Optim Eng (2006) 7:297–316

DOI 10.1007/s11081-006-9973-8

A study using Monte Carlo Simulation for failure

probability calculation in Reliability-Based
Optimization

Dhanesh Padmanabhan · Harish Agarwal ·

John E. Renaud · Stephen M. Batill

Received: 10 June 2004 / Revised: 3 May 2005

C Springer Science + Business Media, LLC 2006


Abstract In this study, a Reliability-Based Optimization (RBO) methodology that

uses Monte Carlo Simulation techniques, is presented. Typically, the First Order Re-
liability Method (FORM) is used in RBO for failure probability calculation and this
is accurate enough for most practical cases. However, for highly nonlinear prob-
lems it can provide extremely inaccurate results and may lead to unreliable designs.
Monte Carlo Simulation (MCS) is usually more accurate than FORM but very com-
putationally intensive. In the RBO methodology presented in this paper, limit state
approximations are used in conjunction with MCS techniques in an approximate MCS-
based RBO that facilitates the efficient calculation of the probabilities of failure. A
FORM-based RBO is first performed to obtain the initial limit state approximations.
A Symmetric Rank-1 (SR1) variable metric algorithm is used to construct and update
the quadratic limit state approximations. The approximate MCS-based RBO uses a
conditional-expectation-based MCS, that was chosen over indicator-based MCS be-
cause of the smoothness of the probability of failure estimates and the availability of
analytic sensitivities. The RBO methodology was implemented for an analytic test
problem and a higher-dimensional, control-augmented-structure test problem. The re-
sults indicate that the SR1 algorithm provides accurate limit state approximations (and
therefore accurate estimates of the probabilities of failure) for these test problems. It
was also observed that the RBO methodology required two orders of magnitude fewer
analysis calls than an approach that used exact limit state evaluations for both test
problems.

Keywords Reliability analysis . Reliability-Based Optimization . Monte Carlo

Simulation . Importance sampling

D. Padmanabhan () · H. Agarwal · J. E. Renaud · S. M. Batill

Department of Aerospace and Mechanical Engineering, University of Notre Dame, Notre Dame,
IN 46556, USA
e-mail: [email protected]
Springer
298 D. Padmanabhan et al.

1 Introduction

In a traditional design optimization, the design variables and other system parameters
are usually considered as deterministic inputs. The design obtained by performing
a deterministic design optimization is usually limited by selected design constraints,
leaving little or no latitude for uncertainties. The resulting deterministic optimum has a
high probability of violating the design constraints, due to the influence of uncertainties
inherently present during the modelling and manufacturing phases of the artifact and
due to uncertainties in the operating environment of the artifact. These uncertainties
include variations in certain parameters (e.g., dimensions, material properties, loads),
and model uncertainties and errors associated with the numerical analysis tools used
for simulation-based design (Oberkampf et al., 2000). In this paper, uncertainties are
modelled as continuous random variables. In applications such as the design of struc-
tures, many of the design constraints represent various failure modes such as yielding,
buckling, fatigue, collapse etc. Designs resulting from deterministic optimization can
be unreliable and might lead to catastrophic failure. Uncertainties in simulation-based
design are inherently present and need to be accounted for in the design optimization
process. Reliability-Based Optimization (RBO) is a methodology that addresses this
problem.
Reliability-Based Optimization (RBO) (Enevoldsen and Sørensen, 1994) is usu-
ally expressed in a form where one is required to minimize a cost function subject to
certain reliability (i.e. probabilistic) constraints and deterministic constraints. The reli-
ability constraints ensure that the probabilities of failure with respect to various failure
modes are below acceptable levels. The probabilities of failure are usually obtained us-
ing standard reliability analysis methods (Haldar and Mahadevan, 2000; Ditlevsen and
Madsen, 1996) like the First Order Reliability Methods (FORM), the Second Order
Reliability Methods (SORM), and Monte Carlo Simulation (MCS) techniques. The
application of most of these reliability methods require the computation of a Most Prob-
able Point (MPP). FORM and SORM are based on linear and quadratic approximations
of the limit state functions at the MPP, respectively. RBO is typically performed with
FORM-based reliability constraints. Variations of FORM-based RBO have been pro-
posed by researchers (Tu and Choi, 1999; Kirjner-Neto et al., 1998; Kuschel and
Rackwitz, 1997) that address the issues of computational cost and poor robustness of
traditional optimization methods used for FORM analysis. Approximation concepts
have also been been used to address the issue of computational cost for FORM-
based RBO (Grandhi and Wang, 1998). FORM has been observed to perform well
for many practical design applications. But the probability of failure estimated using
FORM, can be inaccurate (either under-conservative or over-conservative) and might
lead to unacceptable designs, especially for nonlinear problems with low or medium
reliability.
MCS techniques are typically more accurate than FORM. The major difficulty in
using MCS techniques is that they can be computationally expensive. Moreover, most
MCS techniques cannot be incorporated within a gradient-based optimizer for RBO,
because of the unavailability of analytic sensitivities and non-smoothness of proba-
bility of failure estimates with design changes. In RBO, the probabilities of failure
corresponding to the failure modes have to be computed at a number of designs visited
by the optimizer. Therefore, using MCS techniques directly with the simulation tools,
Springer
A study using Monte Carlo Simulation for failure probability calculation 299

assumed to be expensive themselves, is highly impractical. There are some studies

that incorporate MCS techniques in RBO in an efficient manner using approximation
concepts (Oakley et al., 1998) but they do not address the issue of suitability of MCS
techniques for RBO using gradient-based optimizers. The study presented in this pa-
per deals with the selection of MCS techniques for RBO and the development of RBO
methodologies that incorporate these techniques in a computationally efficient manner
using approximation concepts.

