Proceedings of DETC’04

Proceedings
ASME 2004 Design Engineering Technical of DETC’04
Conferences and
ASME 2004 Design Engineering Technical Conferences and
Computers and Information in Engineering Conference
Computers
September and Information
28-October in Engineering
2, 2004, Conference
Salt Lake City, Utah, USA
Salt Lake City, Utah, September 28 – October 2, 2004

DETC2004-57255
DETC2004/DAC-57255

A SINGLE-LOOP METHOD FOR RELIABILITY-BASED DESIGN OPTIMIZATION

Jinghong Liang Zissimos P. Mourelatos1 Jian Tu
Department of Mechanical Engineering Department of Mechanical Engineering Research & Development Center
Oakland University Oakland University General Motors Corporation
[email protected] [email protected] [email protected]

ABSTRACT stochastic nature of the design variables and model

Reliability-Based Design Optimization (RBDO) can parameters. Variations are represented by standard deviations
provide optimum designs in the presence of uncertainty. It can (typically assumed to be constant) and a mean performance
therefore, be a powerful tool for design under uncertainty. The measure is optimized subject to probabilistic constraints.
traditional, double-loop RBDO algorithm requires nested RBDO can be a powerful tool which can assist in design under
optimization loops, where the design optimization (outer) uncertainty, since it provides optimum designs in the presence
loop, repeatedly calls a series of reliability (inner) loops. Due of uncertainty in design variables/parameters and simulation
to the nested optimization loops, the computational effort can models. For this reason, it has been extensively studied [1-16].
be prohibitive for practical problems. A single-loop RBDO A conventional RBDO formulation is a nested, double-
algorithm is proposed in this paper for both normal and non- loop approach. It consists of an outer design optimization loop
normal random variables. Its accuracy is the same with the in the original space, which repeatedly calls reliability
double-loop approach and its efficiency is almost equivalent to analyses in a series of inner loops. Each reliability analysis is a
deterministic optimization. It collapses the nested optimization separate optimization loop in the standard normal space. A
loops into an equivalent single-loop optimization process by nonlinear transformation is used to map the original space into
imposing the Karush-Kuhn-Tucker optimality conditions of the standard normal space. The Reliability Index Approach
the reliability loops as equivalent deterministic equality (RIA) [1-3] or Performance Measure Approach (PMA) [3,10-
constraints of the design optimization loop. It therefore, 13] can be used in each reliability analysis. In general, PMA is
converts the probabilistic optimization problem into an more efficient and stable than RIA. However, both PMA and
equivalent deterministic optimization problem, eliminating the RIA are based on the First-Order Reliability Method (FORM)
need for calculating the Most Probable Point (MPP) in [17] which is widely used in structural reliability due to its
repeated reliability assessments. Several numerical simplicity and computational speed. However, for problems
applications including an automotive vehicle side impact with nonlinear limit states, FORM can overestimate or
example, demonstrate the accuracy and superior efficiency of underestimate the actual probability of failure. For this reason,
the proposed single-loop RBDO algorithm. second order reliability methods (SORM) have been
developed [18-21], based on a quadratic approximation of the
INTRODUCTION limit state function at the Most Probable Point (MPP). In most
In deterministic design we assume that there is no cases, they improve on the FORM accuracy.
uncertainty in the design variables and/or modeling The nested optimization loop structure makes the
parameters. Therefore, there is no variability in the simulation conventional RBDO formulation computationally very
outputs. However, there exists inherent input and parameter expensive so that it becomes impractical for a variety of
variation that results in output variation. Deterministic practical applications. Although the computational efficiency
optimization typically yields optimal designs that are pushed and algorithmic stability of the inner reliability loop has been
to the limits of design constraint boundaries, leaving little or substantially improved with the introduction of the
no room for tolerances (uncertainty) in manufacturing Performance Measure Approach (PMA) [10-13], the
imperfections, modeling and design variables. Therefore, computational effort of the double-loop (DLP) approach is
deterministic optimal designs that are obtained without taking still prohibitive specially for large-scale applications. For this
into account uncertainty are usually unreliable. Input variation reason, two new classes of RBDO formulations have been
is fully accounted for in Reliability-Based Design recently proposed [22-29]. The first class decouples the
Optimization (RBDO). Probability distributions describe the RBDO process into a sequence of a deterministic design

