You are on page 1of 22

Nat Hazards

DOI 10.1007/s11069-016-2454-2

ORIGINAL PAPER

Slope stability analysis using artificial intelligence


techniques

Shakti Suman1 • S. Z. Khan2 • S. K. Das1 • S. K. Chand3

Received: 21 February 2015 / Accepted: 2 July 2016


 Springer Science+Business Media Dordrecht 2016

Abstract Natural and man-made soil slope failures are complex phenomena and cause
serious hazards in many countries of the world. As a result, public and private property is
damaged worth millions of dollars. It is necessary to understand the processes causing
failure of slopes and prediction of its vulnerability for proper mitigation of slope failure
hazards. Various attempts have been made to predict the stability of slope using both
conventional methods such as limit equilibrium method, finite element method, finite
difference method and statistical methods. Artificial intelligence (AI) methods like artifi-
cial neural networks, genetic programming and genetic algorithms, and support vector
machines are found to have better efficiency compared to statistical methods. The present
study is an attempt to use recently developed AI methods such as functional networks
(FNs), multivariate adaptive regression splines (MARS), and multigene genetic pro-
gramming (MGGP) to predict factor of safety of slope using slope stability data available
in the literature. Prediction model equations are also provided, which can be used to predict
the factor of safety of a slope. The performances of these AI techniques have been eval-
uated in terms of different statistical parameters such as average absolute error, maximum
absolute error, root mean square error, correlation coefficient, and Nash–Sutcliff

& S. K. Das
saratdas@rediffmail.com
Shakti Suman
shaktisuman7@gmail.com
S. Z. Khan
szk1161@yahoo.com
S. K. Chand
skchand2001@gmail.com
1
Civil Engineering Department, National Institute of Technology Rourkela, Rourkela,
Odisha 769008, India
2
Civil Engineering Department, BIET, BPUT, Bhadrakh, Odisha, India
3
Civil Engineering Department, IGIT, Sarang, Odisha 759146, India

123
Nat Hazards

coefficient of efficiency. Of the available models, MARS model had better prediction
performance in comparison with MGGP and FN models in terms of the above statistical
criteria.

Keywords Slope stability  Factor of safety  Artificial intelligence  Multigene genetic


programming  Multivariate adaptive regression splines  Functional network

1 Introduction

Slope stability is an important concept in geotechnical engineering due to its concern for
stability of natural slopes and analysis and design of man-made slopes such as dams and
embankments of highways, railways, open pits, earth dams, retaining walls. The analysis of
stability of slopes involves calculation of factor of safety (FOS) and has closely followed
the developments in computational geotechnics. Various parameters like geometry of
slopes, physical data of the geological materials and their shear strength factors (cohesion
and angle of internal friction), information on pore-water pressures have influential role in
the evaluation of stability of slopes. The stability of slopes is also characterized by many
sources of uncertainty like the soil properties, loading, water pressure. Slope stability is
evaluated based on FOS for various geotechnical models.
Limit equilibrium method (LEM) using method of slices has been one of the earliest
approaches to determine the stability of slope (Fellenius1936; Bishop 1955; Bishop and
Morgenstern 1960; Morgenstern and Price 1965). But, the above methods are based on
many assumptions and complexity of problem increases with a reduction in the number of
assumptions. Though methods like finite element method (Kanungo et al. 2013) and finite
difference method (Singh et al. 2008) are used, it is difficult to consider the complex
nonlinear relationship of all the factors affecting stability of a slope. Hence, still LEM is
the most preferred method by the professional engineers for stability analysis of slopes.
Statistical method was applied by Sah et al. (1994) to predict the factor of safety of slopes
on 46 case histories (29 failed; 17 stable) of circular-type failures and 14 case histories (8
failed; 6 stable) of wedge-type failures. Regression equations based on maximum likeli-
hood method was proposed by them for the above two types of failures. The statistical
models had correlation coefficient (R) of 0.911 and 0.954 for circular and wedge failures,
respectively, but it was not validated with a new database.
In place of statistical methods, artificial intelligence (AI) can be used as an alternate tool
for developing models. Many researchers have found that AI techniques have better pre-
diction capability than statistical methods. Lu and Rosenbaum (2003) used artificial neural
network (ANN) to analyze slope stability based on the dataset of Sah et al. (1994) and Xu
and Xie (1999). They found ANN to have a better accuracy than MLE in prediction of both
the value of FOS and the stability of the slope. Yang et al. (2004) applied genetic pro-
gramming (GP) for developing a model equation based on the same dataset and found it to
be better than the regression model equation as given by Sah et al. (1994). However, the
model equation is found to have poor generalization with less accuracy for testing data.
Sakellariou and Ferentinou (2005) used artificial neural network to find the factor of safety
of the slope and equated it with the statistical method of Sah et al. (1994), in terms of mean
square error (MSE) and found it to be better. Support vector machines (SVMs) were used
by Samui (2008) for the above database and found a higher value of correlation coefficient

