Probabilistic Methods in Geotechnical Engineering, Li & Lo (eds) © 1993 Taylor & Francis.

ISBN 90 54103035

Probabilistic slope analysis - State-of-play

G.R.Mostyn
School ofCivil Engineering, University ofNew South Wales, Kensington, N.S.W, Australia
K.S.Li
University College, University ofNew South Wales & Australian Defence Force Academy, A.c.T., Australia

ABSTRACT: This paper sets out the current "state-of-play" for probabilistic slope analysis as seen by
the authors. In many aspects it does present a state of the art but advances are so rapid in this field that
it would be presumptuous for the authors to try to present "the" state of the art. The advances that the
authors find most pressing are modelling of many soil and rock properties as random fields to take
account of spatial correlation. The vast majority of existing models do not do this and the reliabilities
determined by them may be in error by several orders of magnitude. The paper also includes brief details
regarding the biases present in dealing with data on rock defects, an overview of factors that affect the
reliability and concerns regarding probabilistic modelling of progressive failure.

I INTRODUCTION manner by the choice of material parameters, etc.

In mining engineering and large civil projects
Most slope analysis and design is based on the cost of a "no failures" approach is generally
"deterministic" approaches i.e. a set of design not able to be carried by the project. Thus
parameters are adopted and the loads and designers accept (either knowingly or
resistances are calculated based on these. A factor unknowingly) some risk of failure.
of safety (FOS) is then determined as the available
resistance divided by the applied load.
Usually the factor of safety is selected in view 2 SOURCES OF UNCERTAINTY
of the understanding and knowledge of:
* the material parameters, In slope engineering, as in most areas of activity,
* the problem geometry, and uncertainty may be a result of:
* the consequences of failure.
* Scatter of the data about its mean - which can
Thus there is no single appropriate factor of
be divided into two components. Firstly the
safety, for identical data set~ different designers
variation obtained from measurement of a property
will determine quite different factors of safety.
at a point repeatedly, this is measurement noise
In civil engineering factors of safety are based and is often reduced by increasing the number of
on experience and aimed at giving very reliable measurements. Secondly there is the variation of
structures. In fact failures are usually due to a property from one point to another, that is
mistakes, lack of knowledge, poor detailing or spatial variation. We will discuss this source of
construction and are generally not due to known variability in some detail later.
or "understood" variability of the loads and * Systematic error, this results in the mean of the
resistances. In civil engineering analysis has been measured property varying from the mean of the
based on trying to ensure "no failures" and, in desired property. Thus if we were measuring
general, this has been what society wants and lengths with a tape that had a section removed, we
consequently pays for. The resulting project is would end up with measurements with a
often very expensive. systematic error. There are several sources of
Notwithstanding the above, probabilistic systematic error. As we will see later many biases
methods and reliability analyses are gradually (systematic errors) can creep into measurements of
being accepted in civil engineering because of defect lengths, orientations, etc. Further the actual
economic and theoretical pressures. They have analysis undertaken (e.g. the analytical method
always been taken into account in a de facto adopted) may include systematic errors - this will

89
also be discussed later. friction, f! (i.e. tan <1» are usually taken to be
* Human error, this results when incorrect normally distributed.
mechanisms are modelled or failure occurs due to
mistakes in correctly determining the chosen
model. In many civil engineering applications this 3.1.3 Defects
may be the most significant source of uncertainty.
In slope design this source will often be For rock defects the situation is much more
overshadowed, but not eliminated, by natural complicated as the strength depends on both
uncertainty. surface details and many geometric parameters.
* Finally there is uncertainty introduced by the Nevertheless a Mohr-Coulomb failure criterion is
unknown (or unmodelled) existence of often adopted to model the effects of both the
geotechnical anomalies, such as thin weak layers, basic shear strength and the geometric parameters
faults, etc, which may have a significant effect on with both c and <I> generally being considered as
the reliability. normally distributed.
Many researchers have adopted Barton's (1973)
failure criterion for rock defects:
3 VARIABILITY IN SLOPE ENGINEERING
't = Ci'n tan [JRC 10glO(JCS/Ci'n)+<I>'r] (1)
The variability of the following parameters
generally have to be accounted for in slope Probabilistic treatment in this case has generally
engineering: involved treating <1>'" JRC and JCS as normally
* Materials strength distributed.
* Geometry of surface details and subsurface
boundaries
* Defect properties 3.1.4 Mass
* Pore pressures.

Each will be dealt with in tum.

There are very few models for rock mass strength,
the most commonly adopted (e.g. Priest & Brown,
1983) is the Hoek-Brown empirical failure
3.1 Materials Strength criterion (1980), which has three parameters Cie , m
and s. Treatment of Cie is generally in accordance
3.1.1 Soil with that for intact rock. Parameters m and s are
not directly measurable and thus are often
Where soil materials are being modelled it is most estimated from, say, the RQD (i.e. Rock Quality
common to adopt a Mohr-Coulomb failure Designation). The variability of the relationship
criterion, thus the random variables modelled are between these parameters is unknown but as RQD
the cohesion, c, and either the angle of friction, <1>, is a relatively blunt estimator of much rock
or coefficient of friction, f! or tan <1>. It seems to behaviour any such relationship must be highly
be generally accepted that these can be sensibly variable. To the authors' knowledge this
modelled as either normally or, maybe more variability is not included in any of the analyses
realistically, as beta distributed variates. so far undertaken but it is likely to dominate
resultant slope reliability.
In summary, c, f! and Cie are generally
3.l.2 Rock substance considered normally distributed but the variance of
each may often be chosen over a wide range. The
Generally the intact strength of rock does not resulting probability of failure is very sensitive to
contribute greatly to the uncertainty regarding the the variance and thus different analysts may obtain
stability of a slope. But where it does, it is very different probabilities of failure for the same
usually the unconfined compressive strength, Cie , design and data. Advances are being made in the
and not the frictional component, <1>, that is resolution of this problem but are not yet
modelled as a random variate. commonly adopted. At present the choice of the
Often Cie of a particular rock will be taken as variance of strength properties relies on a
normally distributed; otherwise Cie may be considerable amount of judgement but often is just
assumed to be beta distributed. Occasionally the based on Lumb (1966).
data will indicate a log-normal distribution. When It is worth noting that almost all researchers
a Mohr-Coulomb failure criterion is adopted then assume that c and <I> are independent variables but,
generally both the cohesion and co-efficient of as will be discussed later, this assumption can

90
have a significant impact on the reliability A-A' and 4 short and 2 long for B-B'. An areal
determined for a slope. survey (in lieu of a scanline) would remove this
bias.

3.2 Geometry of unit boundaries

Generally the boundaries of subsurface units

determined during a site investigation are inferred
from relatively sparse information. In design this
is generally accounted for by conservatively
disuibuting the weaker units in space (i.e. making
them larger than they are expected to actually be).
This introduces a conservatve systematic error to
the analysis but the authors are not aware of any
investigation into this bias. It should be noted that Figure I Example of length bias
if it is the analysts intention to predict the mean
strength of the slope then the mean location of
Correcting for (a) and (b) above is relatively
subsurface boundaries should be adopted in
common but much less often is account taken of
analysis (unless a rigorous analysis of the likely
effect (c).
vanallon of the boundaries in space is
Before discussing corrections, etc for these
undertaken).
effects, we will look at the underlying distribution
for length. Almost excusively researchers have
chosen the exponential and log-normal
3.3 Defect properties
distributions to model defect length.
If the truncation level is set fairly high, it is
The most common geometric variability in slope
very difficult to tell the difference between
engineering is that associated with defects. Often
samples drawn from these distributions. This is
estimates are required of length, spacing and
orientation; these are discussed in tum below. illustrated on Figure 2 where it is obvious that if
only long samples are obtained it would be very
difficult to tell these two distributions apart.
3.3.1 Length

Several significant biases are present in scanline or

areal surveys for the length of a discontinuity:
a) Truncation errors occur because fractures
less than a certain length are not recorded. For
large slopes, this truncation level may often be set
at 1m. Hudson & Priest (1983) collected data on
discontinuities down to lOmm long and thus
recorded data for a great number of defects.
b) Censorship errors occur because often one
or both ends of a defect are hidden from view and
thus the true length is not known. Normally the
Figure 2 Tails of the exponential and log-normal
observed length and how many ends are visible
are recorded. distributions
c) Length bias occurs if scanline survey
methods are adopted because the longer joints The authors have generally assumed an
have more chance of intersecting the scanline and exponential distribution for length because it is
thus of being sampled. This can lead to mathematically more tractable but on the balance
discontinuity length appearing to be log-normally the references appear now to favour the
distributed even if in fact it is exponentially log-normal distribution.
disuibuted. An example is given on Figure 1, * Rouleau & Gale (1985) do not decide between
there are 25 defects shown on this area, 23 are exponential and log-normal; their data looks
short and 2 are long. Nevertheless the scanlines log-normal but this appearance could be due to
shown, and most others, intersect a much greater length bias.
proportion of long defects; 3 short and 2 long for * Einstein & Baecher's (1983) raw data looks