1
Corresponding author, Phone/Fax: (248)370-2686/370-4416

1 Copyright © 2004 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/30/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 optimization followed by a set of reliability assessment loops demonstrated in comparison with other RBDO methods. A
[22-24]. The series of deterministic and reliability loops is summary and conclusions are given in Section 5.
repeated until convergence. The second class of RBDO
methods converts the problem into an equivalent, single-loop 2. OVERVIEW OF RELIABILITY-BASED DESIGN
deterministic optimization [25-29]. The approach in [25] uses OPTIMIZATION (RBDO) METHODS
a mean value first-order reliability method. However, it is A typical RBDO problem is formulated as
numerically inaccurate or unstable due to a wrong estimation min f (d, µ X , µ P )
d, µ X
of the probabilistic constraints. The single-loop, single vector
(SLSV) approach [27-30] provides the first attempt in a truly s.t. P (Gi (d, X, P ) ≥ 0 ) ≥ R , i = 1,..., n (1)
single-loop approach. It improves the RBDO computational d ≤ d ≤ d , µ ≤ µX ≤ µ
L U L
X
U
X
efficiency by eliminating the inner reliability loops. It first where d ∈ R k is the vector of deterministic design variables,
transfers all the design variables into a reduced, uncorrelated,
X ∈ R m is the vector of random design variables, P ∈ R q is the
and normalized space which is different from the standard
normal space U [27] and subsequently uses the steepest vector of random parameters, f ( ) is the objective function
descent direction at the approximate MPP of the previous and n, k, m and q are the number of constraints, deterministic
iteration. However, it requires a “probabilistic” active set design variables, random design variables and random
strategy for identifying the active constraints, which hinders parameters, respectively. According to the used notation, a
on its practicality. The present work greatly enhances the bold letter indicates a vector, an upper case letter indicates a
SLSV method providing an accurate, efficient and random variable or a random parameter and a lower case letter
computationally stable single-loop RBDO algorithm. indicates a realization of a random variable or random
In the decoupled RBDO methods [22-24], the reliability parameter. The actual reliability level for the ith deterministic
analyses are performed separately from the design constraint is denoted by
optimization. The Sequential Optimization and Reliability Ri = 1 − p f i = 1,......, n ,
i
(2)
Assessment (SORA) [22] method uses the reliability where p f = P[Gi (d, X, P ) < 0] (3)
information from the previous cycle to shift the violated i

th
deterministic constraints in the feasible domain. This is done is the probability of violating the i deterministic constraint
sequentially until convergence is achieved. SORA appears to which is usually very small. In RBDO, the actual probability
be similar to the safety-factor approach reported in [30]. of failure must be less than the target probability of failure
Although SORA improves the computational efficiency of the which is usually approximated by the following first-order
RBDO process, its convergence rate is slow and not relation [32,33]
guaranteed. Another decoupled approach to perform RBDO p f ≈ Φ (− β t ) (4)
has also been proposed in [23,24]. However, it is restricted to where β t is the target reliability index and Φ is the standard
deterministic design variables in the design optimization loop. normal cumulative distribution function. Combining Eqs (1)
For the reliability requirement, a “ball” constraint of radius through (4) yields,
equal to the reliability index is satisfied. P (Gi (d, X, P) ≤ 0) = FG (0) ≤ Φ (− β t ), i = 1,K, n (5)
In this work, a single-loop RBDO method is presented. It i i

imposes the Karush-Kuhn-Tucker (KKT) optimality where the cumulative distribution function FG ( ) is described
i

conditions of the reliability loops as an equivalent as

deterministic equality constraint in the design optimization FG (0) = ∫G ( d , X , P ) ≤ 0 ...∫ f X , P (x, p)dxdp i = 1,2,..., n (6)
loop. It therefore, converts the probabilistic optimization
i
i

formulation into an equivalent deterministic optimization and f X , P (x, p) is the joint probability density function of all
formulation. The proposed single-loop method does not need random variables and random parameters.
to calculate the MPP for each probabilistic constraint in order Before the proposed single-loop RBDO method is
to assess the reliability eliminating therefore, the reliability described in Section 3, a brief overview is presented of the
optimization loop of the conventional double-loop RBDO double-loop method with percentile formulation (DLP/PMA)
approach. If needed, the MPP’s can be easily calculated after and the Sequential Optimization and Reliability Assessment
the successful termination of the single-loop process. In (SORA) [22] method. This overview provides the basic ideas
general, the proposed single-loop RBDO approach has of both methods as they compare with the proposed single-
comparable efficiency with a deterministic optimization and loop RBDO method. It also describes the percentile
the same accuracy with the conventional double-loop RBDO formulation of the reliability assessment which is used in this
method. In contrast, the accuracy of the decoupled RBDO work.
methods is in general worse than the accuracy of the double-
loop approach. 2.1. Double-Loop RBDO Method
Section 2 of this report provides an overview of the The classical double-loop RBDO method employs two
RBDO methods. Subsequently, Section 3 describes the nested optimization loops; the design optimization loop (outer)
proposed single-loop RBDO method in detail. In Section 4, and the reliability assessment loop (inner). The latter is needed
the performance of the single-loop method is assessed using a for the evaluation of each probabilistic constraint in Eq. (1).
mathematical example, a beam example with two probabilistic For this reason, the double-loop RBDO method is
constraints and a vehicle side impact example. The efficiency computationally very expensive and therefore, almost
of the proposed single-loop method is compared with the impractical for large-scale design problems. Based on Eq. (5),
efficiency of deterministic optimization and its accuracy is there are two different methods for the reliability assessment;

