Struct Multidisc Optim

DOI 10.1007/s00158-016-1418-1

RESEARCH PAPER

An efficient procedure for structural reliability-based robust

design optimization
Renato de S. Motta1 · Silvana M. B. Afonso1

Received: 6 February 2015 / Revised: 5 November 2015 / Accepted: 5 February 2016

© Springer-Verlag Berlin Heidelberg 2016

Abstract This paper focuses on the development of an Measure Approach (PMA) for reliability constraint assess-
optimization tool with the aim to obtain robust and reli- ment. Finally, Normal-Boundary Intersection (NBI) or
able designs in short computational time. The robustness Normalized Normal-Constraint (NNC) multiobjective opti-
measures considered here are the expected value and stan- mization techniques are employed to obtain fast and even
dard deviation of the performance function involved in the spread Pareto robust designs. To illustrate the application
optimization problem. When using these robustness mea- of the proposed tool, optimization studies are conducted for
sures combined, the search of optimal design appears as a a linear (benchmark) and nonlinear trusses problems. The
robust multiobjective optimization (RMO) problem. Reli- nonlinear example consider different loads level, explor-
able design addresses uncertainties to restrict the structural ing the material plasticity. The integrated tool prove to be
probability of failure. The mathematical formulation for very effective reducing the computational time by up to
the reliability based robust design optimization (RBRDO) five orders of magnitude, when compared to the solutions
problem is obtained by adding a reliability based constraint obtained via classical standard approaches.
into the RMO problem. As both, statistics calculations and
the reliability analysis could be very costly, approxima- Keywords Robust optimization · Probabilistic collocation
tion technique based on reduced-order modeling (ROM) method · Reliability-based robust design optimization ·
is also incorporated in our procedure. The selected ROM Multiobjective optimization · Reduced-order modelling ·
is the proper orthogonal decomposition (POD) method, Proper orthogonal decomposition
with the aim to produce fast outputs considering structural
non-linear behavior. Moreover, to obtain RBRDO designs
with reduced CPU time we propose others developments 1 Introduction
to be added in the integrated tool. They are: Probabilis-
tic Collocation Method (PCM) to evaluate the statistics of With the rapid growth of computational power, remark-
the structural responses and, also, an approximated reli- able progress has been made in the research area of
ability constraints procedure based on the Performance optimal structural design (Simpson et al. 2001; Arora
and Wang 2005; Spillers and MacBain 2009; Haftka and
Sobieszczanski-Sobieski 2009; Rao and Savsani 2012;
Cavazzuti 2012), mainly under consideration of determin-
istic models and deterministic parameters. However, in the
 Renato de S. Motta
[email protected] real world, uncertainty and randomness are prevalent in
the context of structural design due to the lack of accurate
Silvana M. B. Afonso
[email protected] data in the structural design, manufacturing process, mate-
rial models and properties, among other uncertainty sources
1 Department of Civil Engineering, UFPE. Rua Acadêmico (Bucher 2009). Commonly such uncertainties are included
Hélio Ramos, S/N, Cidade Universitária, Recife, Brazil in the design process by introducing simplified hypothesis
 R. D. S. Motta, S. M. B. Afonso

and safety or design factors. This tendency has been chang- Reliability based Constraints could be dealt by a reliabil-
ing due to the increase in the use of adequate numerical tools ity index approach (RIA) or by a performance measure
for dealing with this class of problems. approach (PMA) (Youn et al. 2003; Youn and Choi 2004a).
It is well known (Marczyk 2000; Beyer and Send- The first approach, needs the computation of reliability
hoff 2007; Schuëller and Jensen 2008; Doltsinis and Kang index or probability of failure for each new design, they are
2004) that optimization under a deterministic approach obtained here by MC or FORM methods.
generally leads to a final design whose performance may The RBRDO task can be computationally intense as
degrade significantly and or to a design that can violate it involves statistics calculations to compose the objective
constraints because of perturbations arising from uncer- function and constraints, together with reliability methods
tainties. Moreover, the optimum design obtained by the to evaluate the reliability constraints. Moreover, different
deterministic approach may not achieve the desired goal design solutions have to be obtained in order to compose
or may become infeasible due to the structural behaviour the Pareto frontier. The case of nonlinear structures analysis
scatter. with a consider number of degrees of freedom, makes the
In this scenario a better target is the one that explore RBRDO problem a very computational costly procedure. To
the uncertainties effect on the design. That is a feasible overcome that approximation techniques based on reduced-
design which is relatively invariant with respect to changes order modeling (ROM) approach are also incorporated in
in uncertain parameters. The process of finding such opti- our procedure via proper orthogonal decomposition (POD)
mum is referred to as Robust Design Optimization (RDO) method, for fast outputs of structural analysis with nonlinear
(Beyer and Sendhoff 2007; Schuëller and Jensen 2008; behavior.
Doltsinis and Kang 2004; Papadrakakis et al. 2005), in This paper focuses on the main issues related to the above
which feasibility improvement and variability reduction in described topics used to build a computational efficient inte-
the performance are the targets. In the other hand reliable grated tool to carry out robust and reliable structure designs.
designs address uncertainties to restrict the probability of The developed tool can be applied to a wide range of struc-
failure. The latter consideration into the problem mathemat- tures. Here, 2D and 3D truss problems are considered to
ical formulation leads to the Reliable based Robust Design demonstrate the efficiency of such tool.
Optimization (RBRDO) task (Arora et al. 2007; Youn and The remaining of this paper is organized as follows:
Xi 2009; Lee I 2008; Rathod et al. 2011; Yadav et al. 2010). in Section 2 robust and reliable optimization is described,
Several robustness measures have been proposed in the presenting its mathematical formulation and methodologies
literature, in particular, the expected value and standard employed for the required statistics and, also, the reliability
deviation are considered here (Bucher 2009). When these methods used here to calculate the reliability based con-
two robustness measures are combined it follows that the straints; Section 3 briefly describes the governing equations
mathematical formulation of the RBRDO problem emerges of the problems to be solved, considering both, high-fidelity
as a multiobjective optimization problem, in which, both (full structural analysis) and low-fidelity (reduced order
the expected value and the standard deviation of the out- methods); Section 4 presents the MO procedure considered
put of interest have to be improved. In addition to that, the in this work; Section 5 presents the proposed procedure to
robustness in terms of feasibility conditions is also taken combine all these methods in a single efficient computa-
into account, considering the variability for some of the tional tool to obtain suitable solutions of RBRDO structural
constraints. engineering problems; in Section 6 two model examples
As general purpose optimizers do not directly solve MO illustrate the application of the proposed methodology and
problems, Pareto concept (Collette and Siarry 2004) based the performance of the developed tool in which the dif-
methodologies are used here to obtain robust designs sub- ferent strategies are discussed and compared; finally, in
jected to several constraints. For that, efficient techniques Section 7 the main conclusions of the present paper are
such as NBI (Normal-Boundary Intersection) (Das and drawn.
Dennis 1998), and NNC (Normalized Normal-Constraint)
(Messac and Mattson 2004) are used following the modifi-
cation proposed in Motta et al. (2012). Apart from these two 2 Robust and reliable optimization
strategies, two other approaches commonly used in the liter-
ature, weighted sum method and min-max method, are also As already mentioned, RBRDO considers problem uncer-
considered (Hwang et al. 1980; Huang et al. 2008). tainties to obtain a reliable design less susceptible to vari-
To compute the objective functions of the RBRDO prob- ability.
lem two methods for uncertainty propagation analysis are In this work, two objective controls will be considered:
considered: the Monte Carlo (MC) Method and the Prob- the mean and the standard deviation of a selected output
abilistic Collocation Method (PCM)(Ramamurthy 2005). function. Under such consideration, it implies that when the
An efficient procedure for structural reliability-based robust design optimization

