Cite this article Research Article Keywords: computational mechanics/

Kabasi S, Roy A and Chakraborty S Paper 2000220 mathematical modelling/risk & probability
Reliability analysis of structures by iterative sequential sampling based response surface. Received 29/08/2020; analysis
Proceedings of the Institution of Civil Engineers – Structures and Buildings, Accepted 05/01/2021
https://doi.org/10.1680/jstbu.20.00220
ICE Publishing: All rights reserved

Structures and Buildings

Reliability analysis of structures by iterative

sequential sampling based response surface
Sounak Kabasi BTech, MTech Subrata Chakraborty PhD
Research Scholar, Department of Civil Engineering, Indian Institute of Professor, Department of Civil Engineering, Indian Institute of Engineering
Technology Bombay, Mumbai, India (Orcid:0000-0001-6466-7371) Science and Technology, Shibpur, Howrah, India
Atin Roy MTech (Orcid:0000-0002-4792-3844)
Research Scholar, Department of Civil Engineering, Indian Institute of
Engineering Science and Technology, Shibpur, Howrah, India
(Orcid:0000-0002-7353-3903) (corresponding author:
atin.3222@yahoo.com)

In the metamodelling-based Monte Carlo simulation (MCS) framework for reliability analysis of structures, the
training samples used to construct the response surface should be as close to the failure plane as possible to
ensure sufficient accuracy in reliability estimates. For this, an algorithm based on the maximin distance criterion
combined with an error norm based on leave-one-out cross-validation (CV) is proposed to construct a moving
least-squares based response surface for improved reliability estimate. The algorithm hinges on the fact that the MCS
points having predicted responses less than the maximum absolute error obtained by the leave-one-out CV approach
are likely to have the maximum effect on the accuracy of the reliability estimate. Relying on this, two points are
added at each iteration, ensuring that the new data points are sufficiently close to the actual limit state and
adequately far from existing data points. The improved reliability estimation capability of the proposed algorithm
considering the direct MCS-based results as the benchmark is elucidated numerically by considering two example
problems.

S matrix of training samples

Notation
Sl input vector of lth training point
A1 section area of top radial bars of space dome truss
W diagonal matrix of the weight function
A2 section area of peripheral bars of space dome truss
wl weight function corresponding to lth training point
A3 section area of bottom inclined bars of space dome truss
X input vector
c free parameter for weight function
xi ith variable
D N
p matrix of Euclidean distances
y output vector for actual responses
d Euclidean distance between points
yl actual response at lth training point
dij Euclidean distance between ith candidate point X i
ŷlCV response at lth point approximated by the training set
and jth data point S j
of a CV approach
(d min)i minimum value of the ith row of the matrix D
ŷðXÞ approximated response
E Young’s modulus of material of space dome truss
β vector of unknown coefficients
eCV vector of cross-validation (CV) error magnitudes
l β̂ estimate of coefficients vector
eCV CV error magnitude at lth point
β0, βi, βii unknown polynomial coefficients
emax
CV maximum value of elements of vector eCV
Δallow allowable maximum displacement of the top node of
F design matrix
the space dome truss
f(X) vector of basis functions
ΔzP1 maximum vertical displacement of the node under
fy yield stress
load P1
g limit state function
σeq equivalent von Mises stress
k free parameter for weight function
Ω set of candidate points for adaptive sampling
N number of candidate points in the set Ω
n number of input variables
P1 load at centre node of space dome truss
P2 load at each of the six nodes of the middle hexagon 1. Introduction
of the space dome truss Uncertainty is introduced in a structure due to uncertainty in
Pf,i probability of failure estimated at ith iteration the parameters required to characterise it (e.g. geometry,
p number of training points material properties, boundary conditions etc.) and loads acting
p0 internal pressure on it. Structural reliability analysis (SRA) is a theoretical fra-
R radius of hypersphere of influence mework that accounts for the effect of such parameter uncer-
r0 internal radius of hollow sphere tainties. The primary task of SRA is to estimate the
r1 external radius of hollow sphere probability of failure (PoF); this requires the computation of a