91
like it is log-normally distributed and definitely the truncation level and 11 is the population mean
not exponentially but raw data obtained without length) and thus any correction would not be
correcting for length bias results in a log-normal significant. If a higher level is adopted for
sample no matter whether the population is truncation and the data is from an exponential
exponential or log-normal. Thus this data is not distribution then 11 = Il,amp'e-a. This is easy to
definitive. derive, see also Baecher & Lanny (1978) and
* Kulatilake & Wu (1984) obtained their data via Rouleau & Gale (1985).
areal sampling, so there is only small length bias,
and they found an exponential distribution for the Censorship
data. Probably this is good information but they
sampled only small joints (length less than O.2m Correction for censorship bias is complicated by
in a window of approximately 1m). the fact that variable limits for censorship are
* Baecher et al (1977) stated that length is often adopted. If a fixed censorship level is
log-normally distributed but do not appear to have adopted, say as in a tunnel, then Pahl (1981) is an
tested for the exponential distribution. Further excellent source of corrections for data grouped
they undertook scanline surveys but did not into individual sets.
correct for length bias; so they would not have
been able to tell if it were exponential or not. L" = H (l+m)/(1-m) (2)
Note in a review of the literature they find about
half favour the exponential distribution and the where L" is the estimated mean length
other half the log-normal distribution. H =(I h)/(1 cos 9 + h sin 9)
In summary it would appear that one could I is the length of the outcrop,
adopt either the exponential or log-normal h is the width of the outcrop,
distributions for length of defects; the authors 9 is the inclination of the joint set
would suggest the exponential distribution has to the h dimension,
some advantages. m = (n, - nJ/(n + 1),
n, is the number of defects with
one end visible,
3.3.2 Area n2 is the number of defects with
two ends visible,
The following points should be noted: n is the total number of defects.
* If length is exponentially distributed then so is This expression is distribution free. Pahl also
area; if length is log-normal then area will have a derives the variance for length based on an
more complicated distribution. exponential distribution for spacing.
* Most references give strike length and dip Note also that Pahl's work indicates that if
length as equal but all appear to be based on the length is log-normally distributed and measure­
work of Robertson (1970) and not really on any ments are truncated at a then L = L"(l - 2a/h).
independent data. Alternatively if it is exponentially distributed and
* McMahon (1982) uses dip length equal to 0.6 truncated then L = L" (1-a/h).
of the strike length. The above referenced work by Pahl is excellent,
From the above we can probably define the other sources that may be useful are:
range of dip length as between one half of and * Kulatilake & Wu (1984a,b) appear to extend
equal to the strike length but there is not much Pahl's work to several defect sets but is extremely
information available on this. complicated.
It would be useful in a mine to wash off the * Priest & Hudson (1981) is a very good paper
floor and measure both distributions, that is strike but complicated; includes analysis of semi-trace
length in the floor and dip length on faces, but lengths as well.
this is very rarely done. * Einstein & Baecher (1983) give the following
for the mean length, L, of joints which are
exponentially distributed (as used by Rouleau &
3.3.3 Correction of biases in sampling length Gale, 1985):

Truncation L = (r./)/n (3)

Priest & Hudson (1981) and Einstein & Baecher where r.1 is the sum of all trace lengths and n is
(1983) recommend ignoring this correction but the number of defects for which both ends are
truncate at very low levels (i.e. a « 11; where a is visible.

92
 In summary the results of Einstein & Baecher would be a distribution free approach.
(1983) and Pahl (1981) appear the most
promising.
3.4 Modelling of porewater pressure
Length bias
The porewater pressure distribution has an
Priest & Hudson (1981) provide good information important influence on the performance of
on length bias. Most workers, including Rouleau geotechnical structures, especially soil and rock
& Gale in their major work, do not correct for this slopes. Most probabilistic analyses of slopes do
bias which is accentuated by line sampling. Areal not consider the time element at all, but only the
sampling does not suffer from as great a problem. geometric and material parameters as random
variables. This is in spite of the fact that
porewater pressure is one of the most important
3.3.4 Spacing determinants of the stability of most slopes. The
porewater distribution in the field depends on the
Priest & Hudson (1976), Baecher et al (1977), rainfall, ground water regime, infiltration rate and
Einstein & Baecher (1983) and Brooke (1978, for the spatial variation of the hydraulic properties
15 of 16 fault sets) report spacing as exponentially within the soil. Any realistic modelling of
distributed. porewater pressure should involve the prediction
Rouleau & Gale (1985), based on a very good of rainfall, the statistical and mechanistic
set of data, and Sen & Kazi (1984) recommend modelling of the infiltration of rainwater into the
the log-normal distribution and infer that the soils and the water flows within the soil. Often
exponential distribution does not work for spacing. the probelm will consist of predicting the
Baecher et al (1977) report several others who maximum piezometric pressures during the life of
recommend the log-normal distribution but they the slope.
disagree. On the balance the authors would Hachich & Vanmarcke (1983) and Lee & Wu
suggest that the literature favours the exponential (1987) have pioneered the research on
distribution. probabilistic modelling of water flows in dams and
slopes. Another approach to probabilistic
modelling the effect of rainfall on slope stability is
3.3.5 Orientation given in Fell et al (1988). There is definitely
ample room for further research into this relatively
It is a fairly simple matter to correct for new, but difficult, area of geotechnical reliability.
orientation bias but this is often not done as it is
not necessary if the data is collected on variably
oriented faces or boreholes. 4 INDICES OF STABILITY
Often designers just draw in approximate
contours and take the standard deviation to equal Variability in the material strengths, water table,
one sixth of the range which will then depend etc gives rise to variability in the loads and
upon the sample size and contour interval. More available resistances in a slope. As the factor of
sophisticated approaches have been investigated. safety is the ratio of two of these variables, it is
Herget (1978) gives a method to fit the elliptical also variable (i.e. a random variable) and depends
probability density function to data. on the values chosen (or that actually obtain) for
Einstein & Baecher (1983) find that for a large the various input parameters. Usually slope
number of data sets none of the usual distributions analysis and design are completed with
fit the data (i.e. Fisher, uniform, elliptical, conservatively assessed values adopted for the
Bingham, bi-variate normal or bi-variate important input values, thus it is likely that the
log-normal). factor of safety for design may well be less than
Rouleau & Gale (1985) do not correct the data the mean factor of safety. Generally regardless of
and give no statement regarding the probability the design factor of safety there is some
density functions that they adopt; their data probability of failure.
suggests no probability distribution function will Thus the factor of safety is really only an index
fit. of stability, i.e. it informs us about the relative
Thus there is no generally accepted distribution stability of a design. It is generally believed that
to fit orientation data. It would be possible to a design of a slope with a high factor of safety
adapt Monte Carlo methods of slope analysis to should be more stable than a different design with
randomly select from the actual data set - this a lower factor of safety. The actual value that we

93
obtain for the factor of safety depends on the where a given design is constructed once and it
parameters we select and on the method of either fails or it does not (i.e. the actual
analysis that we adopt. probability of failure is zero or unity). The
There are indices of stability other than the authors believe that the second interpretation is the
factor of safety. The safety margin is defined as more valuable, this is especially so in view of
the available resistance, R(X), minus the loads or some of the problems that are associated with
action S(X). This is not dimensionless, whereas actually determining the probability of failure .
the factor of safety is, but in a probabilistic It would be good if there was a monotonic
framework it does have some advantages. relationship between factor of safety and
probability of failure but Figure 3 shows this is
not the case. It can be seen that some slopes with
factors of safety of approximately 2 have
probabilities of failure of 10.7 whereas other slopes
with factors of safety of 4 to 5 have probabilities
of failure greater than 10·'. This is further
illustrated in Figure 4, here failure surface 1 may
have the minimum factor of safety of 1.2 but only
a low probability of failure (say 1%) as a
considerable effort may have been spent in
determining the strength of the very soft clay (i.e.
in reducing the uncertainty). Failure surface 2
may have a factor of safety of 1.8 but because not
much effort was placed into reducing the
uncertainty of the strength of the soft and firm
clays, the probability of failure of this surface may
be 10%. This is discussed further in Li, Lee and
Lo (1993).

l Figure 4 Reliability of different failure surfaces

Figure 3 Probability of failure versus factor of 5 PROBABILISTIC MODELLING OF SOIL

safety for some actual cases of record (after Harr, AND ROCK
1987)
5.1 Introduction

The probability of failure, PI' is also an index Modelling of soil and rock properties forms an
of stability and can be considered in two ways (1) important part of probabilistic slope analysis. In
if a slope were to be constructed many times what the past, soil properties have generally been
per-centage of such slopes would fail, or (2) the modelled as random variables with perfect spatial
level of confidence that can be placed in a design. correlation which means that a soil property over
The first interpretation may be relevant in a the entire spatial domain is represented by one
mining environment where often the same slope is single random variable. This would imply that the
constructed many times as in say a moving high realisations of a soil property are the same at
wall but has little relevance in civil engineering every location within the soil profile. As

