding the shape of the slip sur- in the MATLAB platform to solve the optimization problem
face, the slope stability evaluation is based on the Spencer method involved in each tier of analysis. More detailed description of the
valid for general slip surfaces [39] satisfying all conditions of stat- computational procedure for the determination of the probabilistic
ical equilibrium, which is definitely more rigorous than the critical slip surface based on FORM can be found elsewhere
Bishop’s simplified method, especially for non-homogeneous [4,44,45]. The computer program developed and used in the
slopes. referred studies was validated with reference to two benchmark
slope problems. The availability of such a computer program has
been made use of in the present study for the purpose of determi-
2. Adopted methodologies nation of the probabilistic critical slip surface.

2.1. Slope stability analysis – deterministic and probabilistic

2.2. MARS-based MCS for system reliability analysis of slopes
A typical slope, found in various civil engineering projects
including dams, embankments and open cut for highways, is as 2.2.1. Formulation of Multivariate Adaptive Regression Splines (MARS)
shown in Fig. 1. The stability of these slopes along potential failure An efficient mathematical relationship between input parame-
surfaces is of major interest. Slope stability analyses based on the ters and output feature of interest for a system under investigation
limit equilibrium approach have conventionally been performed based on few algorithmically chosen samples can be established
in a deterministic manner and the entire process consists of two with the help of MARS [33]. It is a nonparametric regression proce-

Fig. 1. A typical slope with potential slip surfaces.

214 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228

dure that makes no supposition about the underlying functional and simplicity and it is found suitable methodologies because it
connection between the dependent and independent variables. is more interpretable than most recursive partitioning, neural
MARS algorithm adaptively selects a set of ‘basis functions’ for and adaptive strategies wherein it distinguishes well between
approximating the response function through a forward and back- actual and noise variables. Moreover, the MARS models are
ward iterative approach. The MARS model can be expressed as reported [47] to work satisfactorily in terms of computational cost
X
n irrespective of dimension (low-medium-high) and noise.
Y¼ ak Hkf ðxi Þ ð1Þ
k¼1 2.2.2. Proposed procedure for system reliability analysis using MARS-
based MCS
with Hkf ðx1 ; x2 ; x3 ; . . . ; xn Þ ¼ 1; for k ¼ 1 As already mentioned, an earth slope with uncertain soil
parameters is often characterized by the presence of numerous
where ak and Hkf ðxi Þ are the coefficient of the expansion and the ‘ba-
potential slip surfaces and more importantly these are not known
sis functions’ respectively. Thus the first term in Eq. (1) becomes a1, beforehand and are determined as part of the analysis. Let Ns be the
which is basically an intercept parameter. The ‘basis function’ can number of potential slip surfaces identified in a slope, and Si
be represented as denote the ith one, i = 1, 2, . . ., Ns. In this study to evaluate the sys-
Y
ik
 q tem probability of failure, the slope has been viewed as a series
Hkf ðxi Þ ¼ zi;k ðxjði;kÞ  ti;k Þ tr ð2Þ system in which each potential slip surface is a component and
i¼1 the critical slip surface is the weakest one. A slope as a series sys-
where ik is the number of factors (interaction order) in the k-th tem fails when an event of failure occurs along any of these slip
basis function, Z i;k ¼ 1, xjði;kÞ is the j-th variable, 1  j(i,k)  n, surfaces. The minimum factor of safety (FSmin) determined under
and t i;k is a knot location on each of the corresponding variables. q an uncertain environment is used as a performance indicator of
is the order of splines. The approximation function Y is composed the slope. In Monte Carlo simulation, if the performance function
of ‘basis functions’ associated with k sub-regions. Each multivariate of the jth sample is given by g(Xj), then a safety judgement for each
trial can be provided as follows:
spline ‘basis function’ Hkf ðxi Þ is the product of univariate spline basis
(
functions zi;k , which is either order one or cubic, depending on the 1; if gðX j Þ ¼ FSmin  1 6 0
j
degree of interrelationship of the approximation. The notation ‘‘tr” dðX Þ ¼ ð6Þ
0; if gðX j Þ ¼ FSmin  1 > 0
means the function is a truncated power function.
 q  q   Then, an estimate of the system failure probability can be obtained
zi;k ðxjði;kÞ  ti;k Þ tr ¼ zi;k ðxjði;kÞ  t i;k Þ for ; zi;k ðxjði;kÞ  t i;k Þ < 0 ð3Þ
by
½zi;k ðxjði;kÞ  ti;k Þqtr ¼ 0; otherwise ð4Þ 1X N
pF;s ¼ dðX j Þ ð7Þ
Here each function is considered as piecewise linear with a trained N j¼1
knot ‘tr’ at each xi;k . Now allowing the basis function to bend at the
knots, MARS can model functions that differ in behaviour over the where N is the total number of simulation cycles. An important
domain of each variable. This is applied to interaction terms as well. point to be noted here is that in this proposed procedure the opti-
The interactions are no longer treated as global across the entire mum number of simulations required intuitively indicates the num-
range of predictors but between the sub-regions of every ‘basis ber of potential slip surfaces to be analyzed (i.e., Ns = N) for an
function’ generated. Depending on fitment, the maximum number accurate system reliability analysis. As the value of probability of
of knots to be considered, the minimum number of observations failure obtained from MCS is known to be sensitive to the number
between knots, and the highest order of interaction terms are calcu- of simulations, to assess the accuracy and the efficiency of the pro-
lated. The screening of automated variables occurs as a result of posed procedure, the optimum number of simulations is first iden-
using a modification of the generalized cross-validation (GCV) tified and the COV of the pF,s is then estimated [using Eq. (8)] to
model fit criterion, developed by Craven and Wahba [46]. MARS compare the results obtained from other methodolgies.
identifies the location and number of the needed spline ‘basis func- sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
tions’ in a forward or backward stepwise fashion. It starts by over- ð1  pF;s Þ
COV pF ¼ ð8Þ
fitting a spline function through each knot, and then by removing N  pF;s
the knots that least contribute to the overall fit of the model as
determined by the modified GCV criterion, often completely elimi- As already indicated, using the above procedure based on any deter-
nating the most insignificant variables. Eq. (5) depicts the lack-of-fit ministic slope stability model (LEM or FEM), the computational cost
(Lf ) criterion used by MARS. involved is very high. In this study, therefore, an efficient surrogate
Pn model is used to map the minimum factor of safety as well as the
~ ¼
1
i¼1 ½Y i  Y k~ ðxi Þ2 location of the corresponding critical slip surface as function of
Lf ðY k~ Þ ¼ Gcv ðkÞ n
h i ð5Þ
~ ~ 2 the uncertain soil parameters. The entire procedure is executed in
1  c nðkÞ
two stages. In stage 1, an explicit form of MARS based surrogate
~ ¼ cðkÞ
~ þ M:k.
~ model is constructed to map the implicit relationship between the
where ~cðkÞ
FSmin and the uncertain shear strength parameters. Then in stage
where ‘n’ denotes the number of sample observations, ~cðkÞ ~ is
2, the scheme of MCS explained above is conducted using MARS
~
the number of linearly independent ‘basis functions’, k is the num- based surrogate, in place of the actual deterministic slope stability
ber of knots selected in the forward process, and ‘M’ is a cost for model, by locating deterministic critical slip surface and the corre-
basis-function optimization as well as a smoothing parameter for sponding FSmin. In comparison to the direct MCS, the computational
the procedure. Larger values of ‘M’ result in fewer knots and time needed for this procedure is significantly reduced as only a few
smoother function estimates. The best MARS approximation is numbers of run of the original deterministic stability analysis are
the one having the highest GCV value. Thus MARS is also compared required to set up the MARS based surrogate.
with parametric and nonparametric approximation routines in A flowchart to summarize the proposed procedure for the sys-
terms of its accuracy, efficiency, robustness, model transparency, tem reliability analysis of a soil slope based on the MARS-based
 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228 215