expected value is minimized, a less conservative design is and its variance vf = SD [f (U)]2
found, while when the standard deviation is minimized a  ∞
 2
design with a much smaller range of variation is obtained SD [f (U)]2 = f (U) − f¯ P (U)dU (4)
(worst case scenario) (Beyer and Sendhoff 2007; Schuëller −∞
and Jensen 2008; Motta et al. 2009). A part from determin-
in which SD is the standard deviation.
istic constraints, reliability based constraints are imposed
In the present work two methodologies are employed for
to the problem. The task of RBRDO is therefore to obtain
statistics calculations of several responses. They are Monte
the trade-off between the two above aims, keeping con-
Carlo method and Probabilistic collocation method (Rama-
trol on the probability of failure or in the reliability index.
murthy 2005; Keane and Nair 2005). Both methodologies
Such compromising solutions are obtained by multiobjec-
are described in the following subsections.
tive optimization techniques (Collette and Siarry 2004; Das
and Dennis 1998; Messac and Mattson 2004; Motta et al.
2.2.1 Monte Carlo method
2012; Hwang et al. 1980; Huang et al. 2008).
The MC method is the most popular non-intrusive method
2.1 Problem definition
and can be used for any problem related to uncertainty prop-
agation (Keane and Nair 2005). Given the joint probability
Here, the RBRDO problem is mathematically formulated as
distribution function of the involved random variables, the
MC method can be applied for approximated calculations
min F (x) = {E (f (x, U )) , SD (f (x, U ))} (1) of the statistics response of a particular quantity, including
x its distribution, with an arbitrary error, as long a sufficient
number of samplings points is given. This approach has also
been used as a “benchmark” to validate other techniques
girr (x) ≤ 0 ir = 1, ...mr for statistics calculations. In this method the functions f (U)
g (x) ≤ 0 i = 1, ...m of interest are calculated in several random points Uk ,
subject to: i (2)
hj (x) = 0 j = 1, ... generated taking into account their probability distribution
xk ≤ xk ≤ xku k = 1, ...ndv P (U). Then the integrals of (3) and (4) are respectively
approximated as
where: x is the design variable vector; U is the random
1 
variables vector; F is the set of objective functions to be nmc
minimized; E(∗) is the expected value; SD(∗) is the stan- f¯ ≈ f¯MC = f (U(i) )
nmc
dard deviation and f (x, U ) is the selected output. The i=1
reliability analysis based constraint is girr , which could be 1
SD[f ]2 ≈ SDmc (f )2 =
related either to the probability of failure or to the reliability nmc − 1
n
index. The others constraints are gi (x) and hj (x), respec- mc  
tively the inequality and equality constraints, that could (or × f (U(i) )2 − nmc f¯mc
2
(5)
not) be dependent on U. The lower and upper bounds of a i=1
typical design variable are respectively xkl and xku . Finally, in which nmc is the number of sampling points for MC sim-
mr , m, l and ndv are the number of reliability constraints, ulation, f¯MC is the MC approximation for the mean values
the number of inequality constraints, the number of equality of f (U ) and SDmc (f ) is the MC approximation for the
constraints and the number of design variables, respec- standard deviation.
tively. The MO problem presented above is solved using
the techniques based on the Pareto minimal concept that are 2.2.2 Probabilistic Collocation Method
described in Section 4.
The basic idea of PCM is to approximate the function f (U )
2.2 Statistics calculation by polynomial functions and to evaluate the integrals of
(3) and (4) by Gaussian quadrature (Ramamurthy 2005).
Assuming U as a set of random variable, any function f (U) Gaussian quadrature is based on the concept of orthonormal
will be random, with its specific probability density function polynomials. These concepts are briefly described here.
(PDF) P (U). The expected value of f (U), called mean of In the numerical integration by Gaussian quadrature for
f (U), can calculated as Miller et al. (1990): integrals of the form
 ∞ 
E [f (U)] = f¯ = f (U)P (U)dU (3) F = f (U ) P (U ) dU (6)
−∞
 R. D. S. Motta, S. M. B. Afonso

the function f (U ) is approximated by a polynomial of order in which F is the failure region, defined as:
2n − 1 as follows (Stoer and Bulirsch 1991)
F = {U : G(U) > 0} (11)

n−1 
n−1
The reliability analysis could be performed by several
f (U ) ≈ fˆ(U ) = bi hi (U ) + hn (U ) ci hi (U )
methods. Here the Pf computation will be proceed via
i=0 i=0
Monte Carlo (MC) method and First Order Reliability
(7)
Method (FORM) (Melchers 1999).
for i = 1 . . . n in which bi and ci are the coefficients of
the approximation, to be obtained, and hi (U ) are polynomi- 2.3.1 Monte Carlo
als of order i from a orthonormal basis with respect to the
weight function P (U ). The MC is the simplest method used for reliability analysis.
The statistics evaluations defined in (3) and (4) via PCM The Pf is obtained after the evaluation of the failure func-
is a direct application of Gaussian quadrature in which the tion in a set of samples in the random variable space. These
PDF is the weighting function. Hence, by orthonormality, random points are generated considering the joint probabil-
the approximated Gaussian quadrature integral (6) can be ity distribution function fU (U) of the random variables (U).
expressed as follows (Ramamurthy 2005) The Pf is the proportion of points that lie over the F region,
 in which G(U) > 0.
F ≈ b 0 h0 P (U ) dU = b0 (8) Thus, to compute Pf via MC, the integral presented in
F
(10) is numerically approximated as Melchers (1999)
To find the coefficients bi and ci of the (7) would be
1 
nmc
necessary to evaluate the function f (U ) in 2n points. How- 
PfMC (x) = G(U(i) ) > 0 (12)
ever, as the integral presented in (8) depends only on the nmc
i=1
coefficient b0 , the n roots (U i∗ ) of hn (U ), can be used to
approximate the function f (U ), in this way cancelling the in which nmc is the number of MC sampling and

second part of (7), as hn (U i∗ ) = 0, i = 1 . . . n. For more G(U(i) ) > 0 = 1 if G(U(i) ) > 0
(13)
details concerning coefficients evaluations see Ramamurthy G(U(i) ) > 0 = 0 if G(U(i) ) ≤ 0.
(2005).
Usually, the structural Pf values are very small (for
The orthonormal polynomials are defined for each PDF
instance, O(10−6 )). This requires huge sample sizes, to
and the roots (U i∗ ) of each polynomial hi (U ) are the
avoid situations in which no point will fall over the fail-
quadrature points or integration point. Solving the approxi-
ure region and as a consequence, PfMC will be zero. In
mation of (7) to find b0 , it follows that the mean value and,
this way, is advised to use a sample size of at least nmc =
analogously, the standard deviation of an output of interest
5/Pf (Deb et al. 2009) which usually is a huge num-
are approximated by PCM as
ber (for instance, O(107 )), turning it, as a consequence, in
n
f¯P C = Pi f (U i∗ ) an inappropriate method to be used directly in structural
i=1 (9) optimization problems.
n
SDP C (f )2 = Pi f (U i∗ )2 − f¯P2 C
i=1 2.3.2 Form
in which Pi , : i = 1 . . . n, are the weight coefficients and
U i∗ the integration points, calculated for the PDF of the FORM approximates the problem around the most proba-
random variables (Ramamurthy 2005). ble failure point (MPP) to an equivalent basic problem by
transforming the original random variables space (U) into a
2.3 Reliability measures standard space (V), through an isoprobabilistic transforma-
tion (T ). In this paper the so called Nataf transformation will
The reliability measures needed to compute the g r con- be used. A graphical interpretation of the method is shown
straint in (1) are obtained through reliability analysis. The in Fig. 1. The MPP (or V∗ ) is the shortest distance point
reliability analysis computes the structural probability of from the limit state GV (V) = 0 to the origin of the stan-
failure (Pf ) and its reliability index (β). For a set of random dard space (V). The minimum distance value is called the
variables U with a joint probability distribution function reliability index.
fU (U) and a failure function G(U), the Pf can be computed β = ||V∗ || (14)
as follows:
 The standard space is a space of equivalent standard
Pf = fU (U)dU (10) normal distribution of uncorrelated random variables, i.e.
F fV (V) is the standard normal distribution ( mean, μ = 0 and
An efficient procedure for structural reliability-based robust design optimization

Fig. 1 FORM graphical

interpretation

standard deviation, σ = 1). In the standard space the limit namely, the Reliability Index Approach (RIA) and the Per-
state is linearly approximated around the MPP (Melchers formance Measure Approach (PMA) (Youn et al. 2003;
1999). Youn and Choi 2004a; Valdebenito and Schuëller 2010;
In the standard space Pf is approximated by (−β), Paiva et al. 2014). The main difference between those two
where  is the cumulative distribution function of the stan- techniques is the way that reliability constraints are evalu-
dard normal distribution. The opposite procedure could ated. The most classical one is the RIA, in which reliability
be used to approximate the reliability index from the Pf constraints are written in terms of β (14). It involves a
value (obtained via MC, for instance). To find the MPP high computational effort, as the structural reliability index
an iterative algorithm, called HL-RD from Hasofer, Lind, (β) (or failure probability) with respect to the imposed
Rackwitz and Fiessler (Hasofer and Lind 1974; Rackwitz constraints have to be found.
and Fiessler 1978), is used in the standard space, from which During the optimization process, it is required the com-
the following updated rule is employed. putation of β index, for each updated design. When RIA is
used each reliability constraint is involved with a reliabil-
∇GV (VK )   ity index target (βtar ). The local optimization problem, to
K T K
VK+1 =  2 ∇GV (V ) V − GV (V )
K
obtain β by RIA, can be formulated as:
∇GV (VK )
min V
(15) (16)
subject to: G(V) = 0
The solution of the above problem can be obtained using
in which GV (V) = G(U) and U = T −1 (V).
the FORM procedure described before. The solution of this
In short the main steps of the FORM algorithm are:
problem V∗RI A , as already mentioned, is the
 MPP. Then the
1. Transform current point V = T (U) (Nataf transforma- reliability index is computed (β = V∗RI A ) and compared
tion); with βtar . Thus, in the main reliability based optimization
2. Compute GV (V) = G(U); process, each reliability constraint has the form:
3. Compute ∇GV (V) = JT−1 ∇G(U)  
constraint: βtar − V∗RI A  ≤ 0 (17)
4. Compute the new V, (15)
5. Transform back U = T −1 (V) (Nataf transformation); On the other hand, reliability based optimization problem
6. Verify convergence, if not go to 1; considering PMA does not require the computation of the
7. Compute β = V. reliability index (β) (16). The search region is restricted to
the region that V = βtar , i.e. is the n-sphere of radius
in which JT is the Jacobian matrix of the transformation T βtar on the reduced random space. The local optimization
and generally the initial guess point is U0 = μ problem formulated according to PMA can be expressed as:

