Structural Safety 33 (2011) 145–154

Contents lists available at ScienceDirect

Structural Safety
journal homepage: www.elsevier.com/locate/strusafe

AK-MCS: An active learning reliability method combining Kriging and Monte

Carlo Simulation
B. Echard ⇑, N. Gayton ⇑⇑, M. Lemaire ⇑⇑
Clermont Université, Institut Français de Mécanique Avancée, EA 3867 Laboratoire de Mécanique et Ingénieries, BP 10448, 63000 Clermont-Ferrand, France

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

Article history: An important challenge in structural reliability is to keep to a minimum the number of calls to the numer-
Received 12 March 2010 ical models. Engineering problems involve more and more complex computer codes and the evaluation of
Accepted 25 January 2011 the probability of failure may require very time-consuming computations. Metamodels are used to
Available online 25 February 2011
reduce these computation times. To assess reliability, the most popular approach remains the numerous
variants of response surfaces. Polynomial Chaos [1] and Support Vector Machine [2] are also possibilities
Keywords: and have gained considerations among researchers in the last decades. However, recently, Kriging, orig-
Reliability
inated from geostatistics, have emerged in reliability analysis. Widespread in optimisation, Kriging has
Metamodel
Kriging
just started to appear in uncertainty propagation [3] and reliability [4,5] studies. It presents interesting
Active learning characteristics such as exact interpolation and a local index of uncertainty on the prediction which can
Monte Carlo be used in active learning methods. The aim of this paper is to propose an iterative approach based on
Failure probability Monte Carlo Simulation and Kriging metamodel to assess the reliability of structures in a more efficient
way. The method is called AK-MCS for Active learning reliability method combining Kriging and Monte
Carlo Simulation. It is shown to be very efficient as the probability of failure obtained with AK-MCS is
very accurate and this, for only a small number of calls to the performance function. Several examples
from literature are performed to illustrate the methodology and to prove its efficiency particularly for
problems dealing with high non-linearity, non-differentiability, non-convex and non-connex domains
of failure and high dimensionality.
Ó 2011 Elsevier Ltd. All rights reserved.

1. Introduction The number of evaluations nMC can be extremely important as it

depends on the expected probability of failure. Indeed, a very weak
Reliability aims at determining the probability of failure of a probability of failure requires a large random population to ascer-
system by considering its inputs as random. The performance func- tain the estimation. To quantify the uncertainty on the probability
tion G(x) characterizes the response of the system: a negative value of failure, its coefficient of variation is calculated as:
means that the failure occurs, a positive implies that the system is
sffiffiffiffiffiffiffiffiffiffiffiffiffiffi
safe in this configuration. The border between the negative and po-
1  Pf
sitive domains is called the limit state (G(x) = 0). To assess the C:O:VPf ¼ ð2Þ
Pf nMC
probability of failure, several methods exist. The most common
of them remains the Monte Carlo Simulation. It consists of a ran- Crude Monte Carlo Simulations are easy to implement and are
dom population of size nMC for which the performance function used in a large domain of applications. However, the assessment
is evaluated. An estimation of the probability of failure Pf is then of the probability of failure requires the evaluation of the perfor-
obtained as the ratio between nG60, the number of samples giving mance function for the whole population. This can easily be done
a negative or null G values and the size of the sampling nMC: on analytical systems as the computational demand can be ne-
nG60 glected but it can become impossible with complex computer
Pf ¼ ð1Þ
nMC codes such as Finite Element Models for example. Engineering
models have become extremely time-consuming process despite
the improvements in computer technology. Therefore, it became
⇑ Principal corresponding author. Tel.: +33 (0) 4 73 28 80 00; fax: +33 (0) 4 73 28
necessary to develop methods to approximate the probability of
81 00.
⇑⇑ Corresponding authors. failure with fewer calls to the expensive performance function.
E-mail addresses: benjamin.echard@ifma.fr (B. Echard), nicolas.gayton@ifma.fr FORM and SORM are two elementary approaches [6,7]. They are based
(N. Gayton), maurice.lemaire@ifma.fr (M. Lemaire). on finding the named ‘‘design point’’ and present great advantages

0167-4730/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.strusafe.2011.01.002
146 B. Echard et al. / Structural Safety 33 (2011) 145–154

