You are on page 1of 11

Computers and Geotechnics 36 (2009) 787–797

Contents lists available at ScienceDirect

Computers and Geotechnics


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

Probabilistic stability analyses of slopes using the ANN-based response surface


Sung Eun Cho *
Korea Institute of Water and Environment, Korea Water Resources Corporation, 462-1, Jeonmin-Dong, Yusung-Gu, Daejon 305-730, South Korea

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

Article history: Slope stability analysis is a geotechnical engineering problem characterized by many sources of uncer-
Received 29 May 2008 tainty. Some of these sources are connected to the uncertainties of soil properties involved in the analysis.
Received in revised form 27 December 2008 In this paper, a numerical procedure for integrating a commercial finite difference method into a proba-
Accepted 2 January 2009
bilistic analysis of slope stability is presented. Given that the limit state function cannot be expressed in
Available online 14 February 2009
an explicit form, an artificial neural network (ANN)-based response surface is adopted to approximate the
limit state function, thereby reducing the number of stability analysis calculations. A trained ANN model
Keywords:
is used to calculate the probability of failure through the first- and second-order reliability methods and a
Slope stability
Probabilistic analysis
Monte Carlo simulation technique. Probabilistic stability assessments for a hypothetical two-layer slope
Response surface as well as for the Cannon Dam in Missouri, USA are performed to verify the application potential of the
Artificial neural network proposed method.
Ó 2009 Elsevier Ltd. All rights reserved.

1. Introduction opment of new perspectives concerning risk and reliability that are
outside the scope of conventional deterministic models.
The conventional tools for dealing with uncertainties associated In many practical probabilistic problems, the limit state surface
with soil properties in the field of geotechnical engineering have is not known explicitly. Instead, it may be known only implicitly
been applied under the use of safety factors and by implementing through a numerical procedure. Therefore, the failure domain only
local experience and engineering judgment. However, it has been can be found through repeated point-by-point numerical analyses
recognized that the factor of safety is not a consistent measure of with different input values. To reduce the computational effort re-
reliability, as slopes with the same safety factor value may exhibit quired for the probabilistic analysis, a response surface that
different reliability levels depending on the variability of the soil approximates the limit state function through polynomial regres-
properties [1]. Accordingly, numerous studies have been under- sion for the results of several selected simulations is needed. The
taken in recent years to develop a probabilistic slope stability anal- obtained closed-form expression of the limit state function is then
ysis that deals with the uncertainties of soil properties in a used to calculate the probability of failure.
systematic manner [2,1,3,4]. Detailed reviews of these studies Xu and Low [12] presented a practical procedure for integrating
can be found in Mostyn and Li [5], El-Ramly et al. [6], and Baecher the finite element method (FEM) and the limit equilibrium meth-
and Christian [7]. ods (LEM) into a probabilistic stability analysis for embankments.
A common approach to a probabilistic analysis was to locate the They adopted a polynomial-based response surface to approximate
critical deterministic surface and then calculate the probability of the limit state function for the stability problems and the first-
failure corresponding to this surface. However, the surface of the order reliability method to calculate the reliability index. They
minimum factor of safety may not be the surface of the maximum carried out the sequential refinement of the response surface to
probability of failure [8]. As an alternative, the critical probabilistic calculate the reliability index using Spencer’s method.
surface associated with the highest probability of failure or the In the present study, a practical procedure that combines a
lowest reliability has been considered [9,1,8,10]. commercial numerical analysis method and artificial neural net-
Although probabilistic slope stability analysis methods do not works into a probabilistic stability analysis of slopes is proposed.
consider all of the components of slope design where judgment An ANN technique is adopted to establish a response surface that
needs to be utilized and do not suggest the level of reliability that represents the functional relationship between soil properties
should be targeted [11], working within a probabilistic framework and the factor of safety. Training and test data sets for the model
does imply that the reliability of the system can be considered in a are obtained from slope stability analyses based on the strength
logical manner. Thus, probabilistic models can facilitate the devel- reduction method using the commercial FDM package FLAC [13].
The ANN model is then connected to the first- and second-order
reliability methods or to a Monte Carlo simulation in an effort to
* Tel.: +82 42 870 7632; fax: +82 42 870 7639.
E-mail address: drsecho@hanmail.net predict the probability of failure.

0266-352X/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compgeo.2009.01.003
788 S.E. Cho / Computers and Geotechnics 36 (2009) 787–797