Fig. 2. Flowchart of stochastic analysis using MARS model.

MCS is shown in Fig. 2. The detailed procedure has the following a certain number of slices beyond which this influence becomes
steps. small or insignificant for all practical purposes. In this context it
may be mentioned that the proposed methodology does not put
Step 1: Define input information: Define the necessary input any restriction on the number of slices. However, it is conceivable
parameters needed to describe the slope geometry and soil that greater the number of slices, greater will be the computing
layer boundaries, the parameters which are of deterministic time and, consequently, lower will be the efficiency of the
nature and the stochastic parameters with their statistical prop- method. Keeping this in view, in a given slope problem, based
erties (i.e., mean and COV). on a sensitivity study carried out by increasing the number of
Step 2: Establish the deterministic stability analysis model for slices, an optimum number of slices can be found out beyond
locating the deterministic critical slip surface and the associ- which there is no significant improvement (decrease) in the value
ated minimum factor of safety (as presented in the first para- of FSmin.
graph of Section 2.1). Here the commonly used curvature
constraint on kinematical admissibility (upward concavity of Step 4: Construct the MARS based surrogates to map the mini-
the shape of a slip surface) is imposed. mum factor of safety as well as to predict the location of the
Step 3: Generate training samples by Latin hypercube sampling critical slip surface.
(LHS) and calculate the FSmin values as well as the co-ordinates Step 5: Test the output of the MARS based response surface and
of the corresponding critical slip surface location for each data determine the optimum numbers of training samples based on
set using the LEM based code developed in Step 2. the value of the coefficient of determination (R2).
Step 6: To evaluate the system failure probability pF,s, run the
In slope stability computations, it is widely known that the trial MCS, predict the FSmin surface and the FSmin value using
accuracy of calculation of the values of FSmin (as well as the co- the MARS models developed (in steps 4 and 5) for each sample
ordinates of the discrete points on the corresponding critical slip run and determine the optimum number of simulation (N). As
surface of general shape) is influenced by the slice division in the discussed before, this gives the total number of failure modes
potential slip surface. This influence is generally substantial up to or potential slip surfaces to be analyzed, Ns = N.
216 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228