for very weak probabilities of failure, yet they cannot be applied to samples (n being the number of random variables) in the initial
all problems (especially with high non-linear or complex limit design of experiments defined by using Latin Hypercube sampling
states in the standard Gaussian space). This led the path to the over the bounds ± five standard deviations. From this design, the
use of metamodels. These models aim at approximating the Kriging model is constructed and the point in space which maxi-
performance function with a strategic design of experiments mises the learning function EFF (learning criterion) is searched
(DoE) to obtain, for a less expensive computational demand, a suf- using the DIRECT global optimisation algorithm [17] (this algorithm
ficiently accurate prediction of the performance function’s sign. Re- is preferred to the BRANCH-AND-BOUND [18] one used in the Kriging
sponse surfaces are the most popular methods. They are used for contour estimation method [16] which they see as too expensive).
their speed but they are limited in global interpolation accuracy. This point is then evaluated on the performance function and
To avoid this problem, Polynomial Chaos can be used [1,8]. It cor- added to the design of experiments. This is done until some stop-
responds to a response surface in a particular base (Hermite or ping condition is satisfied. The surrogate model is then considered
other polynomials). However, the definitions of the design of accurate enough to calculate the probability of failure using
experiments and of the polynomial’s degrees are tricky [9]. Fur- Importance Sampling (IS). The results show that the method is
thermore, the accuracy evaluation requires cross validation. More really efficient. However, it can be noticed that in high dimensions
recently, Support Vector Machine methods have also gained con- problems, the size of the initial design of experiments is extre-
siderations with their great effectiveness for reliability studies mely important. Furthermore, the method approximates the limit
[2,10]. state in the whole design space and therefore, even in regions,
Aside from these metamodels, a stochastic approach has been where configurations show very weak densities of probability
intensively investigated: Kriging. Developed for geostatistics in and have negligible effects on the probability of failure.
the fifties and sixties by Krige and then by Matheron [11], this To avoid that, this paper proposes to develop an Active learn-
method gained considerations in the computer experiments field ing method which combines Kriging and Monte Carlo Simulation
in the eighties. It presents several but interesting differences with to separate predicted negative and positive G b values of a Monte
the other metamodels. First, Kriging is an exact interpolation Carlo population. The idea is to perform a Monte Carlo Simulation
method, i.e. the prediction in a point belonging to the design of without evaluating the whole population on G. The sign in each
experiments is the exact value of the performance function in this point is obtained thanks to the predictions thanks to a Kriging
point. Furthermore, thanks to its stochastic propriety, Kriging pro- model based on a few evaluated points. The first stage of the
vides not only predicted values in any points, but also estimations method consists of generating a Monte Carlo population. An initial
of the local variance of the predictions. This variance defines the lo- design of experiments of very small size is selected among this
cal uncertainty on the prediction and is called the Kriging variance. population and a learning function is computed for all the points.
It is understood in the way that the higher the variance, the less The best next point to include in the design of experiments is se-
certain the prediction. Thanks to this variance, Kriging has been lected by this function. The probability of failure is estimated at
intensively used in optimisation problems in the nineties with ac- each step thanks to the Kriging predictions of the Monte Carlo
tive learning methods such as Efficient Global Optimisation (EGO) population. This method enables to evaluate points near the limit
[12]. Active learning means that the Kriging model is updated by state to improve the accuracy of the metamodel and above all, it
adding a new point to the design of experiments, this point being enables to focus only on the points having sufficiently high densi-
selected for its expected improvement on the Kriging model. In this ties of probability to have a significant impact on the probability
domain, EGO was a major step forward. of failure. The approximation of the limit state is then very accu-
The applications of Kriging to structural reliability problems rate in the Monte Carlo population. This method gives at the same
are rather recent. The precursor work seems to have been pro- time a Kriging estimation of the probability of failure and its coef-
posed by Romero et al. [3] who compare Kriging with polynomial ficient of variation without requiring the expensive evaluation of
regression and finite-element interpolation on progressive lattice- the whole Monte Carlo population. It is named AK-MCS as it is
samplings with analytical functions. The results show that none of an Active learning method combining Kriging and Monte Carlo
the methods is more efficient than the others. Following this, Simulation. It must be seen as a modification of the crude Monte
Kaymaz [4] proposes a method to perform structural reliability Carlo Simulation.
analysis and compares it to classic response surface methods. This article is framed in four sections. Kriging theory is pre-
His method is based on the MATLAB toolbox DACE [13,14] and sented Section 2 to show the real interest in using it in computer
consists of finding the design point. His conclusions are that codes. Following this, the proposed method (AK-MCS) is detailed
Kriging metamodel does not greatly improve the reliability results Section 3 as a solution to the current existing methodologies’ prob-
compared to quadratic response surface methods’ results, unless lems. Its concept is defined as different from previous works.
Kriging parameters are well chosen. Following this, structural reli- Section 4 proposes several examples to validate AK-MCS. It is
ability problems are also investigated using Kriging in [15]. These compared to Monte Carlo Simulation and results of the literature.
methods are based on progressive design of experiments but they AK-MCS is shown to be very efficient as the number of calls to
cannot be defined as active learning methodologies. Indeed, the the performance function is relatively weak compared to the other
design of experiments is not improved by learning from all the approaches.
data supplied by Kriging, i.e. Kriging prediction and variance. This
was noticed by Bichon et al. [5] who propose a fully active learn-
ing method to perform reliability. Inspired by EGO and the Kriging 2. Kriging theory: a reminder
contour estimation method of Ranjan et al. [16], this method is
called Efficient Global Reliability Analysis (EGRA). The active Kriging is based on the idea that the performance function G(x)
learning approach is defined thanks to a learning function called is seen as the realisation of a stochastic field GðxÞ [11]. The first
the Expected Feasibility Function (EFF) which provides an indica- step of Kriging consists of defining this stochastic field with its
tion of how well the true value of the performance function in a parameters according to a design of experiments. Then, the Best
point can be expected to satisfy the constraint G(x) = 0. EFF is ex- Linear Unbiased Predictor (BLUP) is used to estimate the value in
pressed as a function of the Kriging data (local prediction and lo- a given point. The model [19] for GðxÞ is given as:
cal variance). More information about it is given Section 3.3. This
method consists of a sequential algorithm starting with ðnþ1Þðnþ2Þ GðxÞ ¼ Fðx; bÞ þ zðxÞ ð3Þ
2
 B. Echard et al. / Structural Safety 33 (2011) 145–154 147