2 Copyright © 2004 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/30/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 the Reliability Index Approach (RIA) [1-3] and the reliability assessment loops which are performed in series. At
Performance Measure Approach (PMA) [3,10-13]. Although the end of a deterministic design optimization, the reliability
either approach can be used, PMA is in general more efficient, of each constraint is assessed. If the reliability of a particular
especially for high reliability problems [10,12,13]. Every time constraint is less than the specified target, a “shifting” vector
the design optimization loop calls for a constraint evaluation, a is calculated which is used to push the constraint boundary in
reliability assessment loop is executed which searches for the the feasible domain. The “shifted” constraints are then used to
MPP in the standard normal space, based on FORM. If the perform a new deterministic design optimization. The series of
random variables are not normal, a nonlinear transformation T deterministic and reliability assessment loops continues until
is needed [17, 34] to map the original space to the standard convergence is achieved; i.e. the objective function is
normal space. In this case, each constraint G (d, X, P ) is minimized and the target reliability of each constraint is met.
mapped into G (d, T (X ), T (P )) = G (d, U ) in the U space. At convergence the shifting distance is zero.
If the RIA-based approach is used, the general RBDO In all decoupled RBDO methods, an equivalent
problem of Eq. (1) is stated as [3] deterministic formulation of the design optimization loop is
needed. In SORA [22] specifically, a percentile formulation of
min f (d, µ X , µ P )
d, µ X each probabilistic constraint is used which leads to such an
s.t β i = (−Φ −1 ( FG (0))) ≥ β , i = 1,..., n (7) equivalent deterministic formulation. An overview of the
percentile formulation is first given before the essentials of the
i

d ≤ d ≤ d , µ ≤ µX ≤ µ
L U L
X
U
X SORA algorithm are described.
where Eq. (5) has been used to transform each probabilistic An equivalent formulation of the general RBDO
constraint to an inequality between the actual reliability index formulation of Eq. (1), using the percentile formulation is
β and the target reliability index β t . The reliability index stated as [10,13,31]
min f (d, µ X , µ P ) (11)
β = −Φ −1 (FG (0)) is equivalently the minimum distance of a d, µX

point U on the limit state G (U ) = 0 from the origin of the s.t. Gi R (d, X, P) ≥ 0, i = 1,L, n
standard normal space. It is therefore, calculated from the where G R is the R-percentile of the constraint G (d, X, P) . It is
following reliability minimization problem
defined as,
β = min U (8)
U P (G (d, X, P)) ≥ G R ) = R (12)
s.t. G (U ) = 0 where R is the target reliability for the constraint. Fig. 1 shows
where the vector U represents the random variables in the a geometric interpretation of G R . It is clear that if G R ≥ 0 ,
standard normal space. P (G (d, X, P ) ≥ 0) ≥ R , Therefore, G R ≥ 0 provides an equivalent
Similarly, the PMA-based RBDO problem, which is deterministic expression of the probabilistic constraints in Eq.
practically the inverse of the RIA-based RBDO problem, is
(1). The R-percentile G R is evaluated using the PMA method
stated as [10-13]
min f (d, µ X , µ P )
of Eq. (10). After the MPP is calculated, the R-percentile is
d, µ X given by
s.t. G p = ( FG −1 (Φ (− β t )) ≥ 0 , i = 1,..., n (9) G R = G (d, X MPP , PMPP ) (13)
i i i

d ≤ d ≤ d , µ ≤ µX ≤ µ
L U L U Therefore, the optimization problem of Eq. (11) becomes
min f (d, µ X , µ P )
X X
(14)
where the performance measure G p is calculated from the d, µX

following reliability minimization problem s.t. Gi (d, X MPP , PMPP ) ≥ 0, i = 1,L, n

i i

G p = min G (U ) (10) where µ X = X mpp − U mpp σ, µ P = Pmpp − U mpp σ (15)

U i i i i

s.t U = βt . and X MPP , U MPP are the Most Probable Point for the i th
i i

As shown from Eqs (7) and (9), both the RIA and PMA- constraint in X and U spaces, respectively and σ is the vector
based RBDO formulations involve nested optimization loops, of standard deviations. Similarly, PMPP is the value of the
i

which may hinder on their computational efficiency and random parameters at the MPP of the i th constraint in the X
convergence properties. To improve the computational space.
efficiency, different optimization formulations have been
recently proposed that either decouple the two nested
Area= P[G ≥ G R ] = R
optimization loops [22-24] or even better, merge them in a
PDF of G
single optimization loop [25, 27] leading therefore, to
significant efficiency and convergence benefits. The work in
this paper presents a significant advancement of [27] in
developing a robust single-loop RBDO method.

2.2. Sequential Optimization and Reliability

Assessment Method (SORA)
A Sequential Optimization and Reliability Assessment 0 GR g
(SORA) method [22] has been recently proposed for
improving the efficiency of RBDO. This method employs a Figure1. Geometric interpretation of R-percentile.
sequence of decoupled deterministic optimization and