2.3.3 Optimization under reliability constraints max G(V)

(18)
V = βtar
There are two main approaches to evaluate the reliability This is an inverse problem from the FORM and is
of a given failure function during an optimization process, sometimes called First-Order Inverse Reliability Analysis
 R. D. S. Motta, S. M. B. Afonso

(Tu et al. 1999). The above problem will be solved here 2.3.4 Approximated reliability assessment
via the Advanced mean-value (AMV) method (Youn et al.
2003; Youn and Choi 2004a; Wu et al. 1990) and the solu- Another way to check feasibility of reliability constraints
tion is the V∗P MA point. The AMV method is known to be of a given structural design is through the use of statistical
unstable for concave failure function, but it wasn’t the case moments of the constraints outputs. This is usually used in
for the problems treated in here. Then the value of the failure approaches like the 6-sigma and the approximate moment
function G is evaluated in V∗P MA and, in the main RBRDO approach (AMA), as can be seen in Youn and Choi (2004b)
process, each reliability constraint has the form: and Koch et al. (2004), where Monte Carlo and Taylor
series are used to evaluate the statistical moments of the out-
constraint: G(V∗P MA ) ≤ 0 (19) puts. This is particularly advantageous in RBRDO structural
problems, in which the statistics of the structural output is
already computed for the objective functions of the RBRDO
Figure 2 shows the graphical interpretation of both
problem.
approaches (RIA and PMA) in a general standard space of
As shown, to calculate the robust measures for the objec-
two random variables (V). In this generic case, the reliabil-
tive functions of a RDO problem, a sampling procedure
ity constraint is satisfied β > βtar and G(V∗P MA ) < 0.
is used in which the outputs of structural responses are
Note that in cases where the reliability constraint is active
calculated. In the presented work we take advantage of
β = βtar , both solutions are equal (V∗RI A = V∗P MA ) and
the structural outputs already computed at the samplings
G(V∗P MA ) = 0.
to approximate the reliability index. This has an impact
An important aspect in calculation of β is its sign, as
in reducing the total number of full reliability analysis to
according to (14) it is always positive. However, there are
be conducted. This approach is considered here, which is
situations in which β should be negative. These are the cases
explained below.
when the origin of the standard space of the random vari-
For a failure function, given by G = f − flim ≤ 0,
ables (mean values) is within the failure region (G(U0 ) >
in which f is the constraint related output and flim is the
0). In this work, when such situation is detected, a negative
upper limit constraint. The reliability index of G will be
sign is applied to (14). In literature (Lin et al. 2011) there is
approximated by (β̂), using the statistical moments of f as,
another procedure to compute β index that aim to overcome
this problem.
E(f ) + β̂ SD(f ) = flim
Moreover, in the RIA methodology, the search for the
or (20)
MPP point (in which G = 0) will never converge in cases
β̂ = (flim − E(f )))/SD(f )
when all the space of random variables is feasible (i.e G >
0), the same is true if the whole space of random variables
This approximation considers that the failure function
is in the failure region (i.e G < 0) (Tu et al. 1999). Those
distribution can be represented by a Gaussian pdf. If it is not
problems does not occur when PMA is used. Even in general
the case and a log-normal pdf is used to represent failure
situations, PMA is said to be more robust and effective than
function distribution, the β̂ value is obtained directly from
RIA, especially when applied to ’very infeasible’ or ’very
the probability of failure (Pf ). This is done using the statis-
feasible’ designs (Youn and Choi 2004a; Deb et al. 2009;
tical moments of the output to obtain the parameters of the
Valdebenito and Schuëller 2010).
log-normal cdf (cumulative distribution function). The Pf
is obtained evaluating the log-normal cdf in the constraint
limit value, flim .
To illustrate the above approach, suppose that the mean
and standard deviation of some output f were computed
equal to 2 and 1, respectively. The log-normal cdf and pdf
of the output f can be created using the given mean and
standard deviation (2 and 1). Figure 3 shown the cdf and the
pdf generated. The figure shows also, the limit value of flim .
The Pf is the hashed area on the pdf that can be computed
as 1 − LN (4) = 0.0442, where LN is the cdf of the
log-normal distribution of the output.
The Pf and the β̂ value is obtained, respectively, by

Pf = 1 − LN (flim )
(21)
Fig. 2 PMA and RIA graphical interpretation β̂ = −−1 (Pf )
An efficient procedure for structural reliability-based robust design optimization

element method (FEM), and the non linear system of equa-

tions are solved in a incremental/iterative procedure using
Newton-Raphson (Crisfield 2000).
In this sense, for each load step, the solution field (dis-
placements) is obtained such that the internal forces equals
the external forces. The nonlinear discrete equations can be
written as
Fi (u) = Fe (22)
The iterative procedure generally used to obtain solution
of plastic analysis is the Newton-Raphson (NR) method.
The NR method iteratively approximates the nonlinear
equation by a linearization in the current point (solution). In
the plastic analysis it can be formulated for the kth iteration
as

Fig. 3 Cdf and pdf of a log-normal distribution with mean 2 and SD 1

dFi
Fi (uk ) + (uk ) uk+1 = Fe (23)
du
or,

in which, −1 is the inverse cdf for the standard normal Kk uk+1 = rk (24)
distribution. In the above example β̂ = 1.7.
in which Kk = dF i
du (uk ) is the tangent stiffness matrix, and
In the other hand, the solution β̂ based on (20) will
rk = Fe − Fi (uk ) is the residual load vector.
give β̂ = −(2 − 4)/2 = 1, that is quite different from
The iterative technique on its own can only provide a sin-
the one when the log-normal solution is considered. As
gle ’point solution’. In practice, we will often prefer to trace
can be noticed, β̂ can be obtained very fast with no extra
the complete load/deflection response (equilibrium path).
simulation, using the statics of the structural outputs.
To this end, it is useful to combine the incremental and
The pdf shape of an output can be verified for a given
iterative solution procedures. The ’tangential incremental
design through its histogram. In general, the output distri-
solution’ can then be used as a ’predictor’ which provides
bution won’t be nor normal nor log-normal and, as conse-
the starting solution, for the iterative procedure. A good
quence, the β̂ obtained is a poor approximation of its true
starting point can significantly improve the convergence of
value.
iterative procedures and increases the possible incremental
Because of that, in the present work, β̂ value will be used
load step. Indeed it can lead to convergence where otherwise
only to check the need of a full reliability analysis (RIA
divergence would occur (Crisfield 2000).
or PMA). The reliability constraints in which the approxi-
A constant increment is considered here, in which the
mated reliability index (β̂) are within the limits βl ≤ β̂ ≤ βu
increment is proportional to the load at the yield stress.
are the ones required to be computed properly (via RIA
besides that, an increment cut is used in case of convergence
or PMA). The limits values βl and βu are the lower and
fail. The increment (proportional factor) used in the analysis
upper bound of the approximated reliability index. In this
procedure will be specified in the application section.
work, they will be considered respectively, 1 and 5 (1 ≤
β̂ ≤ 5). Note that this prevents some RIA problems stated
3.2 Low fidelity model
previously (end of Section 2.3.3).
The large scale models require high computational costs.
The reduced order method attempt to approximate high
3 Governing equations dimension system in a low dimension space. Most of ROMs
accomplish this by projecting the high dimension system
3.1 High fidelity model
into a lower dimension space (Afonso et al. 2010).
The nonlinear system (22) can be project into a subspace
The high fidelity response will consider nonlinear elasto-
W of base Z ∈ Rndof,w with w < ndof , where ndof is
plastic static analysis with a small strain formulation. An
the total number of degrees of freedom. First we define the
elastoplastic von Mises constitutive model with associative
approximation of our displacement field u.
flow rule and a isotropic strain hardening will be used.
The weak equilibrium equations are approximated via finite Zα = uW ≈ u (25)
 R. D. S. Motta, S. M. B. Afonso

where α is the linear coefficient of the approximation uW . Thus, number m is the sum of the loads steps of the analysis
As a result, the problem (22) becomes of all the samples. The number m can be obtained through
an convergence parametric study.
Fi (Zα) = Fe (26)
The POD basis can be obtained from the eigenvectors
The iterative Newton-Raphson equation is then obtained of the correlation matrix. It can be found directly from a
dFi du eigenvalue decomposition of C, CV = λV. Alternatively,
Fi (Zα k ) + α k+1 = Fe (27) the POD basis can be obtained directly through a SVD of
du dα
the snapshot matrix X.
that can be rearranged as
Kk Zα k+1 = Fe − Fi (Zα k ) (28) X = SDVT where XV = SD and XT S = VDT (34)