1
Downloaded by [ Indian Institute of Technology Madras] on [03/09/23]. Copyright © ICE Publishing, all rights reserved.
 Structures and Buildings
multi-dimensional integral over the unsafe domain, involving
the joint probability distribution function (PDF) of the related
random input variables. The joint PDF of the input variables
is rarely available. Moreover, the exact computation of such an
integral is often computationally demanding. Several methods
have been developed to estimate the PoF, which can be
classified into two groups – approximate analytical techniques
and methods based on statistical simulation (Ditlevsen and
Madsen, 1996; Faravelli, 1989; Haldar and Mahadevan, 2000;
Kwon and Elnashai, 2006). Approximate analytical techniques
are based on the second-moment method, which has attracted
criticism with regard to the accuracy of the estimated reliability
index. Moreover, such approaches require evaluation of not
only the limit state function (LSF) but also its gradients
during iterations, which is computationally challenging in the
case of the implicit LSF of a large complex structure.
The most accurate and conceptually straightforward means
of SRA is based on the Monte Carlo simulation (MCS) tech-
nique (Au and Beck, 2001; Dymiotis et al., 1999; Melchers,
1999; Shinozuka, 1983). This approach is preferred as it
does not require an assumption about the shape of the failure
surface. However, such a full simulation approach requires a
large number of repetitive evaluations of the LSF in order to
achieve acceptable confidence in the reliability estimates. If the
performance function is explicitly available in closed form, the
numbers of performance function calls do not play an impor-
tant role. However, the performance behaviour of real large
complex structures is usually defined by an implicit form of
the LSF because an explicit form is often unavailable.
Although reliability analysis of structures involving an implicit
LSF can be performed by the direct MCS technique, each
performance function evaluation typically requires analysis of
a finite-element model, which involves high computational
cost, especially for large complex structures (Ditlevsen and
Madsen, 1996; Haldar and Mahadevan, 2000). The compu-
tational involvement for the analysis of structures to extract
the necessary responses for statistical analysis was studied by
Kwon and Elnashai (2006). The number of simulations
required may be of the order of several thousand for an accep-
table estimate of reliability, depending on the function being
evaluated and the magnitude of the PoF (Faravelli, 1989). As
an effective solution to such problems, the metamodelling
approach based on the polynomial response surface method
(RSM) is widely used to overcome the computational chal-
lenges of MCS-based reliability analysis of large complex
structures involving an implicit LSF (Bucher and Bourgund,
1990; Faravelli, 1989; Liu and Moses, 1994; Rajashekhar and
Ellingwood, 1993). The RSM simplifies the simulation process
by fitting a response surface model to replace the implicit LSF
approximately. The response surface is an approximated poly-
nomial function of random variables. The coefficients of each
term in the polynomial can be obtained by the least-squares
method (LSM) using the actual structural responses at a lesser
(but sufficient) number of times. Thereby, the computational
2
Downloaded by [ Indian Institute of Technology Madras] on [03/09/23]. Copyright © ICE Publishing, all rights reserved.
Reliability analysis of structures by
iterative sequential sampling based
response surface
Kabasi, Roy and Chakraborty
involvement of the analysis can be reduced dramatically. The
applications of the LSM-based RSM in reliability analysis of
structures are based on a global approximation of scatter posi-
tion data. However, the LSM is one of the major sources of
error in prediction by the RSM and needs a considerable
number of training points to ensure the necessary accuracy. It
thus becomes computationally intensive for practical engineer-
ing problems involving too many variables.
To improve the efficiency and accuracy of the RSM, the
applications of various adaptive metamodelling approaches
are notable – for example, the polynomial-based moving least-
squares method (MLSM) (Kang et al., 2010; Kim et al.,
2005), artificial neural networks (ANNs) (Elhewy et al., 2006;
Lagaros et al., 2009), kriging (Kaymaz, 2005), polynomial
chaos expansion (Blatman and Sudret, 2011), support vector
machines (SVMs) (Ghosh et al., 2018; Li et al., 2006; Roy
et al., 2019) and so on. The MLSM-based RSM is noted to
be the simplest among these. In this regard, it is important to
note that, for accurate estimation of reliability, the design of
experiment (DoE) points should be as close to the LSF as
possible. Based on this fact, Bucher and Bourgund (1990) pro-
posed a sampling method to construct an improved LSM-
based response surface for SRA. The approach was further
modified by Rajashekhar and Ellingwood (1993) for iterative
improvement of the RSM. Such iterative improvements are not
only limited to LSM-based RSMs (Farag and Haldar, 2016;
Gaxiola-Camacho et al., 2017) but have also been studied
in the framework of adaptive metamodelling approaches – for
example, by metamodels based on the MLSM (Goswami
et al., 2016) and SVMs (Richard et al., 2012). However, these
approaches require the complete replacement of DoE sets after
each iteration step. The number of iterations required may be
large in many cases and the replacement of DoE sets may
demand much computational time, especially when the LSF
is implicit and highly non-linear. Thereby, such approaches
reduce the efficiency of the RSM significantly. The iterative
improvement by adding a new data point to the existing train-
ing dataset after each iteration step to improve the reliability
estimate is appealing in this regard (Echard et al., 2011). The
approach selects adaptive sampling points based on the uncer-
tainty in the prediction and the magnitude of the predicted
LSF by the kriging method. However, such an adaptive sampl-
ing procedure can be employed in metamodelling only
when the prediction variance is available. However, most of the
metamodels, not being Gaussian process based regression, are
unable to provide the prediction variance. Sequential adaptive
sampling for such metamodels to improve MCS-based SRA
seems to be attractive in this regard.
Roussouly et al. (2013) attempted a sequential adaptive sampling
for the polynomial RSM by searching a hypercube with the
most probable failure point as its centre and the hypercube con-
sidered as a reduced space. A similar sequential adaptive
sampling was proposed by Guimarães et al. (2018) with a few
 Structures and Buildings Reliability analysis of structures by
iterative sequential sampling based
response surface
Kabasi, Roy and Chakraborty