94
emphasized time and again by the authors (Mostyn (5)
G(X)=F(X)-l
& Soo, 1992; Mostyn & Small, 1987; Li, 1991b,
1992a, 1992b), such an approach is erroneous
because soil and rock properties do vary from where R(X) is the 'resistance' and S(X) is the
point to point even within a statistically 'action', and F(X) is the factor of safety. The
homogeneous profile. format of Eq.5 is more common for historical
Cornell (1971), Lumb (1975), Alonso & Krizek reasons because the safety of slopes is traditionally
(1975), Vanmarcke (1977) have pointed out that characterised by the factor of safety. The format
soil properties should be modelled as spatially of Eq.4, however, is preferable because it has a
correlated random variables or 'random fields'. lower degree of non-linearity. Consequently,
The random field model (Vanmarcke, 1977) is methods based on it for calculating the reliability
now widely used to correctly model the correlation index, discussed later, are more accurate.
structure of soil properties. The use of perfectly The performance function of a slope is usually
correlated soil properties gives rise to unreal­ formulated using the limit equilibrium method.
istically large values of failure probabilities for Early studies adopted simplified methods such as
geotechnical structures (Li & Lumb, 1987; Mostyn the friction-circle method (Stoyan et ai, 1979);
& Soo, 1992). Forster and Weber, 1981), the ordinary method of
Unfortunately, most researchers still model soil slices (e.g. Yucemen et aI, 1973; Harr, 1977,
and rock properties as perfectly correlated random Vanmarcke, 1980; Lee et ai, 1983; Ramachandran
variables (e.g. Wolff, 1991; Chowdhury & Xu, & Hosking, 1985), simplified Bishop's method
1992). (e.g. Alonso, 1976; Tobutt & Richards, 1979;
The probabilistic characterisation of the Tobutt, 1982; Moon, 1984; Bergado & Anderson,
properties of a rock mass is far more complicated 1985), simplified Ianbu's method (e.g. McPhail &
than that for soils. The spatial variability of rock Fourie, 1980; Priest & Brown, 1983; Rama­
mass properties is influenced by the variability of chandran & Hosking, 1985).
intact rock, the spatial structure of rock joints and These simplified methods are generally adopted
the properties of the materials along rock joints. for one of two reasons. Firstly they enable an
The task of formulating a realistic probabilistic explicit performance function to be formulated in
model for rocks is formidable. A good terms of the soil parameters which allows the
probabilistic model for rock should be able to ready evaluation of the derivatives of the
describe the variability of rock joint strength and performance function which are generally required
more importantly the spatial variability of the for calculating the reliability index. Secondly they
length, spacing and orientation of rock joints. allow rapid evaluation of the performance function
Most current models describing the spatial which speeds up Monte Carlo simulations which
variability of a rock mass are poor, and there is a require many such evaluations. Sometimes the
great deal of room for further research. A review simplified methods are used purely for
of spatial correlation in rock masses is contained convenience. With the development of the unified
in Yu & Mostyn (1993) in this volume. solution scheme proposed by Li & White (1987d)
and Li (l992c), an explicit performance function
can now be formulated for rigorous stability
5.2 Perfornumce function models such as Morgenstern & Price's method (Li
& Lumb, 1987; Li & White, 1987c).
The performance function of slopes, denoted by Other formulations of performance function
G(X) where X is the collection of random input have also been used. Kraft & Mukhopadhyay
parameters, is a function which defines the failure (1977) formulated the performance function in
or safety state of a slope. The function is defined terms of the movement at the crest of the slope.
in such a way that failure is implied when G(X)<O A slope is deemed to have failed when the
and safety by G(X»O. The boundary defined by movement exceeds the allowable limit. A
G(X)=O separating the safety and failure domains formulation based on the finite element method is
is called the limit state boundary. also presented by Ishii & Suzuki (1987).
The performance function for a slope is usually
taken as one of the following formats.
6 DETERMINATION OF THE RELIABILITY
G(X)=R(X) -SeX) (4)
If UX) is the joint probability density function
(PDF) of X, the probability of failure of a slope is
or given by the integral of fiX) over the entire

95
failure domain defIned by G(X)<O, i.e. from the mean value and standard deviation of the
performance function which are often obtained
P! =f·· ( f
JG(X)<o x
(X) dx 1dx ···dx
2 n
(6) using the fIrst-order-second-moment (FOSM)
method. By using a Taylor's series expansion of
G(X) about its mean the following estimates are
Reliability calculations can be categorised into obtained for the first two moments.
Levell, 2 and 3 analyses depending on the rigour.
~G=G(~1'~2''''~ n) (8)

6.1 Level J analysis

2 " " aG aG
crG=E E _o_ocr;ocr/p u (9)
In a Level I analysis, the safety of the slope is ;=1 }=I ax; ax}
characterised by the factor of safety without due
regard to the variability of the input parameters. where crj is the standard deviation of Xj; Pi} is the
This approach is still widely used for the design of coefficient of correlation between X; and JS; and
slopes. As discussed earlier, the factor of safety is the partial derivatives of G(X) are evaluated at
not a consistent measure of risk. X=(~I'~2""~")' Eq.8 & Eq.9 are exact only if G(X)
is a linear function. The derivatives of G(X) can
be evaluated analytically or by various numerical
6.2 Level 2 analysis differentiation techniques, but a first order finite
difference approximation is equally acceptable.
In a Level 2 analysis, the safety of a slope is This is because the FOSM method is exact only
measured by the reliability index rather than the for a linear function. The accuracy of a first order
factor of safety. Two definitions of reliability finite different approximation for the derivatives is
index are commonly used for slopes as discussed compatible with that of the FOSM method. Using
by Li & White (l987c). The first definition of an analytical technique or a more sophisticated
reliability index, denoted by p, is given as numerical technique such as the technique of
rational approximation (Li, 1992d; Li & White,
P=~ (7) 1987e; Chowdhury & Xu, 1990) will not improve
cr G the model accuracy of Eq.8 and Eq.9 nor the
accuracy of the computed value of p.
where ~G and crG are respectively the mean value Define GlX)=G(~l'~""~H,X;'~;+I>"'~")' Li
and standard deviation of the performance (1991c) suggested that the partial derivatives of
function. This definition suffers from the G(X) be approximated by the following
shortcoming that the numerical values of p depend expression.
on the format of the performance function.
Different, although equivalent formats, of G(X) aG G;(~ ; +0.5cr)-Gl~;-0. 5cr) b i (10)
can yield different values of p (Li & Lumb, aX j
°i cr;
1987).
Another definition of reliability index which is where b;=G;(~j+0.5cr;) - G;(~j-0.5crJ Using
free from this limitation is that proposed by Eq.lO, Eq.9 can be re-written as
Hasofer & Lind (1974). The reliability index,
denoted by ~L' is defined as the distance between " n
(11)
the origin and the limit state boundary in a cr~=E E b ;ob/P u
;=1 j=l
transformed uncorrelated parameter space.
Detailed discussion of the reliability index of PHL The use of Eq.l0 is motivated by two reasons.
for slopes is given by Li & Lumb (1987). If the First, this enables a simple equation to be
performance function is linear, it can be proved developed for crG2 (i.e. Eq. lI). Second, a central
that the two definitions of reliability index give finite difference approximation is more accurate
identical results. The proof can be obtained in than the forward or backward finite difference
many references such as Leporati (1979) and Li & scheme. If Eq.8 and Eq.ll are used for
White (1987).
calculating ~G and crG, the performance function
will need to be evaluated (2n+ I) times to enable
the reliability index P to be computed where n is
6.2.1 Determination of P the number of variables in X.
An alternative technique to the FOSM method
The reliability index P is usually determined

96
are point estimate methods, PEM. The most calculating ~HL' although in a less compact form.
common point estimate method was developed by To enable proper convergence in calculating
Rosenblueth (1975), and more fully detailed by ~HL> it is preferable to compute the derivative of
Rosenblueth (1981) and Harr (1987). This method G(X) to a high degree of accuracy. An exact
is very popular and is preferably called procedure is available for a performance function
Rosenblueth's method to distinguish it from other formulated using Morgenstern & Price's method
point estimate methods. The method requires 2" (Li & Lumb, 1987). The procedure can easily be
evaluations of G(X) in order to obtain the mean extended to other methods based on the method of
value and standard deviation of G(X). The slices, by adopting the unified solution scheme
method becomes impracticable when n is large. developed by Li & White (1987d) and Li (1992c).
In realistic slope stability analysis, the number of A number of numerical differentiation techniques
random variables involved are usually large are available, including the technique of rational
because soil properties have to be modelled as approximation (Li & White, 1987e; Li, 1992d),
spatially correlated random variables. If the slope Ly's method (1991) and Lagrange's formula
is analysed using the method of slices, for (Abramowitz & Stegun, 1970). To ensure that the
instance, the soil properties for each slice should required accuracy is attained, an adaptive
be treated as separate but correlated random procedure is recommended. The following simple
variables. procedure is suggested by Li (199lb).
Rosenblueth's method has been used frequently 1. Define ho = PREY, where PRE is the specified
for calculating the reliability index of slopes (e.g. tolerance.
Moon, 1984; McGuffey et ai, 1982; Nguyen & 2. Calculate an initial estimate of derivative Go' by
Chowdhury, 1985; Chowdhury & Zhang, 1988; a 2-point formula.
Wolff, 1991). In all these studies, soil properties
are modelled as spatially perfectly correlated G '= Gj(llj+ho)-Gj(llj-ho) (12)
random variables. As a result, the number of o 2h
random variables can be limited to a small value
and the use of Rosenblueth's method becomes 3. Set h j = 0.5 hi-! and calculate G/, the ith
feasible. However, as is discussed later, one estimate of the derivative G', by the following 5­
would doubt the usefulness of an analysis based point Lagrange formula (Abramowitz & Stegun,
on spatially perfectly correlated soil properties. 1970).
Li (199ld, 1992a) has recently proposed a new
point estimate method for calculating the
reliability of slopes. The method requires G.' _ G/ Il j- h j_ )-8Gj(1l j-h)+8G/ll j+h j )-G/ll j+h j_
1 1)
(n2+3n+2)/2 evaluations of G(X) for n random
I 6h j _ 1 (13)
variables. This point estimate method has the
same order of accuracy as Rosenblueth's method,
but is much more efficient for large n. Li's
Note that Lagrange's 5-point formula involves
method gives an exact expected value for a
G(Il), but the multiplier for G(llj) is zero so that
quadratic random function. Therefore, it will yield
the 5-point Lagrange formula reduces to a 4-point
a more accurate mean value, IlG, of G(X) than
formula.
Eq.8, but a standard deviation of the same order of
4. Compare the two most recent estimates.
accuracy as Eq.9 or Eq.l1.
Terminate the calculation if the required accuracy
is achieved. Otherwise, repeat step 3.
Note that only two additional evaluations of the
6.2.2 Determination of ~HL
performance function are required for each
iteration. Although the Lagrange formula is
The calculation of ~L is more involved than
marginally less accurate than the technique of
that of~. It involves an iterative procedure and
rational approximation (Li, 1991 b), it enables a
requires the calculation of the derivatives of the
much simpler adaptive procedure to be developed.
performance function for each iteration. A
An example by Li (1991) shows that a tolerance
number of iterative algorithms are available in the
of less than 10-7 can be achieved using the above
literature, but Li & Lumb (1987) and Li & White
adaptive scheme after two iterations, involving a
(1987b,c) observed that the algorithm proposed by
total of 6 evaluations of G(X).
Parkinson (1978) is most convenient for
Li & White (1987b,c) and Li (199lc) observed
probabilistic slope analysis. Chowdhury & Xu
that the two definitions of reliability index gave
(1992) have also presented an algorithm identical
practically the same results for both formats of
to that proposed by Parkinson (1978) for
performance function described in Eq.4 and Eq.5.