3 Copyright © 2004 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/30/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Fig. 2 illustrates the sequence of deterministic and min f (d, µ X , µ P ) (17)
d, µ X
reliability assessment loops in SORA. At each cycle, a
deterministic optimization is first performed followed by a s. t. Gi (d, X, P) ≥ 0 i=1,2,……n
series of reliability assessments. The number of reliability d ≤ d ≤ d , µ ≤ µ X ≤ µUX
L U L
X
assessments is equal to the number of probabilistic constraints. where the vectors X and P are evaluated at the MPP. The
For the k th cycle, the following “shifting” vector mean objective function is minimized subject to constraints
sik = µ kX−1 − X kMPP
−1
i
which are evaluated in the X space. It is therefore, necessary
is used to shift the boundary of the i th probabilistic constraint to have a consistent relationship between vectors d, µ X , µ P , and
in the feasible domain. The equivalent deterministic vectors d, X, P. The traditional DLP/PMA method calculates
optimization for the k th cycle then becomes the MPP of each constraint based on the current d, µ X , µ P
min f (d, µ X , µ P ) vectors and subsequently, transfers these MPP’s into the X
d, µ X
space in order to calculate the d, X, P vectors. The proposed
s. t. Gi (d, µ X − sik , PMPP
i
k −1
) ≥ 0 , i = 1,..., n
i
(16) single-loop method uses the Karush-Kuhn-Tacker (KKT)
d ≤ d ≤ d , µ ≤ µX ≤ µ
L U L U optimality conditions of the reliability loops as equality
X X
constraints of the design optimization loop in order to relate
Note that each probabilistic constraint has its own
the d, µ X , µ P and d, X, P vectors.
“shifting” vector because it has its own MPP. Also for the k th
cycle, the random parameters of the i th probabilistic constraint
Using the PMA method, an inner reliability loop of the
are evaluated at the MPP of the previous cycle. For the
double-loop method solves the following optimization
reliability assessment in SORA, either of the RIA or PMA
problem in the standard normal space U
approaches of Eqs (8) and (10), respectively can be used.
min G (U ) (18)
Detailed information on SORA is provided in [22]. U

SORA decouples the deterministic optimization and s.t. || U ||= β

reliability assessment loops. However, it still searches for the where G(U) is one of the constraints of the original RBDO
MPP’s in the reliability assessment loops. For this reason, the problem, and βt is the required (target) reliability index for this
computational effort may still be high for practical RBDO
constraint. At the optimal point, the following KKT optimality
applications although it is much lower than that of the
condition [35,36] is satisfied
traditional double-loop approach.
∇G ( U ) + λ ∇H ( U ) = 0 (19)
Cycle1 Cycle2 where H (U ) =|| U || − βt is an equality constraint and λ is the
RA1 RA1
Opt1 Opt2 corresponding Lagrange multiplier. Eq. (19) states that the
Opt1 …… ……
gradients ∇G (U) and ∇H (U) are collinear at the optimum.
Opt2
Since H (U) is an equality constraint, there is no restriction on
RAm RAm the sign of λ. u 2

G(U)
Figure 2. Schematic of SORA method; Opt: Deterministic
optimization; RA: Reliability assessment. G (U ) ≥ 0
∇G
Beta ∇H
3. SINGLE-LOOP METHOD FOR RBDO PROBLEMS A
In the traditional double-loop RBDO formulation, a
reliability optimization (inner loop) is performed when the ∇G u1
G (U) ≤ 0
design optimization (outer loop) calls for a probabilistic Decreasing Direction of G
constraint evaluation. This can drastically increase the B
∇H
computational cost. In this section, a single-loop RBDO
G (U ) = 0
algorithm is presented based on an equivalent deterministic Figure 3. Geometric explanation of optimal solution for one
optimization formulation, which eliminates the need for inner
active constraint.
reliability loops or sequential deterministic and reliability
loops. It also does not increase the number of design variables
It should be noted that in X space, the entire sphere must
as is the case with some other proposed single-loop algorithms be in the feasible domain which according to Eq. (17) is
[25,26]. It is an accurate and computationally very efficient defined by a set of non-negative constraints. The gradient of
RBDO method. It has the same accuracy with the double-loop
each constraint points inside the feasible domain. Also, the
approach and the number of required function evaluations is gradient of the entire hyper-sphere points outwards. Since the
comparable to that of a deterministic optimization.
whole hyper-sphere in X space must be in the feasible domain,
In principle, it converts the probabilistic optimization the gradients of a constraint and the corresponding beta hyper-
problem into an equivalent deterministic optimization problem
sphere must point in opposite directions. For this reason, the
by enforcing the Karush-Kuhn-Tucker optimality conditions
[35,36] of the reliability loop as a constraint of an equivalent Lagrange multiplier λ in Eq. (19) must be non-negative (i.e.
deterministic design optimization formulation. λ ≥ 0 ), indicating that the gradients of G (U) and H (U) are
Using the R-percentile formulation of Eq. (14), the collinear and they point in opposite directions. This is shown
RBDO problem can be expressed as, geometrically in Fig. 3 for one constraint in two dimensions.

4 Copyright © 2004 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/30/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 As G (U) moves in its decreasing direction, it can become min f (d, µ X , µ P )
d, µ x
tangent to the stationary beta circle at points A or B. Point A is
not the MPP because gradients ∇G (U) and ∇H (U) point in s. t. Gi (d, Xi , Pi ) ≥ 0 i=1,2,……n (26)
the same direction indicating that the beta circle is outside the where Xi = µ X − σ * β t * α i ;i
Pi = µ P − σ * β t * α i
i

feasible domain G (U) ≥ 0 . However, point B is the MPP. α i = σ ∗ ∇Gi X, P (d, Xi , Pi ) / || σ ∗ ∇Gi X, P (d, Xi , Pi ) ||
u2
G1 = 0 d L ≤ d ≤ dU , µ XL ≤ µ X ≤ µUX
and β t is the target reliability index for the ith constraint, α i
i