2 Reliability-Based Optimization

Reliability-Based Optimization (RBO) is a methodology of finding optimum designs,

characterized by a low probability of failure. A typical RBO formulation involves the
minimization of a cost function (e.g., weight of a structure) subject to reliability con-
straints and certain deterministic constraints. This can be mathematically represented
by

min f (d, p), (1)

d

s.t. grbo (d, p) ≥ 0, (2)

g Dj (d, p) ≥ 0 for j = 1, . . . , M D , (3)
Bounds: dl ≤ d ≤ du , (4)

where d is the vector consisting of unknown design variables and p is the vector con-
sisting of fixed parameters of the optimization problem. f is the cost function, g Dj is
the jth deterministic constraint and M D is the number of deterministic constraints. De-
terministic constraints arise from other design considerations such as cost, marketing
etc. grbo represent constraints on reliability indices with respect to various component
failure modes or a single constraint on reliability index for an overall system failure
mode. The overall system failure mode can be a series, parallel or a combination of se-
ries and parallel of all the component failure modes (Enevoldsen and Sørensen, 1993).
There are also other system failure criteria, for example a case where a system is said
to have failed if K out of N components fail. The discussion in this paper will be re-
stricted to reliability constraints with component failure modes alone. The component
failure modes are represented by limit state functions, giR , i = 1, . . . , M R , where M R
is the number of component failure modes. Examples of limit state functions are stress
constraints, frequency constraints and structural displacement constraints. If giR < 0,
failure has occurred with respect to the ith component failure mode and vice-versa, if
giR > 0 the failure has not occurred. giR = 0 is usually referred to as the limit state.
In a deterministic design optimization, where uncertainties are not considered, these
limit state functions are treated as deterministic constraints. The component reliability
constraints are constraints on the reliability indices corresponding to the component
failure modes. The reliability techniques used to compute these reliability indices are
discussed in the next section. The component reliability constraints are given by the
Springer
300 D. Padmanabhan et al.

following equation:

girbo = βi − βreqd,i for i = 1, . . . , M R , (5)

where βi is the reliability index and βreqd,i is the required reliability index correspond-
ing to the ith failure mode. The reliability index is directly related to the corresponding
probability of failure as Pi = (−βi ), where  is the one-dimensional Gaussian cu-
mulative distribution function. The reliabilty index is inversely proportional to the
probability of failure.

2.1 Reliability analysis

Standard reliability techniques are required to estimate the probabilities of failure

(Haldar and Mahadevan, 2000; Ditlevsen and Madsen, 1996). In these techniques,
the uncertainties are modelled as continuous random variables, X 1 , X 2 , . . . , X n with
known joint probability density function. The uncertain variables can be denoted by
the vector X, where X = {X 1 , X 2 , . . . , X n }. x = {x1 , x2 , . . . , xn } denotes an instanti-
ation of random vector X and equivalently x1 , x2 ,. . . , xn denote joint instantiations of
random variables X 1 , X 2 ,. . . , X n respectively. The joint probability density function
(PDF) of all the random variables is given by f X (x; θ), where θ is a vector whose
elements are distribution parameters such as means, modes, standard deviations and
coefficients of variation of X. θ consist of variable distribution parameters, θ d , that the
designer is allowed to vary, and fixed distribution parameters, θ p . The elements of the
design vector, d, consist of the variable distribution parameters and other variables, η.
p consist of the fixed distribution parameters and other fixed parameters, pm . Mathe-
matically, the relationships among θ, d, p and η can be represented by the following
equations:

θ = {θ d , θ p }, (6)
d = {θ d , η}, (7)
p = {θ , p }.
p m
(8)

The ith limit state function is denoted as giR (X, η). The probability of failure due to
the ith component failure mode is Pi = P(giR (X, η) < 0). The subscript for g R will
be dropped in the remainder of this section and next section for convenience. It has to
be noted that the following discussions are valid for each limit state function.
In FORM, SORM, and some of the MCS techniques, the following procedures are
typically performed, which involve the calculation of a Most Probable Point (MPP)
in a transformed random variable space:

1. The original random variable vector, X is transformed to a standard normal

independent random variable vector, U, by a one-to-one transformation T (i.e.
U = T(X, θ)). A Rosenblatt transformation is typically used to transform the
random variables (Rosenblatt, 1952). The limit state function in the u space will
be denoted as g R (u, η).
Springer
A study using Monte Carlo Simulation for failure probability calculation 301

Fig. 1 Illustration of original to standard space transformation and MPP, FORM reliability index and
FORM and SORM approximations in standard space

2. A Most Probable Point (MPP) of failure, u∗ , is found by solving the following

optimization problem:

min uT u, (9)
u

s.t. g R (u, η) = 0. (10)

u∗ is the point on the g R (u, η) = 0 surface that is closest to the origin. The
g R (u, η) = 0 surface is usually referred to as a limit state surface and its evalu-
ation requires the computation of the inverse transformation, T−1 .
Figure 1 illustrates the transformation of the limit state function and the associated
failure domain from original space to standard space and also illustrates the MPP in
standard space for an example with two random variables, X 1 and X 2 . Various algo-
rithms exist to perform the MPP searches (Liu and Der Kiureghian, 1991). One of
the approaches is the Hasofer-Lind and Rackwitz-Fiessler (HL-RF) algorithm that is
based on a Newton-Raphson root solving approach. Variants of the HL-RF method
exist that use additional line searches. The family of HL-RF algorithms can exhibit
poor convergence for highly nonlinear and badly scaled problems, since they are based
on a first order approximation of the limit state function. A Sequential Quadratic Pro-
gramming (SQP) approach is often a more robust approach since it is based on a
second order approximation of the limit state function. Moreover, SQP has been suc-
cessfully applied to various nonlinear optimization problems and is readily available
in many optimization packages. Hence, SQP has been chosen for the implementation
studies presented in this paper. FORM and SORM are based on first and second order
approximations of the limit state surface respectively. The magnitude of the FORM
reliability index equals the distance of the MPP from the origin. Figure 1 shows the
approximations corresponding to FORM and SORM in the standard space and also
the FORM reliability index, β.
Springer
302 D. Padmanabhan et al.