Apparently, in step 6, for prediction of the critical slip surface to the measuring process such as operator error or a faulty device,
locations (FSmin surfaces) there is no need for imposing the curva- test imperfections, soil specimen disturbance, limited size of spec-
ture constraints mentioned in step 2. However, the MARS model, imens, differences between in-situ and laboratory stress condi-
being a response surface equation (RSM), cannot be expected to tions, use of imperfect empirical correlations, error in modeling
make 100% accurate predictions. Thus, it is quite likely that a few and computer simulation. These are also termed as epistemic
predicted slip surfaces remain kinematically inadmissible. uncertainties by some authors [48,49]. Therefore, one should first
Therefore, it is required also to check whether all the predicted identify the presence of any random testing error or noise and
critical slip surface locations are acceptable or not by imposing that try to eliminate it. However, it is very difficult to fully eliminate
constraint. If, for a certain sample, the predicted surface is unac- these sources. Therefore, the performance of any modern uncer-
ceptable, then that sample should be discarded from the run and tainty propagation algorithm should be studied with the influence
the next sample is to be tried. of noisy datasets.
An analysis to explore the effect of simulated noise on MARS
Step 7: Calculate the system failure probability, pF, s and the based uncertainty quantification scheme for soil slopes is taken
COV associated with it (COVPF). A COV value of more than 10% up in this section. Gaussian white noise with a specific variance
is generally considered as unacceptable. level (p) is introduced in the set of output responses (the minimum
factor of safety) [51] which is subsequently used for MARS model
In this context, it is useful to mention the guidelines for deter- formation, as in Eq. (9).
mining the sample size, as given below.
F iM ¼ f i þ p  ni ð9Þ
(i) Guideline for determining the sample size in establishing
the MARS model: where f denotes the minimum factor of safety and the subscript i is
the sample number in the design point set. nij is a function that gen-
In establishing the MARS model, one commonly used guideline erates normally distributed random numbers. Subscript M is used
for determining the sample size is that the value of the coefficient here to indicate the minimum factor of safety including the effect
of determination (R2) should be greater than a pre-assigned quan- of noise.
tity (0.95). It may be mentioned that Liu and Cheng [32] have also Subsequently, the simulated noisy dataset (i.e. the sampling
used this guideline. However, a more direct guideline is to make a matrix for MARS model formation) is created by introducing
plot of R2 versus number of samples and pick that value of abscissa pseudo random noise in the responses, while the input design
as the sample size at which the fluctuation in the ordinate has points are kept unchanged. Then for each dataset, MARS based
become negligibly small. In this study the latter guideline has been MCS is carried out to quantify uncertainty of soil slopes, as
made use of. described in Fig. 3b. Effect of noise are found to be accounted for
in several other studies in the available literature [52,53] dealing
(ii) Guideline for determining the sample size in evaluating the with deterministic analysis. Assessment of any surrogate based
system probability of failure: uncertainty propagation algorithm under the effect of noise in
the field of geotechnical engineering is the first attempt of its kind
In evaluation of system probability of failure, one commonly to the best of the authors’ knowledge. It is necessary to take into
used guideline for determining the sample size is that the coeffi- account such simulated noise considering the presence of other
cient of variation (COV) of the probability of failure (pF, s) should sources of uncertainty such as error in measurement of responses,
be less than a pre-assigned quantity (0.15). It may be mentioned error in modeling and computer simulation and various other epis-
that Liu and Cheng [32] have also used this guideline. However, a temic uncertainties involved with the system. Thus the kind of
more direct guideline is to make a plot of pF,s versus number of analysis carried out here will provide a comprehensive idea about
samples and pick that value of abscissa as the sample size at which the robustness of the MARS based model under noisy data.
the fluctuation in the ordinate has become negligibly small. In this
study the latter guideline has been made use of.

3. Illustrative examples and results

2.3. Effect of simulated noise on MARS based uncertainty
quantification scheme To elucidate the methodology presented in the preceding sec-
tion, two example problems are selected from the literature. Exam-
Amongst the various sources of parameter uncertainty [Fig. 3a], ple 1 concerns a simple slope in a 2-layered soil with horizontal
random testing error or noise is one of the important sources in layer boundary and Example 2, a complex slope in a multilayer
many problems. Random testing errors arise from factors related (c, /) soil with arbitrary layer boundaries.

Parameter Uncertainty

Data Scatter Systematic Error

Real Spatial Variability Random Testing Error Statistical Error Measurement Bias

Fig. 3a. Sources of parameter uncertainty (after [50]).

 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228 217

Fig. 3b. Flowchart for analyzing the effect of noise on the MARS based uncertainty quantification algorithm.

3.1. Example 1 step 3 under Section 2.2.2, the number of slice division is selected
as 12.
A simple slope in a two-layer soil: An embankment underlain by Initially, taking the value of cohesion of the soft clay foundation
soft clay foundation [54] as 25 kPa (equal to its mean value), the deterministic critical slip
Fig. 4 shows an embankment underlain by a soft clay founda- surface of general shape has been located (Fig. 5) and the corre-
tion, taken from Ji et al. [54]. The undrained shear strength of the sponding minimum factor of safety (FSmin) is obtained as 1.311
soft clay c2 is considered as a normally distributed random variable using the Spencer method [39] coupled with the SQP method of
with a mean value of 25 kN/m2 and a coefficient of variation of optimization. A FSmin value of 1.462 was reported by Ji et al. [54]
0.25. All the other strength parameters of the two layers are using the Spencer method with a circular slip surface. It is there-
assumed to be deterministic with values as given in Fig. 4. It may fore observed that the value of the minimum factor of safety (FSmin)
be pointed out that previously Ji et al. [54] analyzed this problem as obtained in the present analysis is lower (nearly 10%) than that
assuming slip surfaces to be of circular shape. In the present study, reported by Ji et al. [54].
as mentioned before, there is no a priori assumption regarding the
geometry of the slip surface. In other words, slip surfaces of general
shape have been considered. Further, based on the discussion in

Fig. 5. Slope section and the different critical slip surface locations (namely, the
deterministic critical slip surface, the probabilistic critical slip surface and the
Fig. 4. Slope section in example 1 [54]. 30,000 potential slip surfaces predicted by MARS) in example 1.
218 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228