This system can be solved projecting it into a subspace in above equation D is related to the eigenvalues√of C, such
W, pre-multiplying it by ZT . The final equation is as diag(D) = [ 1 , 2 , . . . m ] and i = λi , V are
eigenvectors of the autocorrelation matrix C and S are the
k α k+1 = rk
KW W
(29)
eigenvectors of the matrix XXT .
in which The size w of the POD basis is determined counting the
first w eigenvalues (in descending order and normalized)
k = Z Kk Z∈R
KW rk = ZT rk ∈ Rw
T w,w W
(30)
until the required error level, tolλ . This procedure can be
and expressed as find the smallest w such that

k+1 = uk + Zα k+1

uW W
(31) 
w
λi
1− λ̄i < tolλ , in which λ̄i = (35)
where the vector α ∈ Rw ,is the linear coefficient of i=1
λj
j
the approximation of the displacement increment uW ∈
Rndof . Thus, the space W, and, consequently, the base vec- In the equation (35), the tolλ parameter is the tolerance
tors Z are fundamentals to the efficiency of the method and related to the energy error in the POD approximation. Note
they commonly represent the differences between ROMs. that [w ≤ m ≤ ndof ]. After the number w of signifi-
cant singular components be determined, the POD basis is
3.3 Proper orthogonal decomposition computed

POD is a ROM that, basically, project the problem into Z = Sw (36)

a subspace formed by a optimum orthonormal basis func-
tions, in the sense that it considers the most significant shape where the upper w index, indicate the first w vectors
(greatest variance) of the output (u) subspace. A practical (column) of the matrix S.
way to obtain these vectors is computing a set of u vectors As can be seen, the POD has equivalence to the principal
for various system configurations (times, design variables, component analysis (PCA), the singular value decomposi-
parameters, loads steps, etc.) then, a singular value decom- tion (SVD) and Karhunen-Love (KL) decomposition, and
position (SVD) is performed to compute the eigenvalues of has been widely and successfully applied in various dis-
the covariance matrix of the outputs (Burkardt et al. 2006; ciplines, including fluid mechanics, static and dynamic
Liang et al. 2002; Tan et al. 2001). This procedure is related structural mechanics, oceanography, statistics, economics,
to the method of snapshots. The method of snapshots was image processing, etc (Tan et al. 2001).
introduced by Lawrence Sirovich in 1987 (Sirovich 1987)
as a easy way to obtain the autocorrelation matrix C
4 Multiobjective optimization solution
C = XT X (32)
in which the snapshot matrix X can be written as: The Pareto optimality concept is used here to obtain MO
  solutions. The Pareto minima are points x p such that no
X = u1 , u2 , . . . , um (33) other point x exist that satisfies
in which the number of snapshots m is assumed to be suf- fk (x) ≤ fk (x p ), k = 1, . . . , nobj
ficiently large to represent the correlation matrix, (32). The (37)
fj (x) < fj (x p )
number of snapshots m depends not only on the number of
samples, but also on the number of load steps from each for one objective function fj at least.
non-linear analysis in the offline stage. It means that the dis- A detailed discussion about this concept can be found
placement field of all load steps are considered as snapshots. elsewhere (Collette and Siarry 2004; Das and Dennis 1998;
An efficient procedure for structural reliability-based robust design optimization

Messac and Mattson 2004; Motta et al. 2012; Hwang et al. 4.2 The NNC method
1980; Huang et al. 2008). The solution of the problem
defined in equations (1) and (2) is very difficult to obtain The NNC method, introduced by Messac et al. (Messac
because in general the objective functions conflict with each and Mattson 2004), works in a similar manner as the NBI
other. Using the Pareto concept, the designer has to iden- method and its graphical representation can be seen in
tify as many Pareto minima points as possible. These points Fig. 5, which illustrates the feasible space and the corre-
can be used to construct a point-wise approximation to sponding Pareto frontier for a bi-objective case. The utopia
the Pareto curve or surface. There are several techniques line indicated in Fig. 5 (analogous to the CHIM in NBI
to obtain the set of Pareto minima. In this work, we uti- method), is the line joining the two individual minima points
lize the weighted sum (WS) method, the min-max method (or end points of the Pareto frontier). To obtain the Pareto
(Hwang et al. 1980), the normal boundary intersection points for a bi-objective case, a set of points Xpj are created
(NBI) method (Das and Dennis 1998) and the normal- in the utopia line. Through an interactive process using a
ized normal-constraint (NNC) method (Messac and Mattson pre-selected point Xpj , a normal line is computed to reduce
2004). Currently, in the literature, the latter two strategies the feasible space as indicated in Fig. 5. Minimizing f¯2
are shown to have more success in obtaining the Pareto results in the Pareto point f¯∗ , consequently, after translat-
curves (bi-objective problems). For problems with more ing the normal line from all points Xpj , the whole set of
than three objective functions a novel scheme proposed in Pareto solutions will be found.
(Motta et al. 2012) appears as an adequate choice. The
main feature of NNC and NBI methodologies are described
next. A modified version of both techniques can be found 5 The computational approach
elsewhere (Motta et al. 2012).
As can be noticed, the optimization task of finding a set of
4.1 The NBI method optimal structural designs considering the uncertainties and
non-linearities involved, ensuring reliability in terms of fail-
The NBI method (Das and Dennis 1998) is based on the ure (constraint) and performance (objectives), encompasses
parameterization of the Pareto front and produces an evenly several difficulties. Apart from the integration of differ-
spread point distribution. The geometric representation of ent methodologies into a unique tool to produce RBRDO
the NBI method is show in Fig. 4 which illustrates the designs, there is the total computational effort (time) draw-
objective function space and its feasible region. The Pareto back required to evaluate a large amount of structural
points are obtained in the intersection of the quasi-normal iterations and the search of the trade-off Pareto solutions.
lines emanating from a Convex Hull of Individual Minima Considering that the structural iterations are the main usage
(CHIM) and the boundary of the feasible objective func- of CPU during the entire optimization process, the total
tion space (δF ). The figure illustrates a set of quasi-normal CPU time consumed to solve the RBRDO problem, stated
lines and each line is associated with a specific coefficient here, can be estimated as
vector and, as can be seen, different (intersection) solutions
tCP U = nmo ∗nopt ∗(nstats +nreliab )∗(nnls ∗tnls +tsens ) (38)
are obtained (in most cases, Pareto points). A modifica-
tion of this method, presented in (Motta et al. 2012), which
allows points not in the quasi-normal lines by means of an
inequality constraint, will be considered.

Fig. 4 The geometric representation of the NBI method for bi- Fig. 5 Graphical representation of the NNC method for bi-objective
objective problems (Das and Dennis 1998) problems (Messac and Mattson 2004)
 R. D. S. Motta, S. M. B. Afonso

where: of iteration of the optimization process (nopt ). Because

of that, such methodologies become less efficient to our
– tCP U is the total CPU time of the optimization process;
purposes, due to the additional computations of the statis-
– nmo is the number of scalar optimizations required for
tics (objective functions) needed in each iteration of the
the multiobjective problem, i.e. nmo is equal to the
robust optimization process. Also, the computational effort
number of Pareto points required.
to compute the statistics can be greater than the computa-
– nopt is the average number of iterations in the scalar
tional effort to compute the reliability constraint. This is
optimizations, i.e. the average number of evaluations of
more evident when using the PMA.
the objective and constraint functions.
As can be seen, the total CPU time consumed (tCP U )
– nstats and nreliab are the average number of non-linear
can easily become impractical for complex structural prob-
structural analyses required to compute the statistics
lems. In order to solve a RBRDO problem, for the design
and the reliability of the structure, respectively.
of structures with non-linear behaviour, in an efficient way,
– nnls is the average number of non-linear structural
it is necessary to develop a practical, robust and reliable
iterations, including all load steps.
methodology that attempts to minimize the parameters in
– tnls is the average CPU time consumed for each struc-
(38). These are the aims of the developed tool.
tural iteration.
Thus, the proposed methodology accomplish these
– tsens is the average CPU time consumed for the compu-
requirements by:
tation of the structural sensitivities for the design vari-
ables and random variables (considering that it doesn’t – Using a MO method that obtain evenly distributed
depend on the structural history behaviour). Pareto points efficiently can lead to a reduction in the
number of Pareto points needed (nmo ).
If the Monte Carlo method is used to handle the structural
– A reduced value of nopt is achieved, using an effi-
uncertainties, the statistics and reliability of the structure
cient gradient based optimization technique, as the SQP
are calculated at the same time (nstats + nreliab = nmc ).
(Nocedal and Wright 2006).
But, as mentioned, to evaluate such quantities properly it
– The use of the PCM ensures a small value of nstats and
is required a huge amount of Monte Carlo simulations, and
improve the accuracy to compute the statistics of the
consequently, a huge number of structural evaluations. In
structural response, for problems with small number of
literature (Robinson and Atcitty 1999; Engelund and Rack-
random variables.
witz 1993), some alternatives have been proposed to circum-
– The reliability constraints can be efficiently evaluated
vent MC drawback as the use of the quasi-MC method and
using PMA, reducing the nreliab value.
the consideration of importance sampling technique. How-
– Finally, tnls and tsens are reduced with the use of the
ever, the reduction in the simulation number achieved by
POD as a model reduction technique. Although it is
those techniques are not enough for direct use in complex
an invasive procedure, the changes in the structural
structural optimization. There are also invasive procedures
analysis code are easy to handle.
to deal with uncertainties, like the Stochastic Finite Element
Method (SFEM) (Sudret and Der Kiureghian 2000; Doltsi- The proposed integration of the methods described to
nis and Kang 2004). Those methods have the drawback that solve RBRDO structural problems is done by an off-line
is the need for several new implementations and changes in stage, to compute the POD data, and an on-line stage with
the code for the structural analysis, that can be non-trivial a two level loops to obtain each Pareto solutions. Two
and laboriously for complex FE analysis codes. inner loops are required to perform the statistic analysis
There are methodologies that attempt to decoupling the (using PCM) and the reliability analysis (by PMA), in which
iterations of the reliability analysis and the optimization for both, the function evaluations are computed via POD.
process. Some schemes perform both iterative procedures The outer loop is the optimization procedure in which the
in a serial way. Others perform the reliability analysis in search for the optimum design variables is carried out. A
parallel with the iteration of the optimization process. A detailed information of the computational implementation
briefly review of the methodologies developed for optimiza- is presented on the algorithm in Table 1.
tion based on reliability constraints can be found in Lopez Remarks of algorithm, Table 1:
and Beck (2012).
In those methodologies a reliability analysis is performed a) The initial designs for the scalar optimizations (3.1) are
before and after an optimization process. This considera- chosen from the previously Pareto solutions found, if
tion may lead to a significant increase in the total number there are any;
An efficient procedure for structural reliability-based robust design optimization