123
Nat Hazards

(R) between the observed and predicted FOS in comparison with model proposed by Sah
et al. (1994). Das et al. (2011) applied ANN with different training algorithm to determine
FOS of slopes and found that differential evolution neural network (DENN) has better
prediction accuracy in comparison with Levenberg–Marquardt (LMNN) and Bayesian
regularization neural network (BRNN). Samui and Kumar (2006) used artificial neural
network as a substitute method to upper limit bound analysis for predicting the stability of
multilayered slope. Back propagation neural network (BPNN) was implemented by Wang
et al. (2005) for estimating the FOS of Yudonghe landslides in China and found the FOS to
be approximately 1.1. Caniani et al. (2008) and Ramakrishnan et al. (2013) used ANN to
create a landslide susceptibility map (LSM) of municipal area of Potenza, Southern Italy,
and Tawaghat area, Kumanon Himalaya, India, respectively. Alavi and Gandomi (2011)
used multi-expression programming (MEP), linear genetic programming (LGP), and gene
expression programming (GEP) to assess the FOS based upon dataset of Wang et al. (2005)
and found LGP to be the best among MEP, GEP, and LGP.
Nourani et al. (2014) used GP, frequency ratio (FR), logistic regression (LR), and ANN
to create LSM for Zonouz Plain, Iran, and found GP to have the best prediction accuracy.
Based on another dataset containing 103 samples, Manouchehrian et al. (2014) proposed a
model equation to predict FOS for circular failure of slopes using genetic algorithm (GA)
and found it to be superior than the GP model (Yang et al. 2004). Liu et al. (2014) used
extreme learning machine (ELM) technique for the evaluation and prediction of stability of
slopes and found ELM to be more efficient than generalized regression neural networks
(GRNN), GP (Yang et al. 2004), and GA (Manouchehrian et al. 2014) models.
Though ANN and SVM are no more considered as ‘black box’ systems, due to the
development of model equation using weights and biases (Das and Basudhar 2006, 2008;
Das et al. 2011; Erzin and Ecemis 2015; Erzin and Turkoz 2015), ANN suffers with poor
generalization on account of instances of attainment of a local minima in the learning step.
SVM does not have the problem of generalization, but approximation of the parameters
C and e to get the best results is a heuristic process. As per the capacity to explain the
practical form of the link between the input parameters, mathematical models are divided
into three types—white box, gray box, and black box (Giustolisi et al. 2007). The models
of the white box type are centered around the physical laws, where the input parameters
and factors are identified from which a comprehensive mathematical model underlying the
physical relationship of inputs and outputs can be obtained. In black box techniques, the
relationship between the input and output is not physical, but is based upon the data. Thus,
it is not easy to get a comprehensible mathematical model. Gray box techniques are related
to conceptual models.
The present study is related to the development of model equations for FOS of slopes
using three AI procedures—functional networks (FNs), multivariate regression splines
(MARS), and multigene genetic programming (MGGP). The dataset used by Man-
ouchehrian et al. (2014) has been used to develop prediction models using FN, MARS, and
MGGP. The results of present study have been compared with the results of GP (Yang
et al. 2004) and GA models proposed by Manouchehrian et al. (2014).
FN is a recently developed prediction tool which follows the physical world structure
and has been applied in various areas of science and engineering such as structural
engineering (Rajasekaran 2004), real-time flood forecasting, transportation engineering
(Attoh-Okine 2005), mining, function approximations (Castillo et al. 1999), petroleum
engineering (El-Sebakhy et al. 2007), pattern recognition, medicine (El-Sebakhy et al.
2006), signal processing, and bioinformatics. Castillo and Ruiz-Cobo (1992), Castillo
(1998), and Castillo et al. (2000) have introduced functional networks. The use of FN in

123
Nat Hazards

geotechnical engineering was not found in the literature to the best knowledge of the
authors.
Friedman (1991) developed MARS (white box technique) to solve linear and nonlinear
regression problems. It is built on the principle of physical laws where prior assumption of
relationships between the input and output is eliminated. Therefore, it is applied in
numerous parts of economics, science, and technology for data mining. Yet in the field of
geotechnical engineering, implementation of MARS algorithm is less (Samui et al. 2011;
Muduli et al. 2013).
MGGP is a modification of GP, a ‘gray box’ technique that provides with a simplified
formula as compared to ANN. Problems related to geotechnical engineering have been
solved using MGGP in the recent past (Gandomi and Alavi 2011, 2012; Muduli and Das
2013a, b, 2014a, b; Muduli et al. 2013, 2014) and are found to be efficient in the devel-
opment of prediction models.
The following sections describe MGGP, MARS, and FN briefly as their application in
geotechnical engineering is limited.

2 Artificial intelligence models

2.1 Functional networks

Functional network (FN) is a recent technique which is being used as an alternate tool to
ANN. In FN, network’s preliminary topology is derived and centered around the modelling
properties of the real domain, or in other words, it is related to the problems of the domain
knowledge, whereas in ANN, by the use of trial-and-error approach, the required number
of hidden layers and neurons is determined, so that a good fitting model to the dataset can
be obtained. After the availability of initial topology, functional equations are utilized to
reach at a much simpler topology. Therefore, functional networks eliminate the problems
of artificial neural networks by utilizing together the data knowledge and the domain
knowledge from, which derives the topology of the problem. With the help of domain
knowledge, FN determines the network structure, and from the data, it estimates the
unknown neuron function. Initially, arbitrary neural functions are allocated with an
assumption that the functions are of multi-argument type and vector-valued in nature.
A FN is derived by learning in two stages as enumerated below:
1. Structural learning: In this stage, preliminary topology of the network is built on the
assets obtainable to the designer and further simplifying is done with the help of
functional equations.
2. Parametric learning: In this stage estimation of the neuron function is based on the
combination of functional families, which is provided initially, and then from the
available data, the associated parameters are estimated. It is similar to the estimation of
the weights of the connections in artificial neural networks.

2.1.1 Working with functional networks

The main elements around which a functional network is built on are shown in Fig. 1. It
can be itemized as:

123
Nat Hazards

1. Storing units
• The inputs, x1, x2, x3, … require 1 input layer of storing unit.
• The outputs, f4, f5, … require 1 output layer of storing unit.
• Processing units contain one or several layers, which evaluates the input from the
preceding layers to deliver it to the succeeding layer, f6 .
2. Computing unit’s layer, f1 ; f2 ; f3 : In this computing unit is the neuron which evaluates
the inputs coming from the preceding layer to deliver the outputs to the succeeding
layer.
3. Directed links set: Intermediary functions are not random in nature, but they depend on
the framework of the networks, such as x7 = f4(x4, x5, x6), as indicated in Fig. 1.
All the elements as shown in Fig. 1 and described above together form the functional
network architecture. The network architecture defines the topology of the functional
network and determines the functional capabilities of the network.
The steps for working with the functional network are as follows:
Step 1: physical relationship of inputs with outputs.
Step 2: preliminary topology of the functional network depends on the dataset of the
problem. Artificial neural network selects the topology by trial-and-error approach,
whereas functional networks select the topology on the properties of the data, which
ultimately leads to a solo network structure.
Step 3: functional equation simplifies the initial network structure of FN. It is done by
constantly searching for a simpler network in comparison with the existing one, which will
predict the same output from the same set of inputs. Once a simpler network is found the
complicated network is replaced with the simpler one.
Step 4: a single neuron function is selected for the specific topology, which yields a set
of outputs.
Step 5: in this step data are collected for the training of the network.
Step 6: on the basis of the data, which are acquired from Step 5, and a blend of the
functional families, the neuron function is estimated. Learning stage of the network can be
linear or nonlinear in nature, which directly depends on the linearity of the neuron function.
Step 7: once a model has been developed it is checked for error rate and also it is
validated with a different set of data.