RBO with FORM-based reliability constraints can be easily implemented using a

standard gradient-based optimizer because the reliability indices and their sensitivities
can be easily computed at a given design, d, as given in the following steps:
1. The FORM reliability index is computed using the following equation:

−∇u g R ∗
β= u . (11)
∇u g R 

2. The analytic sensitivities of β with respect to θ and η is computed using the

following equations:

∂β ∇u g R (u∗ ) ∂ T (x∗ , θ)
=− , (12)
∂θ ∇u g R (u∗ ) ∂θ
∂β 1 ∂g R (u∗ , η)
= . (13)
∂η ∇u g R (u∗ ) ∂η

The analytic sensitivities of FORM reliability index and the corresponding reli-
ability constraint with respect to d is given by the following equation:
 
∂g rbo ∂β ∂β ∂β
= = , . (14)
∂d ∂d ∂θ d ∂η

The probability of failure obtained from FORM can be inaccurate for cases where
the limit state surfaces are highly nonlinear due to nonlinearity of the transformation,
T. The inaccuracy in a FORM estimate usually tends to be larger for problems with
low or medium reliability (β ≈ 2–3). One of the extreme cases is the inaccuracy
of FORM reliability index due to the existence of multiple MPPs for a given limit
state. Another extreme case involves the counter-intuitive sensitivities of the FORM
reliability index with respect to standard deviations of random variables for certain
asymmetrical distribution types (Sørensen and Enevoldsen, 1993). One can obtain
better accuracy with SORM in some of the cases, but in general MCS techniques are
the most accurate of these techniques.

3 Monte Carlo Simulation techniques

A general expression for the probability of failure, P, is given by the multidimensional

integral

P= f Z (z) dz, (15)
f

where  f is the failure region, Z is the random or uncertain variable vector in either
the original space or the standard space, i.e. Z is either X or U and f Z is the joint
probability density function.  f is given by g R (z, η) < 0 for a single failure mode
Springer
A study using Monte Carlo Simulation for failure probability calculation 303

or component. MCS techniques estimate P given by Equation (15) by randomly

generating samples according to some sampling density function. A traditional way
to estimate P would be to use f Z as the sampling density function. The estimate of P,
denoted as P̂, obtained using such a traditional approach can be given by the following
equation:

1 N
P̂ = I (g R (zi ) < 0). (16)
N i=1

In the above equation N is the number of sample points, I is the indicator function
(1 if g R < 0, 0 otherwise) and zi for i = 1, . . . , N are the sample points generated
according to f Z . The number of samples required to estimate P using such an approach
with a reasonable accuracy is usually very high because P is typically in the range
10−2 − 10−6 . P can be efficiently estimated using Quasi Monte Carlo techniques
where sampling is done in important regions in z. The important regions include regions
in the failure domain that contribute significantly to the probability of failure. Example
of an important region is a region centered around the MPP of a limit state function.
Quasi MCS techniques can be classified as indicator-based MCS and conditional-
expectation-based MCS. In the following, a brief description of these two classes of
MCS techniques will be provided and some remarks on obtaining analytic sensitivities
of probabilities of failure will be made.

3.1 Indicator-based MCS

In indicator-based MCS techniques (Engelund and Rackwitz, 1993; Bjerager, 1989),

random samples of a simulation variable, V, is generated according to a sampling
density h V such that the samples are generated in important regions in z space that
have major contributions to P. Hence P and its estimate, P̂, for a single failure mode
or component can be written as follows:

f Z (z(v))
P= h V (v) dv, (17)
g R <0 h V (v)

1 N
  f Z (z(vi ))
P̂ = I g R (z(vi )) < 0 . (18)
N i=1 h V (vi )

In the above equation, vi denote the ith random sample of the random vector V and
z(vi ) denote the corresponding sample transformed to the z space or coordinates. An
example of an indicator-based importance sampling method involves sampling around
the MPP of g R (u, η) according to some sampling density function h V .
In an RBO driven by a gradient-based optimizer, sensitivities of P are required
with respect to the distribution parameters, θ d , and deterministic parameters, η, that
form the design variable vector, d. Sensitivities of P using finite differences will
require repetitions of the MCS and is computationally expensive. It is also not easy to
obtain accurate sensitivities using finite differences because of the discontinuities of
probability of failure that arise with changes in design. This issue will be explained
Springer
304 D. Padmanabhan et al.

later in this section. For indicator-based MCS, the sensitivity of P with respect to
θ can be easily obtained in X space since the failure domain is independent of θ
in this space. This can be obtained by analytically differentiating Equation (15) and
evaluating the resulting multidimensional integral using the same sampling density
function and samples. On the other hand, the sensitivity of P with respect to η is given
by the surface integral


∂P ∂g R fZ
=− dz. (19)
∂η g R (z,η)=0 ∂η ∇z g R 

The above integral is performed on the limit state surface g R (z, η) = 0. This cannot be
performed using the existing samples in most of the indicator-based MCS approaches,
since the samples do not lie on this surface. The adaptive importance sampling tech-
nique (Maes et al., 1993) is an exception where the limit state surface is approximated
as a paraboloid and the sampling densities are picked in such a way that the surface
integral can be performed in the curvilinear co-ordinates of the paraboloid. It should
be noted though that the sensitivity is based on an approximation of the limit state
surface.
A major issue that arises in applying these techniques in an RBO is that there are
discontinuities in the probability of failure estimates with changes in design variables.
This will present problems when performing an RBO using a gradient-based opti-
mizer. The discontinuities of probability of failure can be reduced to some extent by
eliminating the variations due to a varying number of sample points or varying seed
of the random number generator by fixing the number of sample points and the seed
to some initial state. But significant discontinuities in probability of failure will still
arise due to changes in curvature of the limit state surface with design changes. This
results in shifting of sample points across the limit state surface that causes discrete
jumps or drops in probability of failure estimates. Figure 2 illustrates the shifting of
sample points arising due to change in curvature of the limit state surface, for a case
where the sampling is performed at the MPP of the limit state surface and the number
of simulations and seed is not changed. To facilitate the illustration, the z co-ordinates

Fig. 2 Illustration of shifting of

importance-sampling based
MCS samples across limit state
that changes curvature with
change in design. The z
co-ordinates at the perturbed
design, d + δd has been
translated such that the MPPs
and samples coincide with the
original design d

Springer
A study using Monte Carlo Simulation for failure probability calculation 305

at the perturbed design, d + δd has been translated such that the MPPs and samples
coincide with the MPPs and samples at the original design d.