where: putable analytical function. These properties are interesting in reli-

ability studies and metamodels as the Kriging variance represents
 F(x, b) is the deterministic part which gives an approximation of a good index to improve a design of experiments.
the response in mean. It represents the trend of Kriging and cor-
responds to a regression model that can be written:
3. Proposed method: AK-MCS
Fðx; bÞ ¼ fðxÞt b ð4Þ
with f(x)t = {f1(x), . . . , fk(x)} the basis functions and 3.1. Principles of the methodology
bt = {b1, . . . , bk} the vector of regression coefficients. In this paper,
ordinary Kriging is selected which means that F(x, b) is a scalar to The method that we propose here is named AK-MCS. It consists
determine: F(x, b) = F(x, b) = b. All the following equations are of an Active learning reliability method combining Kriging and
now based on ordinary Kriging theory [11]. Monte Carlo Simulation. The active learning principles for reliabil-
 z(x) is a stationary Gaussian process with zero mean and covari- ity with Kriging are proposed in the previous works [5,16]. How-
ance between two points of space x and w defined by: ever, the concept here is different. In fact, instead of
approximating the limit state in the whole space and therefore
covðzðxÞ; zðwÞÞ ¼ r2z Rh ðx; wÞ ð5Þ making expensive evaluations of G for points with very weak den-
sities of probability, it is preferred here to focus on a Monte Carlo
where r is the process variance and Rh the correlation function
2
z
population. AK-MCS classifies a Monte Carlo population of nMC
defined by its set of parameters h.
points without evaluating nMC times the true performance func-
tion. To avoid these evaluations, Kriging is used for its exact inter-
Several models exist to define the correlation function. How-
polation characteristic and for its interesting proprieties in active
ever, in this paper, the anisotropic squared-exponential function
learning methods. Indeed, as shown in introduction for EFF, its sto-
(also called anisotropic Gaussian model) is selected. The anisotropy
chastic nature enables to define a learning function to select
aspect is defined by the fact that h is a vector of length n which cor-
among the nMC points the best next point to evaluate on the perfor-
responds to the number of random variables. The correlation mod-
mance function. In this section, AK-MCS, which must be seen as a
el can be formulated by:
modification of the classic Monte Carlo Simulation, is introduced.
Y
n
Rh ðx; wÞ ¼ exp½hi ðxi  wi Þ2  ð6Þ Then, two learning functions are presented : EFF coming from
i¼1 the EGRA method [5] and a more suitable learning function for
where xi and wi are the ith coordinates of the points x and w and hi is AK-MCS that we propose to name U.
a scalar which gives the multiplicative inverse of the correlation
length in the ith direction. An anisotropic correlation function is pre- 3.2. AK-MCS
ferred here as in reliability studies, the random variables are often
of different natures. The method proposed is given Fig. 1. It consists of 10 stages:
Given a design of experiments [x(1), . . . , x(p)], with xðiÞ 2 Rn the ith
experiment, and Y with Y ðiÞ ¼ GðxðiÞ Þ 2 R, the scalars b and r2z are 1. Generation of a Monte Carlo population in the design space. It is
estimated according to [12] by: named S and is composed of nMC points. At this stage, none of
  them is evaluated on the performance function. They repre-
^ ¼ 1t R1 1 1 1t R1 Y
b ð7Þ
h h sent candidate points to be evaluated if the active learning
2 1 t 1 requires it. This population remains the same during the
rz ¼ ðY  b1Þ Rh ðY  b1Þ
^ ð8Þ
whole process of learning in AK-MCS unless Stage 9 is
p
reached.
where Rhi;j ¼ Rh ðxðiÞ ; xðjÞ Þ is the matrix of correlation between each
2. Definition of the initial design of experiments (DoE). To per-
pair of points of the design of experiments and 1 the vector filled
form Kriging, a design of experiments is required. Here, it
with 1 of length p.
^ and r^z 2 in Eqs. (7) and (8) depend on the corre- consists of a random selection of N1 points among the pop-
However, as b
ulation S. These points are evaluated on the performance
lation parameters hi through the matrix Rh , it is first required to ob-
function and are used as the initial design of experiments
tain them using maximum likelihood estimation:
for the Kriging model. According to our experience, a dozen
1
h ¼ arg minðdet Rh Þp r^z 2 ð9Þ of points are enough. The initial design of experiments is
h
preferred to be defined small and to add step by step only
The next step consists of predicting the response of G(x) in a gi- the point that is improving the most the metamodel. In
ven unknown point. At such a point as x, the Best Linear Unbiased the previous cited papers [5,16], the initial design of exper-
b
Predictor (BLUP) GðxÞ of GðxÞ is computed [11]: iments is much larger as they grow with the dimension of
b
GðxÞ ¼ b þ rðxÞR1 the problem.
h ðY  b1Þ ð10Þ
where r(x) = {Rh(x, x(i))}i = 1, . . . , p. 3. Computation of the Kriging model according to the design of
The Kriging variance r2 ðxÞ is defined as the minimum of the experiments. This stage is performed with the toolbox DACE,
b
G as this tool has been used in several Kriging reference arti-
mean squared error between GðxÞ b and G(x). It can be expressed cles [4,20–22]. The correlation model is chosen Gaussian
as the following analytical function: and the regression model is taken constant (Ordinary
  1 
rb2 ðxÞ ¼ r2z 1 þ uðxÞt 1t R1 uðxÞ  rðxÞt R1 Kriging), as mentioned Section 2.
h 1 h rðxÞ ð11Þ
G 4. Prediction by Kriging and estimation of the probability of fail-
where uðxÞ ¼ 1t R1 b ðiÞ Þ for i = 1, . . . , nMC with
ure. First, Kriging predictions Gðx
h rðxÞ  1.
Kriging is an exact interpolation method. The prediction Gðxb ðiÞ Þ Eq. (10) are obtained thanks to DACE. Then, the probability
in a point x(i) of the design of experiments is exact, i.e. of failure is estimated with the signs of these predictions.
b ðiÞ Þ ¼ GðxðiÞ Þ. Therefore, the Kriging variance is null in these
Gðx It is obtained by the ratio of the points in the population S
points and it becomes important in unexplored areas. This enables with a negative or null Kriging prediction and the total num-
to quantify the uncertainty of local predictions with an easily com- ber of points in S following Eq. (1):
148 B. Echard et al. / Structural Safety 33 (2011) 145–154

Fig. 1. AK-MCS flowchart.

nb 7. Update of the previous design of experiments with the best

c
Pf ¼ G 60 ð12Þ
nMC point. If the stopping condition Stage 6 is not satisfied, the
learning carries on and the best point x⁄ in S is evaluated
5. Identification of the best next point in S to evaluate on the per-
on the true performance function. Following this, it is added
formance function. This stage requires the implementation of
to the design of experiments: Ni+1 = Ni + 1. The method,
what we name the learning function. This function is com-
then, goes back to Stage 3 to compute the new Kriging
puted for each point of S. As it depends on the Kriging pre-
model with the updated design of experiments composed
diction and variance, the Kriging variances r2 ðxðiÞ Þ for
b
G of Ni+1 points.
i = 1, . . . , nMC are first computed with Eq. (11). In this article, 8. Computation of the coefficient of variation of the probability of
we use two learning functions. EFF is proposed in [5] and we failure. If the stopping condition Stage 6 is satisfied, the
implement a new function named U. The best next point x⁄ learning is stopped and the metamodel is said to be accurate
in S to evaluate on the performance function is then identi- enough on the performance function’s signs of the nMC
fied according to the learning function values observed for points. The next stage consists of checking if the Monte Carlo
all the points of S. The criterion to identify x⁄ is named the population S is sufficiently large to obtain a low coefficient of
learning criterion as it depends on the choice of the learning variation on the Kriging estimation of the probability of fail-
function. It is discussed Section 3.3 for each learning ure (Stage 4). Here, a coefficient of variation below 5% is seen
function. as acceptable. It is calculated the same way than Eq. (2) but
6. Stopping condition on learning. Once the best next point x⁄ in on the Kriging results:
S to evaluate has been identified with the learning criterion,
its corresponding learning function’s value is compared to a vffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
stopping one. This is referred as the stopping condition. Once u1  c Pf
C:O:V b ¼ t ð13Þ
again, it depends on the choice of the learning function dis- Pf c
Pf nMC
cussed Section 3.3.
 B. Echard et al. / Structural Safety 33 (2011) 145–154 149