Fig. 1 Typical layout of a functional network

123
Nat Hazards

Step 8: if the model is found to be satisfactory in the cross-validation process, it is ready


to be used.
In FN the learning method is selected on the basis of the neural function, which depends
on the type of data U = {Ii, Oi}, {i = 1, 2, 3, 4, …, n}. Learning procedure involves
minimization of the Euclidean norm of the error function, and it is represented as:
1X n
E¼ ðOi  FðiÞÞ2 ð1Þ
2 i¼1

Estimated neural functions fi(x) can be arranged in the following order:


X
m
fi ðxÞ ¼ aij /ij ðXÞ ð2Þ
j¼1

where / is the shape function, having algebraic expression (1, x, x2, x3, … xn), exponential
function (1, e x, e2x, …, enx), and/or trigonometry function
½1; sinðxÞ; cosðxÞ; sinð2xÞ; cosð2xÞ; sinð3xÞ; cosð3xÞ. A set of linear or nonlinear algebraic
equation is obtained with the help of associative optimization functions.
Previous information about the functional equation is vital for working with functional
network. The functional equation can be defined as a set of functions, unknown in nature,
which excludes the integral and differential equations. Cauchy’s functional equation is the
most common instance for the functional equations, and it is as follows:
f ðx þ yÞ ¼ f ðxÞ þ f ðyÞ; x; y 2 R ð3Þ

2.1.2 Associativity functional network

In this research, the associativity functional network (a simple type of FN) is used. Castillo
and Ruiz-Cobo (1992) and Castillo et al. (2000) proposed that by using the elementary
theory of functional equation, any multi-input network can be converted to an associative
network. A typical associative FN can be represented as given below,
X
m
fs ðxs Þ ¼ asi /si ð4Þ
i¼1

where s = 1, 2, 3 …., m and, /si is polynomial or exponential or trigonometric or any other


admissible function and is known as shape function.

Fig. 2 An associative functional network

123
Nat Hazards

Figure 2 shows an associative functional network with 2 inputs and 1 output; the 2
inputs are x1 and x2, and the output is x3. The function f3 can be presented as:
X
m
f3 ðx3 Þ ¼ a3i /3i ð5Þ
i¼1

The output function in terms of input function can be represented as:


f3 ðx3 Þ ¼ f1 ðx1 Þ þ f2 ðx2 Þ ð6Þ

Therefore, error in the jth data is obtained from the following equation,
ej ¼ f1 ðx1 Þ þ f2 ðx2 Þ  f3 ðx3 Þ ð7Þ

For estimating the coefficients, ai, i = 1, 2, 3, …, m, we minimize the sum of squared


errors as:
!20
Xn X m
E¼ ai ½/i ðx1j Þ þ ð/i ðx2j Þ  /i ðx3j ÞÞ ð8Þ
j¼1 i¼1

Subject to,
X
m
f ðx0 Þ  ai /i ðx0 Þ ¼ a ð9Þ
i¼1

where a = real constant.


With the help of Lagrangian multiplier method, an auxiliary function can be constructed
and it is as given below:
!2 !
Xn Xm X m
Ek ¼ ai bij þk ai /i ðx0 Þ  a ; ð10Þ
j¼1 i¼1 i¼1

where
bij ¼ /i ðx1j Þ þ /i ðx2j Þ  /i ðx3j Þ

Therefore, Eq. (10) can be minimized with the help of Eqs. (11) and (12), which is
given below:
!
oEk X n X m
¼2 ai bij brj þ k/r ðx0 Þ ¼ 0; r ¼ 1; 2 . . . m; ð11Þ
oar j¼1 i¼1

oEk X m
¼ ai /i ðx0 Þ  a ¼ 0 ð12Þ
ok i¼1

In the above group of equations, there are (m ? 1) equations and (m ? 1) unknowns,


which can be solved to get the coefficients, ai, i = 1, 2, 3,…., m:
The matrix form is

123
Nat Hazards

0 1
0
  
BBT /0 aT B0C
¼B C
@ ... A ð13Þ
/T0 0 k
0
where B = coefficient matrix; bij; and a = a1, a2, a3, … , am. On simplification the above
matrix can be presented as:
½Bfug ¼ fvg ð14Þ
Equation (14) can be solved to find the unknowns with respect to v and thus u can be found
out from which the coefficients a = a1, a2, a3, … , am can be derived. By taking m = 1 to
a, the solution can be written in the form of a simple functional equation, which is given
below:
½f3 ðx31 Þ  a31 
f3 ðx3i Þ ¼ f1 ðx1i Þ þ f2 ðx2i Þ ¼ a31 þ a32 or; x31 ¼ ð15Þ
a32

2.2 Multivariate adaptive regression splines

MARS correlates between a set of input variables to an output variable through adaptive
regression method. In MARS a nonlinear, nonparametric approach is used to develop a
prediction model without any prior assumption of any relationship between the input
(independent variables) and the output (dependent variable). MARS algorithm creates
these relationships by using sets of coefficient and basis functions from the dataset as
discussed above. Due to this, MARS is favorable over other learning algorithms where the
number of inputs (independent variables) is more in number.
The backbone of MARS algorithm is founded on divide-and-conquer strategy in which
the dataset is split to a number of groups of piecewise linear segments known as splines,
which varies in gradient. MARS is comprised of knots, which are basically the end points
of splines and the functions (piecewise linear function/piecewise cubic function) between
these knots are called as basis function (BF). In this paper for the case of simplicity of the
model, only piecewise linear basis functions are used.
MARS algorithm proposed by Friedman (1991) is a two-step process to fit data and is
explained below:
1. Forward stepwise algorithm: Basis functions are added in this step. First the model is
developed only with the help of an initial intercept known as b0 . Then in each
successive step, a basis function which shows the greatest decrease in the training error
is annexed. Like this the whole operation is continued until the number of basis
functions reaches its maximum value which has been predetermined beforehand. As a
result, an over-fitted model is obtained. Searching of knots among all the variables are
done by the adaptive regression algorithm
2. Backward pruning algorithm: Elimination of the over-fitting of data is done in this
phase. The terms in the model are snipped (one by one removal of the terms) in this
operation. The best viable submodel is obtained by removing the least effective term.
Then the subset of models is equated among themselves by means of the generalized
cross-validation (GCV) process. GCV for N samples is obtained as per the equation
given below:

123
Nat Hazards

PN
1
½Yi  f ðXi Þ2
GCV ¼ N  i¼1 2 ð16Þ
dðM1Þ

1 N2

where M = number of BF; d = penalizing parameter; N = number of data samples; and


f(Xi) = prediction values of the MARS model. As the complexity of the model increases,
variance also increases, which directly depends on the denominator of GCV. The number
of knots is represented by (M - 1)/2 term of the GCV denominator. So the BFs and the
knots are penalized by the GCV.
For better understanding of MARS algorithm, examine a dataset, which contains an
output y for a set of inputs X = {X1, X2, X3, … , Xp}, which consists of p input variables.
MARS generates a model of the form:
y ¼ f ðX1 ; X2 ; X3 ; . . . ; Xp Þ þ e ¼ f ðXÞ þ e ð17Þ
where e = distribution of error; f(x) = a function which is approximated by BFs (piece-
wise linear function/piecewise cubic function).
For the case of simplicity, only piecewise linear functions have been discussed in this
paper for its easy interpretability. The piecewise linear function is represented as
maxð0; x  tÞ where t is location of the knot. Its mathematical form is,
maxð0; x  tÞ ¼ fx  t; if x [ t or 0; otherwiseg ð18Þ
And finally, f(x) = linear combination of BFs, And interactions between them is defined
as,
X
M
f ðXÞ ¼ b0 þ bm km ðX Þ ð19Þ
i¼1

where km = basis function, which is a single spline or product of 2 or more than 2 splines;
b = coefficients of constant values calculated by least square method.
Graphical representation of MARS using piecewise linear functions for fitting data is
presented in Fig. 3. An illustration containing 22 data samples as inputs with an output is

Fig. 3 Example of how MARS uses piecewise linear splines to fit a dataset

123
Nat Hazards

taken. Random numbers between one and twelve comprised the input matrix {X} with a
single output {Y}, which is obtained as per the equation given below:
1 1
Yi ¼  ð20Þ
sinðXi Þ cosðXi Þ

Also the data samples are normalized in the range of 0 to 1 and MARS analysis is
conducted. The MARS model developed for this dataset is represented as:
^
Y ¼ 0:413 þ 4:066BF1  5:336BF2 þ 1:852BF3 ð21Þ
^
where Y = predicted values;
BF1 ¼ maxð0; Xi  0:40Þ; ð22Þ

BF2 ¼ maxð0; Xi  0:65Þ; and ð23Þ

BF3 ¼ maxð0; 0:65  Xi Þ ð24Þ

In this MARS model the knots are situated at, x = 0.65 and x = 0.40. The R value for
this model is 0.805, which is the correlation coefficient between the actual value and the
predicted value of the MARS model. Proper care should be taken to use normalized values
of Xi (Eq. 22–24), and the denormalized values of the predicted Yi can be obtained as per
Eq. 25.
Y^denorm ¼ Y^norm þ ðXiðmaxÞ  XiðminÞ Þ þ XiðminÞ ð25Þ

Therefore, models developed using MARS algorithm has not only better efficiency but
also simplifies the complex equations just like Eq. 20 to a simple linear equation.

Fig. 4 Example of linear combination of GP tree used in MGGP

123
Nat Hazards

2.3 Multigene genetic programming

Multigene genetic programming (MGGP) is a variation of GP where a model is built from


the combination of several GP trees. Each tree is composed of genes, which represents a
lower nonlinear transformation of input variables. The output is created from weighted
linear combination of these genes and is termed as ‘multigene.’ Example of a MGGP
model where output is a linear combination of two nonlinear models, Gene 1 and Gene 2 is
shown in Fig. 4.
For a MGGP model, the model complexity and accuracy can be controlled by controlling
the maximum depth of GP tree (dmax) and the maximum allowable number of genes (Gmax).
With decrease in Gmax and dmax values, complexity of the MGGP model decreases, whereas
its accuracy is hampered. Thus, there exist optimum values of Gmax and dmax which gives
fairly accurate results with relatively compact model (Searson et al. 2010) for a given
problem. The linear coefficients (c1 and c2) termed as weights of the gene and bias (c0) of the
model are obtained by ordinary least square method on the training data.
First, population initialization is done by creating a number of randomly evolved genes
with lengths varying from 1 to Gmax. Then, for each generation, new population is chosen
from the initial population as per their merit and then implementation of reproduction,
followed by crossover, followed by mutation operations are performed on the function and
terminal sets of the selected GP trees. In subsequent runs population is generated with
addition and deletion of genes using traditional crossover mechanisms from GP and special
MGGP crossover mechanisms. Few distinctive MGGP crossover mechanisms (Searson
et al. 2010) are briefly described below.

2.3.1 Two-point high-level crossover

The process of mating between two individual parents to swap genes between them is
called as a two-point high-level crossover. Suppose there are two trees having four genes
and three genes, respectively, marked by Gi to Gn. Assume that the Gmax value for the
model is five. A crossover point represented by {…} is selected for each individual.
[G1, {G2, G3, G4}], [G5, G6, {G7}]
Genes enclosing the crossover points are interchanged, and thus, two new offspring are
formed as shown below.
[G1, {G7}], [G5, G6, {G2, G3, G4}]
The number of genes in any individual is not allowed to be more than Gmax. But if it
exceeds then randomly genes are selected and eliminated till each individual has Gmax
genes. This process leads to creation of fresh genes for both the individuals, as well as the
deletion of some genes.

2.3.2 Two-point low-level crossover

Standard crossover of GP subtrees in MGGP algorithm is known as two-point low-level


crossover. First, a gene is arbitrarily chosen from each of the individual and then
exchanging of the subtrees under the selected nodes is done. The newly created trees swap
the parent trees in otherwise unchanged individual in the subsequent generation. There are
six types of mutations, which can be performed in this stage (Gandomi and Alavi 2012).

123
Nat Hazards

For the achievement of best MGGP model probability of reproduction, crossover and
mutation have to be given, such that the sum of the probability of these operations should
not exceed 1.