Decreasing
∇H is the normalized gradient of the ith constraint and σ is the
Direction Active standard deviation vector of random variables X and random
of Constraint 1 Constraint 1 parameters P.
A
∇ G1
In the single-loop RBDO formulation of Eq. (26), the
∇G2
u1 objective function is evaluated at the mean point d,µ X , µ P and
Decreasing
Direction
B Inactive the constraints are calculated at the d, X, P point. The
∇ H Constraint 2
of Constraint 2 relationship of Eqs (25) is used to evaluate the constraints
consistently with the values of the design variables. The
G2 = 0
single-loop method does not search for the MPP of each
Figure 4. Geometric explanation of optimal solution for two constraint at each iteration. Instead, the MPP’s of the active
constraints. constraints are correctly identified at the optimum. This
dramatically improves the efficiency of the proposed single-
Fig 4 shows a hypothetical RBDO problem with two loop method without compromising the accuracy.
constraints plotted in the U space. For simplicity, both
constraints are assumed to have the same target reliability and Initialize 0
d 0 , µ X , µ P , β, σ, lb, ub
therefore, the same beta circle. As the design optimization
(outer) loop progresses, the two constraints have a tendency to
Assign X 0 = µ X 0 , P 0 = µ P
move towards the beta circle. At the optimum of this
hypothetical case, the first constraint is tangent to the beta
circle and therefore, active; the second constraint is not active. Calculate αi 0 = σ ∗ ∇Gi ( X,P) (d0 , X0 , P0 ) / || σ ∗ ∇Gi (X, P) (d0 , X0 , P0 ) ||
For both constraints however, ∇G (U) and ∇H (U) are
collinear and point in opposite directions. Point A is the MPP k=0
for the first constraint while point B is not the MPP for the Calculate f (d , µ kX , µ P )
k

second constraint.
Considering that ∇H (U) = 2 * U , Eq. (19) yields
U = −[|| ∇G || /(2 * λ )] * (∇G / || ∇G ||) (20) No
µ kX Change?
where ∇G is the length of ∇G . Since the length of vector U
Yes, k=k+1
is equal to the beta hyper- sphere radius β t and ∇G / || ∇G || is
Calculate αi = σ ∗∇Gi(X,P) (d , Xi , Pik−1) / || σ ∗∇Gi(X,P) (dk , Xik−1Pik−1) ||
k k k −1
a unit vector, the positive quantity || ∇G || /(2 * λ ) in Eq. (20)
must be equal to β t . Therefore, Eq. (20) becomes
U = −βt ∗ α (21) Calculate Xi k = µXk − βt ∗ αi k ∗ σ, Pi k = µP − βt ∗ αi k ∗ σ
i i

where α = ∇ G U ( d, X, P ) / || ∇ G U ( d, X, P ) || (22)
is the constraint normalized gradient in U space. Calculate G i ( d k , X ik , P ik )
The transformation of X space to U space yields the
following relationships between X, P and U,
X = σ * U + µX , P = σ * U + µP . (23) No Is f Yes Stop
Also, using the chain rule, minimized?
∇GU (d, X, P)= σ * ∇GX,P (d, X, P) (24)
Combining Eqs (21) through (24) yields the following Figure 5. Flowchart of the single-loop RBDO method.
relationship between X, P and µ X , µ P ,
3.1. Single-Loop Method Flowchart
X = µ X − σ * βt ∗ α, P = µ P − σ * βt ∗ α (25) Fig. 5 shows the flowchart of the proposed single-
where α = σ ∗ ∇ G X, P ( d, X, P ) / || σ ∗ ∇ G X, P ( d, X, P ) || loop method. First, the initial point d 0 , µ 0X , µ P is specified as
Eqs (25) hold for each constraint of the RBDO well as the target reliability index vector β t for all constraints,
formulation of Eq. (17). the standard deviation vector σ for the random variables and
Using Eqs (25), the RBDO formulation of Eq. (17) is random parameters and the upper and lower bound vectors (ub
transformed to the following single-loop, equivalent and lb, respectively) for all deterministic and probabilistic
deterministic optimization problem design variables. The initial point d 0 , X0 , P 0 needed to evaluate