Table 1 General algorithm for the solution of the RBRDO problem d) The multiple function evaluations required by the
using POD optimizer, statistics calculation and reliability analy-
sis (On-line stage) are calculated using the POD based
surrogate.

6 Examples

The tool developed in this work is used in this section

to obtain a reliable optimum design with reduction in the
structural performance variability.
The first example is a robust design optimization (RDO)
problem of a linear 3D truss, taken from the literature to
validate the implemented proposed tool. Then, for compar-
ison purposes, the same problem is solved considering the
RBRDO formulation.
In the second example, before presenting the RBRDO
results, sampling selection studies are performed in two sce-
narios. In the first, the sampling (in which the HF model
(FEM) is used) to built POD approximations has to be deter-
mined. Thus, variations in the number of samplings are
conducted to check POD basis accuracy. In the second sce-
nario, the sampling for the statistics calculation has to be
obtained. Responses based on POD model will be calcu-
lated on these. Again, in order to provide accurate statistics
a convergence study over the samples for both MC and PCM
methods is carried out.
Also, in this section, in order to compare the results from
the different MO methods, a parameter called Evenness
(Messac and Mattson 2004) will be used. This measures
the uniformity of the Pareto points distribution. Here, this
is calculated considering a procedure presented by Messac
and Mattson (2004). According to them, this parameter is
calculated as follows:

Evenness = SD(D)/E(D) (39)

in which SD is the standard deviation and E is the arith-

metic mean of D, beside that

b) The scalar optimization method used in this work was 

the Sequential Quadratic Programming (SQP) from D = dlow ; dup , (40)
MATLAB optimization toolbox (MathWorks 2014);
c) The multiobjective optimization process was briefly In the above equation, dlow and dup are two vectors con-
described in Section 4 and follows the procedure taining distances related to each point. The parameter dlow
described in (Motta et al. 2012). In this, suboptimiza- refers to the minimum distance of each point to another. The
tion problems can be solved after each individual opti- parameter dup states the maximum radius of a sphere touch-
mum solution has been found, for more details refer to ing each point, with no other points inside the sphere, for
(Motta et al. 2012). more details see Messac and Mattson (2004).
 R. D. S. Motta, S. M. B. Afonso

Table 2 Random variables

Number RV Mean SD COV

1-5 EI − EV 107 2 × 105 -

6 EV I 107 15 × 105 -
7, 8 pCx , pDx 500 50 -
9-14 AI − AV I x1 − x6 − 0.05

To quantify the variation of a function of interest, due to

a given variability of a random variable in a specific design
point, a sensitivity analysis of the objective functions (mean
Fig. 6 3D truss example - structure definition and standard deviation of the compliance) with respect to
the random variables was performed. The gradient obtained
was then multiplied by the respective standard deviation of
It is also worth to emphasize that all function evaluations the random variables to obtain the importance factor of each
required on the problem formulation are computed by the random variable, as shown in Fig. 7. After this procedure,
POD (surrogate model). the random variables considered were just the set: RVi , i =
[6, 14, 13, 11].
6.1 Benchmark problem - 3D truss In the following subsection, firstly, for validation pur-
poses, this structure will be considered for a robust (design)
This problem was proposed by Doltsinis and Kang (2004), optimization as in the reported literature (Doltsinis and
in which the robust optimization considering the structural Kang 2004). Next, it will be conducted a reliability-
compliance of a 25-bar truss structure (see Fig. 6) is to be based robust design optimization study using the PMA
carried out. All units considered in this example are con- methodology.
sistent. The design variables are cross-sectional areas of
six sets of bars. The six independent design variables are 6.1.1 Robust optimization study
selected by linking various member sizes as shown in Fig. 6.
The material mass density is 0.1. All four nodes of the base In the RO formulation the constraints considered are the
are fixed in x, y and z directions. Four nodal forces with val- stresses (σ ) on each bar (as an reliability-like constraint),
ues pAy = pBy = pAx = pBx = −104 are imposed at the the mean weight of the structure (E (w)) and the limits of
first and the second nodes (A and B). Additionally, forces
with random values are applied to the nodes C and D along
x-direction.
Reference Doltsinis and Kang (2004) considers 14 ran-
dom variables numbered in Table 2 as follows: the two nodal
forces applied at the nodes C and D, the six Young modulus
and the six cross-sectional areas for the grouped bars as indi-
cated in Fig. 6. For each random variable (RV) its statistics:
mean, standard deviation (SD) and coefficient of variation
(COV), are shown in Table 2. Although not directly stated in
the original paper, here, lognormal distributions were used
for all random variables, except the loads (RV 7 and 8) in
which normal distributions was adopted. Due to the small
number of degrees of freedom of the structure (ndof = 18),
the solutions of this problem were based in full FEM.
In Motta et al. (2015), a procedure to rank the importance
of each random variable was performed on the initial design,
x = [2.3, 2.3, ..., 2.3]. The most irrelevant ones were con-
sidered as deterministic variables and as consequence, the
number of random variables was reduced. Fig. 7 Importance factor of each random variable, Table 2
An efficient procedure for structural reliability-based robust design optimization

the design variables (x). The optimization problem is then Table 3 3D truss example - Multiobjective optimization perfor-
stated as follows mances: RO approach

Method Time (s) nopt Evenness

min [E (C(x, U)) , SD(C(x, U))] WS 20.9 918 0.572

x∈R 1x6 MinMax 8.8 488 0.482
subject to: NBI 13.8 566 0.311
E(σi ) + 3SD(σi ) ≤ 5000 i = 1, 2, ...., 25 (41)
NNC 11.3 516 0.311
E (w(x)) ≤ 750
0.05 ≤ xk ≤ 10 k = 1, ..., 6