modifications. However, a reduced space with a regular hyper-
cube shape includes a huge amount of unimportant regions.
Xiao et al. (2018a) developed three learning functions for select-
ing the most suitable sample point at each step of iteration for
all types of metamodels. However, the relative weights for these
learning functions are heuristic and require experiences. Xiao
et al. (2018b) proposed another learning function to select
sequential training samples combining the cross-validation (CV)
method, the weighted Euclidean distance and the weights of
sample qualities in the input parameter space. The CV method
was employed to estimate the average probabilistic classification
error function on the candidate sample point. However, the
probabilistic classification function requires both the prediction
and its variance at the candidate sample point. Thus, application
fundamentals of the usual LSM- and MLSM-based response
surfaces are first discussed in this section before presentation
of the proposed algorithm in Section 3.
2.1 LSM-based RSM
The RSM is used to uncover an unknown analytical relation-
ship (an empirical model) between several inputs and outputs.
The reduced quadratic polynomial model (without cross-
terms) is mostly employed to replace the unknown LSF for
reliability analysis due to its trade-off between simplicity and
accuracy (Wong et al., 2005). The reduced quadratic poly-
nomial RSM can be expressed as
X
n X
n

of this learning technique is also limited to regression methods 1: ŷðXÞ ¼ β0 þ β i xi þ βii x2i ¼ f ðXÞT β
i¼1 i¼1
based on kriging or other Gaussian processes. Recently, Roy and
Chakraborty (2020) developed a sequential sampling for support
vector regression based on the maximin distance criterion. where ŷðXÞ is the approximated response for input vector
X consisting of n variables, xi denotes the ith variable, β0, βi
In the present study, an algorithm based on the maximin and βii are unknown polynomial coefficients, f(X) = {1, …, xi,
distance criterion combined with a leave-one-out CV based …, x2i , … }T is the vector of basis functions and β = {β0, …, βi,
error norm is proposed to construct a MLSM-based response …, βii, …}T is the vector of 2n + 1 unknown coefficients. The
surface for an improved estimate of reliability. Specifically, unknown coefficients are obtained by the usual LSM. The
the proposed algorithm attempts to select training samples to training data are obtained by constructing a DoE and evaluat-
construct a response surface as close to the failure plane as ing the corresponding actual responses of the structure. The
possible to ensure sufficient accuracy in the reliability estimate. LSM minimises the sum squared error (SSE) at all the sample
It hinges on the fact that the MCS points having predicted data points considered for training to estimate the polynomial
responses less than the maximum absolute error obtained coefficients. The error norm (SSE) can be expressed as
by the leave-one-out CV approach are likely to have the
!2
maximum effect on the accuracy of the reliability estimate. In X
p X
n X
n
2
this regard, it can be realised that misclassification of a point 2: SSE ¼ y  β0 
l
βi xli  βii ðxlii Þ
l¼1 i¼1 i¼1
can only occur if the error in approximating the LSF at any
sample point is more than the magnitude (irrespective of the ¼ ðy  FβÞT ðy  FβÞ
sign) of the approximated LSF. Hence, it can be intuitively
anticipated that the points corresponding to a magnitude of where y = {y 1, …, y l, …, y p}T is the output vector of actual
approximated LSF less than the value of the noted absolute responses at p number of training data points, S ¼
error obtained by the leave-one-out CV approach are most fS 1 ; . . . ; S l ; . . . ; S p g, F is the design matrix composed of p rows
likely to be misclassified. Hence, the accuracies of MLSM- containing f(X)T for each training point. The LSM-based esti-
based metamodels in approximating the LSF at these points mate of the coefficients vector ( β̂) is obtained as
are of paramount interest for an improved estimate of the PoF.
Thus, a specific error norm based on the leave-one-out CV is  1
3: β̂ ¼ FT F FT
used to decide a reduced input space that is concise and con-
tains only the important regions. Relying on this, two points
are added at each iteration from the reduced space by ensuring
that the new data points are sufficiently close to the actual Once the vector β is obtained, the LSF at any point from the
limit state as well as adequately away from the existing data input space can be approximated by Equation 1.
points to avoid clustering effects. The improved reliability esti-
mation capability of the proposed algorithm considering the 2.2 MLSM-based RSM
direct MCS-based results as a benchmark is elucidated numeri- The MLSM-based RSM is a weighted LSM having various
cally by considering two example problems. weights with respect to the position of approximation. Thus,
the coefficients of the response surface function change with
2. RSMs the change of approximation point of interest. This procedure
The present study deals with the sequential updating of the is interpreted as a local approximation. To fit the polynomial
MLSM-based RSM for improved estimates of the reliability of function to scattered data, the LSM evaluates the unknown
structures. To explain the proposed algorithm, the coefficients of the function by minimising the SSE. However,

3
Downloaded by [ Indian Institute of Technology Madras] on [03/09/23]. Copyright © ICE Publishing, all rights reserved.
 Structures and Buildings Reliability analysis of structures by
iterative sequential sampling based
response surface
Kabasi, Roy and Chakraborty

in the MLSM approach, the SSE is defined as the sum of based error norm to construct a MLSM-based response
weighted errors as surface model for an improved reliability estimate.