5 Copyright © 2004 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/30/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 the constraints is taken equal to d 0 , µ 0X , µ P ; i.e. X0 = µ x 0 and The proposed single-loop RBDO algorithm has excellent
efficiency because it is based on an equivalent deterministic
P 0 = µ P . At this point, the initial normalized gradient vector α
optimization. Its computational efficiency is better than the
for the i th constraint is calculated as decoupled methods which usually have a slower rate of
0
α i = σ ∗ ∇Gi ( X , P ) (d 0 , X 0 , P 0 ) / || σ ∗ ∇Gi ( X , P ) (d 0 , X 0 , P 0 ) || (27) convergence due to the sequence of deterministic and
reliability analyses. A variety of numerical examples
At the k iteration of the optimization loop, the objective
th
presented in the next section, provide evidence on this.
function is calculated at the d k , µ kX , µ P point. For the evaluation The consistency between the design variable vector
of the constraints, the algorithm checks if the optimizer has d , µ X , µ P and the vector d, X, P needed to evaluate the
changed the design vector µ kX compared with the previous constraints makes the single-loop algorithm very robust. A
iteration; if not, the current gradient vector αk is used to variety of numerical examples have been easily solved with
different initial points without any problem. This is not
calculate X k = µ kX − σ * βt ∗ α k and Pk = µp − σ * βt ∗αk for each
however, the case with earlier attempts to develop a robust
constraint in order to finally evaluate the constraints. If µ kX has single-loop RBDO algorithm [27-30] due to inconsistent
changed from the previous iteration, the normalized gradient initial µ X and X vectors. Sensitivity of the SLSV algorithm
k
vector α k is updated before it is used to calculate X and P
k performance and the final optimum to the initial point has
been reported in [29].
which are needed for the constraint evaluation. This is an
Finally, the proposed single-loop RBDO algorithm does
essential step for keeping the design variable vector µ X and the
not require an active constraint set as is the case with the
X, P vectors consistent, having therefore, a robust and stable SLSV algorithm. The active constraint set is simply identified
algorithm. Furthermore, it greatly improves the efficiency by the algorithm. This is a major advantage which simplifies
since the algorithm avoids unnecessary gradient evaluations. the implementation of the proposed single-loop algorithm and
enhances its robustness and efficiency. Furthermore, the
3.2. Single-Loop Method with Non-Normal Random single-loop method does not update the constraint gradients
Variables unnecessarily. The gradients are updated only when the design
For problems with non-normal distribution, it is mean values have changed, improving therefore, the efficiency
necessary to transform the non-normal distribution variables considerably.
into equivalent normal distribution variables. Here, the
Rackwitz-Fiessler two-parameter equivalent normal method 4. APPLICATIONS
[31] is employed to estimate the parameters σ XN , µ XN of an This section demonstrates the performance, accuracy
and efficiency of the proposed single-loop RBDO method.
equivalent normal distribution by imposing two conditions.
Comparisons are made with existing RBDO methods and
The cumulative distribution functions and the probability
deterministic optimization. In all comparisons, the same initial
density functions of the actual non-normal variables and the
point and similar convergence criteria are used for all
equivalent normal variables should be equal at the current
methods. First, a non-linear mathematical example and a
approximate MPP ( x1* , x2* ,LL xn* ) . This yields, cantilever beam example are used with normal and non-
σ XN = ϕ{Φ −1[ FX ( xi* )]} / f X ( xi* ) normal random variables. Subsequently, a vehicle side impact
(28) problem is used to demonstrate the benefits of the proposed
µ XN = xi* − Φ −1[ FX ( xi* )]σ XN
method in practical reliability-based design.
where σ XN , µ XN are the equivalent standard deviation and mean
values and FX ( xi* ), f X ( xi* ) are the CDF and PDF of the actual 4.1. A Mathematical Example
non-normal variables at the current approximate MPP. A A non-linear mathematical example, first introduced in
similar idea has been also used in [28]. [10] is used. There are two random variables x1 , x2 which are
When non-normal variables are used in the single-loop normally distributed and three nonlinear constraints G1 , G2 , G3 .
RBDO method, the equivalent σ XN and µ XN are calculated The objective function is simply the sum of the means of the
using Eq. (28) every time the optimizer updates the mean two random variables. A problem with only two random
values during the optimization process. variables is first used so that we can show graphically the key
characteristics of the proposed single-loop algorithm. The
3.3. Advantages of the Single-Loop Method RBDO problem is stated as
The main advantage of the single-loop method is the min f = µ1 + µ2
µ ,µ
elimination of the repeated reliability loops. Instead of 1 2

performing nested design optimization and reliability loops, it s.t. P(G j ( X) ≥ 0) ≥ R j j =1~ 3
solves an equivalent single-loop deterministic optimization 2
G1 ( X) = x1 * x2 / 20 − 1
problem. Therefore, its efficiency is vastly superior to existing
G2 ( X) = ( x1 + x2 − 5) 2 / 30 + ( x1 − x2 − 12) 2 / 120 − 1
double-loop approaches. It is also more efficient than the
2
decoupling RBDO methods [22-24] which rely on solving G3 ( X) = 80 /( x1 + 8 * x2 + 5) − 1
successive deterministic and reliability optimization problems. 0 ≤ µi ≤ 10 i =1~ 2
Furthermore, the single-loop method has the same accuracy
σ 1 = σ 2 = 0.3 β j = 3 for j=1,2,3
with the double-loop methods.

6 Copyright © 2004 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/30/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 where µ1 , µ2 and σ 1 ,σ 2 are the mean values and standard convergence criteria. It must be noted that now the β –circle
deviations respectively, of the two design random variables x1 is stationary and centered at the origin of the U-space, while
and x2 . For demonstration purposes, the same target reliability the constraint boundaries move from iteration to iteration. If at
index β = 3 is used for all three constraints. In general, a least one constraint becomes tangent to the β –circle, the
different target reliability index may be used for each algorithm stops and returns the constraint optimum. The figure
constraint. shows the shape evolution for each constraint. As constraints
x2 G 1 and G 2 approach the β –circle, constraint G 3 moves away
10 from it. Finally at the optimum, constraints G 1 and G 2 are
G3 = 0 active and constraint G 3 is not.
8 Increasing G3
Increasing G 1 u2
1st Iter
G1 = 0 10 2nd Iter
6 Initial Mean pt G3 4th Iter

Beta G3

Opt. Circle 5
4 G3
Mean pt G1

Final G2 = 0 Increasing G2 -10 -5 0 5 10 15 u1

2 X pt
G1 G2
for G 1
-5 G2
0 2 4 6 8 10 x1
Figure 6. Progress of optimization process for the G2
-10
mathematical example in X space. G1