97
A similar conclusion can also be drawn from the * Decide which input parameters are to be
study by Chowdhury & Xu (1992). This suggests modelled probabilistically and quantify their
that the performance functions of slopes are variability.
reasonably linear. The reliability index ~ is * Select input values for each parameter;
therefore recommended for general use because it deterministic parameters are fixed at their design
is easier to compute than ~HL. values; probabilistic parameters are randomly
selected from their distributions or the raw data.
* Solve the problem with the selected set of input
6.3 Level 3 analysis parameters to determine the value of the
performance function, G(X).
A Level 3 analysis is the most complete method * Complete the preceeding two steps (i.e.
of risk analysis. All the random variables are parameter selection and problem solution)
represented by their joint probability density continually and build up a distribution of G(X). *
function and the failure probability is evaluated Use the distribution of the result to determine the
directly by multiple integrals extended over the probability of failure.
entire failure domain. Thus the Monte Carlo procedure consists of
In practice, the joint probability density solving a deterministic problem many times to
function is usually not known. Various build up a statistical distribution of G(X). It is a
assumptions, such as statistical independence, the simple and versatile method but consumes a great
form of the statistical distribution for soil deal of computer time.
properties (i.e. normal, log-normal, etc), are often A good introduction to Monte Carlo methods of
made to furnish a Level 3 analysis. Even if the analysis is provided by Priest and Brown (1983).
joint PDF of the random variables are known, the The basis of the model is the simplified lanbu
multi-dimensional integral of Eq.6 is usually method modified to include the Hoek-Brown
difficult to evaluate. empirical failure criterion. The method only treats
Few attempts have been made to date to the strength as a random variable and analyses a
analyse the stability by slope using a Level 3 given failure geometry. The technique can be
analysis. Tobutt & Richards (1979) performed a used to estimate the expected volume of failure.
Level 3 analysis for a c'-<j>' slope. Further This is an important paper from the rock
analytical solutions to simple slope problems have mechanics point of view and useful background
been derived by McMahon (1975), Young (1977) for anyone interested in the probabilistic analysis
and Marek & Savely (1978). However, in all of rock slopes. It should be noted, however, that
these cases, the material properties were modelled spatial correlation of the material properties is
as spatially perfectly correlated random variables. ignored in this method and thus, as will be seen,
The value of such analyses is therefore the resulting probabilities of failure are likely to
questionable. Matsuo & Suzuki (1983) presented be much higher than would be obtained with a
a Level 3 analysis for cohesive slopes. The more realistic random field model.
performance function for a cohesive slope is A case study analysed using the method of
relatively simple, making a Level 3 analysis Priest & Brown (1983) implimented on a micro­
tractable. computer indicated that a slope with a central
factor of safety of 3.3 could still have an 8%
probability of failure (Rosenbaum and Jarvis,
6.4 Monte Carlo simulation 1985). It should be noted that the standard
deviation of the strength that was input to this
To avoid the difficulty of evaluating the multi­ analysis was very high. A factor of safety as high
dimensional integral, many attempts have been as this would not normally be associated with such
made to estimate the failure probability using a high probability of failure, but it does serve to
Monte Carlo simulation. Implementations of highlight some of the problems of using the factor
Monte Carlo methods are given in Nguyen & of safety alone as an index of slope stability.
Chowdbury (1984), Glynn et al (1978), Kim et al There are two major drawbacks to the Monte
(1978) and Major et al (1978). Carlo simulation of slopes. Firstly, if the
The computation of probabilities by Monte probability of failure is very small, then it can
Carlo simulation is a procedure commonly adopted take a great many simulations to determine the
to solve problems that are not readily solved by probability of failure to an acceptable level of
analytical means. The procedure is: confidence. Thus if the probability of failure were
* Choose a deterministic solution procedure, such approximately 10.3 then on average the
as Bishop's method. deterministic model would need to be evaluated

98
one thousand times just to obtain one simulated 6.5 System reliability
failure event. To obtain any real confidence the
model would need to be evaluated at least ten The probability of failure of the most critical slip
thousand times. Thus the simplest versions of surface is commonly used as the estimate of the
Monte Carlo analyses consume a great amount of system failure probability (i.e. the probability of
computer time, and this explains why simple failure of the slope as a whole). This probability
deterministic models are generally adopted for is a lower bound for and mayor may not be close
such analyses. to the system failure probability depending on the
One way to overcome the above limit is to degree of correlation between the random input
complete a given number of simulations, assume a parameters for different slip surfaces.
probability density function for G(X), then It is worth noting that the failure surface with
estimate the parameters of the assumed probability the minimum factor of safety, say as found in an
density function from the sample generated from iterative Bishop analysis may not be the failure
the simulations and then check the goodness of fit surface with the maximum probability of failure.
of the assumed dsitribution and the sample. This The studies by Moria Catalan & Cornell (1974)
approach is not statistically robust but for and Oka & Wu (1990) show that the system
moderate probabilities of failure should be as failure probability can be significantly higher than
accurate as FOSM methods. the failure probability of the most critical slip
Recently there have been many other surface. However, these studies do not consider
improvements in using Monte Carlo simulation to the influence of the common assumption of
evaluate a probability of failure. A good outline spatially perfectly correlated random variables. In
of these is given by Melchers (1988). Most of practice, a common set of design parameters and
these "smart" methods rely on evaluating the the same method of analysis are employed for
performance function about some point closer to different slip surfaces. Suppose that the sample
the failure region than the mean. These methods mean values of the strength parameters estimated
are worth investigating in more detail. from the data are below the true mean value, the
The second major drawback to the approach to strength parameters for all slip surfaces will be
Monte Carlo simulation of slopes is the soil underestimated by the same amount. Also, all slip
properties are generally modelled as spatially surfaces are likely to have a similar magnitude of
perfectly correlated random variables and the model error if they are all analysed using the same
results are therefore of doubtful value. With this method. The presence of 'common' random
model the Monte Carlo simulation procedure variables will produce a significant statistical
becomes deceptively simple. To perform a proper correlation between different slip surfaces.
simulation analysis involving spatially correlated Therefore, the failure probability of the most
random variables, it is necessary to generate critical slip surface may be, but is not necessarily,
random numbers which will simulate the much closer to the system failure probability than
correlation structure of sailor rock properties. that indicated by existing studies. Further research
Research on simulation of spatially correlated is required to provide an answer to this important
random fields is rather limited to date. Fenton & question.
Vanmarcke (1991) have recently proposed a
method for generating Gaussian random fields
using the technique of local average simulation. 7 FACTORS AFFECTING TIlE PROBABILITY
A simpler algorithm proposed by Fishman (1978) OF FAILURE
for generating multivariate normal distribution also
has great potential to be used for generating There are several factors that can have a
Gaussian random fields. These simulation significant impact on the actual value obtained for
techniques can be used for probabilistic analysis of the probability of failure of a particular problem.
slopes, although no such attempt has yet been Some of the more important of these are:
made. * The probability density function assumed for
Although it is envisaged that a Level 3 analysis the probability of failure.
will not be used for routine work, it remains the * The deterministic model adopted as the basis of
only tool for checking the accuracy of a Level 2 the probabilistic solution.
analysis. * Human and gross errors.
* Effects of various forms of correlation.
These will be discussed in turn.