!2
X
p X
n X
n
2
The algorithm starts with an initial DoE. As the failure
4: SSEðXÞ ¼ wl y  β0 
l
βi xli  βii ðxlii Þ plane is not known beforehand, training samples for the initial
l¼1 i¼1 i¼1
DoE should be uniformly selected from the entire input space.
T
¼ ðy  FβÞ WðXÞðy  FβÞ Thus, a single-shot DoE is obtained by Latin hypercube
sampling considering the upper and lower boundaries of all
where SSE(X) is the weighted SSE at the training points, which the related random variables. The algorithm hinges on the fact
depend on the location of the point of interest X, wl is the weight that the training points close to the failure plane should be
corresponding to the lth training point and WðXÞ is the diagonal added to the DoE for improved accuracy of approximation of
matrix of the weight function for point of interest X. The weight the LSF close to the failure region. This, in turn, is expected
matrix WðXÞ is constructed using the weighting function in the to improve the accuracy of estimated reliability by the RSM-
diagonal terms. It may be obtained by using weight functions based MCS approach. Once an approximate LSF is con-
such as constants, linear, quadratic, higher-order polynomial, structed by the MLSM based on the initial DoE, a reduced
exponential functions and so on (Kang et al., 2010; Kim et al., domain containing candidates for adaptive samples can be
2005). In the present study, the weight matrix WðXÞ is obtained identified. To provide a measure of fit of a model to a dataset,
using a Gaussian weighting function of the form the CV approach in regression analysis is frequently applied
(Kohavi, 1995; Roy et al., 2019; Roy and Chakraborty, 2020;
2k 2k Xiao et al., 2018b). A fitted model having been constructed,
eðd=cRÞ  eð1=cÞ
5: wðdÞ ¼ 2k
each data point, in turn, is held out and the model is recon-
1  eð1=cÞ structed using the remaining data. The process of evaluating
  the error at each training point by excluding that point
where d ¼ X  S l  is the Euclidean distance between the from the entire training dataset to build a model based on the
point of interest X and the lth training point S l and R is the remaining data is referred to as the leave-one-out CV
approximate radius of hypersphere of influence for X. In approach.
Equation 5, the parameter k is taken as unity to ensure the
Gaussian nature of the weighting function. The value of the par- Let the entire training dataset be represented by fS; yg with
ameter c is taken as 0.4 (Taflanidis and Cheung, 2012). It can be p training points, where S ¼ fS 1 ; . . . ; S l ; . . . ; S p g and the
noted from Equation 5 that the weight associated with a particu- corresponding actual responses y = fy1 ; . . . ; yl ; . . . ; yp gT. In
lar sampling point S l decays as the point X moves away from S l. the leave-one-out CV approach, the lth data pair fSl ; yl g is
The weighting function has its maximum value (1.0) at a nor- held out from the training dataset fS; yg and a metamodel is
malised distance d/R = 0 and minimum value (0) outside the constructed with the remaining data to predict the response at
influence hypersphere (i.e. w(d/R > 1) = 0). The function value Sl (say, ŷlCV is the approximated response). The CV error mag-
l l

 absolute value) eCV at S can be expressed as eCV ¼

decreases smoothly from 1.0 to 0.0. The value of R is so chosen
in order to secure a sufficient number of neighbouring exper-
imental points to avoid singularity. To calculate the weights
using Equation 5, the radius of hypersphere of influence R is
taken as the distance between X and the farthest training point
from it. Now, the coefficient vector β (X) can be obtained as
 1
6: β̂ðX Þ ¼ FT WðX ÞF FT WðX Þy
It can be noted from the above equation that the coefficient
β (X) is a function of the location or position of X. Thus, the
procedure to calculate β (X), being a local approximation and
a moving process, is to perform a global approximation over
the entire design domain.
3. Proposed adaptive MSLM algorithm
for SRA
The aim of this study is to explore an algorithm based on the
maximin distance criterion combined with a leave-one-out CV
l
nitude (i.e.
y  ŷ . Thus, a vector of CV error magnitudes eCV ¼
l l
CV
fe1CV ; . . . ; eCV ; . . . ; epCV gT at all training points is obtained.
l
The maximum absolute error at data points obtained by the
leave-one-out CV approach is the maximum value of elements
of the vector eCV, which can be expressed as emax CV = max{eCV}.
In this regard, it is noted that the CV error magnitude is
usually higher than the magnitude of the actual error in
the prediction by the metamodel. Intuitively, it is unlikely that
the actual prediction error magnitude at any MCS point will
be greater than emax CV . Thereby, the sign of the LSF can only
be misrecognised if the error magnitude is higher than the
magnitude of the approximated LSF. Hence, it can be assumed
that the MCS points for which the predicted absolute values of
the approximated LSF are less than emax CV are the most likely to
be on the other side of the actual limit state as compared
to that predicted by the metamodel. Such points are expected
to have the maximum effect on the accuracy of reliability
estimates by a metamodel. Thus, the accuracy of the predic-
tions at these MCS points should be improved. These MCS
points are thus considered as candidates for adaptive sampling