Next, considering the single mode of failure, the probabilistic

critical slip surface has been determined based on the FORM and
the associated minimum reliability index is obtained as
bmin = 1.006 [the corresponding value of pF = U (bmin) is obtained
as 1.57  101]. It may be noted that the value of bmin obtained in
the present study is significantly lower (nearly 24%) than that
reported by Ji et al. [54] (i.e., 1.32) using circular slip surface.
Fig. 5 presents the location of the probabilistic critical slip surface
alongside the deterministic critical slip surface, which shows that
these two critical slip surfaces are significantly different.
In order to study the system reliability analysis of this slope
example, at first, some training samples are generated by Latin
hypercube sampling (LHS) and the MARS based surrogate model
is then constructed using these training samples to approximate
the minimum factor of safety functional. To test the MARS model,
the obtained FSmin values are compared for another 100 samples
and the value of the coefficients of determination (R2) is deter-
mined. The optimum number of training samples is determined
as 128 by varying the sample size from 16 to 1024 (powers of 2)
(Fig. 6) and the corresponding value of R2 equals to 0.994. Fig. 7
presents the comparison between the values of FSmin predicted
Fig. 7. Comparison between the values of FSmin predicted by different MARS models
by different MARS models and those calculated by the Spencer
and those calculated by the Spencer method coupled with the SQP method in
method coupled with the SQP method, which indicates very good example 1.
fitting and predictive capability of the MARS with 128 samples.
The probability density function plots of the minimum factor of
safety as presented in Fig. 8 also shows a negligible deviation
between MARS model and original LEM model based on Spencer
method indicating validity and high level of precision for the pre-
sent surrogate based approach further.
Then, following the same procedure, MARS models are also con-
structed to predict the locations of the critical slip surface with
numbers of slices equals to 12. For 10 arbitrary set of samples, pre-
dicted locations of critical slip surfaces are compared with the cal-
culated locations using the Spencer method coupled with the SQP
method in Fig. 9, which again indicates very good predictive capa-
bility of the MARS model.
The COV of the minimum factor of safety due the randomness
added to slope system due to the presence of c2 as random is
obtained as 22.1% using the MARS based surrogate. Another study
based on the MARS based surrogate model to show the variation of
the minimum factor of safety with variation of the random
strength parameters in a normalised scale is depicted in Fig. 10,
which shows that the nature of this variation is best fitted by a
Fig. 8. Probability density function for minimum factor of safety based on the LEM
polynomial of order 3. (Spencer method) as well as those based on the constructed MARS model in
example 1.

As the value of probability of failure obtained from MCS is

known to be sensitive to the number of simulations, trial simula-
tions have been conducted using the constructed MARS-based
RSM by varying the sample size from 2500 to 50,000. Fig. 11 shows
the sensitivity of the probability of failure to the sample size when
the seed number is taken as 28061987 (an arbitrary value). It is
seen that beyond a sample size of 30,000 the probability of failure
pF is not sensitive to the sample size, and hence, the same has been
adopted as the optimum number of simulations. The correspond-
ing value of system failure probability is obtained as 1.73  101.
In Fig. 5, the 30,000 potential slip surfaces predicted by MARS
are plotted corresponding to the system failure probability deter-
mined above. Figs. 12a and 12b present the histograms with the
corresponding normal fit for the abscissas and ordinates of vertices
1, 4, 7, 10 and 13 respectively.
Table 1 presents a comparison of the failure probabilities
obtained in the present analysis and that reported by Ji et al.
Fig. 6. Relationship between sample size and coefficient of determination (R2) [54]. It is observed that the value of the system failure probability
(example 1). (pF,s) obtained in the present analysis is higher than those reported
 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228 219

Fig. 9. Comparison between the locations of the critical slip surfaces determined by the Spencer method coupled with the SQP method and those predicted by MARS for 10
arbitrary set of samples in example 1.

in Ji et al. [54] using circular slip surfaces. It is also noted that sys- 3.2. Effect of noise
tem failure probability value is higher than the value of probability
of failure associated with the probabilistic critical slip surface in Representative results presented in Fig. 13 show the effect of
the present study (assuming single mode of failure), which is as noise on the minimum factor of safety. Gaussian white noise with
per expectation. It may be noted that the obtained system failure a specific value of variance (p) in the range of 0.01–0.5 is intro-
probability value of 17.3% indicates the performance level of the duced in the set of minimum factor of safety, which is used for
slope as ‘hazardous’ according to the [55]. MARS model formation. The results furnished in this article are
220 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228

obtained by using 500 such noisy datasets, which comprises the

formation of MARS model and thereby carrying out MCS for each
dataset using corresponding surrogate models [refer Fig. 3b]. The
results presented for different values of variance are compared
with the probability density function of noise-free case to provide
a comprehensive idea about the performance of MARS in uncer-
tainty quantification under the influence of simulated noise for
minimum factor of safety of earth slopes.