The main purpose of this study is to verify the application po- FLAC conducts the slope stability analysis with simulations
tential of integrating an ANN technique with a commercial package based on the shear strength reduction method by progressively
to assess the stability of slopes in probabilistic terms, which has reducing the shear strength of the material to bring the slope to
not been attempted previously by other researchers. a state of limiting equilibrium. The factor of safety is defined
according to the equations:
c
2. Slope stability analysis by the strength reduction method ctrial ¼ ð1Þ
F trial
s
!
Slope stability problems are commonly analyzed using limit tan /
/trial ¼ arctan ð2Þ
equilibrium methods of slices. The failing soil mass is divided into F trial
s
a number of vertical slices to calculate the factor of safety, defined
as the ratio of the resisting shear strength to the mobilized shear A series of simulations are made using trial values of the factor
stress to maintain static equilibrium. The static equilibrium of F trial
s to reduce the cohesion, c, and friction angle, u, until slope fail-
the slices and the mass as a whole are used to solve the problem. ure occurs. If the slope is initially unstable, c and u will be in-
However, all methods of slices are statically indeterminate and, creased until the limiting condition is found.
as a result, require assumptions in order to solve the problem. Lim- The factor of safety at which collapse occurs is found by brac-
it equilibrium methods require that the critical failure surface be keting and bisection method [20]. The lower limit of bracket repre-
determined as a part of the analysis and it is well known that many sents the current F trial
s where the simulation converges. The initial
local minima may exist in addition to the global minimum. This upper limit is any F trial
s where the solution does not converge. The
makes it difficult to locate the critical failure surface using classical next value is selected by bisecting this interval (new F trial s ). If the
optimization methods [14]. solution converges, the lower bracket is replaced by the new value.
The strength reduction method (SRM) is typically applied to cal- Otherwise, it replaces the upper bracket. The process is repeated
culate the factor of safety by progressively reducing or increasing until the difference between the upper and lower bracket values
the shear strength of the material to bring the slope to a state of lim- becomes smaller than 0.005 [13].
iting equilibrium [15]. The definition of the factor of safety is iden- To be consistent with the conventional limit equilibrium meth-
tical to that adopted by the conventional limit equilibrium method. ods, the calculations are performed within the framework of the
Zienkiewicz et al. [15] used the SRM for a slope stability analy- elastic–perfectly plastic Mohr–Coulomb model with no hardening
sis; later the method was applied by Griffiths [16], Naylor [17], or softening. To avoid excessive dilation a nonassociated flow rule
Matsui and San [18], Ugai and Leshchinsky [19], Dawson et al. governed by a dilation angle W = 0 is considered. This volume pre-
[20], Griffiths and Lane [21], Hammah et al. [22], and others. Smith serving response has been assumed, even though the dilation angle
and Griffiths [23] first published source code on the strength is known to be less important for the drained modeling of slopes,
reduction method applied to slope stability analysis. as has been carried out in this paper.
The main advantages of the SRM are that it requires no assump-
tion on the inter-slice force distribution. The critical failure surface 3. Probabilistic analysis
is determined automatically from the shear strain arising from the
application of gravity loads and from the reduction of shear 3.1. Limit state function
strength. Furthermore, it is applicable to many complex conditions
and can provide data regarding stresses, and movements, which The problem of the probabilistic analysis is formulated by a vec-
are not obtainable with the limit equilibrium method [14]. tor, X ¼ ½X 1 ; X 2 ; X 3 ; . . . ; X n , representing a set of random variables.
According to Griffiths and Lane [21] and Dawson et al. [20], the From the uncertain variables, a limit state function g(X) is formu-
strength reduction method is usually not sensitive to the dilation lated to describe the limit state in the space of X. In n-dimensional
angle, soil moduli or the solution domain size and the choice of hyperspace of the basic variables, g(X) = 0 is the boundary between
these parameters is not critical in the analysis. However, if a slope the region in which the target factor of safety is not exceeded and
stability analysis involves pore pressure changes, then the dilation the region in which it is exceeded. The probability of failure of the
angle would be very important [24]. slope is then given by the following integral [7]:
One of the main disadvantages of the strength reduction meth- Z
od is the judgment of failure. Generally, nonconvergence is widely Pf ¼ P½gðX 6 0Þ ¼ fX ðXÞ dX ð3Þ
gðXÞ60
used in strength reduction technique to calculate the factor of
safety. However, as pointed out by Krahn [25], the difficulty in where fX(X) represents the joint probability density function, and
using nonconvergence as a failure criterion is that many factors the integral is carried out over the failure domain.
can cause numerical instability, such as incremental load-step size, For slope stability problems, direct evaluation of the n-fold inte-
gravity loading procedures, low confining stresses near the ground gral is virtually impossible. The difficulty lies in that complete prob-
surface, purely frictional materials with no cohesion and initial abilistic information on the soil properties is not available and the
in situ stress conditions. domain of integration is a complicated function. Therefore, approx-
Hammah et al. [22] and Cheng et al. [14] gave a detailed imate techniques have been developed to evaluate this integral.
description of the performance of the SRM and the LEM. The limit state function for the slope stability is usually defined
FLAC uses an explicit, time marching method to solve the gov- as
erning field equations in which every derivative is replaced by an
gðXÞ ¼ F s  1:0 ð4Þ
algebraic expression written in terms of the field variables at dis-
crete points in space. The convergence criterion for FLAC is the no- The strength reduction method is used to describe the above
dal unbalanced force ratio defined as the ratio of the unbalanced limit state function by calculating the factor of safety, Fs.
force, which is the net force acting on a grid point, to the mean
absolute value of force exerted by each surrounding zone. If the 3.2. First- and second-order reliability methods
maximum unbalanced force ratio for all the grid points in the mod-
el is less than the convergence limit set by the user, then the sys- The first-order reliability evaluation of Eq. (3) is accomplished
tem is in equilibrium [13]. by transforming the uncertain variables, X, into uncorrelated stan-
S.E. Cho / Computers and Geotechnics 36 (2009) 787–797 789

dard normal variables, Y. The primary contribution to the probabil- statistical properties evaluated after the process of simulation,
ity integral in Eq. (3) comes from the part of the failure region such as mean, variance, coefficient of skewness, probability density
(G(Y) 6 0, for which G(Y) is the limit state function in the trans- functions, and cumulative probability distribution functions, can
formed normal space) closest to the origin. The design point is de- provide a broader perspective and a more comprehensive descrip-
fined as point Y* in the standard normal space located on the limit tion of a given slope.
state function (G(Y) = 0) with the maximum probability density at- Despite the fact that the mathematical formulation of the
tached to it. Therefore, the design point, which is the nearest point Monte Carlo simulation is relatively simple and the method has
to the origin in the failure region, is an optimum point at which to the capability of handling practically every possible case regardless
approximate the limit state surface. The probability approximated of its complexity, this approach has not received overwhelming
at the design point is acceptance due to the excessive computational effort required
[31]. To improve the computational efficiency of the method, sev-
P½gðXÞ 6 0  UðbÞ ð5Þ
eral sampling techniques known as variance reduction techniques
where b is the reliability index defined by the distance from the ori- have been developed. A detailed review of these can be found in
gin to the design point and U is the standard normal cumulative Baecher and Christian [7]. Among them, Latin hypercube sampling
density function. can be viewed as a stratified sampling scheme designed to ensure
In the first-order reliability method, a tangent hyper plane is fit- that the upper or lower ends of the distributions used in the anal-
ted to the limit state surface at the design point. Therefore, the ysis are well represented.
most important and demanding step in the method is finding the In this study, the Latin hypercube sampling technique is used to
design point. The design point is the solution of the following non- generate the random properties of soil. In particular, the method
linear constrained optimization problem: proposed by Stein [32] for sampling dependent variables based
on the rank of a target multivariate distribution is implemented
min kYk subject to GðYÞ ¼ 0 ð6Þ
in Matlab. In addition, Monte Carlo simulation method with
Several algorithms have been proposed for the solution of this approximated limit state functions based on the ANN model is
problem [26–28]. adopted to reduce computational cost.
As a by-product of the first-order reliability method, it is possi-
ble to evaluate the measures of sensitivity of the reliability index
with respect to the basic random variables. These sensitivity mea- 4. The ANN-based response surface
sures identify the random variables that have the greatest impact
on the failure probability. The ANN technique is an information processing technique
The unit vector a normal to the limit state surface at the design based on the way biological nervous systems, such as the human
point is the most fundamental sensitivity measure that describes brain, process information. This technique has the capability to re-
the sensitivity of the reliability index with respect to variations spond to input stimuli, produce the corresponding response, and
in each of the standard variates. adapt to the changing environment by learning from experience
a ¼ rY  b ð7Þ [33]. As in nature, the network function is determined largely by
the connections between elements. A neural network is trained
As a does not necessarily imply the same sensitivity ordering of to perform a particular function by adjusting the values of the con-
the random variables in the original space for statistically depen- nections (weights) between elements.
dent variables, a more meaningful sensitivity measure is intro- There has been considerable recent research activity in explor-
duced, as in Haukaas and Der Kiuregian [29]: ing the use of ANN in geotechnical engineering. Recent applica-
aT JY ;X D tions include Ghaboussi et al. [34], Goh [35], Deng et al. [36], and
cT ¼ ð8Þ Juang et al. [37,38]. Juang and his co-workers [37,38] have success-
kaT JY ;X Dk
fully used neural networks to emulate the limit state surface for
where J is the Jacobian and D is the diagonal matrix of the standard reliability analysis of liquefaction potential of soils. The popularity
deviation. of neural networks stems from the ability to identify relationships
The second-order reliability method (SORM) aims at enhancing in numerical data through a mapping of the relationship between
the precision of the results obtained by FORM. Instead of replacing various input variables and one or more output variables [39].
the real limit state surface with a hyperplane, a quadratic (second- In order to predict the failure probability of a slope, the geotech-
order) approximation in the vicinity of the design point is em- nical responses must typically be estimated using a numerical pro-
ployed [30]. cedure such as the limit equilibrium method, finite element
For the FORM/SORM analysis, FERUM (an open-source Matlab method, or finite difference method (FDM). This brings another le-
toolbox for reliability analysis, http://www.ce.berkeley.edu/FER- vel of complexity to the reliability analysis because the limit state
UM) developed by Haukaas and Der Kiureghian [29] is combined function is not available as an explicit, closed-form function of the
with the ANN model that specifies the limit state function. input variables. For a reliability analysis with an implicit limit state
function, several computational approaches, including a Monte
3.3. Monte Carlo simulation Carlo simulation, a response surface method, and a sensitivity-
based analysis can be utilized [40].
An alternative means of evaluating the multi-dimensional inte- With the response surface method, a closed-form limit state
gral of Eq. (3) is the use of a Monte Carlo simulation. In a Monte surface is constructed artificially using polynomial regression.
Carlo simulation, discrete values of the component random vari- However, these methods become computationally impractical for
ables are generated in a fashion consistent with their probability problems involving a large number of random variables and non-
distribution, and the limit state function is calculated for each gen- linear limit state functions, particularly when statistically depen-
erated set. The process is repeated many times to evaluate the dent random variables are involved [40,39].
probability of failure by determining whether the limit state func- As an alternative to improve the modeling of the limit state sur-
tions are exceeded. However, the Monte Carlo simulation method face, the use of the back-propagation neural network algorithm is
is not limited to the calculation of the probability of failure. Various proposed to determine the limit state surface. Recent studies have
790 S.E. Cho / Computers and Geotechnics 36 (2009) 787–797