Fig. 6 shows the progress of the single-loop RBDO -15

process in the X space. The algorithm starts with an initial

Figure 7. Progress of optimization process for the
point µ 0X = (µ1 , µ2 ) = (5,5) . The corresponding initial β –circle
mathematical example in U space.
is also shown. It should be noted that the β –circle in the X-
space retains its circular shape because the two random 7.5
variables are normally distributed. Initially, it is assumed that 7
(x1, x2 ) = (5,5) for all three constraints and the initial normalized
Obj. Function

6.5
gradient vector α 0 is calculated using Eq. (27). For the initial 6
constraint evaluation, three points X = (x1 , x 2 ) , one for each 5.5
5
constraint, are calculated using the relation X0 = µ0X − βt * α0 * σ . Single-Loop
4.5 Det. Opt
These three points are shown with open-circle symbols on the 4
initial β –circle. Fig. 6 shows the position of the β –circle 1 2 3 4 5
for the first three iterations. The center of each β –circle Iteration No.

indicates the value of the design vector µX = (µ1, µ2 ) . Also, on Figure 8. Comparison of single-loop method with
deterministic optimization.
each β –circle the three X = (x1, x2 ) points are indicated with
open-circle symbols. At the optimum, these points are the Fig. 8 and Table 1 illustrate the accuracy and efficiency
MPP’s of the corresponding active constraints. As shown, the of the proposed single-loop RBDO method. Fig. 8 compares
optimum µ X = (3.4391, 3.2864) has been practically reached the efficiency of the single-loop method against deterministic
after three iterations. Constraints G1 and G 2 are active and optimization. Both need four iterations to converge. The final
constraint G 3 is inactive. It must be noted that physically the value of the objective function is 5.1765 and 6.7255 for the
deterministic and single-loop methods, respectively. As
β –circle moves within the feasible domain defined by the indicated in Table 1, both the deterministic and the single-loop
three constraints, until it either becomes tangent to at least one methods required 19 function evaluations. In comparison, the
constraint (active constraint) without penetrating the fixed SORA and the DLP/PMA methods required 151 and 1566
boundaries of the feasible domain (constrained optimum) or function evaluations, respectively. In this study, the evaluation
the objective function is minimized without any constraints of the objective function and all constraints is considered as a
being active (unconstrained optimum). In the latter case, the single function evaluation. In this example, the efficiency of
β –circle is within the feasible domain without touching its the single-loop method is the same with that of deterministic
boundaries. optimization. In general, the efficiency of the single-loop
Fig. 7 shows the progress of the single-loop RBDO method is comparable to the deterministic optimization. It
method in the U-space for the 1st, 2nd and 4th iterations. The should be noted that the efficiency of any RBDO method can
algorithm converged in four iterations based on the specified not theoretically exceed the efficiency of the deterministic
optimization. The last two columns of Table 1 compare the

7 Copyright © 2004 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/30/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 double-loop and single-loop optimal results for a non-normal the random yield strength, Z and Y are mutually independent
distribution case. Although both optima are similar, the single- random loads in the vertical and lateral directions respectively,
loop is considerably more efficient. and E is the Young modulus. A reliability index β = 3 is used
for both constraints.
Table 1. Comparison of RBDO methods for the mathematical
example. 10

9.5
Det. DLP/ SINGLE DLP/ SINGLE

Obj. Function
SORA 9 Det. Opt
Opt. PMA LOOP PMA LOOP
8.5 Single-Loop (Beta=3)
Design x1 =5, Normal x 1 =3.5,Lognormal
8
DLP/PMA
Variables x 2 =5, Normal x 2 =3.5, Normal 7.5