9. Update of the population. If the coefficient of variation is Table 1

found to be too high, S is increased with new points coming Definition of the learning criterion (Stage 5) and the stopping condition (Stage 6) for
the two learning functions proposed. They are given for x 2 S.
from another Monte Carlo population (generated like Stage
1). AK-MCS goes back to Stage 4 to predict the new popula- Learning function EFF U
tion and the active learning method carries on until the stop- Learning criterion max(EFF(x)) min(U(x))
ping condition is met again. It is important to note that no Stopping condition max(EFF(x)) 6 0.001 min(U(x)) P 2
information about previous performance function’s evalua-
tions is lost.
10. End of AK-MCS. If the coefficient of variation of c Pf is low Stage 5 of AK-MCS, the best next point x⁄ in S to evaluate has to
enough, AK-MCS stops and the last estimation of the proba- be identified by the learning criterion. It is mentioned Section 3.2
bility of failure is considered as the result of the method. that it depends on the choice of the learning function. In the case
of EFF, the best next point to evaluate is the point among S having
The advantage of AK-MCS is that it enables to focus only on the the maximum EFF value [5]. The learning criterion is then
configurations which are likely to happen for a certain level of fail- max(EFF(x)) for x 2 S.
ure. It separates the negative and the positive predicted values like Stage 6, the learning function value of x⁄ is compared to a stop-
a Monte Carlo Simulation would do on the performance function. ping one. In [5], it is defined as 0.001. Therefore, once the maxi-
Furthermore, the points that are added to the design of experi- mum of the expected feasibility function for the points in S is
ments are only among those already included in the initial Monte below 0.001, AK-MCS stops the learning and goes to Stage 8. Table 1
Carlo population and the probability of failure is computed at each sums up the different parameters for EFF, i.e. the learning criterion
iteration. It becomes easier to follow the convergence of the prob- and the stopping criterion.
ability of failure in terms of the number of calls to the performance EFF reflects a balance trend between the search in the vicinity of
function. Moreover, the DIRECT [5] or BRANCH-AND-BOUND [16] algo- the limit state and a more global search in the design space [5]. It is
rithms do not need to be implemented as the learning is not per- important to note that this learning function is designed for EGRA
formed on the whole design space but only among the that is to say for a method that approximates the whole limit state.
population. Even if the number of predictions by Kriging in the Here, we classify points which are obtained from a Monte Carlo
present method can be important as the whole Monte Carlo popu- population. It is a different concept which does not guarantee
lation is estimated, the computation time of the predictions can be EFF to perform well with AK-MCS. For this reason, we propose
neglected compared to the time required to evaluate the perfor- the more suitable learning function that we name U.
mance function (particularly if the performance function is calcu-
lated by a complex computer code). Furthermore, the DIRECT and
3.3.2. Learning function U
BRANCH-AND-BOUND algorithms get expensive to perform in high
Coming from the statement that in Monte Carlo Simulations
dimensions. Here, high dimension is not a problem as the Monte
only the sign of the performance function is important, a learning
Carlo population remains the same throughout iterations and only
function based on a different concept than EFF can be proposed. In-
depends on the probability of failure to reach.
deed, in AK-MCS, accuracy on the sign is needed only among the
points of the Monte Carlo population and the limit state can be
3.3. Learning functions roughly approximated in configurations with low densities of
probabilities according to the distributions of the random vari-
The learning functions are essential to AK-MCS. They are used to ables. The points with a high potential risk of crossing the pre-
determine the best next point to evaluate on the performance b
dicted separator GðxÞ ¼ 0 have to be added to the design of
function. They represent the active learning aspect of AK-MCS. In experiments and therefore evaluated on the performance function.
introduction, EFF is mentioned. More information about it is given In fact, the uncertainty on these points can cause them to change
in the first paragraph of this section. Following this, a more suitable from positive to negative (or negative to positive) predicted values.
function named U for AK-MCS is proposed. This leads then to a modification of the probability of failure. The
potentially ‘‘dangerous’’ points can present three characteristics:
3.3.1. Expected feasibility function (EFF) to be close to the limit state, to have an important uncertainty
EFF comes from the EGRA method [5]. Its general formulation (high Kriging variance) or both at the same time. To identify them,
provides an indication of how well the true value of the perfor- the learning function U(x) is proposed:
mance function in a point x is expected to satisfy the equality con-
b
j GðxÞj  UðxÞ rbG ðxÞ ¼ 0 ð15Þ
straint G(x) = a over a region defined by a ± . It reads:

b
EFFðxÞ ¼ ð GðxÞ
It indicates the distance in Kriging standard deviations between
" ! ! !# the prediction and the estimated limit state. It represents a reliabil-
b
a  GðxÞ b
ða  Þ  GðxÞ b
ða þ Þ  GðxÞ
 aÞ 2U U U ity index on the risk of making a mistake on the sign of G(x) con-
rbG ðxÞ rbG ðxÞ rbG ðxÞ sidering G(x) with the same sign than GðxÞ. b This index U(x) can
" ! ! !#
b
a  GðxÞ b
ða  Þ  GðxÞ b
ða þ Þ  GðxÞ be related to the lower confidence bounding (lcb) function pro-
 rb ðxÞ 2/ / / posed by Cox and John [23] for optimisation. In fact, they define
G rbG ðxÞ rbG ðxÞ rbG ðxÞ
" ! !# lcb as:
b
ða þ Þ  GðxÞ b
ða  Þ  GðxÞ
þ U U
rbG ðxÞ rbG ðxÞ b
lcbðxÞ ¼ GðxÞ b rbG ðxÞ ð16Þ
ð14Þ
First, they select b as equal to 2 or 2.5. This choice depends on
where U is the standard normal cumulative distribution function the wish of a local improvement b = 2 or a more global one
and / the standard normal density function. b = 2.5. Then, lcb is computed and minimised. Here, it works the
In the case of reliability, the threshold a is 0. In EGRA, the opposite way as it is a reliability problem and not an optimisation
expected feasibility function is built with  ¼ 2r2 . The same  is one. lcb is not computed as only the sign of the prediction matters,
b
G
selected for AK-MCS+EFF. therefore it is defined as lcb = 0. Then, U(x) which is equivalent to b
150 B. Echard et al. / Structural Safety 33 (2011) 145–154