demonstrated the feasibility of this integrated approach [41– Neural networks are iteratively adjusted, or trained, in order to
43,39]. minimize the mean-squared error between the real response sur-
Gomes and Awruch [42] have compared the polynomial-based face and that predicted from the neural network model. Back-prop-
response surface method with ANN-based response surface meth- agation has generally been the most popular method used to train
od in terms of CPU time and the number of limit state function nonlinear, multi-layered neural networks to perform function
evaluation. Elhewy et al. [40] also compared two methods and approximation. The term back-propagation refers to the manner
showed that the ANN-based response surface method is more effi- in which the gradient is computed for nonlinear multi-layer net-
cient and accurate than the conventional response surface method. works. A number of variations on the basic algorithm that are
According to Schueremans and Van Gemert [44], neural networks based on other standard optimization techniques, such as conju-
have the advantage of their flexibility to adapt to more complex gate gradient and Newton methods, have been developed to in-
limit state functions that might not be represented well by means crease the speed of convergence. In the present study, the
of a low order polynomial. Levenberg–Marquardt (LM) algorithm is adopted for its efficiency
It is difficult to develop a global response surface over the entire in training networks. A more detailed description on the ANN mod-
variable domain in many problems, especially in nonlinear prob- el is given in Appendix A.
lems. If a limit state function is complicated and a whole-area fit-
ting is required, a simple first- or second-order polynomial may not 4.2. Data preparation
be adequate. As a developing and promising technology, the capa-
bility of an artificial neural network to cope with uncertainty in ANN model is good at interpolating data but not extrapolating.
complex situations has been seized upon for wide ranging applica- Therefore, to achieve a valid ANN model, the data selected for
tions in recent years [45]. training should cover the full range of input random variables.
The number of sampling points required to accurately model
4.1. ANN architecture and training algorithm the limit state surface is dependent on the number of random vari-
ables and the nonlinearity of the problem considered. However,
A typical neural network structure used in this study is shown there are no precise guidelines or theories for the selection of the
in Fig. 1. As shown in figure, the ANN structure consists of an input sampling points. Several designs have been proposed to estimate
layer, one hidden layer, and an output layer. Each layer has its cor- an accurate low order polynomial response surface, such as a cen-
responding neurons and weight connections. It has been shown tral composite design or a fractional factorial design [48]. The main
that ANNs with one hidden layer can approximate any function gi- drawback of these techniques is the number of limit state evalua-
ven that sufficient degrees of freedom are provided [46]. The input tions required to build a sufficiently accurate response surface. For
represents the random variables X in the input layer. The output a higher number of random variables, the required number of eval-
consists of a single neuron representing the response surface. To uations of the real limit state function grows exponentially. To
determine the number of neurons in the hidden layer is a very meet this drawback, iterative procedures have been established.
important part, which determines the accuracy of ANN models. These alternatives focus on finding a good quality response surface
Unfortunately there is not a rule (or algorithm) to determine this near the region where the limit state function equals 0 or in the
number. It is normally determined by a trial-and-error process, neighborhood of the design point. This region contains the major
so this number will be optimized during the training process. contribution to the failure probability that is sought.
The initial number is picked by guesswork and experience. If the According to Hurtado and Alvarez [43], the sampling method
network has trouble learning, further neurons can be added to for generating the training population has no effect in the learning
the hidden layer and the process is repeated until the performance phase of the ANN-based response surface method as long as the
of the trained model is acceptable [47]. In this study, the logistic resulting sample is reasonably dense. They recommended the gen-
sigmoid transfer function is used to transfer the values of the input eration of a sample set with compact uniform distributions. In this
layer neurons to the hidden layer neurons, whereas the linear paper, the sampling points were selected randomly to fall within
transfer function is adopted to transfer the values from the hidden the range of lX ± 3rX with uniform distributions, which has been
layer to the output layer. successfully used in the previous studies [39,41].
The procedure of the probabilistic slope stability analysis using
Input layer Hidden layer Output layer the ANN-based response surface is presented in Fig. 2.