4
Downloaded by [ Indian Institute of Technology Madras] on [03/09/23]. Copyright © ICE Publishing, all rights reserved.
 Structures and Buildings
in the proposed algorithm. It is desirable to include a
minimum number of adaptive training points into the DoE
for efficiency of the metamodelling approach. To fill the
subdomain with adaptive samples effectively, avoiding cluster-
ing, the minimum inter-distance of the training samples can
be maximised similarly to the widely used maximin distance
criterion (Johnson et al., 1990). If p data points are already
in the DoE then, to add data point p + 1, a set of can-
didate points Ω is first selected. An N
p matrix can be
obtained as
0 1
d11 ... d1p
B . .. .. C  
7: D¼B
@ .. .
C
. A; dij ¼ X i  S j 
dN1  dNp
where N is the number of candidate points in the set Ω and dij
is the Euclidean distance between the ith candidate point X i
and the jth data point S j.
The minimum value of the ith row of the matrix D (i.e. (d min)i-
= min{di1···dip}) is the Euclidean distance between the ith can-
didate point and the nearest data point. The candidate point
having the maximum value of d min is the next best choice for
the new data point and it is included in the DoE. At each iter-
ation step, two new data points are added. The MCS points
situated in the approximated unsafe domain, having the magni-
tude of the approximated LSF less than the value of emax CV , are
considered as candidate samples to add as the first point.
Based on the mentioned maximin distance criterion, one point
from this domain is added to the existing DoE. To add the
next point, the MCS points having the magnitude of the
approximated LSF less than the value of emax CV are selected
from the approximated safe domain as candidate samples.
Thus, the latest two adaptive samples (one having the predicted
LSF greater than zero and the other less than zero) are
included in the DoE. The actual responses corresponding to
the two new points are evaluated to update the training
dataset. Subsequently, approximation of the LSF by the
MLSM is updated with the current MLSM-based response
surface. Now, a new reduced domain of candidate samples is
obtained based on the current approximation of LSF and the
modified error norm obtained by the leave-one-out CV
approach. Subsequently, two new points can be included in the
DoE. This enrichment of the DoE is continued until a suitable
stopping criterion is satisfied. In the present study, the stopping
criterion is decided based on the convergence in the estimated
PoF as
 
Pf;i1  Pf;i 
8:  0:05
Pf ;i
where Pf,i and Pf,i−1 are the PoFs estimated at iterations i and
i − 1, respectively.
Downloaded by [ Indian Institute of Technology Madras] on [03/09/23]. Copyright © ICE Publishing, all rights reserved.
Reliability analysis of structures by
iterative sequential sampling based
response surface
Kabasi, Roy and Chakraborty
The algorithm is summarised as follows.
& Step 1. An initial single-shot DoE is constructed by Latin
hypercube sampling over the entire physical domain of the
input variables.
& Step 2. The metamodel is trained by all the data points of
the DoE.
& Step 3. Using the leave-one-out CV approach, note the
absolute maximum error emax CV .
& Step 4. The LSFs at all the MCS points are predicted by
the metamodel to estimate the PoF.
& Step 5. The reduced space is built by the MCS points
having the absolute value of the predicted response less
than emax
CV .
& Step 6. Two points from the reduced space are selected
based on the maximin distance criterion, each one on
either side of the approximate limit state.
& Step 7. The actual responses corresponding to the new
points are evaluated and these two data points are added
to update the DoE set.
& Step 8. Go to step 2 until convergence of the PoF is
observed.
& Step 9. The final reliability index is calculated
corresponding to the converged PoF.
A flowchart explaining implementation of the proposed algor-
ithm is provided in Figure 1.
4. Numerical study
The effectiveness of the proposed adaptive MLSM approach
based on sequential updating of the training dataset is now
elucidated numerically using two examples.
4.1 Example 1: Sphere subjected to internal pressure
This first example is to approximate the von Mises stress of a
hollow sphere under internal pressure (Figure 2). The material
is homogenous without spatial variety. The LSF (g) is defined
based on the equivalent von Mises stress (σeq)
  
3p0 r31
9: g ¼ fy  σ eq ; σ eq ¼
2 r1  r30
3
where fy is the yield stress of the material of the hollow sphere,
p0 is the internal pressure and r0 and r1 are the internal and
external radii, respectively. p0, r0 and r1 are assumed to follow
a lognormal distribution with mean values of 1200 MPa,
50 mm and 100 mm, respectively. The coefficients of variation
of all the random parameters are assumed to be 7.5%. The
physical limits of the random variables are taken as the
mean ± 3SD (SD being the standard deviation).
To study the effectiveness of the proposed adaptive MLSM,
the response of the sphere under internal pressure was
5
 Structures and Buildings Reliability analysis of structures by
iterative sequential sampling based
response surface
Kabasi, Roy and Chakraborty

Perform leave-one-out CV
Train the metamodel
Initial DoE approach to note
based on the current DoE
the absolute maximum error

Predict the LSF at all MCS

Evaluate the actual responses points by the metamodel
corresponding to the new point sand add
these two points to the existing DoE

Estimate the PoF

Select two MCS points from the reduced

space based on maximin criterion, each
one on either side of the limit state
No
Converged?

Build the reduced space from the MCS

points having the approximated LSF Yes
value less than the noted CV error
Obtain the final PoF
and the corresponding
reliability index