in (16) is computed (Stage 5) on the whole population. The best Table 2

point to evaluate becomes the one in S which minimises U(x). Example 1, k = 6, reliability results – comparison of AK-MCS (population size:
nMC = 106) with a Monte Carlo Simulation on the same population and with
The learning criterion (Stage 5) is then min(U(x)) for x 2 S. metamodels from [26]. Legend: Ncall, number of calls to the performance function;
Stage 6 requires the stopping condition. For U, it is defined as c
P f , the probability of failure; C:O:V b , its coefficient of variation for the Monte Carlo
Pf
min(U(x))P2 for x 2 S. As a reliability index, U = 2 corresponds to Simulation; b, the corresponding reliability index and b, its percentage error in
a probability of making a mistake on the sign of U(2) = 0.023. comparison with the reference reliability index (the Monte Carlo Simulation result for
Cox and John [23] propose this value of 2 for their optimisation AK-MCS and the reference value mentioned in [26] for the other metamodels). If the
predicted c
Pf and b are similar to the reference value, ⁄ is mentioned in the last
algorithm and moreover, it is seen on the following validation
column.
examples that such a stopping condition guarantees great accu-
racy. Table 1 sums up the learning parameters for U. Method Ncall c
P f ðC:O:V b Þ b b (%)
Pf
The learning function U is seen to give more weight to the
Monte Carlo 106 4.416  103 (1.5%) 2.618 –
points in the close neighbourhood of the predicted limit state AK-MCS+U 126 4.416  103 2.618 ⁄
rather than further ones with high Kriging variance like EFF can AK-MCS+EFF 124 4.412  103 2.619 0.004
do. This learning function is therefore slower than EFF in estimat- Directional Sampling (DS) 52 4.5  103 2.61 0.04
ing roughly the probability of failure but faster in converging to- DS + Response Surface 1745 5.0  103 2.57 1.91
wards an accurate probability of failure at the stopping DS + Spline 145 2.4  103 2.82 7.63
condition. The choice of this learning function U is rather natural. DS + Neural Network 165 4.1  103 2.64 0.76
It is easy to understand that the U value is directly linked to the Importance Sampling (IS) 1469 4.9  103 2.58 1.53
Kriging variance and prediction. Aside from its use in optimisation IS + Response Surface 1375 4.5  103 2.62 ⁄
IS + Spline 428 4.5  103 2.62 ⁄
in a different way, a close idea is proposed by Juang et al. [24] in
IS + Neural Network 52 5.7  103 2.53 3.44
the science of environment to improve the reliability of Kriging-
based delineation of contaminated soils.

Table 3
4. AK-MCS academic validation Example 1, k = 7, reliability results – comparison of AK-MCS (population size:
nMC = 106) with a Monte Carlo Simulation on the same population and with
AK-MCS efficiency is illustrated on several examples which cov- metamodels from [27]. Legend: Ncall, number of calls to the performance function;
c
P f , the probability of failure; C:O:V b , its coefficient of variation for the Monte Carlo
er a wide variety of limit states: first, examples of dimension 2 are Pf
Simulation; b, the corresponding reliability index and b, its percentage error in
tested to observe the method’s behaviour. They were selected for comparison with the reference reliability index (the Monte Carlo Simulation result for
their high non-linearity and rather complex limit state. Following AK-MCS and the reference value mentioned in [27] for the other metamodels). If the
this, a non-linear analytical function with a moderate dimension predicted c
Pf and b are similar to the reference value, ⁄ is mentioned in the last
(<10) is tested. It corresponds to the dynamic response of a non- column.
linear oscillator. To finish, the method is performed on a high Method Ncall c
P f ðC:O:V b Þ b b (%)
Pf
dimension (40, 100) analytical example. These examples are com-
pared to crude Monte Carlo Simulation and results from literature. Monte Carlo 106 2.233  103 (2.1%) 2.843 –
AK-MCS+U 96 2.233  103 2.843 ⁄
AK-MCS+EFF 101 2.232  103 2.843 <104
4.1. Example 1: series system with four branches
Directional Sampling (DS) 9192 2.6  103 2.79 2.11
DS + Response Surface 830 1.0  103 3.03 6.32
The first example consists of a series system with four branches DS + Spline 107 1.0  103 3.00 5.26
which has been proposed in [25–27]. The random variables xi are DS + Neural Network 67 1.0  103 3.05 7.01
standard normal distributed random variables. The performance DS + Ordinary Kriging 107 1.5  103 2.95 3.51
function reads: Importance Sampling (IS) 4750 2.2  103 2.84 0.35
8 9 IS + Response Surface 3877 2.0  103 2.84 0.35
>
> 3 þ 0:1ðx1  x2 Þ2  ðx1pþxffiffi2 2 Þ ; > > IS + Spline 724 2.0  103 2.81 1.40
< = IS + Neural Network 125 2.9  103 2.76 3.16
Gðx1 ; x2 Þ ¼ min 3 þ 0:1ðx1  x2 Þ2 þ ðx1pþxffiffi2 2 Þ ; ð17Þ IS + Ordinary Kriging 146 2.0  103 2.88 1.05
>
> >
>
: ;
ðx1  x2 Þ þ pkffiffi2 ; ðx2  x1 Þ þ pkffiffi2

with k taking two values (6 then 7) to fit literature examples. the nearly radial shape of the limit state is an adequate problem
AK-MCS is compared with numerous metamodels proposed in to solve with crude directional sampling [26]. About IS+NN, it must
[26,27]. The results are summarised Tables 2 and 3. The informa- be seen that the reliability index is not very accurate. An error of 3%
tion is given as the number of calls to the performance function is obtained in comparison to the reference reliability index [26].
Ncall, the number of points nMC in the Monte Carlo population S, The same remark can be made for the second application with
the probability of failure, the corresponding reliability index and k = 7 (Table 3), where Directional Sampling + Neural Network
the error percentage compared to the reference reliability index. (DS+NN) needs less calls but remains less accurate (7% error on
The proposed method called AK-MCS is tested with the two differ- the reliability index). A method with Ordinary Kriging (OK) is also
ent learning functions defined previously (U and EFF) on the same proposed in [27] for k = 7. This method is found to require more
Monte Carlo population and initial design of experiments. calls to the performance function than AK-MCS and its accuracy re-
mains very poor too.
4.1.1. Reliability results AK-MCS shows great effectiveness and it must be noticed that
First, the results (Tables 2 and 3) show that the AK-MCS is more AK-MCS+U gives a prediction of the probability of failure similar
efficient than the other metamodels found in literature. Indeed, the to the one obtained by Monte Carlo Simulation on the true perfor-
number of calls to the performance function is lower than most of mance function for the same population for both k. This can be
the metamodels proposed in [26,27] and the prediction of the linked to Fig. 2 which shows that an accurate separator is defined
probability of failure is very accurate. For k = 6 (Table 2), only crude in regions, where Monte Carlo sampling points are situated.
Directional Sampling and Importance Sampling + Neural Network Locations with extremely weak densities of probability are
(IS+NN) require less calls to the performance function. However, badly approximated as they do not present any interest in the
 B. Echard et al. / Structural Safety 33 (2011) 145–154 151

For both learning functions, the probability of failure is well

approximated with only 70 calls to the performance function for
k = 7, the 30 following calls being only required to satisfy the stop-
ping condition (Fig. 3). Depending on the aim of the analysis, it is
possible to select a stopping condition like U > 2 or a more ‘‘visual’’
one by plotting the evolution of the probability of failure in terms
of the number of calls to the performance function. However, this
is riskier as for this example, one using AK-MCS+U could have been
tempted to stop at the first branch that is to say about 30 calls.
To finish, the positions of the added points to the design of
experiments can be compared for U and EFF. It is seen that the
learning criterion of EFF is likely to add more points in regions
which are not of interest (Fig. 4), whereas for U, it really focuses
on the limit state (Fig. 5).