3.3. Example 2

A complex slope in a multilayer (c,/) soil with arbitrary layer

boundaries (ACADS Study [56]).
This example is taken from the ACADS study [56] and is also
considered by Ji and Low [14], Zhang et al. [16], Kang et al. [57]
and Liu and Cheng [32] in their analysis. Fig. 14 shows the geom-
etry of the multilayered slope that consists of three layers of differ-
ent materials, and the layer boundaries are not entirely horizontal.
Table 2 presents soil parameters for this slope. Strength parame-
Fig. 10. Variation of the FSmin with c2 in example 1. ters of layer 2 and layer 3 are considered as random variables
and all these variables are assumed to be normally distributed. In
the present study, as mentioned in example 1, slip surfaces of gen-
eral shape have been considered, and, further, based on the discus-
sion in step 3 under Section 2.2.2, the number of slice division is
selected as 12.
Initially, like the example 1, assuming the soil properties to be
deterministic with values equal to their mean values (Table 2),
the deterministic critical slip surface of general shape has been
located and is as shown in Fig. 15. The associated minimum factor
of safety based on the Spencer method [39] is obtained as
FSmin = 1.35. The recommended solution by the experts in the
ACADS study is 1.39. Table 3 presents a comparison of the values
of FSmin obtained in the present analysis and that reported by pre-
vious investigators. It is observed that the value of the minimum
factor of safety (FSmin) obtained in the present analysis is lower
(nearly 4%) than those reported in the previous studies.
Like example 1, assuming the single mode of failure, the prob-
abilistic critical slip surface has been determined based on the
FORM and the associated minimum reliability index is obtained
as bmin = 2.24 [the corresponding value of pF = U (bmin) is obtained
as 1.25  102]. Fig. 15 presents the location of the probabilistic
Fig. 11. Relationship between sample size and probability of failure (example 1). critical slip surface alongside the deterministic critical slip surface,

Fig. 12a. Histogram density plot of the abscissas of vertices in example 1.

 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228 221

Fig. 12b. Histogram density plot of the ordinates of vertices in example 1.

Table 1
Summary of results of probabilistic analyses for example 1.

Studies Probabilistic method Probability of failure, pF (%)

Ji et al. [54] Failure probability of the most critical slip surface based on the FORM (circular slip surface) 9.34% (bmin = 1.32)
Present study Failure probability of the most critical slip surface based on the FORM (slip surface of general shape) 15.7% (bmin = 1.006)
System failure probability based on MCS using the MARS based RSM (30,000 samples) 17.3% (COV = 1.71%)

Fig. 13. Effect of noise on MARS based uncertainty quantification for minimum factor of safety of example 1.
222 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228

Fig. 14. Slope section in example 2.

Table 2
Statistical properties of soil parameters for example 2.

Layers Cohesion, c (kN/m2) Friction angle, / (°) Unit weight, c (kN/m3)

Mean COV (%) Mean COV (%)
1 0.0 – 38.0 – 19.5
2 5.3 30 23.0 20 19.5
3 7.2 30 20.0 20 19.5

Fig. 15. Slope section and the different critical slip surface locations (namely, the deterministic critical slip surface, the probabilistic critical slip surface and the 40,000
potential slip surfaces predicted by MARS) in example 2.

which shows that these two critical slip surfaces are markedly dif- ers of 2) (Fig. 16) and the corresponding value of R2 equals to
ferent in shape and location. 0.99. Fig. 17 presents a comparison between the values of FSmin
As has been done for the example 1, for the system reliability predicted by different MARS models and those calculated by the
analysis, at first, some training samples are generated by Latin Spencer method coupled with the SQP method of optimization,
hypercube sampling (LHS) and the MARS based surrogate model which indicates very good fitting and predictive capability of
is then constructed using these training samples to approximate the MARS with 512 samples. Like example 1, the probability den-
the minimum factor of safety functional. To verify the MARS sity function plots of the minimum factor of safety as presented
model, the obtained FSmin values are compared for another 100 in Fig. 18 also shows a negligible deviation between MARS model
samples and the value of the coefficient of determination (R2) is and original LEM model based on Spencer method indicating
determined. The optimum number of training samples is deter- validity and high level of precision for the present surrogate
mined as 512 by varying the sample size from 16 to 1024 (pow- based approach further.
 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228 223

Table 3
Summary of results of deterministic analyses for example 2.

Studies Slip surface shape Methodology used Minimum Factor

of safety FSmin
Limit equilibrium Method of solving the factor of Optimization technique to
method safety equations for a specific slip surface search for the critical slip surface
Ji and Low [14] Circular Spencer (1967) Not available Not available 1.406
Zhang et al. [16] Circular Bishop (1955) Not required Not available –
Kang et al. [57] Circular Bishop (1955) Not required Not available 1.405
Liu and Cheng [32] Circular Bishop (1955) Not required Not available 1.405
Present study General Spencer [39] Sequential quadratic programming (SQP) Sequential quadratic 1.350
programming (SQP)

Fig. 18. Probability density function for minimum factor of safety based on the LEM
based on Spencer method as well as the constructed MARS model in example 2.

Fig. 16. Relationship between sample size and coefficient of determination (R2)
(example 2).
with the SQP method in Fig. 19, which again indicates very good
predictive capability of MARS model.
A MARS based sensitivity analysis by comparing the COV of the
minimum factor of safety due the randomness added to slope sys-
tem due to presence of specific random soil parameter is shown in
Fig. 20(a). It is observed that /3 has the most significant influence
to the system failure probability, followed by c3. Another sensitiv-
ity study based on the MARS based surrogate model by comparing
the variation of the minimum factor of safety with variation of the
random strength parameters in a normalised scale is depicted in
Fig. 20(b), which corroborates the observation made from Fig. 20
(a).
Like example 1, as the value of probability of failure obtained
from the MCS is known to be sensitive to the number of simula-
tions, trial simulations have been conducted using the constructed
MARS-based RSM by varying the sample size from 2500 to 50,000.
Fig. 21 shows the sensitivity of the probability of failure to the
sample size when the seed number is taken as 28,061,987 (an arbi-
trary value). It is seen that beyond a sample size of 40,000 the
probability of failure pF is not sensitive to the sample size, and
hence, the same has been adopted as the optimum number of sim-
ulations. The corresponding value of system failure probability is
obtained as 1.80  102. In Fig. 15, the 40,000 potential slip sur-
Fig. 17. Comparison between the values of FSmin predicted by different MARS faces predicted MARS are plotted corresponding to the system
models and those calculated by the Spencer method coupled with the SQP method probability of failure determined above. Figs. 22a and 22b present
in example 2. the histogram with the corresponding normal fit for the abscissas
and ordinates of vertices 1, 4, 7, 10 and 13 respectively.
Table 4 presents a comparison of the failure probabilities
Following the same procedure as in example 1, MARS models obtained in the present analysis and that reported by previous
are also constructed to predict the locations of the critical slip sur- investigators using different methodologies. It is observed that
face with numbers of slices equals to 12. For 10 arbitrary set of the value of the system failure probability (pF,s) obtained in the
samples, predicted locations of critical slip surfaces are compared present analysis is higher than those reported in the previous stud-
with the calculated locations using the Spencer method coupled ies using circular slip surfaces. It may therefore be concluded that
224 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228