99
7.1 Probability density junction when compared with equivalent three dimensional
models. Thus if the appropriate 3-D model and
Many of the methods of evaluating the probability adequate computing resources were available, the
of failure rely on first determining the moments of computed mean factor of safety may well be 1.6
the distribution of G(X). Even if the mean and with a standard deviation and distribution similar
standard deviation of G(X) are known, the simple to the cases discussed above. This is also
choice between various probability density illustrated on Figure 5. In this case the estimated
functions to model G(X) can lead to large probability of failure would be 0.1 %.
variations in the estimated probability of failure.
Baecher (1987) illustrates this by an example
giving the probability of failure for various values
of the mean and co-efficient of variation for each
of three different probability density functions
adopted to model the factor of safety (linearly
related to G(X)). In the example given a factor of
safety with a mean of 2 and a co-efficient of
variation of 0.2 (i.e. standard deviation of 0.4) the
following approximate probabilities of failure were
found:
* 2x 10. 3 if the factor of safety is assumed to be Figure 5 Effect of detetwinistic model
normally distributed,
* 2x10·4 if log-normally distributed, and
* 2x 10.4 if gamma distributed. Thus a change in deterministic model producing
Thus this "simple" assumption can lead to an a 30% reduction in the factor of safety has led to
order of magnitude variation in the estimate of the a 300% increase in the estimated probability of
probability of failure. It should be noted that failure and a 500% decrease in the reliability, ~.
generally a symetrical distribution, like the The variation in the factor of safety is not really a
normal, is assumed. problem because the acceptable factor of safety is
to a large extent conditioned by the choice of
deterministic model but there is not yet the
7.2 Deterministic model experience to do this with probabilities of failure.
Another case showing the impact of the
The effect of the choice of the deterministic model deterministic model is illustrated on Figure 6
on the estimated probability of failure is illustrated taken from Li and White (1987b) which shows the
in Figure 5. Say a slope has been analysed using variation of the estimated probability of failure for
a Monte Carlo simulation with an extremely high moment and force solutions to a stability problem.
number of simulations; to improve the efficiency We can see that for this case if a force solution
of the computation many such simulations are only is adopted, common in rock engineering, then
based on the ordinary method of slices which may there is two orders of magnitude variation in the
well result in a distribution of the factor of safety estimated probability of failure purely from the
similar to that shown on the left in Figure 5. In choice of A. in a Spencer' s method of solution
this case the factor of safety could well have a (equivalent to Morgenstern-Price solution with fix)
mean of 1.1 and a standard deviation of 0.2 and equal to a constant). As can be seen the variation
be approximately normally distributed; this would for Gm (i.e. moment based solution) is very much
result in an estimated probability of failure of less; this parallels the case for deterministic
about 30%. solutions.
But if a slightly more complicated deterministic
model was adopted, say Bishop's simplified
method, then it is well known that this generally 7.3 Human and gross errors
gives slightly higher factors of safety than the
ordinary method. Thus it is likely that the mean Structures such as large bridges and dams are
of the factor of safety may well be higher at, say, generally very well designed using well accepted
1.3 but the standard deviation and distribution are methods and detailed site investigations, yet
unlikely to change significantly. In this case the research indicates that these still have a
estimated probability of failure would be 7%. probability of failure of about 10. 3; this is due to
Further it is well known that two dimensional human or gross error, i.e. not knowing about a
limiting equilibrium methods are conservative mode of failure, etc.

100
 * correlation between parameters,
* correlation between failure surfaces, and
* correlation of a parameter with itself in space.
These are discussed in turn below.

7.4.1 Correlation between strength parameters

Many researchers have based their probabilistic

method on the assumption that the shear strength
parameters (c and <I> or 11) are independent
(A-Grivas & Asaoka, 1982; Tobutt & Richards,
1979; Chowdhury, 1984; Felio, 1984; Oboni et ai,
1984; Bergado & Anderson, 1983; Vanmarcke,
Figure 6 Effect of model on probability of failure 1980; Alonso, 1976). This assumption is often
(from Li & White, 1987) made to simplify the analysis, many of the writers
refer to the paper of Alonso (1976) in support of
the assumption, others offer no support.
This component of the probability of failure The evidence offered by Alonso is not strong
should presumably be larger, even considerably so, but does indicate a weak negative correlation for
for slopes. The probabilistic models discussed unsaturated residual soils. A-Grivas &
earlier do not include this component and thus it is Harrop-Williams (1979) state that cohesion and
somewhat meaningless to talk about computed friction are negatively correlated (i.e. a high angle
probabilities of failure much less than 10.3 . of friction is associated with a low cohesion) and
It is argued by Melchers (1987) that the design provide one example. Indeed it should be
point (i.e. choice of design for say minimum expected that for any set of strength data derived
expected cost) is not greatly influenced by gross for a particular soil, the values of c and <I> derived
error rates. Discussion of the following equation therefrom would be negatively correlated. Some
illustrates this: writers state that correlation can be ignored
because the probability of failure for a negative
Cr = CjPE) + (P E + PG).CF (14) correlation is higher than it is under the
assumption of independence, thus this assumption
where: Cr is the expected total cost which is a is conservative (A-Grivas & Asaoka, 1982;
function of PE; PE is the estimated probability of Chowdhury, 1984; Bergado & Anderson, 1983; Li
failure; CjPE ) is the estimated initial cost as a and White, 1987b; Vanmarcke, 1980). This is
function of P E ; P G is the probability of failure due supported by the cases analysed by Cherubini et al
to a gross error; and CF is the present value of the (\983).
cost of a failure. Recent work by Yu & Mostyn (to be published)
The designer' s problem is not to correctly has indicated that the degree of conservatism in
estimate the probability of failure but to select the ignoring this correlation may be more than is
design that will have the minimum cost (i.e. to desirable for such an elaborate method (i.e.
minimise Cr ). From the above fOtmula it can be compared with deterministic methods). Thus for a
seen that the minimisation of Cr will not be very typical slope with say ox=oy=1O (see later for
sensitive to PG as long as PG « 1 and CF « C/, explanation), Figure 7 shows the reduction in
these conditions will not always be met in slope probability of failure varying from 10-6 to less than
design. 10-9 for Pc~ varying from + 1 to -1. Even over a
Further consideration of the above equation also sensible range from say the often assumed Pc~ of 0
shows that we cannot simply use PE to compare to, say, -0.5 there is morc than an order of
different designs if P E< P G which will often be the magnitude difference in the probability of failure.
case for slopes in civil engineering construction. If perfect spatial correlation (say /):1000 in all
directions) is assumed then the effect of Pc~
becomes less but is still up to a factor of four.
7.4 Effects of correlation Athanasiou-Grivas & Harr (1979) present one
analysis in which the probability of failure is 8.8%
Several forms of correlation can effect the when the strength parameters are negatively
probability of failure of a slope; these include: correlated and is only 4.8% when they are
assumed independent. This indicates that the

101
assumption of independence is not always the slope in Figure 8 is 10% then the overall
conservative and this assumption needs critical probability of failure of the slope is also 10% if
evaluation if this result is correct. the probabilities of failure for all the planes are
The methods outlined by Young (1985) and perfectly correlated, i.e. if one fails then all of
Tabba (1984) take correlation between the strength them would fail, if one is stable then all are
parameters into account. It is also easy to take stable. But if the probability of failure on each
into account using the approximate method of Li plane is independent of the probability of failure
(1 992a). of the other planes, then the overall probability of
failure is the probability that not all planes are
stable (Le. I - (I - 0.1)\0) which equals 65%.
7.4.2 Correlation between failure surfaces Thus the probability of failure is greatly
influenced by correlation of various modes of
The matter of most interest is usually the failure. This leads naturally to a discussion of the
probability of failure of a slope as a whole (Le. effects of correlation of a parameter with itself in
system reliability) rather than the more commonly space (i.e. spatial correlation or autocorrelation).
detennined probability of failure of a palticular
surface. The system failure probability can be
greatly influenced by the effects of cOlTelation 7.4.3 Autocorrelation
between various failure surfaces. Reference to
Figure 8 will illustrate this. It is intuitively obvious that many geotechnical
If the probability of failure for a single plane of properties are spatially autocorrelated. For
example, jf the cohesion of a certain material at a
given location is known to be, say, Co then we
suspect that the cohesion of the same material at a
nearby point is more likely to be close to Co rather
than to the value of the cohesion at some distant
point. In spite of this most probabilistic models
have assumed (generally unstated) that each soil
property is perfectly correlated with itself for an
infinite distance. Thus for any given realisation if
the cohesion for a soil is say 40 units at one
location it is 40 units at all locations. Thus
cohesion for the soil is a single random variable
which applies unifonnly throughout the soil mass.
Although autocorrelation may be extremely
important it is quite difficult to take into account.
La Pointe (1980, 1981) shows the importance
of autocorrelation by numerically modelling the
stresses in a mass due to a foundation load.
About 15% of the entire mass has failed when the
Figure 7 Effect of correlation between c and cp on distance over which strength is correlated is small;
probability of failure the failed zone increases to over 80% of the mass
when the strength is correlated over large
distances. Thus just taking account of the
autocorrelation of strength dramatically influences
predicted behaviour. Li & White (1987c)
comment that ignoring autocorrelation can lead to
very large predicted probabilities of failure of, for
example, 40% for a factor of safety of 1.2 and that
"this kind of scary prediction has no doubt
aroused scepticism and an unwillingness to usc a
probabilistic approach by practising engineers".
An example given indicates that the probability of
failure may reduce by three orders of magnitude if
auto-correlation is taken into account.
Probabilistic models including autocorrelation
Figure 8 Correlation between failure sUIfaces in I, 2 or 3 dimensions have been described by