3 Database and preprocessing

In this study, two different types of models, Model 1 and Model 2 are developed for
prediction of FOS of slope using associative FN, MARS, and MGGP with the database
available in the literature as compiled in Manouchehrian et al. (2014). The input data
consist of parameters unit weight, c (kN/m3), cohesion, c (kPa), internal friction angle,
u(), slope angle, b(), height of slope, H(m), and pore pressure parameters, ru. Table 1
provides the statistical parameters of the inputs and outputs value of the discussed database
in terms of minima, maxima, mean and standard deviation. It can be seen that the database
consists of a wide range of values. The first model, Model 1, consists of six inputs and one
output, i.e., FOS of the slope and separate models are developed using FN, MARS, and
MGGP. The second model, Model 2, consists of six inputs and one output, i.e., the stability
or failure of the slope in terms of failed (0) or stable (1) (Das et al. 2011). Model 2 checks
the accuracy of an AI technique in terms of its ability to predict whether a slope is stable or
failed. Out of 103 samples, 75 data were randomly selected and used in the training phase
and the remaining 28 data were used in the testing phase for both Model 1 and Model 2.
The training and testing dataset were normalized between 0 and 1 to develop the FN and
MARS model. However, no preprocessing was done to develop the MGGP model. Pre-
dictive equations are also presented for FN, MARS, and MGGP.
The FN, MARS, and MGGP algorithms were implemented with the use of MATLAB
(Mathworks 2005). The results from this study have been compared with the previous
results, existing in the literature based on statistical factors such as maximum absolute
error (MAE); root mean square error (RMSE); average absolute error (AAE); Nash–
Sutcliff coefficient of efficiency (E); and coefficient of correlation (R).

4 Results and discussion

4.1 Model 1: prediction of value of FOS

4.1.1 Functional network

Using the prescribed basis function and degree of the selected function, the model in
functional network was developed. Though increase in degree leads to more accurate
results, at the same time the complexity of the model is increased, and hence, a trade-off

Table 1 Statistical parameters


c c / b H ru FOS
of input and output used in this
study
Min 12.00 0.00 0.00 16.00 3.60 0.00 0.63
Max 28.44 50.00 45.00 53.00 214.00 0.50 2.31
Avg. 19.93 10.45 25.83 33.05 41.76 0.20 1.28
Std. dev. 3.64 9.98 10.96 9.34 44.44 0.17 0.40

123
Nat Hazards

Fig. 5 Associative functional network used in this study

was made in the present study. FN model with five degrees and exponential basis function
was found to give an optimum value. Figure 5 shows the associative functional network for
the prediction of FOS. The general equation for predicting the outcomes can be represented
as:
n X
X m
y ¼ a0 þ fi ðxj Þ ð26Þ
i¼1 j¼1

where n = no. of variables; m = degree of the variable. Here, n = 6 and m = 5, and the
above Eq. (26) can be presented in prolonged form as:
a0 ¼ 0:5159 ð27Þ
2 3 4
f1 ðx1 Þ ¼ 2:2502ec  6:2786ec þ 4:2406ec ð28Þ
3 4
f2 ðx2 Þ ¼ 0:6885ec  4:0893ec þ 3:7255ec ð29Þ
2 3 4 5
f3 ðx3 Þ ¼ 12:4460e;  44:8180e; þ 56:7434e;  23:8521e; ð30Þ
3 4 5
f4 ðx4 Þ ¼ 7:6993eb þ 15:2296eb  7:7928eb ð31Þ

123
Nat Hazards

2 3 4 5
f5 ðx5 Þ ¼ 18:4014eH þ 83:8029eH  122:8152eH þ 57:3517eH ð32Þ
3 4 5
f6 ðx6 Þ ¼ 4:7699eru  13:2319eru þ 8:8308eru ð33Þ

Fig. 6 Plot of measured and predicted factor of safety for MGGP, MARS, and FN for a training and
b testing data

123
Nat Hazards

The input parameters, which are used in Eqs. (28)–(33), are normalized in the range of
0–1. The summation of Eqs. (27)–(33) gives the normalized value of the output, say
FOSnorm. FOS obtained can be denormalized with the help of Eq. (34), as given below, to
get the actual predictions.
FOS ¼ 1:68FOSnorm þ 0:63 ð34Þ

It can be seen from Fig. 6a, b that the scattering of the data points for FN is inside the
80 % prediction limit. The values of different statistical parameters obtained from FN
model are also presented in Table 1. The R (Table 2) values for training and testing were
found to be 0.910 and 0.907, respectively. As per with previous experience, it is identified
that the R value is a biased estimate (Das and Sivakugan 2010), therefore another statistical
parameter known as Nash–Sutcliff coefficient of efficiency (E) (Das and Basudhar 2008)
can be used for assessing the efficiency of a model. Table 1 also presents the E value for
MGGP and other prediction methods. The E value is defined as
E1  E2

E1
where
X
N
E1 ¼ ðFOSm  FOSrm ave Þ2
t¼1
X
N  2
E2 ¼ FOSp  FOSrm
t¼1

and FOSm, FOSm ave, FOSp are the measured, average of the measured, and predicted FOS
values, respectively. Similarly, the P50 and P90 values for FN as shown in Table 3 were
found to be 0.983 and 1.218, respectively. For 50 % cumulative probability (P50), if the
calculated value is less than one, then the developed model is said to be an underpredicted
model, and if it is more than one, then it is an overpredicted model. When the P50 value is
approximately one, then it can be termed as the ‘best’ model. Also the P90 value, which is
the 90 % cumulative probability value, represents the deviation in the ratio of FOSp/FOSm
for the total number of observations. The model with FOSp =FOSm close to 1.0 is the better
model. Thus, the prediction was observed as satisfactory at 50 % cumulative probability
but there was an overprediction at 90 % cumulative probability.

Table 2 Statistical performance of FN, MARS, and MGGP for Model 1


Reference Coefficient of correlation (R) Coefficient of efficiency (E)

Training Testing Training Testing

FN 0.910 0.907 0.828 0.802


MARS 0.923 0.922 0.853 0.849
MGGP 0.898 0.904 0.806 0.816

123
Nat Hazards

Table 3 Cumulative probabilities depending on sorted FOSm/FOSm for MGGP, MARS, and FN for Model
1
Prediction model P50 P90

FN 0.983 1.218
MARS 0.999 1.180
MGGP 0.988 1.247

Fig. 7 Comparison of MAE, AAE, and RMSE for MGGP, MARS, and FN for Model 1 for a training and
b testing