Fig. 19. Comparison between the locations of the critical slip surfaces determined by the Spencer method coupled with the SQP method and those predicted by MARS for 10
arbitrary set of samples in example 2.

system reliability analysis assuming the shape of the slip surfaces 3.5. Effect of noise
as circular underestimates the system failure probability. It is also
noted that system failure probability value is higher than the value Like example 1, Fig. 23 shows some representative results
of probability of failure associated with the probabilistic critical describing the effect of noise on the minimum factor of safety. Sim-
slip surface (assuming single mode of failure), which is as per ilarly, Gaussian white noise with a specific value of variance (p) in
expectation. the range of 0.01–0.5 is introduced in the set of minimum factor of
 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228 225

Fig. 20. Sensitivity studies based on the MARS model in example 2.

safety, which is used for MARS model formation and the results
provided in this article are obtained by using 500 such noisy data-
sets. The results presented for different values of variance are again
compared with the probability density function of noise-free case
to provide a comprehensive idea about the performance of MARS
in uncertainty quantification under the influence of simulated
noise for minimum factor of safety of earth slopes.

4. Conclusions

Development of a surrogate based algorithm for system reliabil-

ity analysis of earth slopes considering general slip surfaces in a
Monte Carlo simulation framework is considered. The geomechan-
ical parameters of the slope system have been treated as random
variables. A multivariate adaptive regression splines (MARS) based
approach is proposed under the framework of limit equilibrium
method of slices to map the variation of minimum factor of safety
caused due to uncertain input parameters. Similar procedure has
also been used to predict the locations of critical slip surfaces.
Fig. 21. Relationship between sample size and probability of failure (example 2). Spencer method valid for general slip surfaces [39] satisfying all

Fig. 22a. Histogram density plot of the abscissa of vertices in example 2.

226 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228

Fig. 22b. Histogram density plot of the ordinates of vertices in example 2.

Table 4
Summary of results of probabilistic analyses for example 2.

Studies Probabilistic method Probability of failure, pF (%)

Ji and Low [14] System reliability bounds based on the FORM 1.08–1.30%
System reliability bounds based on the SORM 1.35–1.53%
System failure probability based on MCS using the stratified RSM (50,000 samples) 1.34%
Zhang et al. [16] Failure probability of the most critical slip surface based on MCS (50,000 samples) 1.01% (COV = 4.4%)
System failure probability based on MCS (50,000 samples) 1.33% (COV = 3.8%)
System failure probability calculated based on the to the representative slip surfaces with MCS (50,000 samples) 1.08% (COV = 4.3%)
Kang et al. [57] System failure probability based on MCS using the GPR based RSM (10,00,000 samples) 1.59% (COV = 0.8%)
Liu and Cheng [32] System failure probability based on MCS using the MARS based RSM (1,00,000 samples) 1.28% (COV = 2.78%)
System failure probability based on MCS using the quadratic response surface method (QRSM) (1,00,000 samples) 1.55% (COV = 2.52%)
Present study Failure probability of the most critical slip surface based on the FORM 1.25%
System failure probability based on MCS using the MARS based RSM (40,000 samples) 1.80% (COV = 3.69%)

conditions of static equilibrium is used as a slope stability model in mum factor of safety functional and the location of the critical slip
combination with the sequential quadratic programming (SQP) as surface with the variation of the uncertain soil strength parame-
a nonlinear programming technique of optimization. In order to ters. No assumptions are made on the shape of the slip surface
compare the results of the reliability analysis, considering the sin- (i.e. general slip surface is considered for the analyses), which is
gle mode of failure, the FORM based probabilistic critical slip sur- the first ever attempt in conjunction with any surrogate model to
face has also been located and the associated minimum the best of authors’ knowledge. A novel paradigm is proposed to
reliability index (and the corresponding maximum probability fail- account for the effect of simulated noise that can be implemented
ure) is determined. The effectiveness of the proposed procedure for to other surrogate based analyses in the field of slope stability.
system reliability analysis has been demonstrated by two bench- Some of the key observations include:
mark example problems comprising an embankment underlain
by a soft clay foundation and a complex slope in a multilayer (c, 1. The numerical results show that the system reliability analysis
/) soil with arbitrary layer boundaries. Subsequently, the perfor- with circular slip surfaces can significantly underestimate the
mance of MARS based uncertainty propagation algorithm under system failure probability of the soil slope.
the effect of simulated noise is also investigated. For each of the 2. The value of the system failure probability is higher than the
two examples studied in this paper, once the MARS model is con- value of the probability of failure associated with the proba-
structed and verified, the computational time needed for calculat- bilistic critical slip surface (assuming a single mode of failure),
ing the system failure probability based on the Monte Carlo which is as per expectation.
simulation coupled with the MARS based surrogate is found to 3. The accuracy of the proposed approach is critically dependent
be negligible. Therefore, total computational expense is reduced on the number of training samples and the number of simula-
by (128/30,000  1/234) times for the Example 1 and tions. In-depth analyses are presented for determination of
(512/40,000  1/78) times for the Example 2 with respect to the the optimum number of training samples as well as the opti-
direct Monte Carlo simulation without compromising the accuracy mum number of simulations with a reasonable accuracy.
of results. Thus, the proposed approach is an efficient tool for sys-
tem reliability analysis. Future works could include the proposed MARS based system
Novelty of the present study includes the application of data reliability analysis algorithm to be implemented within probabilis-
driven MARS model to approximate the relationship of the mini- tic finite element analysis of slopes.
 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228 227