4.2. Example 2: modified Rastrigin function

Fig. 2. Example 1, Monte Carlo population estimated by AK-MCS+U for k = 7.
The next example consists of a modified version of the Rastrigin
function [28]. This modification is made to obtain negative (failure)
computation of the probability of failure. Consequently, this meth- and positive (safe) values. The idea, here, is to define a highly non-
od certifies a correct approximation of the probability of failure linear performance function involving non-convex and non-connex
with a minimum number of calls to the performance function. domains of failure (i.e. ‘‘scattered gaps of failure’’). The random
variables xi are standard normal distributed random variables.
The performance function reads:
4.1.2. Comparison of the learning functions
According to the results, the choice of the learning function does X
2
 
not seem to have a real impact on this example. However, the use Gðx1 ; x2 Þ ¼ 10  x2i  5 cosð2p xi Þ ð18Þ
of the function U tends to give better results as the probability of i¼1

failure is similar to the one estimated by Monte Carlo Simulation The method is compared Table 4 with Monte Carlo Simulation,
on the performance function. AK-MCS+EFF is found to give a Subset Simulations [29] performed with Phimeca-Soft [30] and a
slightly less accurate probability than AK-MCS+U for approxi- passive Kriging metamodel, that is based on a fixed Latin Hyper-
mately the same number of calls to the performance function. To cube design of experiments which size is equal to the number of
improve accuracy, the stopping condition has to be more restric- calls required by AK-MCS+U.
tive but this would require more evaluations of G. By plotting the This example shows that AK-MCS can be used on highly non-
evolution of the probability of failure in terms of the number of linear limit states and on problems involving non-convex and
calls to the performance function for k = 7 (Fig. 3), it can be seen non-connex domains of failure. Its performance is incomparable
that AK-MCS+EFF converges earlier towards the right probability. to Subset Simulation or even to passive Kriging combined with
AK-MCS+U is slower to converge but the fastest to satisfy its stop- Latin Hypercube sampling. Indeed, the probability of failure esti-
ping condition. Four levels (one for each branch) can be observed in mated by AK-MCS with only 400 calls is found to be the same than
its evolution Fig. 3. These levels are coming from the way the learn- the one obtained by Monte Carlo Simulation for the same popula-
ing criterion works. Indeed, as the limit state has 4 different tion (Table 4). Furthermore, in this case, AK-MCS+U and
branches, the learning criterion of U focuses on one of them first. AK-MCS+EFF perform the same. However, it must be noticed
Once the metamodel is accurate enough in this region, the learning Fig. 6 that AK-MCS+EFF gives an earlier convergence than
criterion goes to another branch and carries on. AK-MCS+U. Fig. 7 shows the design of experiments required by
AK-MCS+U to satisfy the stopping condition and the similar prob-
ability of failure than the one estimated by a classic Monte Carlo
Simulation on the same population. It is seen that the points are

Fig. 3. Example 1, evolution of c P f normalised by Pf obtained by Monte Carlo

Simulation in terms of Ncall for k = 7. Each level 1–4 corresponds to the learning step
of U for each branch of the limit state. Fig. 4. Example 1, approximation by AK-MCS+EFF for k = 7.
152 B. Echard et al. / Structural Safety 33 (2011) 145–154

Fig. 5. Example 1, approximation by AK-MCS+U for k = 7. Fig. 7. Example 2, approximation by AK-MCS+U.

Table 4
Example 2, reliability results – comparison of AK-MCS (population size:
nMC = 6  104) with a Monte Carlo Simulation on the same population, with passive
Kriging and Latin Hypercube Sampling and Subset Simulation using Phimeca-Soft
[30]. Legend: Ncall, number of calls to the performance function (expensive); c
Pf , the
probability of failure; C:O:V b , its coefficient of variation for the Monte Carlo
Pf
Simulation; b, the corresponding reliability index and b, its percentage error in
comparison with the reference reliability index (Monte Carlo Simulation result). If the
predicted c
Pf and b are similar to the reference value, ⁄ is mentioned in the last
column.

Method Ncall c
P f ðC:O:V b Þ b b (%)
Pf

Monte Carlo 6  104 7.34  102 (1.5%) 1.45 –

AK-MCS+U 416 7.34  102 1.45 ⁄
AK-MCS+EFF 417 7.34  102 1.45 ⁄
Passive Kriging+LHS 416 6.73  102 1.50 3.44
Subset 5000 7.65  102 1.43 1.45

Fig. 8. Example 2, Monte Carlo population estimated by AK-MCS+U.

4.3. Example 3: dynamic response of a non-linear oscillator

The following example is a problem with a moderate number of

random variables. It consists of a non-linear undamped single de-
gree of freedom system (Fig. 9). This example is also studied in
[26,31,32]. The performance function is given as:
 2 
2F 1 x0 t1
Gðc1 ; c2 ; m; r; t1 ; F 1 Þ ¼ 3r  sin ð19Þ
m x0 2 2
qffiffiffiffiffiffiffiffiffi
with x0 ¼ c1 m þc2
. The six random variables are listed
Table 5. AK-MCS is compared with a Monte Carlo Simulation and

Fig. 6. Example 2, evolution of c

P f normalised by Pf obtained by Monte Carlo
Simulation in terms of Ncall.

concentrated in the vicinity of the limit state. Furthermore, Fig. 8

shows that the limit state is well approximated, where the Monte
Carlo population is located. Indeed, the approximation coincides
exactly with the limit state in the region covered by the Monte Car-
lo population. Then, the interpolation worsens, where no Monte
Carlo points are observed. Once again, this is not important as it
corresponds to locations which present extremely weak densities
of probability and therefore have a negligible effect on the proba-
bility of failure. Fig. 9. Example 3, non-linear oscillator [26].
 B. Echard et al. / Structural Safety 33 (2011) 145–154 153

Table 5
Example 3, random variables [26].