102
many writers. The interested reader should refer of probabilistic methods of slope analysis. Most
to some of the following: Vanmarcke (1977b), methods (i .e. most implementations of Monte
Tabba & Yong (l981a,b), Christakos (1987), Li & Carlo, FOSM and PEM) implicitly assume that
White (1987b,c), Schultze (1979a), Kulatilake & Ox=Oy=oo (i.e. perfect spatial correlation), thus a
Varatharajah (1986), Asaoka & A-Grivas (1982), single value of each parameter is taken to apply
Bergado & Iu (1986), Nguyen & Baafi (1986), along the entire failure surface for each realisation.
Alonso & Krizek (1975) and Bergado & Chang The effect of this assumption has been
(1986). The work of the first four groups is of illustrated by Mostyn & Soo (l992) who re­
most interest and may provide the basis of a analysed several cases from the literature which
suitable practical model. Yu & Mostyn (1993) in had previously been analysed assuming perfect
this volume provide a summary of approaches (but unrealistic) spatial correlation. They were re­
adopted to modelling rock joints. analysed allowing for spatial correlation. One
To analyse the autocorrelation of geotechnical result is shown on Table I for a case previously
data generally requires considerably more data reported by Chowdhury (1987) and shown on
than are collected even in a detailed site Figure 10. As can be seen for 0 = 10000 (i.e.
investigation. Thus analysis is generally based on approximating the reported analysis) the
assumptions regarding the autocorrelation structure probability of failure is 1.9xlO·2 but analysis
of the materials. allowing for a typical Ox = 10m and Oy = 1m
Typical values of autocorrelation distances, 0, results in a reduction in the probability of failure
for soils are given by Li & White (l987a) and for to 2xlO· 5 • Thus ignoring auto correlation (i.e.
rock joints in Yu & Mostyn (1993). Typical setting Ox = 00) is conservative but the three orders
correlation distances are Oy in the range 0.1 to 5m of magnitude conservatism is probably
and Ox from 2 to 30m. considerably more than desired. Reference to
Thus if we need an estimate of the average and Figure 7 will also show over five orders of
variance of a parameter over a line, surface or magnitude variation due to autocorrelation. These
volume as for probabilistic slope analysis, then: analyses were completed using a program
* if the volume (or line, etc) is small relative to developed by Li (1987) and described by Li &
the correlation distance, the variance of the Lumb (1987). Further examples are given in
parameter will equal the point valiance and be Mostyn and Soo (1992).
relatively large. The results of Mostyn & Soo (1992) are
* if the volume is large relative to the correlation summalised in Figure II. A number of features
distance then each sample volume will include the evident from this figure are:
full range of the parameter so the variance of the * The solid line represents equality between the
parameter for that volume will be relatively small two analyses, the broken lines below this are
and depend only on the support for tbe sample contours at ratios of two orders of magnitude
mean. between the answers obtained assuming the usual
This valiance reduction is shown on Figure 9 perfect spatial correlation and a more realistic
from Baecher (1987) which shows the valiance 0= 10m.
reduced to 30% of the point value for L = 20. * For a Pf of 50% there is no difference as
Given that autocorrelation exists it is autocorrelation only reduces the variance.
worthwhile to see what effect it has on the results * Above a Pf of 50% ignoring autocorrelation is
not conservative but such a high Pf would always
represent a problem.

Figure 9 Variance reduction due to autocorrelation Figure 10 Slope analysed by Chowdhury (1987)

103
Table I Effect of autocorrelation Pockets and thin layers of softer or stronger
material may be present within a uniform soil
b, (m) by (m) Pf matrix. The presence of a soft spot in a slope
10000 10000 1.9E-2 may significantly affect the stability of the slope.
100 10 1.3E-3 Hilem & Tang (1991) have recently considered the
10 10 6.6E-5 influence of the presence of anomalies on the
10 1 2.0E-5 reliability of slopes. The conditional failure
1 1 8.4E-6 probabilities of slopes given the position of the
0.001 0.001 6.0E-6 anomalies are computed and the failure probability
is then obtained by combining the conditional
probabilities and the probability distribution for
the location of the anomalies.

8.2 Progressive slope failure

There has been continued reseach into progressive

slope failure over the past decades (Chowdhury,
1992; Chowdhury & A-Grivas, 1982; Chowdhury
& Derooy, 1985). However, as pointed out by Li
(1991a,1992b), the current probabilistic approaches
to progressive slope analysis are fraught with
fundamental flaws.
Modelling of soil properties for residual
strength is incorrect and can lead to the somewhat
ridiculous conclusion that a soil can be stronger
after it has failed. Stress redistribution, which is
the most fundamental aspect that should be
considered in analysing the progressive failure of
slopes, is not considered in most, if not all, of the
Figure 11 Effect of autocorrelation on probability existing models. This could lead to the erroneous
of failure conclusion that a slope must fail after a local
failure has occurred anywhere along the slip
surface (Li, 1988, 1991a, 1992b). The fonnulation
* As the P f decreases from 50% the ratio between of the perfonnance function for progressive slope
the two increases rapidly to 105 for Pf (b=oo) equal failure is also erroneous as discussed in Li
to 1%. (1991a). In addition an incorrect equation has
* A line of best fit to the logrithmic data is been used to calculate the probability of
shown as a dashed line. It can be seen from this progressive failure, as demonstrated by the simple
that for realistic values of Pf (b=oo) of say 10 to counter examples discussed in Li (1988,199Ia).
0.1 % there is between two and six orders of This is an important area of research as most
magnitude difference between the results generally slopes tend to fail progressively. It has been
reported and the more realistic results obtained pointed out as early as 1955 by Bishop that yield
taking account of autocorrelation. These zones can occur in a slope even if the factor of
artificially high values of Pf are likely to explain safety is as high as 1.8. Some fundamental
some practising engineers reluctance to adopt improvements to current methods are needed
probabilistic analyses. before one can confidently use a probabilistic
* The conservatism of the usual results is such approach for analysing progressive slope failure.
that they should not be used for economic
analyses but are, at the best, an index.
9 CONCLUSIONS

8 RECENT DEVELOPMENTS Probabilistic slope analysis is a developing field

with various models including different effects. A
8.1 Slopes with anomalies complete model for probabilistic analysis of a soil
slope is still some way off and probably even
Soil profiles are seldom entirely homogeneous. more remote for rock slopes. Nevertheless even a

104
poor probabilistic analysis can add valuable Athanasiou-Grivas D. & Harr M.E. (1979) A
information to guide the slope designer in decision reliability approach to the design of soil slopes,
making. Proc 7th European Conference on Soil
Hasofer (1982) reports that Schultz states that Mechanics and Foundation Engineering, Design
calculated probabilities of failure are often Parameters in Geotechnical Engineering, V I,
considerably greater than those observed in pp.95-99.
practice. Hasofer concludes "it is, however, Baecher G. (1987) Statistical quality control of
necessary for the potential users of these new tools engineered embankments, u.S. Army Engineer
(i.e. multivariate probabilistic techniques) to be Waterways Experiment Station, Vicksburg,
well aware of the assumptions underlying the 167p.
techniques as well as of their limitations. In the Baecher G.B. & Lanney N.A (1978) Trace length
final instance statistical analysis can be a useful biases in joint surveys, Proc 19th US
evaluation tool but it cannot replace sound Symposium on Rock Mechanics, Pt I, pp.
engineering experience and common sense". In a 56-65.
similar vein, Chowdhury (1984) concludes his Baecher G.B., Lanney N.A. & Einstein H.H.
state of the art report on probabilistic methods (1977) Statistical description of rock properties
with advice that "one must, therefore, be wary of and sampling, Proc 18th US Symposium on
attaching overwhelming importance to the absolute Rock Mechanics, pp. 5Cl-l - 5C 1-8.
values of calculated magnitude of risk. It is the Barton N. (1973) Review of a new shear strength
comparison of calculated risk for different criterion for rock joints, Engineering Geology,
alternatives that is really important". These V8, pp. 287-332.
thoughts should be kept well in mind when Bergado, D.T. & Anderson, L.R. (1985) Stochastic
adopting any probabilistic analysis. analysis of pore-pressure uncertainty for the
From the foregoing it is obvious that of the probabilistic assessment of the safety of earth
matters that affect the probability of failure the slopes. Soils and Foundations, 25(2):87-105.
dominant tractable one is spatial correlation. Thus Bergado D.T. & Anderson L.R. (1983) Stochastic
the authors feel that it is now time to abandon analysis of earth slopes, Proc 4th International
models that do not take account of the spatial Conference on the Application of Statistics and
correlation of data and to move forward to models Probability to Soil and Structural Engineering,
based on random fields. Florence, pp. l377-1388.
Bergado D.T. & Chang J.e. (1986) Application of
an inter-active computer-aided probabilistic
REFERENCES slope stability assessmetn for embankments on
soft Bangkok clay, Proceedings of a
A-Grivas D. & Asaoka A. (1982) Slope safety Symposium on Computer Aided Design in
prediction under static and seismic loads, Geotechnical Engineering, AlT, pp. 184-246.
AS.e.E., J. Geot Eng Div, V108 , GT5, May, Bishop, AW. (1955). The use of the slip circle in
pp.713-729. the stability analysis of slopes. Geotechnique,
A-Grivas D. & Harrop-Williams K. (1979) Joint 5:7-17.
distribution of the components of soil strength, Brooke J.P. (1978) Geomathematical investigation
Proc 3rd Int Conf on the Application of of fault populations at selected locations, Proc
Statistics and Probability to Soil and Stmctural 19th US Symposium on Rock Mechanics, Pt 1,
Engineering, Sydney, pp. 189-197. pp.23-29.
Abramowitz, M. & Stegun, LA (Eds) (1970) Cherubini c., Cotecchia V., Renna G. & Schiraldi
Handbook of Mathematical Functions. New B. (1983) The use of bivariate probability
York:Dover Publication. density functions in Monte Carlo simulation of
Alonso, E.E. (1976) Risk analysis of slopes and its slope stability in soils, Proc 4th International
application to slopes in Canadian sensitive clay. Conference on the Application of Statistics and
Geotechnqiue, 26:453-472. Probability to Soil and Structural Engineering,
Alonso, E.E. & Krizek, R.J. (1975) Stochastic Florence, pp. 1401-1411.
formulation of soil profiles. Proc. ICASP2, Chowdhury R.N. (1984) Recent developments in
Aachen, 9-32. landslide studies: Probabilistic methods,
Asaoka A. & A-Grivas D. (1982) Spatial State-of-the-art-report Session VII(a),
variability of the undrained strength of clays, International Symposium on Landslides, pp.
AS.e.E., J. Geot Eng Div, V108, GT5, May, 209-228.
pp. 743-756. Chowdhury R.N. (1987) Practical aspects of
probabilistic studies for slopes, Soil Slope