Fig. 23. Effect of noise on MARS based uncertainty quantification for minimum factor of safety of example 2.

Acknowledgment [10] Ching J, Phoon KK, Hu YG. Efficient evaluation of reliability for slopes with
circular slip surfaces using importance sampling. J Geotech Geoenviron Eng
2009;135(6):768–77.
This research has been supported by the Newton-Bhabha PhD [11] Huang JS, Griffiths DV, Fenton GA. System reliability of slopes by RFEM. Soils
Placement Grant 2015-16 jointly funded by the British Council Found 2010;50(3):343–53.
[12] Wang Y, Cao Z, Au SK. Practical reliability analysis of slope stability by
(United Kingdom) and the Department of Science and Technology
advanced Monte Carlo simulations in a spreadsheet. Can Geotech J 2011;48
(DST), Govt. of India for a collaborative research project at the Zien- (1):162–72.
kiewicz Centre for Computational Engineering (ZCCE), Swansea [13] Zhang J, Zhang LM, Tang WH. New methods for system reliability analysis of
soil slopes. Can Geotech J 2011;48(7):1138–48.
University, UK. This support is gratefully acknowledged. TM
[14] Ji J, Low BK. Stratified response surface for system probabilistic evaluation of
acknowledges the financial support from the Swansea University slopes. J Geotech Geoenviron Eng 2012;138(11):1398–406.
through the award of Zienkiewicz Scholarship during the period [15] Cho SE. First-order reliability analysis of slope considering multiple failure
of this work. SA acknowledges the financial support from the Royal modes. Eng Geol 2013;154:98–105.
[16] Zhang J, Huang HW, Juang CH, Li DQ. Extension of Hassan and Wolff method
Society of London through the Wolfson Research Merit award. for system reliability analysis of soil slopes. Eng Geol 2013;160:81–8.
[17] Li L, Wang Y, Cao Z. Probabilistic slope stability analysis by risk aggregation.
Eng Geol 2014;176:57–65.
[18] Low BK, Zhang J, Tang WH. Efficient system reliability analysis illustrated for a
retaining wall and a soil slope. Comput Geotech 2011;38(2):196–204.
References [19] Ditlevsen O. Narrow reliability bounds for structural systems. J Struct Mech
1979;7(4):453–72.
[1] Li KS, Lumb P. Probabilistic design of slopes. Can Geotech J 1987;24:520–35. [20] Jiang SH, Huang JS. Efficient slope reliability analysis at low-probability levels
[2] Hassan AH, Wolff TF. Search algorithm for minimum reliability index of earth in spatially variable soils. Comput Geotech 2016;75:18–27.
slopes. J Geotech Geoenviron Eng, ASCE 1999;125(4):301–8. [21] Li L, Wang Y, Cao Z, Chu X. Risk de-aggregation and system reliability analysis
[3] Bhattacharya G, Jana D, Ojha S, Chakraborty S. Direct search for minimum of slope stability using representative slip surfaces. Comput Geotech
reliability index of earth slopes. Comput Geotech 2003;30(6):455–62. 2013;53:95–105.
[4] Metya S, Bhattacharya G. Probabilistic critical slip surface for earth slopes [22] Jiang SH, Li DQ, Cao ZJ, Zhou CB, Phoon KK. Efficient system reliability analysis
based on the first order reliability method. Indian Geotech J 2014;44 of slope stability in spatially variable soils using Monte Carlo simulation. J
(3):329–40. Geotech Geoenviron Eng (ASCE) 2015;141(2):04014096.
[5] Metya S, Bhattacharya G, Chowdhury R. Reliability analysis of slopes in strain- [23] Li L, Chu XS. Multiple response surfaces for slope reliability analysis. Int J
softening soils considering critical slip surfaces. Innovative Infrastruct Solut Numer Anal Methods Geomech 2015;39(2):175–92.
2016;1:35. http://dx.doi.org/10.1007/s41062-016-0033-8. [24] Zhang J, Huang HW, Phoon KK. Application of the Kriging-based response
[6] Li DQ, Tang XS, Phoon KK. Bootstrap method for characterizing the effect of surface method to the system reliability of soil slopes. J Geotech Geoenviron
uncertainty in shear strength parameters on slope reliability. Reliab Eng Syst Eng (ASCE) 2013;139(4):651–5.
Saf 2015;140:99–106. [25] Yi P, Wei K, Kong X, Zhu Z. Cumulative PSO-Kriging model for slope reliability
[7] Cornell CA. Bounds on the reliability of structural systems. J Struct Div 1967;93 analysis. Probab Eng Mech 2015;39:39–45.
(1):171–200. [26] Cho SE. Probabilistic stability analyses of slopes using the ANN-based response
[8] Oka Y, Wu TH. System reliability of slope stability. J Geotech Eng 1990;116 surface. Comput Geotech 2009;36(5):787–97.
(8):1185–9. [27] Kang F, Li J. Artificial bee colony algorithm optimized support vector
[9] Chowdhury RN, Xu DW. Geotechnical system reliability of slopes. Rel Eng Syst regression for system reliability analysis of slopes. J Comput Civ Eng
Saf 1995;47(3):141–51. 2015:04015040.
228 S. Metya et al. / Computers and Geotechnics 87 (2017) 212–228