X1 5. Example analysis

In this section, application of the presented procedure is illus-


X2 trated through example problems. To construct response surface
data sets that relate the sampled random variables to the factor
of safety, a series of analyses of FDM was performed. A FISH (the
Factor of safety built-in programming language of FLAC) function was written to
X3
manipulate input/output for batch processing.

As the strength of the soil is a spatially distributed random var-


iable, soil strength should be modeled as random field considering

spatial correlation [1]. A combination of random field and finite


element (or finite difference) analysis method would enable more
realistic failure mechanisms to seek out the critical path through
Xn the soil. However, the influence of spatial uncertainties is not con-
sidered in this study due to the limitation in incorporating random
Error Propagation field modeling into the response surface method.
A three-layered feed-forward neural network is established to
Fig. 1. Typical structure of an ANN model. approximate the mapping between the factor of safety and the ba-
S.E. Cho / Computers and Geotechnics 36 (2009) 787–797 791

START

Define the Problem


Specify random variables and their
probabilistic characteristics
Define the limit state function

Prepare the data sets to form response surface


Factor of safety corresponding to input

Slope stability analysis


SRM

Construct the response surface


Determine the ANN architecture
Training and testing of the ANN model

Is the ANN model


acceptable? Add neuron in hidden layer
No

Yes

Select the probabilistic


Generate random data sets
analysis method
Latin hypercube sampling

FORM Calculate the limit


SORM Monte Carlo simulation
state function from
the trained ANN

Calculate Statistica l response


Reliability index Probability density
Sensitivity measure Cumulative probability
Probability of failure Probability of failure

Add No
Is the accuracy of
training Pf acceptable?
data set

Yes
END

Fig. 2. The procedure of the probabilistic slope stability analysis using the ANN-based response surface.

sic random variables. The number of hidden neurons is selected by the dimensionless coefficient of variation (COV), defined as
trial and error. VX = rX/lX, the mean and standard deviation of the underlying nor-
Random variables are assumed to be characterized statistically mal distribution of ln X are then given by
by a normal or lognormal distribution defined by a mean lX and a
standard deviation rX. The lognormal distribution ranges between qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
zero and infinity, skewed to the low range, and is therefore partic- rln X ¼ lnf1 þ V 2X g ð9Þ
ularly suited for parameters that cannot take on negative values.
lln X ¼ ln lX  0:5r2ln X ð10Þ
Once the mean and standard deviation are expressed in terms of
792 S.E. Cho / Computers and Geotechnics 36 (2009) 787–797

15m

2
c1
1 9.14m

6.1m
c2, φ2

15m 33.28m

Fig. 3. Example 1: cross-section [8].

Table 1 sented in Fig. 5a. The arrows in Fig. 5 are velocity vectors, which
Example 1: statistical properties of soil parameters (based on Hassan and Wolff [8]). indicate the pattern of motion at the initiation of failure and show
lX COV a well-defined failure surface.
The neural network was composed of three input parameters
c1 (kPa) 38.31 0.2
c2 (kPa) 23.94 0.2 (c1, c2 and u2) in the input layer of three neurons, a hidden layer
u2 (°) 12 0.1 of three neurons and an output layer of one neuron.
The ANN model was trained by increasing the number of train-
ing data sets, starting with 15 data sets. Fig. 4 compares the predic-
5.1. Example 1: application to a two-layered slope tions of the final ANN model with the finite difference analyses for
the training and the test. As shown, a good correlation can be ob-
In this example, a series of probabilistic studies are performed served between the predictions of the ANN model and the finite
on a two-layered slope with a cross-section, as shown in Fig. 3. This difference analysis results. This suggests that the model can be
example is taken from Hassan and Wolff [8]. The basic soil param- used to represent a response surface with small deviations. In this
eters related to the stability of the slope, including the friction an- study, the model is used for the subsequent probabilistic analyses.
gle and cohesion, are considered as random variables. The unit Table 2 presents the results obtained from the probabilistic
weight of the slope is assumed as 19 kN/m3. Table 1 summarizes analyses, i.e., the sensitivities, the reliability index and the proba-
the statistical properties of soil parameters for the slope, and the bility of failure. According to the results obtained from the Monte
analysis is performed for two cases assuming that all random vari- Carlo simulation, with greater than 20 data sets the ANN model
ables are characterized by either a normal or lognormal seems able to train adequately and leads to an acceptable accuracy.
distribution. The sensitivities show the relative importance of the uncer-
There are 1236 zones and 1647 grid points in the mesh for anal- tainty in each random variable. Therefore, sensitivity analysis can
ysis. The forces generated by the self-weight of the soil are applied. be used to identify random variables whose uncertainty has insig-
Movement is allowed vertically on both lateral boundaries, while nificant influence on the reliability index and which can be re-
the constraint is imposed in both directions of the base boundary. placed by deterministic values. The plot of the shear strain rate
The minimum factor of safety calculated from a deterministic corresponding to the design point obtained from FORM shows that
analysis based on the mean values of the soil properties is 1.60, the probabilistic failure surface passes through the upper layer
which is slightly lower than 1.646 from the finite element method where the variation of the shear strength is larger than that of
and 1.634 from the Spencer’s noncircular slip surface in Xu and the lower layer; this surface is associated with the maximum prob-
Low [12]. The plot of the shear strain rate shows that the failure ability of failure (Fig. 5b). The reliability index calculated by FORM
surface passes through both the upper and lower soil layer, as pre- is 2.189 for the case of a normal distribution and 2.812 for a log-

a 2.5 b 2.5

2 2
Simulation result(FS)

Simulation result(FS)

1.5 1.5

1 1

0.5 0.5
0.5 1 1.5 2 2.5 0.5 1 1.5 2 2.5
Target result(FS) Target result(FS)

Fig. 4. Example 1: comparison between the ANN model and stability analysis using: (a) training data and (b) test data.
S.E. Cho / Computers and Geotechnics 36 (2009) 787–797 793

Table 3
Example 1: design point in original space.

Input parameter lX  3rX Design point, X* lX + 3rX


c1 (kPa) 15.324 21.55 61.296
c2 (kPa) 9.576 24.32 38.304
/2 (°) 8.4 12.05 15.6

X*is the obtained design point in original space X.