Figure 1. Flowchart of proposed algorithm

3.8
DMCS
r1 MLSM-adaptive_initial
3.6
MLSM-adaptive_final
p0 MLSM-equivalent_DoE
3.4
r0
Reliability index

fy 3.2

3.0

Figure 2. Schematic diagram of hollow sphere under internal 2.8

pressure
2.6

approximated by the proposed MLSM-based response surface.
To perform the MCS study, 105 random samples were gen-
erated according to the assumed PDF for each variable. The
reliability results were obtained for varying values of fy. To
study the effect of the size of the initial DoE, reliability results
were obtained using the proposed MLSM-based response
surface starting with three different sizes of the initial DoE
(11, 14 and 17 samples); the results are shown in Figures 3–5,
respectively. In these figures, the results obtained by the
MLSM-based metamodel using the initial DoE and those
obtained from the proposed MLSM-based metamodel with
converged DoE after necessary iterations are denoted as
MLSM-adaptive_initial and MLSM-adaptive_final, respect-
ively. For a meaningful comparison, the reliability results were
also obtained by the MLSM trained by an equivalent single-
shot DoE with the same number of training points as used to
obtain the final converged results by the proposed adaptive
MLSM (denoted as MLSM-Equivalent-DoE in Figures 3–5).
The results of the most accurate direct MCS (denoted DMCS)
2.4
267 268 269 270 271 272 273 274
fy
Figure 3. Comparison of estimated reliability indices with varying
fy considering 11 initial training samples
are considered as the benchmark. It can be readily seen that
the reliability results obtained by the proposed approach were
much better than those obtained by the equivalent single-shot
DoE with the same number of training points as required
by the proposed approach. It is important to note that the
improved performance of the proposed MLSM-based
approach was observed for all the DoE configurations, which
shows the robustness of the proposed algorithm.
Figure 6 shows the absolute percentage errors in estimating
reliability indices for each step of the proposed MLSM
approach (denoted as MLSM-adaptive_i for the ith step)

6
Downloaded by [ Indian Institute of Technology Madras] on [03/09/23]. Copyright © ICE Publishing, all rights reserved.
 Structures and Buildings Reliability analysis of structures by
iterative sequential sampling based
response surface
Kabasi, Roy and Chakraborty

3.8 18
DMCS MLSM-equivalent-DoE
16 MLSM-adaptive_1 MLSM-adaptive_2
3.6 MLSM-adaptive_initial
MLSM-adaptive_3 MLSM-adaptive_4

Error in reliability index: %

MLSM-adaptive_final 14
3.4 MLSM-equivalent_DoE
Reliability index

12
3.2 10

3.0 8
6
2.8
4
2.6
2
2.4 0
267 268 269 270 271 272 273 274 267 268 269 270 271 272 273 274
fy fy

Figure 4. Comparison of estimated reliability indices with varying Figure 6. Comparison of absolute percentage error in estimating
fy considering 14 initial training samples reliability indices for each iteration step of the proposed MLSM
approach considering 11 initial training data points with varying fy

3.8
DMCS

50 m
3.6 MLSM-adaptive_initial
MLSM-adaptive_final 15
3.4 16
MLSM-equivalent_DoE
Reliability index

17 14
3.2 8

25 m
21.65 m
3.0 3 2 7
9 13
18
2.8 4 1 x

2.6 10 5
6 12 24
19
2.4
267 268 269 270 271 272 273 274 11
20 23
fy 22
21
Figure 5. Comparison of estimated reliability indices with varying
fy considering 17 initial training samples 12.5 m
25 m
y 43.3 m

starting with 11 initial training samples and the corresponding P1

MLSM-Equivalent-DoE. It can be seen that the errors
P2 P2 P2 P2 x
reduced dramatically after the first iteration by the proposed
2m

MLSM-based metamodelling approach. The improved capa-

8.216 m

bility of reliability estimates using the proposed approach with

only four iterations can be readily noted from the figure. The
final estimate of reliability using the proposed approach was
very close to that obtained by the direct MCS technique (a z
deviation of around 1–2%). Comparisons of the absolute per-
Figure 7. Schematic diagram of space dome truss (redrawn from
centage of errors for the other two initial DoEs are not shown
Keshtegar (2017))
here for brevity, but the observations were similar.

4.2 Example 2: A space dome truss

The second example involving an implicit LSF is a space of all the bars, the section areas of the top radial bar numbers
dome truss (Keshtegar, 2017), a schematic illustration of which 1–6 (A1), the peripheral bar numbers 7–12 (A2) and the
is shown in Figure 7. The Young’s modulus (E) of the material bottom inclined bar numbers 13–24 (A3), the point load P1 at

7
Downloaded by [ Indian Institute of Technology Madras] on [03/09/23]. Copyright © ICE Publishing, all rights reserved.
 Structures and Buildings Reliability analysis of structures by
iterative sequential sampling based
response surface
Kabasi, Roy and Chakraborty

the centre node and the point load P2 at each of the six nodes 3.6
DMCS
of the middle hexagon (see Figure 7) were considered as the
MLSM-adaptive_initial
six independent random variables. The statistical properties of 3.2 MLSM-adaptive_final
these variables are shown in Table 1. The implicit LSF is MLSM-equivalent_DoE

Reliability index
defined as 2.8

 
10: g ¼ Δallow  ΔzP1  2.4

2.0
where ΔzP1 is the maximum vertical displacement of the node
under load P1 and Δallow is the allowable maximum displace- 1.6
ment. The Ansys mechanical APDL module was employed to
obtain the maximum displacement ΔzP1, which is necessary for
0.01 0.01025 0.0105 0.01075 0.011 0.01125 0.0115 0.01175
evaluating the LSF.
Δallow

As in example 1 (Section 4.1), three initial DoEs consisting of

Figure 9. Comparison of reliability indices for different values of
20, 25 and 30 training data points were constructed within the Δallow considering 25 initial training samples
physical domain (mean ± 0.3
mean) of the random variables.
For the MCS, 105 random simulation samples were generated.
The reliability results for different values of Δallow estimated by
3.0
the proposed MLSM-based metamodels starting with 20, 25 DMCS
2.8 MLSM-adaptive_initial
and 30 initial samples are compared in Figures 8–10, respect-
MLSM-adaptive_final
ively. The absolute percentage errors in the reliability results 2.6 MLSM-equivalent_DoE
with varying Δallow when compared with the direct MCS using
Reliability index