in which C(x, U) is the structural compliance. The stress functions, i.e. one statistic analysis and one reliability anal-
constraints in the original problem (Doltsinis and Kang ysis, which in turns take several structural analysis (MEF
2004), is set as E(|σi |) + 3SD(σi ) ≤ 5000. But in order to calls). The parameter Evenness, is also present in Table 3
allow the solutions obtained by Doltsinis and Kang (2004), and, as already mentioned, the closer to zero the better.
the stress constraint should be modified (as some of the opti- As can be observed, the MO results obtained by NBI and
mum solutions shown in Doltsinis and Kang (2004) are not NNC agree closely to each other and these two methods
feasible, due to a high stress on the bar number 13). Tak- obtained the Pareto point distribution with lowest evenness
ing this finding into account, we strongly suspect that the value. Also, at least one solution obtained via WS and Min
constraint considered should be E(σi ) + 3SD(σi ) ≤ 5000, Max method were not Pareto optimum (optimization fail to
note that the negative stresses are virtually always feasible. converge).
The multiobjective robust optimization problem was solved Table 4 presents the optimum designs (x1 , x2 , ..., x6 ) for
in Doltsinis and Kang (2004) via WS method and for the the ten Pareto points, obtained via NBI MO technique.
statistics calculation the perturbation based stochastic finite The solutions P1 and P10, are the individual (scalar) opti-
element method was used. mum for the SD and mean of the compliance, respectively.
The multiobjective robust optimization solutions for this The solutions “P* (SD)” and “P* (Mean)” are the optimum
problem, using PCM for the statistics calculation and the of the SD and mean of the compliance, respectively, given
four multiobjective techniques explained in this paper, are by the literature (Doltsinis and Kang 2004).
presented in Fig. 8. The columns E (C) and SD(C) are the values of the
The performance of the MO solutions are given in objectives computed considering the method used in the
Table 3, in which the time, total iterations of the opti- optimization process. In the case of the results from the lit-
mization procedure (nopt) and the evenness parameter erature (“P* (SD)” and “P* (Mean)”) it is considered the
associated to each technique are shown. The nopt is the perturbation based stochastic finite element method. In the
number of calls involved on both objective and constraints case of the current work (remaining lines in Table 4), PCM
is used.
The respective values for the objective functions (Mean
and SD of the compliance) were recomputed through MC
for the twelve designs given in Table 4. The results appears
in columns “E (C) MC” and “SD(C) MC”, which are the
values of the objectives recomputed via MC. Samples sizes
with 3×103 and 5×104 points were considered. The results
for MC with 3 × 103 points used in Doltsinis and Kang
(2004), was not stable enough, so 5 × 104 points were used
in this work, for validation purposes.
As can be observed in Table 4, the results obtained here
are in good agreement to those obtained in the literature.
A 30 % reduction in the compliance variability (SD(C)) is
observed when comparing the values points P1 and P10.

6.1.2 Reliability-based robust design optimization study

Fig. 8 3D truss example - Pareto results for the MO method consid- The same problem is now solved considering reliability
ered: RO approach index based constraints in which the failure function is
 R. D. S. Motta, S. M. B. Afonso

Table 4 3D truss example - Pareto by NBI and reported solutions: RO approach

Solution x1 x2 x3 x4 x5 x6 E (C) SD(C) E (C)MC SD(C)MC

P* (SD) 0.147 0.672 3.465 0.566 0.822 8.048 6196 276 6184 287
P1 (SD) 0.145 0.742 3.567 0.378 1.184 7.023 5997 273 6136 296
P2 0.125 0.652 3.894 0.409 1.192 6.811 5886 278 5957 297
P3 0.05 0.05 4.417 0.606 1.212 6.829 5782 284 5832 301
P4 0.05 0.05 4.803 0.640 1.149 6.672 5682 290 5720 308
P5 0.05 0.05 4.835 0.762 1.141 6.598 5593 298 5622 316
P6 0.05 0.05 4.903 0.921 1.111 6.536 5510 306 5531 325
P7 0.05 0.05 5.001 1.119 1.072 6.453 5438 316 5450 336
P8 0.05 0.05 5.18 1.325 1.034 6.298 5380 327 5387 348
P9 0.05 0.05 5.434 1.459 1.044 5.992 5340 340 5344 363
P10 (Mean) 0.05 0.05 5.678 1.659 1.045 5.679 5324 357 5328 380
P* (Mean) 0.05 0.05 5.74 1.72 1.05 5.574 5328 357 5322 377

* Results from paper (Doltsinis and Kang 2004)

expressed in terms of bars stress. This problem is formulated example had convergence problems, no proper Pareto front
as: was found.
For MO solutions the number of Pareto points chosen
min [E (C(x, U)) , SD (C(x, U))] (42) was ten. The Pareto solutions are presented in Fig. 9 and the
x∈R 1x6
performance of the MO techniques is given in Table 5. In
subject to: this table, apart from the time(s) and the evenness parameter,
β(Gi ) ≥ 3 i = 1, 2, ...., 25 the total number of objective/constraint function evaluation
E (w(x)) ≤ 750 (43) (nopt) is given. Again, the results via NBI and NNC agree
0.05 ≤ xk ≤ 10 k = 1, ..., 6 closely to each other and these two methods obtained the
Pareto point distribution with lowest evenness value.
where β(Gi ) is the reliability index of the stress failure Table 6 presents the results for the ten Pareto points. As
function Gi related to the stress of element i, as follows: the previously problem, the respective values of the objec-
tive functions (Mean and SD of the compliance) obtained
Gi = σi − 5000 i = 1, 2, ...., 25 (44) via NBI and the statistics recomputed through MC with
This means that the stress in each bar involves a different 5x104 points are shown. In this table, the P1 and P10 solu-
reliability based constraint. tions are the individual optimum of the SD and mean of the
A high number of reliability constraints could be a diffi- compliance, respectively.
cult task to deal with in the optimization process. However it
is noticed that many of them are far from the failure surface
(always feasible) and, as a consequence, can be disregarded.
Due to that fact, the solution fields, already computed at
the samples to be used for the statistics calculation, are here
used to approximate the reliability constraint. This strategy
was also used in RO approach (Section 6.1.1) and already
explained in Section 2.3.4. The reliability index for each
stress bar were approximated by (β̂i ), using its statistical
moments as
E(σ ) + β̂i SD(σ ) = 5000
(45)
β̂i = (5000 − E(σ )) /SD(σ )

For the bars which 1 ≤ β̂i ≤ 5, the true true reliability

level is required to be calculated through reliability analysis.
The results shown are for the reliability constraints han-
dled through a PMA methodology. Results by RIA in this Fig. 9 3D truss example - Pareto solutions: RBRDO problem
An efficient procedure for structural reliability-based robust design optimization

Table 5 3D truss example - Performance of the MO methods:

RBRDO approach

Method Time(s) nopt Evenness

WS 338 535 0.52

Min-Max 375 562 0.57
NBI 232 463 0.21
NNC 300 451 0.21

Also the results present in the Table 3 are very close to

those given in Table 5, RO solution. This can also be ver-
ified comparing the two Pareto front of RO and RBRDO
formulation show in Figs. 8 and 9, respectively.
However, the CPU time consuming is much higher when
considering the reliability constraint as seen when Table 5
is compared with Table 3. This is due to the additional relia-
bility analysis (PMA local optimization process) for several
constraints. It is worth to mention that the results could be
very different depending on the resulting pdf of the output
considered in the reliability constraint.
The closer is the output pdf to a normal (Gaussian) pdf,
the better the results using the approximation β̂ in (45).
Fig. 10 2D truss example - Structure and problem definition

6.2 2D truss

A plane truss from (Motta and Afonso 2011) will be opti- the vertical load on the top of the structure (U1 ) and the hori-
mized here, considering material nonlinearity under static zontal load on the top-left side of the structure (U2 ). The first
load conditions. The geometric configuration and boundary one (U1 ) has a normal distribution with mean μ1 = 4KN
conditions are presented in Fig. 10, where the number of and standard deviation SD1 = 2KN, the second random
degrees of freedom is 1210. variable (U2 ) has a log-normal distribution with mean μ2 =
The material considered follows a simple von Mises 2KN and standard deviation SD2 = 1KN. Three designs
plasticity model, formulated for small deformation, and variables are considered, which are the cross section area of
elastoplastic material properties. The material stress-strain the bars of three regions, as shown in Fig. 10. The initial
data considered are given in Table 7. cross section areas (designs variables) are equal to two cm2
Figure 10 shows the random variables (U) and the design and the design variables are bounded by 0.1 ≤ x ≤ 10. The
variables (x). Thus, two random variables are considered: total number o degrees of freedom is ndof = 1210.

Table 6 3D truss example - Pareto by NBI solutions: RBRDO approach

Solution x1 x2 x3 x4 x5 x6 E (C) SD(C) E (C)MC SD(C)MC

P1 (SD) 0,051 0,213 3,666 0,540 1,270 7,150 6033 275 6032 277
P2 0,055 0,320 4,061 0,520 1,162 7,033 5889 276 5888 279
P3 0,050 0,109 4,439 0,578 1,208 6,781 5777 281 5776 284
P4 0,050 0,050 4,857 0,619 1,167 6,592 5678 288 5677 291
P5 0,050 0,050 4,874 0,745 1,158 6,531 5588 295 5587 298
P6 0,050 0,050 4,933 0,908 1,125 6,482 5505 303 5504 306
P7 0,050 0,050 5,025 1,111 1,081 6,413 5433 313 5432 316
P8 0,050 0,050 5,193 1,328 1,038 6,275 5376 324 5375 328
P9 0,050 0,050 5,442 1,466 1,043 5,982 5338 337 5337 341
P10 (Mean) 0,050 0,050 5,673 1,661 1,046 5,679 5323 353 5322 357
 R. D. S. Motta, S. M. B. Afonso