3.2 Conditional-expectation-based MCS

There are other MCS techniques that are based on conditional expectations (Bjer-
ager, 1989; Ditlevsen and Madsen, 1996), for example directional simulation and axis
orthogonal simulation. These methods are typically performed in u-space. In the cur-
rent paper, axis orthogonal simulation was used and hence will be discussed in some
detail. In an Axis-Orthogonal Simulation, the following steps are carried out for the
calculation of failure probability for a single failure mode:
1. MPP of the limit state function is computed.
2. A new co-ordinate system {v, vn } is obtained by a linear or affine transformation
of u-space. This transformation is also used to transform the random variables
U to {V† , Vn }. The simulation variable V is defined in v space (which is n − 1
dimensional) and the sampling density is h V . It should be noted that V† and V are
different random variables. For a single failure mode, the vn axis is picked such
that it coincides with the vector joining the origin and the MPP, see Figure 3. The
linear transformation from u to {v, vn } can be accomplished by a Gram-Schmidt
Orthogonalization process.
3. P is computed by sampling in V according to sampling density h V and by
performing an exact integral in vn conditioned on V = v. P and estimate of P
are given by the following equations:
 
f V† (v)
P= f Vn (vn |v) dvn h V (v) dv, (20)
Rn−1 g R <0 h V (v)
N 
1  f V† (vi )
P̂ = f V (vn ) dvn . (21)
N i=1 g R <0 n h V (vi )

Fig. 3 Illustration of axis

orthogonal simulation in
standard space showing the
{v, vn } co-ordinates and the
projection vn (v) corresponding
to a sample point v

Springer
306 D. Padmanabhan et al.

The integral in vn is a one dimensional integral. This is computed by solving for

the roots of g R ({vi , vn }) = 0 for each sample vi . The root solving can be done
using a Newton-Raphson scheme. The interval in which the root is to be located
can be chosen such that it corresponds to a range of high probability level, say
0 to 6. One systematic way to select the upper bound of the interval is to make
sure that the error obtained in the probability estimate calculation by neglecting
the solutions outside this range is a small fraction, say 0.01, of the probability
corresponding to the required reliability sought. So if κ is the upper bound, κ can
be calculated using the relation (−κ) = 0.01(−βreqd ). It is possible to have
cases with more than a single solution, but only a single solution is assumed in
the current study. The root solving is equivalent to projecting the sample points
along vn axis. If a single root exists for the root solving problem and if V is
chosen as a standard normal random variables then the estimate of probability
of failure can be given by following equation:

1 N
P̂ = (1 −  (vn (vi ))) . (22)
N i=1

Unlike the indicator-based MCS techniques, discontinuities due to “shifting” of

sample points across the limit state surface are absent for this MCS technique, assuming
that the number of samples and the seeds are fixed. But discontinuities can still arise
especially when there is a change in the number of samples that are projectable onto the
limit state surface. This happens when the solution of the 1-D root solving problem, for
a given sample, is also a local minimum for g R along the direction vn . The magnitude
of variation in a probability of failure estimate mainly depends on the magnitude of vn
at which this discontinuity arises. The larger this value, the smaller the variation. These
discontinuities are usually milder than the discontinuities that arise in indicator-based
MCS techniques.
Unlike the indicator-based MCS techniques, all the sample points are strictly pro-
jected onto the limit state surface in the axis orthogonal simulation and consequently
the sensitivities of P̂ can be easily obtained. Hence axis orthogonal simulation was
selected as the MCS technique in this study. The sensitivities of P̂ with respect to d,
∂ P̂
∂d
depend on ∂v∂dn (vi )
and is given by the following equation:

∂ P̂ −1 
N
∂vn (vi )
= φ (vn (vi )) , (23)
∂d N i=1 ∂d

where φ is the one-dimensional Gaussian density function. ∂v∂d n (vi )

depend on the sen-
sitivities of g , with respect to d, evaluated at u(vi , vn ). ∂d is comprised of ∂v∂θ
R ∂v (v
n i ) n (vi )
d

and ∂v∂ηn (vi )

, which are given by the following equations:

∂T
∂vn (vi ) ∇u g R ∂θ d
=  , (24)
∂θ d ∇u g R ∂u(vi ,vn )
∂vn
Springer
A study using Monte Carlo Simulation for failure probability calculation 307

− ∂g
R
∂vn (vi ) ∂η
=  . (25)
∂η ∇u g R ∂u(vi ,vn )
∂vn

It has to be noted that the same sample points are used for computing these sensitivities.
Since the probability of failure and reliability index are related by P̂ = (−β), one
can compute the sensitivities of reliability index and reliability constraints using the
following relation:

∂g rbo ∂β −1 ∂ P̂
= =   . (26)
∂d ∂d φ  ( P̂) ∂d
−1

It has to be noted that the axis orthogonal simulation is based on the assumption that
only one MPP exists for each limit state function. Hence axis orthogonal simulation
will not be very effective in computing probabilities of failure for cases with multiple
MPPs. A directional simulation technique is a better choice for such class of problems.
In directional simulation, the samples are generated in polar co-ordinates in u space
and the samples are projected onto the limit state surface radially. If the samples are
uniformly distributed in polar co-ordinates then one can obtain good estimates of
probabilities of failure for limit states with multiple MPPs. But this technique would
require lot more simulations to be able to get good accuracies in probabilities of failure
than axis orthogonal simulation for limit states with single MPPs.

4 Approximate MCS-based RBO methodology

In this section, a new RBO methodology developed in this investigation is presented.

This method uses approximate MCS techniques. The MCS-based reliability analysis
could use exact limit state function evaluations but this will be very expensive for
problems with many random variables and design variables. Hence the main goal of
the current study is to reduce this computational expense. Approximation concepts
are used in the approach presented herein to address this issue. In this study, approx-
imations of the limit state functions are constructed, which are used to perform an
approximate MCS-based RBO. This is performed in an iterative fashion. The approx-
imate MCS-based RBO methodology presented here serves as a refinement of results
obtained from a FORM-based RBO approach. The FORM-based RBO is used in the
current approach to construct initial limit state approximations. It is important that
the FORM-based RBO be successful which basically requires that the computation
of MPP be performed using robust optimization algorithms (e.g. SQP). It is also as-
sumed that extreme cases such as the existence of multiple MPPs do not occur for
the application problems under consideration. The details of the main steps of the
approximate MCS-based RBO methodology follow.
1. FORM-based RBO: First, a FORM-based RBO, that solves Equations (1–4), is
performed with an initial design, dinit and with reliability constraints on com-
ponent reliability indices based on FORM. This requires computation of MPPs
and sensitivities of reliability indices with respect to d during each iteration. The
Springer
308 D. Padmanabhan et al.