123
Nat Hazards

4.1.2 Multivariate adaptive regression splines

The efficiency of the MARS model basically hangs on the maximum number of basis
functions that can be allowed in the final results and also on the number of basis functions,
which is allowed initially in the forward phase. It can be observed that, as the number of
BFs increases, the accuracy of the model also increases, but it also leads to the increase in
the complexity of the model. Therefore, a trade-off has to be made between the complexity
and the accuracy of the model. The required number of basis functions which were adopted
in this paper is 13 in number. The result of the MARS model can be represented in an
equation form, and it is as follows:
FOS ¼ 0:565 þ 1:474  maxð0; c  0:523Þ  maxð0; 0:6  ru Þ  0:714  maxð0; 0:2  CÞ
þ 48:651  maxð0; 0:2  CÞ  maxð0; 0:021  HÞ þ 1:617  maxð0; ;  0:8Þ
 0:683  maxð0; 0:8  ;Þ  2:045  maxð0; b  0:649Þ þ 0:42  maxð0; 0:649  bÞ
þ 20:746  maxð0; 0:8  ;Þ  maxð0; b  0:784Þ  10:23  maxð0; 0:649  bÞ
 maxð0; b  0:351Þ  0:338  maxð0; H  0:22Þ þ 0:811  maxð0; 0:220  HÞ
þ 6:747  maxð0; ru  0:6Þ  maxð0; ru  0:9Þ  1:173  maxð0; ru  0:6Þ
ð35Þ

In Eq. (35) the input variables/parameters are in a normalized form, which ranges from
0 to 1, and the FOS obtained is also in normalized form. For getting the denormalized value
of the FOS (Eq. 35), Eq. (34) is used to obtain the actual output.
For training and testing the R values of the MARS model are 0.923 and 0.922,
respectively, as shown in Table 2. As per Smith (1986), the predicted values for MARS
model show a strong correlation with the measured values of FOS. It can be noticed from
the plot of measured vs predicted FOS values in Fig. 6 that the scatter of data is well within
the 80 % prediction limit. The E values for training and testing are 0.853 and 0.849,
respectively. In terms of cumulative probability, MARS also showed the same trend as
MGGP with P50 and P90 values of 0.999 and 1.180 as shown in Table 3. However, the
extent of overprediction in MARS is less as compared to MGGP. It can be observed from
Tables 2, 3, and Fig. 7 that MARS has a better prediction capability as compared to MGGP
in terms of comparison of statistical criteria.

4.1.3 Multigene genetic programming

Multigene genetic programming (MGGP) model performance depends on the size of the
population; total number of generations; crossover percentage; mutation percentage; dmax
and Gmax values. In this research, the best GP model obtained was by takingpopulation size
to be 700; number of generations to be 150; reproduction probability as 5 percent; tour-
nament size to be 7; crossover probability as 85 percent; and finally mutation probability as
10 percent. The optimum result was achieved, with Gmax as 4 and dmax as 5. Model
equation obtained from the analysis with the help of MGGP algorithm is presented below:

123
Nat Hazards

FOS ¼ 1:534 þ 0:2429 tanhð2ð;  bÞ þ sin C þ tanh CÞ


þ 0:01426fru ðC  c þ 2bÞ  ð;  bÞ sin b þ ð;  H Þ tanh bg
þ 0:000626 cos sin sinðbÞ  cosðru Þ  ðb  H Þ  ðru  c þ sin cÞ  0:1627  ;  ru2
 cos sinðcÞ  cos sinðc  ru Þ
ð36Þ

Figure 6a, b shows the plot of measured vs predicted FOS for training and testing data,
respectively. The R value between measured and predicted values of FOS for the MGGP
model and other prediction methods is presented in Table 1. The R value according to
MGGP model is 0.898 and 0.904 for training and testing data, respectively. This R value
suggests a strong correlation between the measured and predicted values of FOS as per
Smith (1986). It can be seen from Fig. 6 that the most of the data lie within 80 %
prediction limit. The values of MAE, AAE, and RMSE for training and testing are shown
in Fig. 7a, b, respectively.
Table 2 also presents the E value for MGGP and other prediction methods. The E value
for MGGP in training and testing data is 0.806 and 0.816, respectively. A value of E close
to 1 indicated that the predicted values of the output match perfectly with the observed
variables.
Abu-Farsakh and Titi (2004) proposed that the standard deviation (r) and the mean (l)
of the ratio between FOSp/FOSm are the important gauges for determining the accuracy and
precision of the developed AI models. Cumulative probability of the FOSp/FOSm can also
be considered for the assessment of the model. Figure 8 shows the plot of cumulative
probability vs the ratio of predicted and measured values of FOS for MGGP and other
prediction models. Table 3 provides the P50 and P90 for MGGP and other prediction
methods. The values of P50 and P90 for the MGGP model are 0.988 and 1.247,

Fig. 8 Comparison of cumulative probability distribution for MGGP, MARS, and FN for Model 1

123
Nat Hazards

respectively. Thus, MGGP model has a fair prediction at 50 % cumulative probability but
an overprediction at 90 % cumulative probability.
A comparison of the statistical parameters as provided in Tables 2 and 3 suggests that
among the prediction tools used in this study MARS has the best performance followed by
FN and MGGP. The outcomes of the current research were also compared with the existing
methods as per the literature. Manouchehrian et al. (2014) used genetic algorithm (GA)-
based prediction model and obtained R value of 0.86 in training using the above dataset.
Also, the value of RMSE in training was for GA model was 0.20. A comparison of these
values with the R and RMSE value obtained in the present study suggests that MGGP,
MARS, and FN are more efficient in predicting the FOS of slope than GA model as per
Manouchehrian et al. (2014).
The efficacies of the proposed model are also compared with genetic programming (GP)
model provided by Yang et al. (2004). Seven samples in the dataset were omitted from this
calculation due to having a / value of zero. The R value between the FOS calculated by
LEM and that of GP model by Yang et al. (2004) for the remaining 95 samples was found
to be 0.6. Similarly, the R value for the same set between the FOS calculated by LEM and
that calculated using the GA model provided (Manouchehrian et al. 2014) was found to be
0.79. The respective values for the models derived in this study using MGGP, MARS, and
FN were found to be 0.905, 0.923, and 0.907, respectively. Hence, the developed models
MARS, FN, and MGGP models are found to be efficient compared to existing GP (Yang
et al. 2004) and GA (Manouchehrian et al. 2014) models.

4.2 Model 2: prediction of safety of slope

In the present section, Model 2 was developed to identify the field condition of the slope as
stable or unstable (failed) based on input parameters. Earlier Das et al. (2011) and Samui
(2008) used ANN and SVM, respectively, to predict the field condition of the slopes using
a different set (Das et al. 2011). As discussed in the previous section, MARS model was
found better compared with FN and MGGP. Hence, while developing Model 2, i.e., to
predict the stability of slopes in terms of failed or stable, only MARS and FN techniques
are used. The six inputs used are height of slope, H(m), unit weight, c (kN/m3), cohesion,
c (kPa), internal friction angle, u(), slope angle, b(), and pore pressure parameters, ru,
and one output is the field condition of the slope, i.e., failed(0) or stable(1).
Das et al. (2011) and Samui (2008) used a relaxed criterion for specifying a slope as
stable or failed, i.e., if FOS [ 0.5, then the slope is said to be stable, and if FOS \ 0.5, the
slope is considered as failed. According to these criteria, ANN predicted the training
sample with just one inaccuracy and testing sample with 100 % accuracy (Das et al. 2011).