The reliability index obtained from Xu and Low [12] by the re-
sponse surface based on the finite element method assuming a
normal distribution of all random variables was 2.18. The result
is in good agreement with that obtained using the proposed meth-
od. However, as stated in their paper, they only considered prob-
lems related to a linear limit state function. That makes it
possible to use only 2n + 1 sampling points to construct a response
surface since two points per one random variable can characterize
the linear relationship. For each random variable, its mean value, l,
and two other values, l ± r, are sampled. Consequently, according
Fig. 5. Example 1: results of the stability analysis: (a) failure mode at the mean
point (Fs = 1.60) and (b) failure mode at the design point (Fs = 1.0). to their results, the design point falls outside the range of l ± r,
which implies that the function approximation is based on the
extrapolation method [12]. Due to these limitations, they only con-
sidered small values of the coefficients of variation related to soil
normal distribution of the soil properties. Table 3 shows that the properties. On the contrary, in this study, the design point falls
design point falls within the range of data points sampled by the within the range of sampled data points as shown in Table 3 since
Latin hypercube sampling technique. the used ANN model can construct a global response surface over
A Monte Carlo simulation was performed using a sample of the entire variable domain even though the limit state function is
200,000 data sets based on the statistical information. Fig. 6 pre- complicated.
sents the results calculated from the Monte Carlo simulation. Rel-
atively good agreement with those obtained from FORM and SORM 5.2. Example 2: application to the Cannon Dam
is noted (Table 2).
Fig. 6a shows that the histogram of the factor of safety for a nor- This example concerns the stability for the end-of-construction
mal distribution of random variables has a negative coefficient of stage of the Cannon Dam reported in Hassan and Wolff [8]. A typ-
skewness. The histogram of the factor of safety for a lognormal dis- ical slope section showing the soil profile is presented in Fig. 7. The
tribution of soil properties shows a nearly symmetrical distribution structure consists of two zones of compacted clay: Phase I fill and
(Fig. 6b). Fig. 6c shows the probability distribution of the factor of Phase II fill over layers of sand and limestone. Strength parameters
safety. As expected, the failure probability obtained from the log- of the two clay layers were considered as random variables (c1, u1,
normal distribution of random variables is smaller compared to c2, u2). Statistical moments for these parameters based on UU tests
that obtained from the normal distribution of random variables. of recorded samples from the embankment are shown in Table 5.
Table 4 presents a summary of other probabilistic analyses ob- They show somewhat large variations related to soil properties.
tained from the literature for the same problem. The reliability in- Hassan and Wolff [8] made no reductions in the variance for spatial
dex determined in this study under the assumption of a lognormal correlation; in their study it was conservatively assumed that var-
distribution of soil properties, 2.812, is very close to those deter- iance over a failure zone was potentially as large as the point var-
mined by the MVFOSM (Mean-Value First-Order Second-Moment) iance. As their paper does not provide all the required input
method that assumes a lognormal distribution of the factor of parameters, values for the missing parameters are assumed from
safety. This is because the assumption of a lognormal distribution those selected in other literature [50]. Lognormal distribution
of the factor of safety by the MVFOSM method is fairly reasonable was assumed for all probabilistic input variables in performing
in this case as shown in Fig. 6b. Note that when the soil properties probabilistic analyses in order to avoid negative values.
are normally distributed the assumption of a lognormal distribu- The domain of analysis consists of 1407 zones and 1944 grid
tion of the factor of safety is not correct, as shown in Fig. 6a for this points. The boundary conditions consist of roller boundaries on
example. In conclusion, the assumption that the factor of safety is the left and right sides of the model as well as a fixed base. The
lognormally distributed is not always valid because the MVFOSM gravity loads generated by the self-weight of the soil are applied.
method does not account for the influence of the distributional The factor of safety associated with the mean values of soil
shape of the soil properties. properties is 2.49. Fig. 8a shows the shear strain rate and velocity

Table 2
Example 1: results of the probabilistic analysis.

Number of training data sets Sensitivity (FORM) Reliability index, b Failure probability, Pf
cc1 cc2 c/2 FORM SORM FORM SORM Monte Carlo
Normal 15 0.9974 0.0672 0.0262 2.143 2.152 1.61e2 1.57e2 1.56e2
20 1.0000 0.0004 0.0005 2.167 2.209 1.51e2 1.36e2 1.37e2
25 0.9991 0.0369 0.0219 2.189 2.204 1.43e2 1.38e2 1.38e2
Lognormal 15 0.9940 0.1010 0.0412 2.704 2.733 3.43e3 3.13e3 3.1e3
20 0.9999 0.0102 0.0042 2.770 2.865 2.85e3 2.08e3 2.3e3
25 0.9997 0.0193 0.0114 2.812 2.835 2.5e3 2.3e3 2.3e3
794 S.E. Cho / Computers and Geotechnics 36 (2009) 787–797

a Lognormal(1.570,0.214)
Table 4
Example 1: results of probabilistic analyses from different methods.
μ=1.570, σ=0.214, skew=-0.566
2.5 Method Reliability Failure
index, b probability,
2.0 Pf
Probability density

Hassan and Wolff [8] – MVFOSM, Spencer method, 2.869 –


1.5 noncircular slip surface
Bhattacharya et al. [9] – MVFOSM, Spencer method, 2.861 –
noncircular slip surface by direct search
1.0
Crum [49] 2.23 1.29e2
Xu and Low [12] – response surface, FEM (PLAXIS) 2.18 1.46e2
0.5 Xu and Low [12] – response surface, Spencer 2.06 1.97e2
method, noncircular slip surface
0.0 Xu and Low [12] – spreadsheet-based direct 2.20 1.39e2
implementation of Spencer’s method with search
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3
Factor of Safety for the critical circular slip surface

b Lognormal(1.572,0.194)
μ=1.572, σ=0.194, skew=-0.022 vectors. The shear strain rate contours delineate that the location
2.5
of the failure surface passes through the foundation sand. The fac-
tor of safety obtained in the present analysis is slightly lower com-
2.0
pared to both the value of 2.647 reported by Hassan and Wolff [8]
Probability density

and that of 2.612 by Bhattacharya et al. [9] based on Spencer’s


1.5
method for the noncircular surfaces. This can be attributed to the
error from scaling the dimensions of the dam from the cross-sec-
1.0
tion given in the paper by Hassan and Wolff [8], to the assumed
material parameters, or to the dissimilar analysis methods of the
0.5
LEM and the SRM, among other factors.
0.0
When a large number of variables are considered and the vari-
ations of the variables are large the number of sampling points re-
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3
Factor of Safety quired to accurately construct the response surface increases. As
the number of training data sets to be used is not known in ad-
c 1 vance, the learning process initiates with a relatively small number
of training data sets and gradually increases the number of training
data sets, until the desired convergence is achieved.
0.8 Given that the variation in soil properties is large, the ANN
Cumulative probability density