105
 Instability and Stabilization (ed. Walker B.F. & Halim, I.S. & Tang, W.H. (1991) Reliability of
Fell R.) A.A. Balkema. undrained clay slope considering geologic
Chowdhury, R.N. (1992) Simulation of risk of anomaly. Proc. ICASP6, Mexico, 2:776-783.
progressive slope failure. Canadian Harr, M.E. (1977) Mechanics of Particulate Media.
Geotechnical Journal, 29:94-102. New York:McGraw Hill.
Chowdhury, R.N. & A-Grivas, D. (1982) Harr M.E. (1987) Reliability-Based Design in
Probabilistic model of progressive failure of Civil Engineering, McGraw-Hill Book
slopes. Journal of Geotechnical Engineering, Company, 290p.
ASCE, 108:803-819 Hasofer A.M. (1982) Some multivariate
Chowdhury, R.N. & Derooy, E.R. (1985) probabilistic techniques in geotechnical
Progressive reliability of a strain softening engineering, Math. Scientist, V7, pp. 15-24.
slope. Civil Engineering Transaction, IEAust., Hasofer A.M. & Lind N.C. (1974) Exact and
CE27(1):79-94. invariant second-moment code format,
Chowdhury, R.N. & Xu, D.W. (1990) Application A.S.C.E., J. Engineering Mechanics Division,
of Rational Polynomial Technique in Slope V1OO, EM1, February, pp. 111-121.
Reliability Analysis, Research Report Herget G. (1978) Analysis of discontinuity
No.S090/2, Department of Civil and Mining orientation for a probabilistic slope stability
Engineering, University of Wollongong. design, Proc 19th US Symposium on Rock
Chowdhury, R.N. & Xu, D.W. (1992) Reliability Mechanics, Pt I, pp. 42-50.
index for slope stability assessment - two Hoek E. (1980) An empirical strength criterion
methods compared. Reliability Engineering and and its use in designing slopes and tunnels in
System Safety, 37(2):99-108. heavily jointed weathered rock, Proc 6th
Chowdhury, R.N. & Zhang, S. (1988) Prediction Southeast Asian Conference on Soil
of critical slip surface. Proc. 5th ANZ Engineering, Taipei, pp. 111-158.
Conference on Geomechanics, 451-455. Hudson J.A. & Priest S.D. (1983) Discontinuity
Christakos G. (1987) Stochastic simulation of frequency in rock masses, International Journal
spatially correlated geo-processes, Mathematical for Rock Mechanics, Mining Sciences and
Geology, V19, No 8, pp. 807-831. Geomechanical Abstracts, V20, 2, pp. 73-89.
Cornell, C.A. (1971) First-order uncertainty Ishii, K. & Suzuki, M. (1987) Stochastic finite
analysis of soils deformation and stability. Proc. element methods for slope stability analysis.
ICASP1, 129-144. Structural Safety, 5:111-129.
Einstein H.H. & Baecher G.B. (1983) Probabilistic Kim H-S., Major G. & Ross-Brown D. (1978)
and statistical methods in engineering geology: Application of Monte Carlo techniques to slope
Specific methods and examples - Part I: stability analyses, Proc 19th US Symposium on
Exploration, Rock Mechanics and Rock Rock Mechanics, Pt 2, pp. 28-39.
Engineering, V 16, pp. 39-72. Kraft, L.M. & Mukhopadhyay, J. (1977)
Fell R., Mostyn G., Maguire P. & O'Keefe L. Probabilistic analysis of excavated earth slope.
(1988) Assessment of the probability of rain Proc. 9th ICSMFE., 2: 109-116.
induced landsliding, Proc 5th Australian and Kulatilake P.H.S.W. & Varatharajah P. (1986) A
New Zealand Conference on Geomechanics, weighted regression approach to model and
Sydney, pp. 73-77. estimate spatial variability of soil properties in
Fenton, G.A. & Vanmarcke, E.H. (1990) one dimension, Proceedings of a Symposium on
Simulation of random fields via local average Computer Aided Design in Geotechnical
subdivision. Journal of Engineering Mechanics, Engineering, AlT, pp. 54-65.
ASCE,116(8):1733-1749. Kulatilake P.H.S.W. & Wu T.H. (1984) Estimation
Fishman, G.S. (1978) Principles of Discrete Event of mean trace length of discontinuities, Rock
Simulations. New York:J. Wiley and Sons. Mechanics and Rock Engineering, V17, pp.
Forster, W. & Weber, E. (1981) Influences on the 215-232.
probability of failure of slopes. Proc. 10th Kulatilake P.H.S.W. & Wu T.H. (1984) Sampling
ICSMFE, 1:127-130. bias on orientation of discontinuities, Rock
Glynn E.F., Veneziano D. & Einstein H.H. (1978) Mechanics and Rock Engineering, V17, pp.
The probabilistic model for shearing resistance 243-253.
of jointed rock, Proc 19th US Symposium on Kulatilake P.H.S.W. & Wu T.H. (1984) The
Rock Mechanics, Pt 1, pp. 66-76. density of discontinuity traces in sampling
Hachich, W. & Vanmarcke, E.H. (1983) windows, International Journal for Rock
Probabilistic updating of pore pressure fields. Mechanics, Mining Sciences and
Journal of Geotechnical Engineering, ASCE, Geomechanical Abstracts, V21, 6, pp. 345-347.
109(3):373-387.

106
La Pointe P.R. (1980) Analysis of the spatial Li, K.S. & White, W. (1987c) Reliability index of
variation in rock mass properties through slopes. Proc. 5th International Conference on
geostatistics, Proc 21st US Symposium on Rock Applications of Statistics and Probability in Soil
Mechanics, pp. 570-580. and Structural Engineering, 2:755-762.
La Pointe P.R. (1981) Improved numerical Li, K.S. & White, W. (1987a) Probabilistic
modelling of rock masses through geostatistical Characterization of Soil Profiles, Research
characterization, Proc 22nd US Symposium on Report No.19, Department of Civil Engineering,
Rock Mechanics, pp. 4l6-421. Australian Defence Force Academy, Canberra,
Lee I-M. & Wu T.H. (1987) Groundwater Australia, 54p.
fluctuations in hillside slopes, Proc 5th Li, K.S . & White, W. (1987e) Engineering
International Conference on the Application of Applications of the Technique of Rational
Statistics and Probability in Soil and Structural Approximation, Research Report No.22,
Engineering, Vancouver, B.C., V2, pp. 694-701. Department of Civil Engineering, University
Lee, I.K., White, W. & Ingles, O.G. (1983) College, The University of New South Wales,
Geotechnical Engineering, Boston:Pitman. Australian Defence Force Academy, Canberra,
Leporati, E. (1979) The Assessment of Structural Australia, 67p.
Safety. Research Studies Press. Li, K.S. & White, W. (1987b) Probabilistic
Li, K.S. (1988) Discussion of 'Reliability model Approaches to Slope Design, Research Report
on progressive slope failure'. Geotechnique No.20, Department of Civil Engineering,
38(4):641-644. Australian Defence Force Academy, Canberra,
Li, K.S. (1991a) Fallacies of current probabilistic Australia, 54p.
approaches to progressive slope failure. Proc. Li, K.S. & White, W. (1987d) A Unified Solution
6th International Conference on Applications of Scheme for the Generalized Procedure of Slices
Statistics and Probability in Civil Engineering, in Slope Stability Problems, Research Report
Mexico, 2:784-791. No.2!, Department of Civil Engineering,
Li, K.S. (l991b) Point estimate methods in Australian Defence Force Academy, Canberra,
geotechnics. Proc. Asian Pacific Conference on Australia, 41 p.
Computational Mechanics, Hong Kong, 1:827­ Li, K.S., Lee, I.K. & Lo, S-C.R. (1993) Principles
832. of limit state design in geotechnics. Proc.
Li, K.S. (1991c) Reliability Index for Probabilistic Conference on Probabilistic Methods in
Slope Analysis, Research Report, RIOI, Geotechnical Engineering, Rotterdam:Balkema
Department of Civil and Maritime Engineering, (this volume).
University College, University of New South Lumb P. (1966) The variability of natural soils,
Wales, Australian Defence Force Academy, Canadian Geotechnical Journal, V3, N02, pp.
Canberra, Australia, 24p. 74-97.
Li, K.S. (1992a) A point estimate method for Lumb P. (1975) Spatial variability of soil
calculating the reliability index of slopes. Proc. properties, Proc 2nd International Conference
6th Australia-New Zealand Conference on on the Application of Statistics and Probability
Geomechanics, Christchurch, 448-451. to Soil and Structural Engineering, pp. 397-421.
Li, K.S. (1992b) Some common mistakes in Ly, L. (1990) Numerical differentiation using
probabilistic analysis of slopes. Proc. 6th Gaussian quadrature. Journal of Engineering
International Symposium on Landslides, Mechanics, ASCE, 116:2568-2572.
Christchurch, 1:475-480. Major G., Ross-Brown D. & Kim H-S. (1978) A
Li, K.S. (l992c) A unified solution scheme for general probabilistic method for three
slope stability analysis. Proc. 6th International dimensional wedge failures, Proc 19th US
Symposium on Landslides, Christchurch, 1:481­ Symposium on Rock Mechanics, Pt 2, pp.
486. 45-56.
Li, K.S. (1992d) Numerical differentiation using Marek J.M. & Savely J.P. (1978) Probabilistic
Gaussian quadrature. Journal of Engineering analysis of the plane shear failure mode, Proc
Mechanics, ASCE, 118(3):648-650. 19th US Symposium on Rock Mechanics, Pt 2,
Li , K.S. (1992e) Point estimate method for pp.40-44.
calculating statistical moments. Journal of Matsuo, M. & Suzuki, H. (1983) Use of charts for
Engineering Mechanics, ASCE, 118(7): 1506­ reliability-based design of embankments on
1511. saturated clay layer. Soils and Foundations,
Li, K.S. & Lumb, P. (1987) Probabilistic design of 23(3): 13-26.
slopes. Canadian Geotechnical Journal, McGuffey, V., Grivas, D., lori, J. & Kyfor, Z.
24(4):520-535. (1982) Conventional and probabilistic