Table 4 Accuracy of prediction of FN and MARS for different evaluation criteria


Accuracy of prediction (number of successful prediction)

Models Relaxed criteria Less stringent criteria Stringent criteria

Training (out Testing (out Training (out Testing (out Training (out Testing (out
of 75) of 28) of 75) of 28) of 75) of 28)

FN 74 27 72 20 65 18
MARS 74 28 64 27 52 19

123
Nat Hazards

On the other hand, SVM model (Samui 2008) could predict the training set with 100 %
accuracy but the prediction of testing set was limited to 85.7 % only. In the present study,
MARS and FN models were developed to assess the field condition of the slopes based on
the dataset present in Manouchehrian et al. (2014). A FN model with degree 9 and
polynomial basis function was found to be optimum for this purpose. Also, a MARS model
with 13 BFs was adopted. The accuracy of the prediction in terms of number of successful
prediction is presented in Table 4. Based on the criteria used in the previous literature (Das
et al. 2011; Samui 2008), designated here as relaxed criteria, the developed MARS model
could accurately predict all test cases and failed for just one case from the training set. The
prediction performance of FN model was found to be equally good, and it predicted both
the training and testing cases with just one inaccurate prediction in each set.
Apart from the aforesaid relaxed criteria, two other criteria were also employed in this
study to measure the performance of the developed models in accurately predicting
whether a slope is going to be stable or will fail. They are listed as below:
1. Less stringent criterion: FOS [ 0.65: Stable
FOS \ 0.35: Failed
2. Stringent criterion: FOS [ 0.85: Stable
FOS \ 0.15: Failed
The results from MARS and FN for prediction of slope as stable or unstable were tested
against criteria I and II. The obtained results are presented in Table 4.
Table 4 shows that for less stringent criterion and more stringent criterion, a few of the
cases are inaccurately predicted for both MARS and FN. On comparison between the two,
MARS and FN can be said to be equally good at prediction of the field condition of the
slopes. As per Das and Basudhar (2008) the efficacy of the models should be compared in
terms of testing data. Based on the testing dataset, it was observed that MARS model has
better prediction capabilities than FN model. As such a study for the above database is not
available in the literature, the above results could not be compared. Similar to other AI, it is
imperative to mention here that the prediction equations provided for different models
above can be applied to a new data to calculate the factor of safety of a slope when the
inputs for the new dataset lie between the ranges of inputs used in this study.

5 Conclusion

The importance of accurate determination of the FOS of slopes is undisputed and hence has
attracted many researchers to develop models to predict FOS of slopes based upon both
numerical and analytical analyses. AI techniques have been lately used as a prediction tool
in several fields and have been validated as quite an accurate tool in the field of
geotechnical engineering by many researchers in recent decades. This paper discussed the
use of recently developed AI techniques, MGGP, MARS, and FN as an alternate tool to
predict the value of FOS of slopes and to classify a slope as failed or stable. Data previ-
ously available in the literature were used to train two different models, Model 1 to predict
the value of FOS and Model 2 to assess whether a slope is stable or not.
For Model 1, separate model equations were presented as per MGGP, MARS, and FN
for the prediction of FOS of slope, which can be used to calculate the FOS, based on input
parameters, unit weight, c (kN/m3), cohesion, c (kPa), internal friction angle, u(0), slope
angle, b(0), height of slope, H(m), and pore pressure parameters, ru. The developed models

123
Nat Hazards

are found to be efficient in comparison with GA and GP models available in the literature.
Based on different statistical performances criteria, R, E, MAE, AAE, and RMSE the
developed MARS model is found to better than MGGP and FN models.
For Model 2, as per the existing relaxed criteria for classification of slopes, MARS and
FN were found to accurately predict whether a slope is going to be stable or not. However,
for less stringent criterion and more stringent criterion, proposed in the present study the
percentage of accuracy decreased. Based on accuracy for testing data, MARS model has
better prediction capabilities than FN model.

References
Abu-Farsakh MY, Titi HH (2004) Assessment of direct cone penetration test methods for predicting the
ultimate capacity of friction driven piles. J Geotech Geoenviron Eng 130(9):935–944
Alavi AH, Gandomi AH (2011) A robust data mining approach for formulation of geotechnical engineering
systems. Eng Comput 28(3):242–274
Attoh-Okine NO (2005) Modeling incremental pavement roughness using functional network. Can J Civ
Eng 32:899–905
Bishop AW (1955) The use of the slip circle in the stability analysis of slopes. Geotechnique 5(1):7–17
Bishop AW, Morgenstern NR (1960) Stability coefficients for earth slopes. Geotechnique 10(4):129–150
Caniani D, Pascale S, Sdao F, Sole A (2008) Neural networks and landslide susceptibility: a case study of
the urban area of Potenza. Nat Hazards 45:55–72
Castillo E (1998) Functional networks. Neural processing letter 7:151–159
Castillo E, Ruiz-Cobo R (1992) Functional equations in science and engineering. Marcel Dekker, New York
Castillo E, Cobo A, Gutierrez JM, Pruneda RE (1999) Working with differential, functional and difference
equations using functional networks. Appl Math Model 23:89–107
Castillo E, Cobo A, Manuel J, Gutierrez JM, Pruneda E (2000) Functional networks: a new network-based
methodology. Comput Aided Civ Infrastruct Eng 15:90–106
Das SK, Basudhar PK (2006) Undrained lateral load capacity of piles in clay using artificial neural network.
Comput Geotech 33:454–459
Das SK, Basudhar PK (2008) Prediction of residual friction angle of clays using artificial neural network.
Eng Geol 100(3–4):142–145
Das SK, Biswal RK, Sivakugan N, Das B (2011) Classification of slopes and prediction of factor of safety
using differential neural networks. Environ Earth Sci 64:201–210
Das SK, Sivakugan N (2010) Discussion of Intelligent computing for modeling axial capacity of pile
foundations. Can Geotech J 37(8):928–930
El-Sebakhy EA, Faisal KA, Helmy T, Azzedin F, Al-Suhaim A (2006) Evaluation of breast cancer tumor
classification with unconstrained functional networks classifier. In: Proceeding of the 4th ACS/IEEE
international conference on computer systems and applications, 281–287
El-Sebakhy EA, Abdulraheem A, Raharja P, Azzedin F, Sheltami T, Ahmed M (2007) Functional network
as a novel approach for prediction of permeability in a carbonate reservoir. In Proceedings SPE
conference, 11–14
Erzin Y, Ecemis N (2015) The use of neural networks for CPT-based liquefaction screening. Bull Eng Geol
Environ 74(1):103–116. doi:10.1007/s10064-014-0606-8
Erzin Y, Turkoz D (2015) Use of neural networks for the prediction of the CBR value of some Aegean
sands. Neural Comput Appl. doi:10.1007/s00521-015-1943-7
Fellenius W (1936) Calculation of the stability of earth dams. In: Transactions of the 2nd international
congress on large dams, Washington, USA, 445–462
Friedman J (1991) Multivariate adaptive regression splines. Ann Stat 19:1–141
Gandomi AH, Alavi AH (2011) Multi-stage genetic programming: a new strategy to nonlinear system
modeling. Inf Sci 181(23):5227–5239
Gandomi AH, Alavi AH (2012) A new multi-gene genetic programming approach to nonlinear system
modeling. Part II: geotechnical and Earthquake Engineering Problems. Neural Comput Appl
21(1):189–201
Giustolisi O, Doglioni A, Savic DA, Webb BW (2007) A multi-model approach to analysis of environmental
phenomena. Environ Model Softw 22(5):674–682