[28] Chowdhury R, Rao BN. Probabilistic stability assessment of slopes using high [44] Metya S, Bhattacharya G. Probabilistic stability analysis of the bois Brule levee
dimensional model representation. Comput Geotech 2010;37(7–8):876–84. considering the effect of spatial variability of soil properties based on a new
[29] Li DQ, Zheng D, Cao ZJ, Tang XS, Phoon KK. Response surface methods for slope discretization model. Indian Geotech J 2016;46(2):152–63.
reliability analysis: review and comparison. Eng Geol 2016;203:3–14. [45] Metya S, Bhattacharya G. Reliability analysis of earth slopes considering spatial
[30] Hassan AH, Wolff TF. Closure to ‘Search Algorithm for Minimum Reliability variability. Geotech Geol Eng – An Int J 2016;34(1):103–23.
Index of Earth Slopes’ by Hassan AH and Wolff TF. J Geotechn Geoenviron Eng, [46] Craven P, Wahba G. Smoothing noisy data with spline functions. Numer Math
ASCE 2001;127(02):198–200. 1979;31:377–403.
[31] Hong H, Roh G. Reliability evaluation of earth slopes. J Geotech Geoenviron [47] Crino S, Brown DE. Global optimization with multivariate adaptive regression
Eng 2008;134(12):1700–5. splines. IEEE Trans Syst Man Cybern Part B: Cybern 2007;37(2).
[32] Liu LL, Cheng YM. Efficient system reliability analysis of soil slopes using [48] Bourdeau PL, Amundaray JI. Non-parametric simulation of geotechnical
multivariate adaptive regression splines-based Monte Carlo simulation. variability. Géotechnique 2005;55(2):95–108.
Comput Geotech 2016;79:41–54. [49] Ang AHS, Tang WH. Probability concepts in engineering: emphasis on
[33] Friedman JH. Multivariate adaptive regression splines. Ann Stat 1991;19 applications to civil and environmental engineering. 2nd ed. New York: John
(1):1–67. Wiley & Sons Inc; 2007.
[34] Sudjianto A, Juneja L, Agrawal A, Vora M. Computer aided reliability and [50] Christian JT, Ladd CC, Baecher GB. Reliability applied to slope stability analysis.
robustness assessment. Int J Reliab, Qual Saf 1998;5:181–93. J Geotech Geoenviron Eng, ASCE 1994;120(12):2180–207.
[35] Samui P. Multivariate adaptive regression spline (MARS) for prediction of [51] Mukhopadhyay T, Naskar S, Dey S, Adhikari S. On quantifying the effect of
elastic modulus of jointed rock mass. Geotech Geol Eng 2013;31(1):249–53. noise in surrogate based stochastic free vibration analysis of laminated
[36] Dey S, Mukhopadhyay T, Naskar S, Dey TK, Chalak HD, Adhikari S. Probabilistic composite shallow shells. Compos Struct 2016;140:798–805.
characterization for dynamics and stability of laminated soft core sandwich [52] Nejad FB, Rahai A, Esfandiari A. A structural damage detection method using
plates. J Sandwich Struct Mater. SAGE Publication; 2016. In press. static noisy data. Eng Struct 2005;27:1784–93.
[37] Mukhopadhyay T. A multivariate adaptive regression splines based damage [53] Mukhopadhyay T, Chowdhury R, Chakrabarti A. Structural damage
identification methodology for web core composite bridges including the identification: a random sampling-high dimensional model representation
effect of noise. J Sandwich Struct Mater. SAGE Publication; 2016. In Press. approach. Adv Struct Eng 2016;19(6):908–27.
[38] Chowdhury R, Flentje P, Bhattacharya G. Geotechnical slope [54] Ji J, Liao HJ, Low BK. Modeling 2-D spatial variation in slope reliability analysis
analysis. Balkema: CRC Press; 2010. p. 737. using interpolated autocorrelations. Comput Geotech 2012;40:135–46.
[39] Spencer E. The thrust line criterion in embankment stability analysis. [55] US Army Corps of Engineers. Introduction to probability and reliability
Geotechnique 1973;23(1):85–100. methods for use in geotechnical engineering, in ETL 1110-2-547; 1995.
[40] Duncan JM, Wright SG. The accuracy of equilibrium methods of slope stability [56] Donald IB, Giam PSK. Soil slope stability programs review, ACADS –
analysis. Eng Geol 1980;16(1–2):5–17. Association for Computer Aided Design, Melbourne, Australia, Report U255;
[41] Bhattacharya G, Basudhar PK. A new procedure for finding critical slip surfaces 1989.
in slope stability analysis. Indian Geotech J 2001;31(1):149–72. [57] Kang F, Han S, Salgado R, Li J. System probabilistic stability analysis of soil
[42] Rao SS. Engineering optimization: theory and practice. 4th ed. New slopes using Gaussian process regression with Latin hypercube sampling.
York: Wiley; 2009. Comput Geotech 2015;63:13–25.
[43] Haldar A, Mahadevan S. Probability reliability and statistical methods in
engineering design. John Wiley and Sons; 2000.