Fig. 11 2D truss example - Convergence of the approximated dis-

placement (d) against the sample size (N) Fig. 12 2D truss example - Pareto solutions via different MO meth-
ods: RIA Solutions
The RBRDO problem can be formulated as:
For a total of 120 different cases (considering differ-
min [E (σ (x, U)) , SD (σ (x, U))] , ent values for design variables and random variables) a
x∈R3
1210x144 snapshot matrix (X) was obtained through the
subject to : (46) storage of the load steps from the FE nonlinear analysis
V (x) ≤ V0 of each case. The 120 samples was obtained by a Latin
β (U, x) ≥ 3.3 Hypercube Sample (LHS) algorithm (Stein 1987).
0.1 ≤ xi ≤ 10, i = 1, 2, 3 A reduced model approximation can be obtained using
the full snapshot matrix (X) as the basis (Z = X), this
in which E (σ (x, U)) and SD (σ (x, U)) are the statistical approach is called Reduced Basis Method (RBM) (Afonso
moments of the maximum von misses stress to be mini- et al. 2010). The error of such approach is a lower limit
mized. The current volume should be less or equal to the of the POD error. Thus, to ensure a good accuracy of the
initial volume (V0 = 12371 cm3 ), the reliability index (β) POD model, the use of X matrix must lead to a good
is computed for the following failure function approximation.
g (x, U) = d (x, U) − dmax (47) A convergence study for the structural displacement (d)
for different X matrix dimensions was performed, at the ini-
in which d (x, U) is the horizontal displacement in the top tial design. Then, from the full snapshot matrix, different
left corner of the structure, see Fig. 10, and dmax = 3.5 cm POD approximation were built for tolerances varying form
is its maximum allowed value. 10−2 to 10−8 .
The load increment used was P y/10 during the nonlinear The displacement d error (L2 norm) compared to the
structural analysis via standard FEM and via POD reduced FEM results were calculated using RBM and POD mod-
model, in which P y is the load level that lead to the yield els (with different tolerance values). The results are shown
stress. in Fig. 11. Also in that figure, the POD tolerance values
are indiated by black ’x’s. As can be seen, the conver-
Table 7 2D truss example - Stress-Strain Curve gence of the POD approximation is evident in the figure.
Strain Stress (N/cm2 )
Table 8 2D truss example - Optimization performance considering
0.0025 20700 RIA, with PCM and POD method
0.0037 24840
MO Method Time (sec) nopt Evenness
0.0050 28980
0.0100 33120 WS 1308.41 84 1.3257
0.0175 37260 MinMax 1577.93 106 0.1748
0.0350 41400 NBI 2811.72 163 0.0114
0.0750 45540 NNC 802.75 78 0.0116
An efficient procedure for structural reliability-based robust design optimization

6.2.1 Reliability index approach

The RBRDO problem was solved twice using the two dif-
ferent approaches to handle the reliability constraint. First
the RIA considering the FORM method to evaluate the reli-
ability index is carried out. PCM will be used to evaluate
the statistics of the structure for each design variable set.
The statistics will be also used for the approximation of the
reliability constraint. The nonlinear analysis responses were
obtained using POD reduced order model (for w = 61).
The Pareto points obtained via the various MO methods
cited here, are shown in Fig. 12. Four multiobjective opti-
mization techniques performances, are shown in Table 8.
As in the previously example, in that table, the number of
function evaluations (nopt) is the total number of statis-
tic analysis and reliability analysis evaluated to obtain all
Fig. 13 2D truss example - Pareto solutions via different MO meth-
the Pareto points. The uniformity distribution of the Pareto
ods: PMA Solutions points parameter (Evenness) indicates the quality of the
Pareto points distribution (Messac and Mattson 2004; Motta
et al. 2012).
The WS method obtained a poor representation of the
As expected, as the number of sample points increases the Pareto front, the Min-Max was better. As expected, the
approximated results converge to the FE solution. results via NBI and NNC agree closely, as previous cases. A
For a required tolerance of tolλ = 10−5 , the size of POD better Pareto points distribution was obtained by these two
basis generated was w = 61. For this configuration, the methods.
online POD analysis took 0.29s while the full FE analysis
took 0.98s. The total off-line CPU time was 142.1s. The 6.2.2 Performance measure approach
time required to perform a structural and sensitivities anal-
ysis for the initial design was more than 3 times faster than The RBRDO problem is now solved using PMA to han-
those using FEM. dle the reliability constraints. As the previously examples,
Moreover, for the initial design, the standard deviation PCM was used to evaluate the statistics of the structure and
of the displacement d obtained via PCM for different lev- the nonlinear analysis responses were obtained using POD
els of approximation were compared with the results from reduced order model (for w = 61).
MC simulations using a sampling with 105 points. The rel- The Pareto points obtained via the various MO methods
ative difference between the PCM solutions considering a cited here, are shown in Fig. 13. The four multiobjec-
3x3 integration points and the MC results were about 10− 4. tive optimization techniques performances, are shown in
The computational time required by MC solution was 104 Table 9.
greater than that obtained by PCM. Therefore, for the statis- As can be seen, the performances via PMA (Table 9)
tics required in the optimization, 9 collocation points (3x3) of all methods are highly improved when compared to the
were considered. Also, at the initial design, the CPU time performance via RIA (Table 8). The use of PMA produces
consumed to compute the statistics (via PCM) of a single Pareto frontier more than 2 times faster. In this example, the
design via POD (w = 61) was 4.2s, while via full FEM was WS method results were the poorest one. Although it was
12.2s. faster, WS was still unable to find Pareto solutions in all

Table 9 2D truss example - Optimization performance considering Table 10 2D truss example - Optimization performance considering
PMA, with PCM and POD method PMA, with PCM method and FEM

MO Method Time (sec) nopt Evenness MO Method Time (sec) nopt Evenness

WS 560.57 92 1.5701 WS 2256.37 88 1.4205

MinMax 402.07 90 0.1491 MinMax 2129.68 88 0.2642
NBI 1221.0 180 0.0114 NBI 4217.97 136 0.01155
NNC 333.90 78 0.0116 NNC 2458.30 94 0.01195
 R. D. S. Motta, S. M. B. Afonso

For this optimum design, the probability of failure

obtained via FORM (through the β index) and via MC was
respectively, 0.029 % and 0.040 %. The β index obtained
via FORM and via MC (through the Pf value) was respec-
tively 3.44 and 3.35. As well know, the MC method, unlike
the FORM, is able to obtain the Pf value with a controlled
error, based on the sample size choice. However, the value
of the β index (obtained from the Pf value) has an inherent
error, as the Pf computed via FORM. The summary of the
FORM and MC results are shown in Table 11.

7 Conclusions

In this paper we propose an efficient integrated tool for

RBRDO solutions. The tool encompasses a POD algorithm
Fig. 14 2D truss example - MC results for optimum mean stress to approximate the FEM analysis, PCM to evaluate the
design statistics (1st and 2nd statistical moments) of the response
of the structural applications and, also, to consider approx-
the regions of the Pareto front. Again, the results via NBI imated reliability constraints into a procedure based on
and NNC agree closely and the better Pareto points distribu- the PMA for reliability constraint assessment and NBI
tion was obtained by these two methods. It is important to or NNC multiobjective optimization techniques to obtain
emphasize that Evenness value for these two techniques are Pareto solutions, in order to allow fast robust designs.
the same ones obtained when the constraints were handled The main conclusions based on the present study are:
by RIA (see Table 8). – A good agreement between optimum designs using full
A fully FEM based optimization was performed, consid- FE analysis and POD was achieved.
ering PMA, for comparison purposes. The results are shown – Reanalysis via POD, in the presented applications, was
in Table 10. Again, when NBI and NNC are used, the Even- shown to be 3 times faster than classical FE analysis.
ness parameter are very closed to each other. Also, POD – There is a need of advanced multiobjective optimiza-
results (Tables 8 and 9) are in good agreement to FEM based tion techniques to obtain appropriate Pareto frontier for
results (Table 10). As expected, solutions by POD are much RBRDO problems.
faster. For instance, when NNC (POD) technique is used, – The most efficient MO technique to conduct RBRDO,
solutions were obtained seven times faster than the ones for in terms of computational costs and Pareto points dis-
NNC(FEM). tribution, are NNC and NBI.
– For statistics computations, PCM proved to be reliable
6.2.3 Optimum designs post-process and 5 orders of magnitude faster than MC, for problems
with few significant random variables.
Figure 14 shows the MC results considering 5 × 104 sam- – The quantification of the importance factor of each ran-
pling points for the design related to the upper-left Pareto dom variable with respect to the statistics of the outputs
solution in Fig. 13 (the one that minimize the expected value of interest is very important when techniques such as
of the stress). This sample size was determined based on PCM are employed.
convergence test of the Pf value. In that figure, the cir- – The use of approximated reliability constraints based on
cles in color scale are the displacement d value at each PCM samplings provide extra computational savings by
MC integration point and the ’x’ points represent the failure reducing the number of full reliability analysis.
points. – The PMA proved to be very effective, around two times
faster, when compared to the solutions obtained via
Table 11 2D truss example - Reliability data at optimum design RIA.
Results MC FORM Overall, considering the reanalysis by POD, the use of
statistics via PCM, the use of those statistics to approximate
Pf (x∗ ) 0.040 % 0.029 %
the reliability analysis, the PMA to evaluate the full relia-
β (x∗ ) 3.35 3.44
bility constraints and the use of robust MO techniques, the
n 5x104 9
total computational savings by the presented tool makes the
An efficient procedure for structural reliability-based robust design optimization