123
Nat Hazards

Kanungo DP, Pain A, Sharma S (2013) Finite element modeling approach to assess the stability of debris
and rock slopes: a case study from the Indian Himalayas. Nat Hazards 69:1–24
Liu Z, Shao J, Xu W, Chen H, Zhang Y (2014) An extreme learning machine approach for slope stability
evaluation and prediction. Nat Hazards 41(3):1–18. doi:10.1007/s11069-014-1106-7
Lu P, Rosenbaum MS (2003) Artificial neural networks and grey systems for the prediction of slope
stability. Nat Hazards 30(3):383–398
Manouchehrian A, Gholamnejad J, Sharifzadeh M (2014) Development of a model for analysis of slope
stability for circular mode failure using genetic algorithm. Environ Earth Sci 71:1267–1277
Math Work Inc. (2005) Matlab User’s Manual, Version 6.5. Natick, Massachusetts
Morgenstern NR, Price VE (1965) The analysis of the stability of general slip surfaces. Geotechnique
15(1):79–93
Muduli PK, Das SK (2013a) First order reliability method for probabilistic evaluation of liquefaction
potential of soil using genetic programming. J Geomechanics, Int. doi:10.1061/(ASCE)GM.1943-5622.
0000377
Muduli PK, Das SK (2013b) SPT-based probabilistic method for evaluation of liquefaction potential of soil
using multi-gene genetic programming. Int J Geotech Earthquake Eng 4(1):42–60
Muduli PK, Das PK (2014a) Evaluation of liquefaction potential of soil based on standard penetration test
using multi-gene genetic programming model. Acta Geophys 62(3):529–543
Muduli PK, Das PK (2014b) CPT-based seismic liquefaction potential evaluation using multi-gene genetic
programming approach. Indian Geotech J 44(1):86–93
Muduli PK, Das MR, Samui P, Das SK (2013) Uplift capacity of suction caisson in clay using artificial
intelligence techniques. Mar Georesour Geotechnol 31(4):375–390
Muduli PK, Das PK, Bhattacharya S (2014) CPT-based probabilistic evaluation of seismic soil liquefaction
potential using multi-gene genetic programming. Georisk Assess Manag Risk Eng Syst Geohazards
8(1):14–28
Nourani V, Pradhan B, Ghaffari H, Sharifi SS (2014) Landslide susceptibility mapping at Zonouz Plain, Iran
using genetic programming and comparison with frequency ratio, logistic regression, and artificial
neural network models. Nat Hazards 71(1):523–547
Rajasekaran S (2004) Functional networks in structural engineering. J Comput Civ Eng 18:172–181
Ramakrishnan D, Singh TN, Verma AK, Gulati A, Tiwari KC (2013) Soft computing and GIS for landslide
susceptibility assessment in Tawaghat area, Kumaon Himalaya, India. Nat Hazards 65(1):315–330
Sah NK, Sheorey PR, Upadhyama LW (1994) Maximum likelihood estimation of slope stability. Int J Rock
Mech Min Sci Geomech Abstr 31:47–53
Sakellariou MG, Ferentinou MD (2005) A study of slope stability prediction using neural networks. Geotech
Geol Eng 23:419–445
Samui P (2008) Slope stability analysis: a support vector machine approach. Environ Geol 56:255–267
Samui P, Kumar B (2006) Artificial neural network prediction of stability numbers for two-layered slopes
with associated flow rule. The Electronic Journal of Geotechnical Engineering. http://www.ejge.com/
2006/Ppr0626/Abs0626.htm. Accessed 17 July 2014
Samui P, Das S, Kim D (2011) Uplift capacity of suction caisson in clay using multivariate adaptive
regression spline. Ocean Eng 38(17–18):2123–2127
Searson DP, Leahy DE, Willis MJ (2010) GPTIPS: an open source genetic programming toolbox from
multi-gene symbolic regression. In: Proceedings of the International multi conference of engineers and
computer scientists, Hong Kong, vol 1(3), 77–80
Singh TN, Gulati A, Dontha L, Bhardwaj V (2008) Evaluating cut slope failure by numerical analysis—a
case study. Nat Hazards 47(2):263–279
Smith GN (1986) Probability and statistics in civil engineering: an introduction. Collins, London
Wang HB, Xu WY, Xu RC (2005) Slope stability evaluation using back propagation neural networks. Eng
Geol 80:302–315
Xu W, and Xie S (1999) Slope stability analysis and evaluation with probabilistic artificial neural Network
method. Site Investig Sci Technol 3. http://periodical.wanfangdata.com.cn/peridicals/kckxjs/kckx9903/
990304.htm
Yang CX, Tham LG, Feng XT, Wang YJ, Lee PKK (2004) Two-stepped evolutionary algorithm and its
application to stability analysis of slopes. J Comput Civil Eng 18(2):145–153

123

You might also like