model was trained by increasing the number of training data sets


from the initial 50 sets. Table 6 presents the results obtained from
0.6 probabilistic analyses after this increase. From the table, it can be
seen that increasing the number of training data sets makes a dif-
ference in the failure probability. Artificial neural networks need
0.4 properly chosen training sets and using the increased training data
sets improves the accuracy of approximation. However, consider-
ing the required computational time to develop the training data
0.2 sets, using a large number of training data sets may result in little
normal improvement in accuracy and efficiency. Actually, from the Table 6,
lognormal
it can be seen that the increase in the number of training data sets
0 from 150 to 200 only causes a slight difference in the failure prob-
0 1 2 3 ability. Fig. 9 presents the results obtained from the Monte Carlo
Factor of safety
simulation.
Fig. 6. Example 1: results of Monte Carlo simulation: (a) histogram and lognormal The reliability index determined from FORM is 3.067. The crit-
fit for the calculated factor of safety when a normal distribution of random variables ical probabilistic surface determined by FORM passes through the
is assumed; (b) histogram and lognormal fit for the calculated factor of safety when Phase I fill, as shown in Fig. 8b. An obvious noncircular slip sur-
a lognormal distribution of random variables is assumed and (c) probability
distribution of the calculated factor of safety.
face can be identified from the figure. Therefore, a stability anal-

199.34
3.5 3 Downstream
1 1
186.23
176.17
169.16

166.12

156.97

140.21

Fig. 7. Example 2: cross-section (Cannon Dam, Hassan and Wolff [8], length unit: m).
S.E. Cho / Computers and Geotechnics 36 (2009) 787–797 795

Table 5
Example 2: statistical properties of soil parameters (Cannon Dam, based on Hassan a Lognormal(2.352,0.424)
and Wolff [8]). μ=2.352, σ=0.424, skew=-0.31
1.2
Material Parameter lX COV Correlation coefficient

Probability density
Phase I clay c1 (kPa) 117.79 0.5 +0.1
u1 (°) 8.5 1 0.8
c1 (kN/m3) 22 –
Phase II clay c2 (kPa) 143.64 0.55 0.55
u2 (°) 15 0.6
c2 (kN/m3) 22 – 0.4

Sand filter u (°) 35 –


c (kN/m3) 22 –
Foundation sand c (kPa) 5 – 0.0
u (°) 18 – 0.4 0.8 1.2 1.6 2 2.4 2.8 3.2 3.6 4
c (kN/m3) 20 – Factor of Safety
Fill c (kPa) 5 –
u (°) 35 – b Lognormal(2.352,0.424)
c (kN/m3) 25 – μ=2.352, σ=0.424, skew=-0.31
1

Cumulative probability density


0.8

0.6

0.4

0.2

0
0 1 2 3 4 5
Factor of safety

Fig. 9. Example 2: results of Monte Carlo simulation for a lognormal distribution of


random variables: (a) histogram and lognormal fit for the calculated factor of safety
and (b) probability distribution of the calculated factor of safety.
Fig. 8. Example 2: results of the stability analysis: (a) failure mode at the mean
point (Fs = 2.49) and (b) failure mode at the design point (Fs = 1.03).
Table 7
Example 2: design point in original space.
ysis method that can capture the failure mechanism related to
Input parameter lX  3rX (P0) Design point, X* lX + 3rX
the noncircular slip surface is required to obtain an accurate reli-
c1 (kPa) 0 31.04 294.48
ability index for a layered slope. According to the sensitivities
/1 (°) 0 1.25 34
from FORM, the uncertainty in the cohesion of the Phase I fill is c2 (kPa) 0 99.14 380.65
the most important source of uncertainty, and the friction angle /2 (°) 0 14.25 42
of the Phase I fill is the second significant variable. Table 7 shows
X*is the obtained design point in original space X.
the design point falls within the range of sampled training data
sets (0 6 X 6 lX + 3rX).
The results obtained from the Monte Carlo simulation are in rel- find the minimum reliability index associated with the critical
atively good agreement with those obtained from SORM, but differ probabilistic surface. In their analysis, the reliability index corre-
somewhat from those obtained from FORM (Table 6). This can be sponding to the searched critical probabilistic surface was calcu-
attributed to the assumption of a tangent hyperplane in the first- lated using the MVFOSM method assuming a lognormal
order reliability method. distribution for the factor of safety. According to Hassan and
The example was previously solved by Hassan and Wolff [8] Wolff [8], the minimum reliability index was 2.664 from Spen-
and Bhattacharya et al. [9] to illustrate their search algorithm to cer’s noncircular slip surface. Bhattacharya et al. [9] also obtained

Table 6
Example 2: results of the probabilistic analysis.

Number of training data sets Sensitivity (FORM) Reliability index, b Failure probability, Pf
cc1 c/1 cc2 c/2 FORM SORM FORM SORM Monte Carlo
50 0.8257 0.5147 0.1901 0.1311 2.910 3.166 1.81e3 7.72e4 1.00e3
150 0.7416 0.4554 0.4094 0.2740 2.957 3.199 1.55e3 6.91e4 6.05e4
200 0.8092 0.5354 0.2230 0.0938 3.067 3.219 1.08e3 6.43e4 7.00e4
796 S.E. Cho / Computers and Geotechnics 36 (2009) 787–797