RBRDO structural problem accessible for practical purpose Lee I (2008) Reliability-based design optimization and robust design
(about 6×105 times faster, with an error around 10−4 ). Also, optimization using univariate dimension reduction method. PhD
thesis, University of Iowa
the developed tool could be extended to be used in more
Liang Y, Lee H, Lim S, Lin W, Lee K, Wu C (2002) Proper orthogo-
complex classes of applications. nal decomposition and its applications-part i: theory. J Sound Vib
252(3):527–544
Lin PT, Gea HC, Jaluria Y (2011) A modified reliability index
Acknowledgments The authors acknowledges the financial support
approach for reliability-based design optimization. J Mech Des
given by the Brazilian research agency CNPq and Pernambuco state
133(4):044501
research agency FACEPE to the execution of the present work.
Lopez RH, Beck AT (2012) Reliability-based design optimization
strategies based on form: a review. J Braz Soc Mech Sci Eng
34(4):506–514
References Marczyk J (2000) Stochastic multidisciplinary improvement: beyond
optimization. In: Proceedings of 8th AIAA/USAF/NASA/ISSMO
symposium on multidisciplinary analysis and optimization. Long
Afonso S, Lyra P, Albuquerque T, Motta R (2010) Structural analysis Beach
and optimization in the framework of reduced-basis method. Struct MathWorks (2014) Optimization toolbox user’s guide: MATLAB.
Multidiscip Optim 40(1–6):177–199 MathWorks Inc., USA
Arora J, Wang Q (2005) Review of formulations for structural
Melchers RE (1999) Structural reliability: analysis and prediction, 2nd
and mechanical system optimization. Struct Multidiscip Optim
edn. Wiley, Chichester
30(4):251–272
Messac A, Mattson CA (2004) Normal constraint method with guar-
Arora JS, Messac A, Mullur AA (2007) Optimization of structural and
antee of even representation of complete Pareto frontier. AIAA J
mechanical system. World Scientific Pub Co Inc, University of
42(10):2101–2111
Iowa, USA
Miller I, Freund J, Johnson R (1990) Probability and statistics for
Beyer H-G, Sendhoff B (2007) Robust optimization - a comprehensive
engineers, 4th edn, Prentice-Hall, Englewood Cliffs
survey. Comput Methods Appl Mech Eng 196(33):3190–3218
Bucher C (2009) Computational analysis of randomness in structural Motta RS, Afonso SMB (2011) Optimization of trusses under nonlin-
mechanics: structures and infrastructures book series. Structures ear conditions considering the proper orthogonal decomposition
and infrastructures. Taylor & Francis method. In: XXXII CILAMCE - Iberian Latin American congress
Burkardt J, Gunzburger M, Lee HC (2006) Pod and cvt-based reduced- on computational methods in engineering. Ouro Preto, Brazil
order modeling of navier-stokes flows. Comput Methods Appl Motta RS, Afonso S, Lyra P (2009) Structural robust optimization con-
Mech Eng 196:337–355 sidering reduced-basis method, Master’s thesis, Civil Engineering
Cavazzuti M (2012) Optimization methods: from theory to scientific Department. UFPE, Recife-PE Brazil. in Portuguese
design and technological aspects in mechanics. Springer Science Motta RS, Afonso S, Lyra P (2012) A modified nbi and nc method
& Business Media for the solution of n-multiobjective optimization problems. Struct
Collette Y, Siarry P (2004) Multiobjective optimization: principles and Multidiscip Optim 1:1–21
case studies. Springer Motta RdS, Afonso SMB, Lyra PR, Willmersdorf RB (2015) Devel-
Crisfield MA (2000) Non-linear finite element analysis of solids and opment of a computational efficient tool for robust structural
structures: essentials, vol 1. Wiley, Chichester optimization. Eng Comput 32(2):258–288
Das I, Dennis JE (1998) Normal-boundary intersection: a new method Nocedal J, Wright S (2006) Numerical optimization. Springer Science
for generating the Pareto surface in nonlinear multicriteria opti- & Business Media
mization problems. SIAM J Optim 8(3):631–657 Papadrakakis M, Lagaros N, Plevris V (2005) Design optimization of
Deb K, Gupta S, Daum D, Branke J, Mall AK, Padmanabhan D steel structures considering uncertainties. Eng Struct 27(9):1408–
(2009) Reliability-based optimization using evolutionary algo- 1418
rithms. IEEE Trans Evol Comput 13(5):1054–1074 Paiva RM, Crawford C, Suleman A (2014) Robust and reliability-
Doltsinis I, Kang Z (2004) Robust design of structures using optimiza- based design optimization framework for wing design. AIAA J -
tion methods. Comput Methods Appl Mech Eng 193(23):2221– Special Section Multidiscip Des Optim 52:711–724
2237 Rackwitz R, Fiessler B (1978) Structural reliability under combined
Engelund S, Rackwitz R (1993) A benchmark study on importance load sequences. Comput Struct 9(5):489–494
sampling techniques in structural reliability. Struct Saf 12:255– Rao RV, Savsani VJ (2012) Mechanical design optimization using
276 advanced optimization techniques. Springer Science & Business
Haftka RT, Sobieszczanski-Sobieski J (2009) Structural optimiza- Media
tion: history structural optimization: history, in encyclopedia of Ramamurthy D (2005) Smart simulation techniques for the evalua-
optimization. Springer, pp 3834–3836 tion of parametric uncertainties on black box systems. PhD thesis,
Hasofer A, Lind N (1974) Exact and invariant second-moment code Washington State University
format. ASCE J Eng Mech Div 100:111–121 Rathod V, Yadav O, Rathore A, Jain R (2011) Reliability-based robust
Huang B, Fery P, Xue L, Wang Y (2008) Seeking the Pareto front for design optimization: a comparative study. In: IEEE international
multiobjective spatial optimization problems. Int J Geogr Inf Sci conference on industrial engineering and engineering management
22(5):507–526 (IEEM), 2011, pp 1558–1563
Hwang C-L, Paidy SR, Yoon K, Masud ASM (1980) Mathematical Robinson D, Atcitty C (1999) Comparison of quasi- and pseudo-
programming with multiple objectives: a tutorial. Comput Oper monte carlo sampling for reliability and uncertainty analysis. In:
Res 7(1):5–31 40th structures, structural dynamics, and materials conference and
Keane AJ, Nair PB (2005) Computational approaches for aerospace exhibit, American institute of aeronautics and astronautics
design: the pursuit of excellence. Wiley Schuëller GI, Jensen HA (2008) Computational methods in opti-
Koch P, Yang R-J, Gu L (2004) Design for six sigma through robust mization considering uncertainties–an overview. Comput Methods
optimization. Struct Multidiscip Optim 26(3–4):235–248 Appl Mech Eng 198(1):2–13
 R. D. S. Motta, S. M. B. Afonso

Simpson T, Poplinski J, Koch PN, Allen J (2001) Metamodels for Valdebenito MA, Schuëller GI (2010) A survey on approaches
computer-based engineering design: Survey and recommenda- for reliability-based optimization. Struct Multidiscip Optim
tions. Eng Comput 17(2):129–150 42(5):645–663
Sirovich L (1987) Turbulence and the dynamics of coherent structures, Wu Y-T, Millwater H, Cruse T (1990) Advanced probabilistic struc-
part 1: Coherent structures. Q Appl Math 45(3):561–571 tural analysis method for implicit performance functions. AIAA J
Spillers WR, MacBain KM (2009) Structural optimization. Springer 28(9):1663–1669
Science & Business Media Yadav OP, Bhamare SS, Rathore A (2010) Reliability-based robust
Stein M (1987) Large sample properties of simulations using latin design optimization: a multi-objective framework using hybrid
hypercube sampling. Technometrics 29(2):143–151 quality loss function. Qual Reliab Eng Int 26(1):27–41
Stoer J, Bulirsch R (1991) Introduction to numerical analysis, 2nd edn. Youn BD, Choi K (2004a) An investigation of nonlinearity of
Springer-Verlag, Heidelberg, Berlin reliability-based design optimization approaches. J Mech Des
Sudret B, Der Kiureghian A (2000) Stochastic finite element methods 126(3):403–411
and reliability: a state-of-the-art report. Department of Civil and Youn BD, Choi KK (2004b) Selecting probabilistic approaches
Environmental Engineering, University of California for reliability-based design optimization. AIAA J 42(1):124–
Tan BT, Damodaran M, Willcox K (2001) Proper orthogonal decom- 131
position extensions and their applications in steady aerodynamics. Youn BD, Xi Z (2009) Reliability-based robust design optimization
Master’s thesis, Ho Chi Minh City University of Technology, using the eigenvector dimension reduction (edr) method. Struct
Vietnam Multidiscip Optim 37(5):475–492
Tu J, Choi KK, Park YH (1999) A new study on reliability-based Youn BD, Choi KK, Park YH (2003) Hybrid analysis method for
design optimization. J Mech Des 121(4):557–564 reliability-based design optimization. J Mech Des 125(2):221–232