2.4

Table 1. Details of the random variables of the space dome truss 2.2

Probability distribution 2.0

Random
variable Type Mean Coefficient of variation 1.8

A1 Normal 0.013 m2 0.1 1.6

A2 Normal 0.01 m2 0.1
0.01 0.01025 0.0105 0.01075 0.011 0.01125 0.0115 0.01175
A3 Normal 0.016 m2 0.1
Δallow
E Normal 205 GPa 0.05
P1 Gumbel max. 20 kN 0.15
P2 Gumbel max. 10 kN 0.12 Figure 10. Comparison of reliability indices for different values of
Δallow considering 30 initial training samples

3.6 the proposed adaptive approach with 30 initial data samples

DMCS
MLSM-adaptive_initial after each iteration and the corresponding equivalent single-
3.2 MLSM-adaptive_final shot MLSM-based reliability results are shown in Figure 11.
MLSM-equivalent_DoE
Observations similar to those noted for example 1 can also be
Reliability index

2.8
noted for this example. The errors were dramatically reduced
by the proposed MLSM approach even after the first enrich-
2.4
ment of the DoE.

2.0
5. Summary and conclusions
1.6 A new adaptive MLSM-based RSM based on a maximin
space-filling design criterion combined with a leave-one-out CV
based error norm is proposed for improved reliability estimates
0.01 0.01025 0.0105 0.01075 0.011 0.01125 0.0115 0.01175
of structures. The algorithm relies on two points being added
Δallow
after each iteration. The maximin distance criterion combined
with a CV based specific error norm ensures that the new data
Figure 8. Comparison of reliability indices for different values of
Δallow considering 20 initial training samples points added sequentially are sufficiently close to the actual
limit state and, at the same time, are sufficiently far away from