Variable P.D.F. Mean Standard deviation

m Normal 1 0.05
c1 Normal 1 0.1
c2 Normal 0.1 0.01
r Normal 0.5 0.05
F1 Normal 1 0.2
t1 Normal 1 0.2

metamodels found in literature. It is seen Table 6 that AK-MCS re-

quires less calls to the performance function than the metamodels
proposed in [26]. Furthermore, the probability of failure of AK-
MCS+U is found to be similar to the one estimated by Monte Carlo
Simulation on the true performance function with the same popu-
lation, whereas AK-MCS+EFF, which requires less calls to reach its
stopping condition, is less accurate.
The performance function is well approximated with about 30 Fig. 10. Example 3, evolution of c
P f normalised by Pf obtained by Monte Carlo
calls for the two learning functions (Fig. 10). Among other strate- Simulation in terms of Ncall.
gies proposed in [26], only DS+NN is found to be as accurate as
AK-MCS. However, it requires more calls to obtain the right prob-
ability of failure. This example shows that AK-MCS can be applied Table 7
on moderate dimension problems with great effectiveness. Example 4, reliability results - Comparison of AK-MCS (population size: nMC = 3  105)
with a Monte Carlo Simulation on the same population. Legend: Ncall, number of calls
to the performance function; c Pf , the probability of failure; C:O:V b , its coefficient of
4.4. Example 4: high dimensional example Pf
variation for the Monte Carlo Simulation; b, the corresponding reliability index.

The last example is proposed in [33]. It consists of an analytical Method n Ncall c

P f ðC:O:V b Þ b
Pf
performance function, where the number of variables can be chan-
Monte Carlo 40 3  105 1.813  103 (4.3%) 2.91
ged without modifying significantly the level of failure probability.
100 3  105 1.647  103 (4.5%) 2.94
It reads:
AK-MCS+U 40 112 1.813  103 2.91
pffiffiffi X
n
100 153 1.647  103 2.94
Gðx1 ; . . . ; xn Þ ¼ ðn þ 3r nÞ  xi ð20Þ
i¼1 AK-MCS+EFF 40 124 1.813  103 2.91
100 167 1.647  103 2.94
The random variables are taken as lognormal variables with
unit means and standard deviation of r = 0.2. Two studies are per-
formed on this example: 40 and 100 random variables. The results
of the reliability analyses are summarised Table 7 and are com-
pared with Monte Carlo Simulations of 3  105 points. Less than
200 calls to the performance function are required for both
AK-MCS methods to obtain the similar probability of failure than
the Monte Carlo one on the same population. It is shown that on
this example, increasing the number of random variables on the
same level of failure has little influence on the number of calls
required.

Table 6
Example 3, reliability results – comparison of AK-MCS (population size:
nMC = 7  104) with a Monte Carlo Simulation on the same population and with
metamodels from [26]. Legend: Ncall, number of calls to the performance function; c
Pf ,
the probability of failure; C:O:V b , its coefficient of variation for the Monte Carlo
Pf
Simulation; b, the corresponding reliability index and b, its percentage error in
comparison with the reference reliability index (the Monte Carlo Simulation result for
AK-MCS and the reference value mentioned in [26] for the other metamodels). If the
predicted c
Pf and b are similar to the reference value, ⁄ is mentioned in the last
column.
Fig. 11. Example 4, evolution of c P f normalised by Pf obtained by Monte Carlo
Method Ncall c
P f ðC:O:V b Þ b b (%) Simulation in terms of Ncall for n = 40.
Pf

Monte Carlo 7  104 2.834  102 (2.2%) 1.906 – For 40 random variables, the probability of failure starts to con-
AK-MCS+U 58 2.834  102 1.906 ⁄
verge around 50 calls to the performance function. AK-MCS+EFF
AK-MCS+EFF 45 2.851  102 1.903 0.16
and AK-MCS+U show the same behaviour in this example and none
Directional Sampling (DS) 1281 3.5  102 1.81 5.24
seems to give a faster convergence (Fig. 11). The initial design of
DS + Response Surface 62 3.4  102 1.82 4.71
DS + Spline 76 3.4  102 1.83 4.19 experiments of small size is found to perform well. Indeed, all
DS + Neural Network 86 2.8  102 1.91 ⁄ efforts are concentrated on adding interesting points rather than
Importance Sampling (IS) 6144 2.7  102 1.93 1.04 supplying the metamodel with a large initial design of experiments
IS + Response Surface 109 2.5  102 1.96 2.62 to cover space like in [5]. Furthermore, the use of a Monte Carlo
IS + Spline 67 2.7  102 1.93 1.04 population enables to reduce the number of points to predict. In
IS + Neural Network 68 3.1  102 1.87 2.01
fact, to approximate sufficiently the performance function with
154 B. Echard et al. / Structural Safety 33 (2011) 145–154

the other methods presented before, a large number of predictions Probabiliste pour la conception Robuste en Fatigue), which is
is needed to cover space. Here, once again, the use of a Monte Carlo gratefully acknowledged by the authors.
population is very relevant as it focuses only on configurations
with sufficiently high densities of probability. References