an index of 2.674 from Spencer’s noncircular slip surface using 2ðp  min pÞ
pn ¼ 1 ðA1Þ
the direct search method for a minimum reliability index. To ðmax p  min pÞ
make a comparison between the MVFOSM method and the pro- 2ðt  min tÞ
posed method is not appropriate, since the MVFOSM method only tn ¼ 1 ðA2Þ
ðmax t  min tÞ
requires the mean value and standard deviation of the random
t ¼ 0:5ðtn þ 1Þðmax t  min tÞ þ min t ðA3Þ
variables without incorporating information on distributional
shape of soil properties, and hence does not provide any informa- where, p is the input parameter, t is the output parameter, pn is the
tion for determining the shape of the probability density function normalized input parameter, tn is the normalized output parameter,
of the factor of safety. For this example, however, it seems clear min p is the minimum values of input parameter, max p is the max-
that the obtained distribution of the factor of safety does not imum values of input parameter, min t is the minimum values of
agree with the assumption of the MVFOSM method, a lognormal output parameter, and max t is the maximum values of output
distribution of the factor of safety, as shown in Fig. 9a and b. This parameter.
result occurred partly because the proposed method does not as-
sume a particular failure mechanism; rather, the failure mecha- A.1. Example 1
nism corresponding to the input soil properties is determined A three-layer network, with a tan–sigmoid transfer function in
automatically during the process of analysis. the hidden layer and a linear transfer function in the output layer,
was used. This is a useful structure for function approximation
problems. Three neurons were used in the hidden layer. The net-
6. Conclusions work has one output neuron, because there is only one target value
associated with each input vector. Then, the normalized output can
This paper considered slope stability problems with uncertain be expressed in a matrix form as:
quantities through a practical procedure that combined a commer-
tn ¼ purelin ½C13  tan sigðA33  pn31 þ B31 Þ þ D11  ðA4Þ
cial numerical analysis code and artificial neural networks into the
probabilistic analysis of slope stability. An ANN technique was where A = weight matrix between the input layer and the hidden
adopted to establish a model for the approximation of the limit layer, B = bias vector at the hidden layer, C = weight matrix between
state function. Training and test data sets for the model were ob- the hidden layer and the output layer, D = bias at the output layer.
tained from numerical calculations based on the strength reduc- The used transfer functions are denoted as follows:
tion method. The adoption of the strength reduction method to
2
the probabilistic stability analysis makes it possible to omit the tan sigðXÞ ¼  1; purelinðXÞ ¼ X ðA5Þ
½1 þ expð2XÞ
process of determining the probabilistic critical surface where
the maximum probability of failure is calculated, compared to The input vector p is given as follows:
the use of the limit equilibrium method. The ANN model was then 2 3
c1
connected to a reliability method, in this case the first- and second- 6 7
order reliability method and the Monte Carlo simulation method, p ¼ 4 c2 5 ðA6Þ
to predict the failure probability. /2
Two illustrative examples are presented to validate the appli-
The connection weights and biases for the neural network mod-
cability of the proposed procedure for slope stability problems.
el trained from 25 data sets are exhibited in Table A1.
They provide insight regarding the deterministic and probabilis-
tic slope stability analysis through comparison with other prob-
A.2. Example 2
abilistic analyses obtained from the literature. The obtained
A 4-14-1 structure of the neural network was used. The struc-
results also show that the ANN-based response surface can be
ture has four neurons in the input layer, 14 neurons in the hidden
successfully applied to the problem of probabilistic slope
layers, and one neuron in the output layer. Then, the normalized
stability.
output can be expressed in a matrix form as
Although the finite difference program was used to calculate
the factor of safety of the slope in this study, the proposed method tn ¼ purelin ½C114  log sigðA144  pn41 þ B141 Þ þ D11  ðA7Þ
is independent of the stability analysis and treats the finite ele- 1
ment (or finite difference) model as a black box, as in the case of log sigðXÞ ¼ ðA8Þ
1 þ expðXÞ
a commercial code. Thus any available analysis method can be
used to apply the proposed procedure. The input vector p is given as follows:
2 3
c1
6/ 7
Appendix A 6 17
p¼6 7 ðA9Þ
4 c2 5
Detailed description on the ANN model /2

For a better network performance, the input–output data pairs


are subjected to a scaling process before being used in the network
operation. This is because the compiled raw training data for differ-
Table A1
ent parameters can vary significantly in their actual values. When
Example 1: connection weights and biases for the trained ANN model.
such data are directly used in the training procedure, the network
could exhibit ill-conditioning. The problem can be avoided by a A 1.9118 0.9993 0.3722 B 1.3526
0.0738 0.2189 0.1314 0.0780
proper scaling on the raw data patterns, such that the data are nor-
1.1233 0.0398 0.4827 0.5414
malized to fall within a prescribed bound of (1, 1) using the fol-
C 0.5224 2.1280 0.2529 D 0.2834
lowing equations. The output of the network is subjected to an
inverse scaling to return the normal quantities of the output max t = 2.26, min t = 0.72, max p = [58.74, 36.37, 15.5]T and min p = [15.39,
parameters by Eq. (A3). 10.61, 8.61]T.
S.E. Cho / Computers and Geotechnics 36 (2009) 787–797 797

Table A2
Example 2: connection weights and biases for the trained ANN model.

A 7.1406 6.4321 0.0683 0.1647 B 6.1401


0.1627 0.1467 0.6533 0.4085 14.4936
1.0894 2.0418 5.7563 4.4106 6.6421
1.1011 1.6911 2.8505 1.7217 4.4057
3.8530 6.1375 1.2301 2.0092 7.5621
3.5440 2.4417 8.2401 3.4295 9.8363
0.3794 0.5685 7.1900 3.5831 8.4305
4.4435 3.8604 2.3725 1.3499 8.9199
0.5697 1.1841 22.8579 0.9727 4.4379
1.2740 1.6343 0.4956 0.3193 1.4057
5.1149 3.1564 8.9668 3.9203 10.6560
3.8453 6.1113 1.0558 2.0885 1.2286
4.2537 2.7642 8.4848 3.6224 3.7494
1.5985 1.7782 0.2814 0.1661 4.8290
C C1–7 0.1484 4.9170 0.1848 0.2941 0.6972 3.5347 0.6623 D 0.2081
C8–14 0.4859 1.1904 1.9942 2.8008 0.7063 6.2326 9.7160

max t = 3.72, min t = 0.186, max p = [293.4, 33.78, 378.85, 41.98]T and min p = [1.387, 0.086, 0.564, 0.219]T.