FORM-based RBO is used to provide a good estimate of initial design for the
MCS-based RBO to be performed iteratively in steps 3–5. The result obtained
from FORM-based RBO be denoted as d0 .
2. Initial Limit State Approximations: Next, initial approximations for all limit state
functions are created. Each limit state approximation is constructed at its MPP
in {x, η} space, as presented in Oakley et al. (1998). One of the motivations to
construct the approximations in {x, η} space as opposed to approximations in
{u, η} space is that for applications like structural optimization problems, certain
classes of approximation functions are known to represent structural responses
fairly accurately and hence it may be easier to construct these approximations
in {x, η} space. Moreover, by constructing limit state approximations in {x, η}
space, the effect of non-linearity of the transformation T can be captured. In
the work presented here, the initial limit state approximations are constructed
using the limit state function and sensitivity information generated during each
iteration of the FORM-based RBO. The subscript for g R is dropped here in this
section for convenience. It is implied that the procedures presented in this step
and step 5 need to be performed for each limit state function. The following
steps are carried out to obtain the initial limit state approximations:
a. Start at k = 1. At the kth iteration of the FORM-based RBO, let η k and θ k
be the values of the deterministic design variable and distribution parameter
vectors respectively. Let u∗,k and x∗,k = T−1 (u∗,k , θ k ) be the MPPs in u and
x space, respectively. Let w = {x, η} and wk = {x∗,k , η k }.
b. A quadratic approximation, g̃ R,k (w), of the following form is constructed:

1
g̃ R,k (w) = g R,k + ∇w g R,k (w − wk ) + (w − wk )T Hk (w − wk ).
2

In the above equation, g R,k and ∇w g R,k are the known value and sensitivity
of g R at wk respectively, and Hk is the unknown Hessian of the quadratic
approximation. The Hessian is obtained using a variable metric algorithm,
that is used for updating Hessians in optimization algorithms based on Quasi-
Newton methods. A Symmetric Rank 1 (SR1) update algorithm (Nocedal
and Wright, 1999) is used in this work. SR1 is known to generate Hessians
accurately compared to other variable metric algorithms such as Broyden-
Fletcher-Goldfarb-Shanno (BFGS) algorithm. This update is essentially a
two-point update, where the Hessian is constructed by matching the exact
sensitivities at previous iteration, i.e. ∇w g̃ R,k (wk−1 ) = ∇w g R,k−1 . The update
is given by the following equation:

(qk − Hk−1 sk )(qk − Hk−1 sk )T

Hk = Hk−1 + , (27)
sk T (qk − Hk−1 sk )

T
where qk = ∇w g R,k − ∇w g R,k−1 and sk = wk − wk−1 . The initial guess for
the Hessian, H0 , is taken as zero matrix. The update is performed only if
T
|sk (qk − Hk−1 sk )| > ε1 . The choice of ε1 is important. The choice of a
small value for ε1 could result in large changes in the Hessians for small
Springer
A study using Monte Carlo Simulation for failure probability calculation 309

changes in the design and consequently result in poor approximations. This

was noticed in the second test problem and a choice of ε1 = 10−3 produced
good results for the test cases considered herein. The choice of ε1 can be
made based on the magnitude of sk . A relative version of the update
scheme can be used for this purpose such that the update is performed only if
T
|sk (qk − Hk−1 sk )| > ε1 max(1, sk ) and a typical value of ε1 = 10−6 can
be used. Such an update scheme would prevent the updates from happening
for very small changes in design.
c. Let K be the number of iterations taken by the FORM-based RBO. If k < K ,
repeat steps 2 and 3, else, stop; the approximation for each limit state after
the FORM-based RBO is g̃ R,K .

It has to be noted that the approximation type is not limited to quadratic and
additional sampling can be performed to build these approximations. In the
test problems considered, the SR1 variable metric algorithm was successful in
giving good approximations and hence is presented here. Moreover, the SR1
update is quite simple to implement. It also should be noted that the limit state
approximations being created during every iteration of the FORM-based RBO
can also be used to perform approximate FORM analysis to reduce computa-
tional costs. This was not done in the current study primarily because the main
focus here is to reduce the computational cost associated with the MCS-based
RBO. Then g̃ R,K obtained from steps (a)–(c) is set as the initial approximation
for the following iterative phase of the methodology, which starts at the iteration
counter j = 0.
3. Approximate MCS-based RBO: With d j as the starting point, perform an ap-
proximate MCS-based RBO, that solves Equations (1)–(4), using limit state
approximations, i.e. g̃ R,K + j , for the computation of grbo , and using exact eval-
uations for f and g D . The computations of MCS reliability constraints during
this RBO involve finding MPPs and computing P̂ and ∂∂dP̂ for each limit state
approximation using axis orthogonal simulation. The sampling density function
h V is chosen as standard normal. The seeds of the random number generator and
number of samples are kept fixed throughout this approximate RBO phase to
reduce the discontinuities in probability of failure. Additional move-limit con-
straints can be imposed on d, i.e. replace Equation (4) by − ≤ d − d j ≤ ,
where are the move-limits. Let the solution obtained from the approximate
RBO be denoted as d j+1 .
4. Convergence Check: The convergence is achieved if d j − d j+1  ≤ ε2 . ε2 is
set as 10−3 in the current study. If the convergence is achieved, stop, otherwise
proceed to next step. ε2 needs to be chosen based on the magnitudes of the design
variables. A relative version of the convergence criteria can be used where the
convergence is achieved if d j − d j+1  ≤ ε2 max(1, d j ). Many alternative
convergence criteria involving the relative change of cost function, the norm of
the gradient of the cost function and others that are typically used in gradient-
based optimizers can be used to supplement or replace the current criteria.
5. Update Limit State Approximations: The limit state approximations are updated
in the same fashion as described in step 2. To update the approximations, an exact
MPP search corresponding to each limit state function, g R , is performed at d j+1
Springer
310 D. Padmanabhan et al.

to obtain w K + j+1 , g R,K + j+1 , and ∇w g R,K + j+1 . Using this information and the
previous Hessian, H K + j , one can obtain an updated Hessian, H K + j+1 , from the
SR1 update given by Equation (27). The updated Hessian can be substituted in
Equation (27) to obtain the updated limit state approximation, g̃ R,K + j+1 . After
the approximations have been updated, increment j by 1 and go to step 3.