107
 embankment design. Journal of Geot Eng Div, Pahl P.l (1981) Estimating the mean length of
ASCE, 108(10):1246-1254. discontinuity traces, International Journal for
McMahon B.K. (1975) Probability of failure and Rock Mechanics, Mining Sciences and
expected volume of failure in high rock slopes, Geomechanical Abstracts, V18, pp. 221-228.
Proc 2nd Australian and New Zealand Parkinson, D.B. (1978) Solution for second
Conference on Geomechanics, Brisbane, pp. moment reliability index. Journal of
308-313. Engineering Mechanics Division, ASCE,
McMahon B.K. (1982) Probabilistic methods in 104:1267-1275.
slope design, unpub, 14pp. Priest, S.D. & Brown, E.T. (1983) Probabilistic
McPhail, G.I. & Fourie, AB. (1980) A practical stability analysis of variable rock slopes.
application of probabilistic slope stability Transaction of Institution of Mining and
analysis method. Proc. South African Metallurgy (Section A:Mining Industry), 92:Al­
Geotechnical Conference, Silverton, 65-77. A12.
Melchers R.E. (1987) Structural reliability Priest S.D. & Hudson J.A. (1981) Estimation of
assessment and human error, Proc 5th discontinuity spacing and trace length using
International Conference on the Application of scanline surveys, International Journal for Rock
Statistics and Probability in Soil and Structural Mechanics, Mining Sciences and
Engineering, Vancouver, B.C., VI, pp. 46-54. Geomechanical Abstracts, V18, pp. 183-197.
Melchers R.E. (1988) Smart Monte Carlo methods Priest S.D. & Hudson J.A. (1976) Discontinuity
for probability integration, Proc 11 th ACMSM, spacings in rock, International Journal for Rock
Auckland, pp. 83-88. Mechanics, Mining Sciences and
Moon, AT. (1984) Probabilistic Slope Stability Geomechanical Abstracts, V13, pp. 135-148.
Analysis Using Rosenblueth's Method. Ramachandran, K. & Hosking, I.A (1985)
Unpublished Report 1984/53. Tasmania Reliability approaches to stability analysis of
Department of Mines. soil/rock slopes. Proc. 5th International
Morla Catalan, 1 & Cornell, e.A (1976) Earth Conference in Numerical Methods in
slope reliability by a level-crossing method. Geomechanics, IO( 19-1 )28.
Journal of Geot Eng Div, ASCE, 102(6):591­ Rosenbaum M.S. & Jarvis J. (1985) Probabilistic
604. slope stability analysis using a microcomputer,
Mostyn G.R. & Small J.e. (1987) Methods of Quarterly Journal of Engineering Geology, V18,
analysis, Soil Slope Instability and Stabilization pp. 353-356.
(ed. Walker B.F. & Fell R.) AA Balkema, Ch Rosenblueth E. (1975) Point estimates for
3, pp. 71-120. probability moments, Proc. Nat. Acad. Sci.,
Mostyn G.R. & Soo S. (1992) The effect of USA, vn, 10, pp. 3812-3814.
autocorrelation on the probability of failure of Rosenblueth E. (1981) Two-point estimates in
slopes, Proc Sixth ANZ Conference on probabilities, Applied Mathematical Modelling,
Geomechanics, pp542-546. V5, Oct, pp. 329-335.
Nguyen V.U. & Baafi E.Y. (1986) Site Rouleau A. & Gale J.E. (1985) Statistical
characterisation by geostatlstlcs, Speciality characterisation of the fracture system in the
Geomechanics Sysposium on Interpretation of Stripa Granite, Sweden, International Journal
Field Testing for Geotechnical Design, I.E. for Rock Mechanics, Mining Sciences and
Aust., Adelaide, 18-19 August, pp. 111-117. Geomechanical Abstracts, V22, 6, pp. 353-367.
Nguyen V.U. & Chowdhury R.N. (1985) Schultze E. (1979) Discussion Session 3 ­
Simulation for risk analysis with correlated Statistics, reliability theory and safety factors,
variables, Geotechnique, V35, 1, pp. 47-58. Proc 7th European Conference on Soil
Nguyen, V.U. & Chowdhury, R.N. (1984) Mechanics and Foundation Engineering, Design
Probabilistic study of spoil pile stability in strip Parameters in Geotechnical Engineering, V4,
coal mines - two techniques compared. Int. pp. 101.
Journal for Rock Mechanics, Mining Science Sen Z. & Kazi A (1984) Discontinuity spacing
and Geomechanics, 21(6):303-312. and RQD estimates from finite length scanlines,
Oboni F., Bourdeau P.L. & Bonnard Ch. (1984) International Journal for Rock Mechanics,
Probabilistic analysis of Swiss landslides, Mining Sciences and Geomechanical Abstracts,
International Symposium on Landslides, pp. V2I, 4, pp. 203-212.
473-478. Stoyan, D., Foster, W. & Weber, E. (1979) On the
aka, Y. & Wu, T.H. (1990) System reliability of probability of failure of slopes. Proc. ICASP3,
slope stability. Journal of Geotechnical Sydney, 2:459-465.
Engineering, ASCE, 116(8): 1185-1189. Tabba M.M. (1984) Deterministic versus risk

108
 analysis of slope stability, International
Symposium on Landslides, pp. 491-498.
Tabba M.M. & Yong R.N. (1981) Mapping and
predicting soil properties: theory, A.S.C.E., J.
Engineering Mechanics Division, V 107, EM5,
Oct, pp. 773-793.
Tabba M.M. & Yong R.N. (1981) Mapping and
predicting soil properties: applications,
A.S.C.E., 1. Engineering Mechanics Division,
V107, EM5, Oct, pp. 795-811.
Tobutt, D.C. (1982) Monte Carlo simulation
methods for slope stability. Computers and
Geosciences, 8(2): 199-208.
Tobutt, D.C. & Richards, E.A. (1979) The
reliability of earth slopes. Int. Journal for
Numerical and Analytical Methods in
Geomechanics, 3:323-354.
Vanmarcke, E.H. (1977) Probabilistic modeling of
soil profiles. Journal of Geot Eng Div, ASCE,
103(11):1227-1246.
Vanmarcke, E.H. (1980) Probabilistic stability
analysis of earth slopes. Engineering Geology,
16:29-50.
Wolff, T.F. (1991) Embankment reliability versus
factor of safety: before and after slide repair.
Int. J. Num. & Analy. Methods in Geomech.,
15:41-50.
Young D.S. (1977) Probability analysis of rock
slopes and its application to a pit slope design,
Proc 18th US Symposium on Rock Mechanics,
pp. 5C5-1 - 5C5-6.
Young D.S. (1985) A generalized probabilistic
approach for slope analysis, International
Journal of Mining Engineering, V3, pp.
215-228.
Yu Y.F. & Mostyn G.R. (1993) Spatial correlation
of rock joints, Proc lnt Conf on Prob Methods
in Geot Eng, (this volume)
Yucemen, M.S., Tang, W.H. & Ang, A.H.S.
(1973) A Probabilistic Study of Safety and
Design of Earth Slopes. Technical Report,
University of Illinois, Urbana.

109