The final connection weights and biases for the ANN model [25] Krahn J. The 2001 R.M. Hardy lecture: the limits of limit equilibrium analyses.
Can Geotech J 2003;40(3):643–60.
trained from 200 data sets are shown in Table A2.
[26] Hasofer AM, Lind NC. Exact and invariant second-moment code format. J Eng
Mech – ASCE 1974;100(1):111–21.
References [27] Rackwitz R, Fiessler B. Structural reliability under combined load sequences.
Comput Struct 1978;9(5):489–94.
[1] Li KS, Lumb P. Probabilistic design of slopes. Can Geotech J 1987;24(4):520–35. [28] Liu PL, Der Kiureghian A. Optimization algorithms for structural reliability.
[2] Alonso EE. Risk analysis of slopes and its application to slopes in Canadian Struct Safety 1990;9(3):161–77.
sensitive clays. Géotechnique 1976;26(3):453–72. [29] Haukaas T, Der Kiuregian A. A computer program for nonlinear finite element
[3] Christian JT, Ladd CC, Baecher GB. Reliability applied to slope stability analysis. analysis. In: Proceedings of the eighth international conference on structural
J Geotech Eng – ASCE 1994;120(12):2180–207. safety and reliability. Newport Beach (CA); 2001.
[4] Griffiths DV, Fenton GA. Probabilistic slope stability analysis by finite [30] Der Kiureghian A, Lin HZ, Hwang SJ. Second order reliability approximations. J
elements. J Geotech Geoenviron Eng 2004;130(5):507–18. Eng Mech – ASCE 1987;113(8):1208–25.
[5] Mostyn GR, Li KS. Probabilistic slope stability-state of play. In: Li KS, Lo S-CR, [31] Papadrakis M, Papadopoulos V, Lagaros N. Structural reliability analysis of
editors. Conference on probabilistic methods in geotechnical engineering. The elastic–plastic structures using neural networks and Monte Carlo simulation.
Netherlands: Balkema, Rotterdam; 1993. p. 89–110. Comput Methods Appl Mechan Eng 1996;136(1):145–63.
[6] El-Ramly H, Morgenstern NR, Cruden DM. Probabilistic slope stability analysis [32] Stein ML. Large sample properties of simulations using Latin Hypercube
for practice. Can Geotech J 2002;39(3):665–83. sampling. Technometrics 1987;29:143–51.
[7] Baecher GB, Christian JT. Reliability and statistics in geotechnical [33] Cheng J, Zhang J, Cai CS, Xiao RC. A new approach for solving inverse reliability
engineering. John Wiley & Sons; 2003. problems with implicit response functions. Eng Struct 2007;29(1):71–9.
[8] Hassan AM, Wolff TF. Search algorithm for minimum reliability index of earth [34] Ghaboussi J, Garrett Jr JH, Wu X. Knowledge-based modeling of material
slopes. J Geotech Geoenviron Eng 1999;125(4):301–8. behavior with neural networks. J Eng Mech – ASCE 1991;117(1):132–53.
[9] Bhattacharya G, Jana D, Ojha S, Chakraborty S. Direct search for minimum [35] Goh ATC. Seismic liquefaction potential assessed by neural networks. J
reliability index of earth slopes. Comput Geotech 2003;30(6):455–62. Geotech Eng – ASCE 1994;120(9):1467–80.
[10] Li KS, Cheung RWM. Discussion of ‘‘search algorithm for minimum reliability [36] Deng J, Yue ZQ, Tham LG, Zhu HH. Pillar design by combining finite
index of earth slopes”. J Geotech Geoenviron Eng 2001;127(2):197–8. element methods, neural networks and reliability: a case study of the Feng
[11] D’Andrea R. Discussion of ‘‘search algorithm for minimum reliability index of Huangshan copper mine, China. Int J Rock Mech Mining Sci 2003;40(4):
earth slopes”. J Geotech Geoenviron Eng 2001;127:195–7. 585–99.
[12] Xu B, Low BK. Probabilistic stability analyses of embankments based on finite [37] Juang CH, Chen CJ, Rosowsky DV, Tang WH. CPT-based liquefaction analysis,
element method. J Geotech Geoenviron Eng 2006;132(11):1444–54. part 2: reliability for design. Géotechnique 2000;50(5):593–9.
[13] ITASCA Consulting Group Inc. FLAC fast Lagrangian analysis of continua, [38] Juang CH, Chen CJ, Tien YM. Appraising CPT-based liquefaction resistance
Minneapolis, Minnesota (USA); 2002. evaluation methods-artificial neural network approach. Can Geotech J
[14] Cheng YM, Lansivaara L, Wei WB. Two-dimensional slope stability analysis by 1999;36(3):443–54.
limit equilibrium and strength reduction methods. Comput Geotech [39] Goh ATC, Kulhawy FH. Reliability assessment of serviceability performance of
2007;34(3):137–50. braced retaining walls using a neural network approach. Int J Numer Anal
[15] Zienkiewicz OC, Humpheson C, Lewis RW. Associated and nonassociated visco- Methods Geomech 2005;29(6):627–42.
plasticity and plasticity in soil mechanics. Géotechnique 1975;25(4):671–89. [40] Elhewy AH, Mesbahi E, Pu Y. Reliability analysis of structures using neural
[16] Griffiths DV. Finite element analyses of walls, footings and slopes. In: network method. Probabilist Eng Mech 2006;21(1):44–53.
Randolph MF, editor. Symposium on computer applications to geotechnical [41] Deng J, Gu D, Li X, Yue Z. Structural reliability analysis for implicit performance
problems in highway engineering. Cambridge (UK): Pub. PM Geotechnical functions using artificial neural network. Struct Safety 2005;27(1):25–48.
Analysts Ltd.; 1980. p. 122–46. [42] Gomes HM, Awruch AM. Comparison of response surface and neural network
[17] Naylor DJ. Finite elements and slope stability. In: Martins JB, editor. with other methods for structural reliability analysis. Struct Safety
Proceedings of the NATO advanced study institute, numerical methods in 2004;26(1):49–67.
geomechanics. D. Reidel Publishing Company; 1982. p. 229–44. [43] Hurtado JE, Alvarez DA. Neural network-based reliability analysis: a
[18] Matsui T, San KC. Finite element slope stability analysis by shear strength comparative study. Comput Methods Appl Mech Eng 2001;191(1–2):113–32.
reduction technique. Soils Found 1992;32(1):59–70. [44] Schueremans L, Van Gemert D. Benefit of splines and neural networks in
[19] Ugai K, Leshchinsky D. Three-dimensional limit equilibrium and finite element simulation based structural reliability analysis. Struct Safety 2005;27(3):
analysis: a comparison of results. Soils Found 1995;35(4):1–7. 246–61.
[20] Dawson EM, Roth WH, Drescher A. Slope stability analysis by strength [45] Chau KW. Reliability and performance-based design by artificial neural
reduction. Géotechnique 1999;49(6):835–40. network. Adv Eng Software 2007;38(3):145–9.
[21] Griffiths DV, Lane PA. Slope stability analysis by finite elements. Géotechnique [46] Hornik K, Stinchcombe M, White H. Multi-layer feed-forward networks are
1999;49(3):387–403. universal approximators. Neural Networks 1989;2(5):359–68.
[22] Hammah RE, Yacoub TE, Corkum B, Curran JH. A Comparison of finite element [47] Heaton JT. Introduction to neural networks with Java. Heaton Research, Inc.;
slope stability analysis with conventional limit-equilibrium investigation. In: 2005.
Proceedings of the 58th Canadian geotechnical and 6th joint IAH-CNC and CGS [48] Haldar A, Mahadevan S. Reliability assessment using stochastic finite element
groundwater specialty conferences saskatoon, Saskatchewan, Canada, analysis. John Wiley & Sons; 2000.
September; 2005. [49] Crum DA. Discussion of ‘‘search algorithm for minimum reliability index earth
[23] Smith IM, Griffiths DV. Programming the finite element method. 2nd ed. John slopes”. J Geotech Geoenviron Eng 2001;127:194–5.
Wiley & Sons; 1982. [50] Rocscience Inc., Slide 2D limit equilibrium slope stability for soil and rock
[24] Hicks MA, Onisiphorou C. Stochastic evaluation of static liquefaction in a slopes. Verification Manual; 2006.
predominantly dilative sand fill. Géotechnique 2005;55(2):123–33.

You might also like