5 Implementation studies

The approximate MCS-based RBO methodology and an exact MCS-based RBO

method (for comparison purposes) were implemented for two test problems. The
exact MCS-based RBO method was performed by solving Equations (1–4) with the
optimum design obtained from a FORM-based RBO as the starting design. In the rest
of this section, the approximate MCS-based RBO methodology will be referred to as
the approximate approach and the exact MCS-based RBO will be referred to as the ex-
act approach. The first problem is an analytic test problem and the second problem is a
control-augmented-structure problem. The methodology was implemented in Matlab.
The Sequential Quadratic Programming (SQP) subroutine was used to perform all the
optimizations including the calculation of MPPs. The required component reliability
indices for all the limit state functions for both FORM-based RBO and MCS-based
RBO were set at 3.0 for both the test problems.

5.1 Analytic test problem

There are 2 design variables, d1 , d2 and one parameter, p, in the analytic test problem.
There are two random variables, X 1 and X 2 . There are two limit state functions and
one deterministic constraint. The details of the problem are as follows:

Cost Function: f = d12 + y1 (d1 , p, d2 ) + e−y2 (d1 , p,d2 ) .

Limit state functions: g1R = y1 (X 1 , X 2 , η)/2 − 1,
g2R = X 1 + 1.5η − 0.5X 2 − 2.5.
Constraints: girbo = βi − 3.0 for i = 1, 2,
g = y2 (d1 , p, d2 )/2 − 1,
D

where d1 = μ X 1 , p = μX2 = 0 and d2 = η.

d bounds: 0 ≤ d1 ≤ 10, 0 ≤ d2 ≤ 10.

y1 and y2 are obtained by solving the following system of coupled equations:

y1 (x1 , x2 , η) = x12 + η − 0.2y2 (x1 , x2 , η),

y2 (x1 , x2 , η) = y1 (x1 , x2 , η) + x1 + x2 .

X 1 and X 2 are chosen as independent and normal random variables with standard
deviation 0.5 each. g1R is a nonlinear function and g2R is a trivial linear function.
Springer
A study using Monte Carlo Simulation for failure probability calculation 311

Table 1 The initial design and RBO designs obtained using FORM-RBO, approxi-
mate and exact MCS-RBO approaches for Test Problem 1 are shown

MCS-RBO designs

Initial design FORM-RBO design Approximate Exact

d1 4.00 0.45 0.47 0.47

d2 5.00 2.58 2.54 2.54
f 35.32 2.73 2.71 2.71
gD 3.20 0.00 0.00 0.00
β1 3.0 (MCS 3.4) 3.0 3.0
β2 3.27 (MCS 3.27) 3.17 3.17
Analysis calls 1,300 1,500 62,700

Note. FORM-RBO gives a conservative design. The final designs obtained from ap-
proximate and exact approaches are close and the former requires 40 times fewer
analysis calls.

In both the proposed approximate approach and the exact approach, axis orthogonal
simulations using 100 samples for both limit state functions were performed. The
samples were generated according to standard normal sampling density functions.
The seed for the random number generator used to generate the samples was set to
the same initial state for all the iterations. The initial design, the FORM-based RBO
design and the final designs obtained from the approximate and exact MCS-based RBO
approaches are shown in Table 1. Table 1 also presents the reliabilities for the FORM-
RBO and MCS-RBO designs. In the column representing the FORM-RBO results,
the FORM estimate of the reliability for each limit state function is presented and an
MCS estimate at this design is also provided in parentheses. The values corresponding
to the active constraints are shown in bold.
In both the FORM-based RBO design and the MCS-based RBO design the relia-
bility constraints for g1R and g D were active. Though the FORM-based RBO design
and the final design obtained from the MCS-based RBO appear to be similar, the
MCS reliability indices were significantly different for these two designs. The MCS
reliability index for g1R was 3.4 at the FORM-based RBO design. The probability of
failure corresponding to a reliability index of 3.4 is 4 times lower than the probability
corresponding to the required reliability index of 3.0. Hence the FORM-based RBO
design was conservative and the MCS-based RBO methodology was able to lower
the reliability index to the desired value and was also able to lower the cost function,
f . The FORM result is conservative because the failure domain for g1R is convex in
standard space for the random variable distributions considered. It will be difficult to
say if this will be the case for other distributions. The final design obtained using the
approximate MCS-based RBO approach was identical to the result obtained from the
exact approach. The accuracies of the MCS reliability indices were verified by per-
forming confirmatory runs where the axis orthogonal simulations were performed with
larger number of samples. The exact approach that used exact limit state evaluations
required as many as 62,700 analysis calls whereas the approximate approach required
only 1,500 analysis calls (excluding confimatory runs), of which 1,300 analysis calls
were taken by the FORM-based RBO. The number of analysis calls required was 40
times lower than the exact approach considering both the FORM and the approximate
Springer
312 D. Padmanabhan et al.

Fig. 4 Control augmented

structure

MCS-based RBO. The number of analysis calls was almost 300 times lower than the
exact approach excluding the analysis calls required by the FORM-based RBO.