[1] Ghanem RG, Spanos PD. Stochastic finite elements: a spectral

5. Conclusion approach. Berlin: Springer; 1991.
[2] Hurtado JE. An examination of methods for approximating implicit limit state
functions from the viewpoint of statistical learning theory. Struct Safety
This paper proposes an Active learning method combining Kri- 2004;26(3):271–93.
ging and Monte Carlo Simulation (AK-MCS) to perform structural [3] Romero VJ, Swiler LP, Giunta AA. Construction of response surfaces based on
progressive-lattice-sampling experimental designs with application to
reliability analysis. The proposed strategy is found to be economic
uncertainty propagation. Struct Safety 2004;26(2):201–19.
in the number of calls to the expensive performance function and [4] Kaymaz I. Application of Kriging method to structural reliability problems.
its results are very accurate for the probability of failure. This ap- Struct Safety 2005;27(2):133–51.
[5] Bichon BJ, Eldred MS, Swiler LP, Mahadevan S, McFarland JM. Efficient global
proach combines the advantages of two methods: Kriging meta-
reliability analysis for nonlinear implicit performance functions. AIAA J
model and Monte Carlo simulation. The idea is to perform a 2008;46:2459–68.
Monte Carlo simulation without evaluating the whole population. [6] Ditlevsen O, Madsen HO. Structural reliability methods. John Wiley & Sons;
Indeed, the population is predicted using a Kriging metamodel 1996.
[7] Lemaire M. Structural reliability. ISTE Wiley; 2009.
which is defined thanks to only a few points of the population that [8] Sudret B, Der Kiureghian A. Comparison of finite element reliability methods.
are evaluated on the performance function. This strategy permits Probabilist Eng Mech 2002;17(4):337–48.
to focus on configurations with sufficiently high densities of prob- [9] Blatman G, Sudret B. An adaptive algorithm to build up sparse polynomial
chaos expansions for stochastic finite element analysis. Probabilist Eng Mech
ability. Therefore, the performance function’s sign is not approxi- 2010;25(2):183–97.
mated accurately in the whole design space but only, where [10] Deheeger F, Lemaire M. Support vector machine for efficient subset
Monte Carlo points are located. Moreover, the active learning is simulations: 2SMART method. In: Kanda J, Takada T, Furuta H, editors. ICASP
10 – applications of statistics and probability in civil engineering. Taylor and
preferred to update the Kriging metamodel by adding (and evalu- Francis; 2007.
ating on G) a new point step by step to the design of experiments. [11] Matheron G. The intrinsic random functions and their applications. Adv Appl
This is done thanks to the Kriging variance and prediction which Probab 1973;5(3):439–68.
[12] Jones DR, Schonlau M, Welch WJ. Efficient global optimization of expensive
permit to define a learning function to target the best next point
black-box functions. J Global Optimiz 1998;13(4):455–92.
to evaluate, that is to say the point which is likely to cross the pre- [13] Lophaven SN, Nielsen HB, Sondergaard J. DACE, a matlab Kriging toolbox,
dicted limit state and to disrupt the probability of failure for the version 2.0. Tech. Rep. IMM-TR-2002-12; Technical University of Denmark;
2002a. <http://www2.imm.dtu.dk/hbn/dace/>.
learning function U. This results in reducing drastically the number
[14] Lophaven SN, Nielsen HB, Sondergaard J. Aspects of the matlab toolbox DACE.
of calls to the performance function (and therefore the computa- Tech. Rep. IMM-REP-2002-13; Technical University of Denmark; 2002b.
tion time) without penalising the accuracy on the probability of <http://www2.imm.dtu.dk/hbn/dace/>.
failure. Two learning functions are compared (U, EFF). AK-MCS+U [15] Giunta A, McFarland J, Swiler L, Eldred M. The promise and peril of uncertainty
quantification using response surface approximations. Struct Infrastruct Eng:
is found to perform slightly better in most of the cases. To validate Maint, Manage, Life-Cycle Des Perform 2006;2(3):175–89.
the approach, the method is conducted on several examples which [16] Ranjan P, Bingham D, Michailidis G. Sequential experiment design for contour
cover a wide variety of limit states: high non-linearity, non-convex estimation from complex computer codes. Technometrics 2008;50(4):527–41.
[17] Jones DR, Perttunen CD, Stuckman BE. Lipschitzian optimization without the
and non-connex domains of failure, moderate dimensional prob- Lipschitz constant. J Optimiz Theory Appl 1993;79(1):157–81.
lem (<10) and high dimensional one. AK-MCS performs well in [18] Land AH, Doig AG. An automatic method of solving discrete programming
all these cases as the number of calls to the performance function problems. Econometrica 1960;28(3):497–520.
[19] Sacks J, Schiller SB, Welch WJ. Design for computer experiment. Technometrics
is found to be small and the accuracy on the probability of failure 1989;31(1):41–7.
very high. Even if some approaches such as Directional Sampling [20] Marrel A, Iooss B, Van Dorpe F, Volkova E. An efficient methodology for
can be more efficient for suitable limit states such as the first modeling complex computer codes with gaussian processes. Comput Stat Data
Anal 2008;52(10):4731–44.
example with k = 6, AK-MCS is adapted to singularities. Its ability
[21] Kleijnen JPC. Kriging metamodeling in simulation: a review. Eur J Oper Res
to adapt and its parsimony open a new way to applications imply- 2009;192(3):707–16.
ing Finite Element mechanical model. Furthermore, it is seen on [22] Dellino G, Lino P, Meloni C, Rizzo A. Kriging metamodel management in the
design optimization of a CNG injection system. Math Comput Simul
the high dimensional example that the number of calls to the
2009;79(8):2345–60.
expensive performance function is not really sensitive to the num- [23] Cox DD, John S. SDO: a statistical method for global optimization. In:
ber of random variables. The method can be performed either in Alexandrov MN, Hussaini MY, editors. Multidisciplinary design optimization:
the physical space regardless of the distributions or in the standard state-of-the-art. Philadelphia: Siam; 1997. p. 315–29.
[24] Juang KW, Liao WJ, Liu TL, Tsui L, Lee DY. Additional sampling based on
space by converting the random variables to standard normal ones. regulation threshold and kriging variance to reduce the probability of false
Indeed, as underlined in [5], Kriging is not greatly affected by the delineation in a contaminated site. Sci Total Environ 2008;389(1):20–8.
choice of the space as no assumptions are made on the shape of [25] Waarts PH. Structural reliability using finite element methods: an appraisal of
directional adaptive response surface sampling (DARS). Ph.D. Thesis; TUDelft;
the limit state. Moreover, the use of a Monte Carlo population en- 2000.
ables to perform AK-MCS on problems for which the topology is [26] Schueremans L, Van Gemert D. Benefit of splines and neural networks in
not known. To finish, on the example of AK-MCS, some other meth- simulation based structural reliability analysis. Struct Safety 2005;27(3).
[27] Schueremans L, Van Gemert D. Use of Kriging as meta-model in simulation
ods can be proposed such as AK–IS or AK-SUBSET which define an procedures for structural reliability. In: 9th International conference on
Active learning method combining Kriging with respectively structural safety and reliability, 2005b. p. 2483–90.
Importance Sampling or Subset Simulation. These methods open [28] Törn A, Zilinskas A. Global optimization. Lect Notes Comput Sci 1989:350.
[29] Au SK, Ching J, Beck JL. Application of subset simulation methods to reliability
the way to efficient calculations of very weak probabilities of
benchmark problems. Struct Safety 2007;29(3):183–93.
failure. [30] Lemaire M, Pendola M. Phimeca-soft. Struct Safety 2006;28(1–2):130–49.
[31] Rajashekhar MR, Ellingwood BR. A new look at the response surface approach
for reliability analysis. Struct Safety 1993;12(3):205–20.
Acknowledgements [32] Gayton N, Bourinet JM, Lemaire M. CQ2RS: a new statistical approach to the
response surface method for reliability analysis. Struct Safety
2003;25(1):99–121.
This research was supported by the French Agence Nationale de [33] Rackwitz R. Reliability analysis – a review and some perspectives. Struct Safety
la Recherche – ANR under the project APPRoFi (APproche mécano 2001;23(4):365–95.