8
Downloaded by [ Indian Institute of Technology Madras] on [03/09/23]. Copyright © ICE Publishing, all rights reserved.
 Structures and Buildings
22
MLSM-equivalent-DoE
20 MLSM-adaptive_1
18 MLSM-adaptive_2
Error in reliability index: %
16 MLSM-adaptive_3
14
12
10
8 (2017) A novel reliability technique for implementation of
6 performance-based seismic design of structures. Engineering
4
2 based metamodeling for seismic reliability analysis of structures.
0 Applied Mathematical Modelling 64: 584–602.
0.01 0.01025 0.0105 0.01075 0.011 0.01125 0.0115 0.01175
Δallow
Figure 11. Comparison of absolute percentage errors in
obtaining reliability indices for different values of Δallow
considering 30 initial training data samples
existing data points. The algorithm adds one new data point
from a safe domain and another from an unsafe domain
to reduce the bias in response approximation near the limit
state. The results of numerical studies using two examples
showed the improved reliability estimation capability of the pro-
posed algorithm compared with the conventional approach con-
sidering the direct MCS-based results as the benchmark
solution. The improved capability of estimating reliability with
different sets of training data clearly reveals the robustness of the
proposed approach. It is worth mentioning here that, to satisfy
both criteria, the algorithm does not use any heuristic weighting
scheme, but instead uses a step-by-step approach. The algorithm
can thus be applied generically to any reliability analysis
problem. Although the proposed algorithm was tested for a
MLSM-based adaptive metamodel, it can be readily applied to
other adaptive metamodelling approaches (kriging, ANNs,
SVMs, and so on). This does, however, require further study.
REFERENCES
Au SK and Beck JL (2001) Estimation of small failure probabilities in
high dimensions by subset simulation. Probabilistic Engineering
Mechanics 16(4): 263–277.
Blatman G and Sudret B (2011) Adaptive sparse polynomial chaos
expansion based on least angle regression. Journal of
Computational Physics 230(6): 2345–2367.
Bucher CG and Bourgund U (1990) A fast and efficient response surface
approach for structural reliability problems. Structural Safety 7(1):
57–66.
Ditlevsen O and Madsen HO (1996) Structural Reliability Methods.
Wiley, Chichester, UK.
Dymiotis C, Kappos AJ and Chryssanthopoulos MK (1999) Seismic
reliability assessment of RC frames with uncertain drift and
member capacity. Journal of Structural Engineering 125(9), https://
doi.org/10.1061/(ASCE)0733-9445(1999)125:9(1038).
Echard B, Gayton N and Lemaire M (2011) AK-MCS: an active learning
reliability method combining kriging and Monte Carlo simulation.
Structural Safety 33(2): 145–154.
Downloaded by [ Indian Institute of Technology Madras] on [03/09/23]. Copyright © ICE Publishing, all rights reserved.
Reliability analysis of structures by
iterative sequential sampling based
response surface
Kabasi, Roy and Chakraborty
Elhewy AH, Mesbahi E and Pu Y (2006) Reliability analysis of structures
using neural network method. Probabilistic Engineering Mechanics
21(1): 44–53.
Farag R and Haldar A (2016) A novel reliability evaluation method for
large engineering systems. Ain Shams Engineering Journal 7(2):
613–625.
Faravelli L (1989) Response-surface approach for reliability analysis.
Journal of Engineering Mechanics 115(12): 2763–2781.
Gaxiola-Camacho JR, Azizsoltani H, Villegas-Mercado FJ and Haldar A
Structures 142: 137–147.
Ghosh S, Roy A and Chakraborty S (2018) Support vector regression
Goswami S, Ghosh S and Chakraborty S (2016) Reliability analysis of
structures by iterative improved response surface method.
Structural Safety 60: 56–66.
Guimarães H, Matos JC and Henriques AA (2018) An innovative
adaptive sparse response surface method for structural reliability
analysis. Structural Safety 73: 12–28, https://doi.org/10.1016/
j.strusafe.2018.02.001.
Haldar A and Mahadevan S (2000) Probability, Reliability, and Statistical
Methods in Engineering Design. Wiley, New York, NY, USA.
Johnson ME, Moore LM and Ylvisaker D (1990) Minimax and maximin
distance designs. Journal of Statistical Planning and Inference
26(2): 131–148.
Kang SC, Koh HM and Choo JF (2010) An efficient response surface
method using moving least squares approximation for structural
reliability analysis. Probabilistic Engineering Mechanics 25(4):
365–371.
Kaymaz I (2005) Application of kriging method to structural reliability
problems. Structural Safety 27(2): 133–151.
Keshtegar B (2017) A hybrid conjugate finite-step length method for
robust and efficient reliability analysis. Applied Mathematical
Modelling 45: 226–237.
Kim C, Wang S and Choi KK (2005) Efficient response surface modeling
by using moving least-squares method and sensitivity.
AIAA Journal 43(11): 2404–2411.
Kohavi R (1995) A study of cross-validation and bootstrap for
accuracy estimation and model selection. In Proceedings of the
14th Fourteenth International Joint Conference on Artificial
Intelligence (Mellish CS (ed.)). Morgan Kaufmann Publishers,
San Francisco, CA, USA, vol. 2, pp. 1137–1145.
Kwon OS and Elnashai A (2006) The effect of material and ground
motion uncertainty on the seismic vulnerability curves of RC
structure. Engineering Structures 28(2): 289–303.
Lagaros ND, Tsompanakis Y, Psarropoulos PN and Georgopoulos EC
(2009) Computationally efficient seismic fragility analysis
of geostructures. Computers & Structures 87(19–20):
1195–1203.
Li H, Lü Z and Yue Z (2006) Support vector machine for structural
reliability analysis. Applied Mathematics and Mechanics 27(10):
1295–1303.
Liu YW and Moses F (1994) A sequential response surface method and
its application in the reliability analysis of aircraft structural
systems. Structural Safety 16(1–2): 39–46.
Melchers RE (1999) Structural Reliability Analysis and Prediction, 2nd
edn. Wiley, Chichester, UK.
Rajashekhar MR and Ellingwood BR (1993) A new look at the response
surface approach for reliability analysis. Structural Safety 12(3):
205–220.
Richard B, Cremona C and Adelaide L (2012) A response surface method
based on support vector machines trained with an adaptive
experimental design. Structural Safety 39: 14–21.
9
 Structures and Buildings
Roussouly N, Petitjean F and Salaun M (2013) A new adaptive response
surface method for reliability analysis. Probabilistic Engineering
Mechanics 32: 103–115.
Roy A and Chakraborty S (2020) Support vector regression based
metamodel by sequential adaptive sampling for reliability
analysis of structures. Reliability Engineering & System Safety 200:
106948.
Roy A, Manna R and Chakraborty S (2019) Support vector regression
based metamodeling for structural reliability analysis.
Probabilistic Engineering Mechanics 55: 78–89.
Shinozuka M (1983) Basic analysis of structural safety. Journal of
Structural Engineering 109(3): 721–740.
10
Downloaded by [ Indian Institute of Technology Madras] on [03/09/23]. Copyright © ICE Publishing, all rights reserved.
Reliability analysis of structures by
iterative sequential sampling based
response surface
Kabasi, Roy and Chakraborty
Taflanidis AA and Cheung SH (2012) Stochastic sampling using
moving least squares response surface approximations.
Probabilistic Engineering Mechanics 28: 216–224.
Wong SM, Hobbs RE and Onof C (2005) An adaptive response surface
method for reliability analysis of structures with multiple loading
sequences. Structural Safety 27: 287–308.
Xiao NC, Zuo MJ and Zhou C (2018a) A new adaptive sequential sampling
method to construct surrogate models for efficient reliability
analysis. Reliability Engineering & System Safety 169: 330–338.
Xiao NC, Zuo MJ and Guo W (2018b) Efficient reliability analysis based
on adaptive sequential sampling design and cross-validation.
Applied Mathematical Modelling 58: 404–420.
How can you contribute?
To discuss this paper, please email up to 500 words to the
editor at journals@ice.org.uk. Your contribution will be
forwarded to the author(s) for a reply and, if considered
appropriate by the editorial board, it will be published as
discussion in a future issue of the journal.
Proceedings journals rely entirely on contributions from the
civil engineering profession (and allied disciplines).
Information about how to submit your paper online
is available at www.icevirtuallibrary.com/page/authors,
where you will also find detailed author guidelines.