5.2 Control-augmented-structures problem

Figure 4 shows the control-augmented-structure as presented by Sobieszczanski-

Sobieski et al. (1990). The structure is a 5 element cantilever beam, numbered 1–5
from the free end to the fixed end, as shown in the Figure 4. There are three static
loads T1 , T2 and T3 and a dynamic force f (t), which is a ramp function. Controllers
A and B are designed as an optimal Linear Quadratic Regulator. This is a coupled
problem since the design of the controllers depend on the dynamic characteristics of
the structure, which in turn requires the weight of the controllers that are modelled as
a lumped mass on the structure. The cost function for this problem is the total weight
of the beam and the controllers.
There are 11 design variables in this test problem. The design variables and design
variable bounds for this test problem are given below as,

d = [b1 , b2 , b3 , b4 , b5 , h 1 , h 2 , h 3 , h 4 , h 5 , c]T ,
3 ≤ bi ≤ 36, i = 1, . . . , 5,
3 ≤ h i ≤ 36, i = 1, . . . , 5,
0.01 ≤ c ≤ 0.06,

where bi and h i are the breadth and height of the ith element respectively and c is the
damping matrix to stiffness matrix ratio (scalar).
For this problem, the following limit state functions were selected:

2
dl1
g1R = 1 − ,
50
2
dr1
g2R = 1 − ,
0.2
σi
R
g2+i = 1− , i = 1, . . . , 5,
σa

where dl1 and dr1 are the lateral and rotational static displacements of element 1
respectively, and σi is the maximum static stress in element i. The random variables
for this problem are the ultimate static stress, σa , density, ρ, modulus of elasticity,
E, and loads T1 , T2 and T3 . The distribution properties of the random variables are
given in Table 2. The distribution functions for σa , ρ and E were taken as lognormal
Springer
A study using Monte Carlo Simulation for failure probability calculation 313

Table 2 Test problem 2-random

variables Distribution Mean Std. Dev.

σa (psi) Lognormal 30,000 3,000

ρ (lb/in 3 ) Lognormal 0.1 0.01
E (ksi) Lognormal 10,500 1,050
T1 (lbs) Uniform 1,000 400
T2 (lbs) Uniform 5,000 2,000
T3 (lbs) Uniform 1,000 400

since these are non-zero quantities and the coefficients of variation were set at 0.1.
T1 , T2 and T3 were taken to be uniformly distributed with coefficients of variation set
at 0.4, which introduced considerable non-linearity in the problem that resulted in a
significant difference between the FORM and MCS reliabilities. It has to be noted that
the uncertainties were chosen to be independent for convenience of implementation.
In a real scenario, the uncertainties related to material properties could be dependent.
The purpose of this work was to illustrate the computational efficiency of the proposed
approximate MCS-based RBO methodology over an exact approach, primarily. But it
has to be noted that the reliability techniques such as FORM and the MCS technique
used in this work can be straightforwardly extended to problems with dependent
random variables because the Rosenblatt transformation can handle dependent random
variables (Rosenblatt, 1952). The reliability constraints for this problem are given by
the following equation:

girbo = βi − 3.0 for i = 1, . . . , 7.

In both the proposed approximate approach and the exact approach, axis orthogonal
simulations using 500 samples for all the limit state functions were performed. The
samples were generated according to standard normal sampling density functions.
The seed for the random number generator used to generate the samples was set to the
same initial state for all the iterations in the approximate RBO. The initial design, the
FORM-based RBO design and the final designs obtained from the approximate and
exact MCS-based RBO approaches are shown in Table 3. Table 3 also presents the re-
liabilities for the FORM-RBO and MCS-RBO designs. In the column representing the
FORM-RBO results, the FORM estimate of the reliability for each limit state function
is presented and an MCS estimate at this design is also provided in parentheses. The
values corresponding to the active constraints are shown in bold. The initial design
was selected as the result obtained from a deterministic optimization problem similar
to the RBO problem but with no reliability constraints and all limit state functions
treated as deterministic constraints. At this design, the limit state function g1R was
active and the reliability indices for all the limit state functions were way below the
required reliability index of 3.0 and were close to zero (50% probability of failure) for
certain limit state functions.
The final design obtained using approximate MCS-based RBO approch was similar
to the result obtained using the exact approach. At the FORM-based RBO design and
MCS-based RBO design, the reliability constraints for the static stress constraints,
i.e. giR , i = 3, . . . , 7, were active. It can be seen that the FORM-based RBO design
corresponds to a heavier structure than the final design. In the FORM design, the
Springer
314 D. Padmanabhan et al.

Table 3 The initial design and RBO designs obtained using FORM-RBO, approximate and
exact MCS-RBO approaches for test problem 2 are shown

MCS-RBO designs

Initial design FORM-RBO design Approximate Exact

bi , i = 1, . . . , 5 3.00 3.00 3.00 3.00

h1 3.70 3.74 3.70 3.70
h2 7.04 9.67 9.48 9.48
h3 9.81 13.56 13.30 13.30
h4 12.00 16.58 16.26 16.26
h5 13.84 19.14 18.77 18.76
c 0.06 0.06 0.06 0.06
f 1493.9 1926.6 1894.3 1894.1
β1 4.79 (MCS 5.07) 4.61 4.60
β2 3.33 (MCS 3.63) 3.26 3.26
β3 3.00 (MCS 3.18) 3.00 3.00
β4 , β5 , β6 3.00 (MCS 3.29) 3.00 3.00
β7 3.00 (MCS 3.30) 3.00 3.00
Analysis calls NA ≈ 2, 600 ≈ 273, 000

Note. FORM-RBO gives a conservative design. The final designs obtained from approximate
and exact approaches are close and the former requires 100 times fewer analysis calls.

MCS-computed reliability indices for the static stress constraints were as high as 3.3.
The probability of failure corresponding to a reliability of 3.3 is more than 2.5 times
lower than the probability corresponding to the required reliability index of 3.0. Hence
the design was conservative. As mentioned in the previous subsection, the FORM esti-
mates are probably conservative for this implementation study too, because the failure
domains are convex in standard space for the random variable distributions considered.
This need not necessarily be true for other random variable distributions. It was noted
for this problem that the choice of ε1 in SR1 update scheme greatly influenced the
accuracy of initial limit state approximation for g1R and consequently the prediction of
the corresponding MCS reliability index, which was extremely low. Moreover, the ac-
curacy of the limit state approximation for g1R deteriorated when large design changes
were made during the approximate MCS-based RBO phase, which consequently led
to extremely high reliability estimates. Hence the move-limits for the approximate
MCS-based RBO were reduced to ±0.5% of the design variable bounds, which forced
the approximate MCS-based RBO to make small design changes. For this choice of
move-limit size, the reliability index for g1R did not exhibit large variations between
iterations and smoothly converged to the reliability index obtained from the exact
MCS-based RBO approach. The approximate MCS-based RBO could predict the
reliability indices fairly well for other limit states and the accuracies in the reliability
indices were insensitive to the choice of move-limit size or ε1 . The acccuracies of the
MCS reliability indices were verified by performing confirmatory runs where the axis
orthogonal simulations were performed with larger number of samples. The approxi-
mate MCS-based RBO methodology required about 2,600 finite element analysis calls
(excluding confirmatory runs), whereas the exact MCS-based RBO approach that used
exact limit state evaluations required as many as 273,000 finite element analysis calls.
Springer
A study using Monte Carlo Simulation for failure probability calculation 315

The proposed methodology required 100 times fewer analysis calls than the exact
approach.

6 Conclusions

An approximate MCS-based RBO methodology that uses limit state approxima-

tions has been developed and successfully implemented for two test problems.
A conditional-expectation-based MCS technique is better suited for RBO than an
indicator-based MCS and hence it was used for the approximate RBO in the RBO
methodology presented in this study. For comparison purposes, results using exact
calculations of the limit states are also presented for both these test problems. The
SR1 update algorithm typically provided accurate limit state approximations that cap-
tured the trends of the limit state function. For both of the test problems, the proposed
methodology required nearly two orders of magnitude fewer analysis calls than an
approach that required exact evaluations of the limit states and this dramatically illus-
trates the potential of the methodology.
It is expected that the accuracy of the limit state approximations is very critical
for the success of this methodology. The current methodology provides no way of
determining directly the accuracy or goodness of the limit state approximations that
are constructed using SR1 algorithm. The limit state approximations can be com-
puted using other techniques such as those based on Design of Experiments, where
statistical goodness-of-fit measures such as R 2 can be used. This can be used in tan-
dem with a move-limit management strategy to further increase the robustness of the
methodology. In the proposed methodology, FORM-based RBO uses exact limit state
calculations. One could employ limit state approximations in FORM-based RBO to
further reduce limit state evaluations or analysis calls. The axis orthogonal simula-
tion technique used in the current study does not effectively calculate probabilities
of failure for problems with multiple MPPs since the simulations are performed after
the calculation of an MPP. Other MCS techniques such as directional simulation need
to be explored to address such problems. The current implementation also assumes
that only single solution exists for the projection of each sample on the limit state.
The implementation can be improved to account for multiple solutions that may arise
in non-linear examples. In the current study, it is assumed that the joint probability
distributions are available for the uncertainties. Typically the joint distributions are
difficult to obtain and only partial information such as correlations are available along
with individual marginal distributions. The MCS techniques also need to be extended
for such problems of practical importance. One possible approach to address such
problems would be to use the Morgenstern and Nataf joint distribution models that
can be easily evaluated in terms of known marginals and correlation coefficients (Liu
and Der Kiureghian, 1986).

Acknowledgements The authors would like to acknowledge the support provided by the Center for Applied
Mathematics and Aerospace and Mechanical Engineering Department at University of Notre Dame. The
authors would like to acknowledge the contributions made by Dr. Victor M. Pérez, former Graduate Research
Assistant at University of Notre Dame, currently working at General Electric, Mexico
Springer
316 D. Padmanabhan et al.

References

Bjerager P (1989) Methods for structural reliability computations. Technical report, University of California,
Berkeley. Lecture Notes for the course structural reliability: methods and applications
Ditlevsen O, Madsen HO (1996) Structural reliability methods. John Wiley & Sons, Inc
Enevoldsen I, Sørensen JD (1993) Reliability-Based Optimization of series systems of parallel systems. J
Structural Eng 119(4):1069–1084
Enevoldsen I, Sørensen JD (1994) Reliability-Based Optimization in structural engineering. Structural
Safety 15(3):169–196
Engelund S, Rackwitz R (1993) A benchmark study on importance sampling techniques in structural
reliability. Structural Safety 12:255–276
Grandhi RV, Wang L (1998) Reliability-based structural optimization using improved two-point adaptive
nonlinear approximations. Finite Elements in Analysis and Design 29
Haldar A, Mahadevan S (2000) Probability, reliability and statistical methods in engineering design. John
Wiley & Sons, Inc
Kirjner-Neto C, Polak E, Der Kiureghian A (1998) An outer approximations approach to reliability-based
optimal design of structures. J Optimiz Theor Appl 98(1):1–16
Kuschel N, Rackwitz R (1997) Two basic problems in reliability-based structural optimization. Math Meth
Operations Res 46(3):309–333
Liu P-L, Der Kiureghian A (1991) Optimization algorithms for structural reliability. Structural Safety
9:161–177
Liu P-L, Der Kiureghian A (1986) Multivariate distribution models with prescribed marginals and covari-
ances. Prob Eng Mech 1(2):105–112
Maes MA, Breitung K, Dupuis DJ (1993) Asymptotic importance sampling. Structural Safety 12:167–186
Nocedal J, Wright SJ (1999) Numerical optimization, springer series in operations research. Springer
Oakley DR, Sues RH, Rhodes GS (1998) Performance optimization of multidisciplinary mechanical systems
subject to uncertainties. Prob Eng Mech 13(1):15–26
Oberkampf WL, DeLand SM, Rutherford BM, Diegert KV, Alvin KF (2000) Estimation of total uncertainty
in modeling and simulation. Technical Report SAND2000-0824, Sandia National Laboratories
Rosenblatt M (1952) Remarks on a multivariate transformation. The Annals Math Stat 23(3):470–472
Sobieszczanski-Sobieski J, Bloebaum CL, Hajela P (1990) Sensitivity of control-augmented structure ob-
tained by a system decomposition method. AIAA J 29(2):264–270
Sørensen JD, Enevoldsen I (1993) Sensitivity weaknesses in application of some statistical distributions in
first order reliability methods. Structural Safety 12:315–325
Tu J, Choi KK (1999) A new study on reliability-based design optimization. J Mech Design, Trans ASME
121:557–564

Springer