x1 3.1139 3.4391 3.4409 3.4391 3.4619 3.4179 7

0 2 4 6 8 10
x2 2.0626 3.2866 3.2909 3.2864 3.2939 3.2811
Iteration No.
Objective
f(X) 5.1765 6.7257 6.7318 6.7255 6.7558 6.747 Figure 10. Optimization history for the beam example.
Constraints
G 1 (X) 0 0 0 0 0.0041 0.0557 Fig. 10 and Table 2 show the accuracy and efficiency of
G 2
(X) 0 0 0 0 0.0007 0.0008 the single-loop method. Fig. 10 compares the efficiency of the
(X) 1.5643 0.5 0.5071 0.5097 0.5044 0.5138
single-loop method against deterministic optimization and
G 3
DLP/PMA. Both deterministic optimization and the SLP
No. of 4 4 3
Iterations
4
(Outer) (cycles)
4
(Outer)
2 method converge in six iterations and both calculate the
No. of F. E. 19 1566 151 19 1177 11 constraint optimum with 54 function evaluations. Again, the
efficiency of the single-loop method is the same with
4.2. A Cantilever Beam Example deterministic optimization. The DLP/PMA method needs 8
In this example, a cantilever beam in vertical and iterations to converge (see Fig. 10) and 2490 function
lateral bending [30] is used (see Fig. 9). The beam is loaded at evaluations (see Table 2). Both DLP/PMA and single-loop
its tip by the vertical and lateral loads Y and Z, respectively. methods calculated the same optimum point at which the
Its length L is equal to 100 in. The width w and thickness t of objective function is equal to 9.52 and only the first constraint
the cross-section are random design variables. The objective is is active. The deterministic optimum is equal to 7.66 and both
to minimize the weight of the beam. This is equivalent to constraints are active. The efficiency and accuracy of the
proposed single-loop method is excellent for this example as
minimizing f = w ∗ t , assuming that the material density and
well. The last two columns of Table 2 compare the single-loop
the beam length are constant. and DLP/PMA methods when all four random parameters are
lognormally distributed.
Y
Z Table 2. Comparison of RBDO methods for the beam
t example.
L=100 in w Det.
DLP/PMA
SINGLE
DLP/PMA
SINGLE
Opt. LOOP LOOP
Design 4 lognormal random
4 normal random parameters
Figure 9. Cantilever beam under vertical and lateral. Variables parameters
w = x1 2.047 2.4781 2.4479 2.3280 2.3206
Two non-linear failure modes are used. The first failure t = x2 3.7459 3.8421 3.8892 4.0762 4.1343
mode is yielding at the fixed end of the cantilever; the other Objective
failure mode is that the tip displacement exceeds the allowable f(x) 7.6679 9.5212 9.5202 9.4895 9.5940
value D0 = 2.5" The RBDO problem is formulated as, Constraints
min f = w ∗ t G 1 (X ) 0.0001 0.0038 0 129.8162 221.0026
w, t
G 2 (X ) 0 0.3593 0.3184 0.2799 0.0013
s.t. P(G j (d, P) ≥ 0) ≥ R j =1~ 2 No. of
6 8 6 5(outer) 3
Iterations
600 600
with G1 ( y , Z , Y , w, t ) = y − ( 2 * Y + 2 * Z ) No. of F. E. 54 2490 54 2358 44
wt wt
4 L3 Y Z 4.3. Vehicle Side Impact Example
G2 ( E , Z , Y , w, t ) = D0 − ( )2 + ( 2 )2
Ewt t 2 w Here, a vehicle crashworthiness study is performed under
0 ≤ w, t ≤ 5 a variety of side impact constraints. Uncertainties in structural
design variables, material properties and operating conditions
where G1 and G2 are the limit states corresponding to the two
among others, are very important in automotive vehicle side
failure modes. The design variables w and t are deterministic impact studies. For this reason, RBDO of vehicle
while Y, Z, y and E are normally distributed random crashworthiness has recently gained considerable attention
parameters with Y~ N (1000, 100) lb, Z~ N (500,100) lb, y~ N [15]. This example has been extensively used to test the
(40000,2000) psi and E~ N ( ( 29 * 10 6 ,1 . 45 * 10 6 ) psi; y is

8 Copyright © 2004 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/30/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 accuracy and efficiency of RBDO methods [13,22,29,30] in A total of seven random variables and four random
the past few years. parameters are used. The seven random variables ( x1 ~ x7 )
The side impact is one of the safety requirements an represent dimensions of some vehicle structural parts
automotive vehicle design must meet, according to the including thickness of B-pillar (inner and reinforcement),
National Highway Traffic Safety Administration (NHTSA). thickness of floor side, thickness of cross member, thickness
The side impact procedure based on the European Enhanced of door beam, thickness of door belt line reinforcement and
Vehicle - Safety Committee (EEVC) is used in this example. thickness of roof rail. The four random parameters include the
The performance of the dummy in side impact, in terms of material of B-pillar (inner) x 8 and floor side (inner) x 9 as
head injury criterion (HIC), chest V*C's (viscous criterion)
and rib deflections (upper, middle and lower) must meet well as the barrier height x10 and barrier hitting position x11 .
EEVC requirements. Table 3 shows sequentially the description of the random
The finite element model of the vehicle used in this study variables and parameters and their lower and upper bounds.
and the moving deformable barrier are shown in Fig. 11. The The objective is to reduce the vehicle weight. The vehicle
total number of nodes and shell elements of the model are weight is a function f ( µ X , µ P ) of the random variable vector
approximately 96,000 and 86,000, respectively. A finite X and the random parameter vector P. The RBDO problem is
element model of the dummy is also used. The velocity of B- stated as
Pillar at middle point and the velocity of front door at B-Pillar min f (µ X , µ P )
are considered. The position of the moving deformable barrier µX

is specified according to the EEVC side impact procedure. All s. t. P ( FAL ≤ 1KN ) ≥ R1
nodes of the moving barrier are assigned an initial velocity of P( Dlow ≤ 32cm) ≥ R 2
31 mph.
P ( Dmid ≤ 32cm) ≥ R3
P ( Dup ≤ 32cm) ≥ R4
P (VClow ≤ 0.32cm) ≥ R5
P (VCmid ≤ 0.32cm) ≥ R6
P (VCup ≤ 0.32cm) ≥ R7
P ( Fps ≤ 4.0 KN ) ≥ R8
P (VB − Pillar ≤ 9.9m / s ) ≥ R9
Figure 11. Finite element model for vehicle side impact P (Vdoor ≤ 15.69m / s ) ≥ R10
example. µ iL ≤ µ i ≤ µUi i =1~ 7

In side impact design, the increase of gage design and µ8 , µ9 = 0.345 or 0.192 (material properties)
variables tends to improve the dummy performance. However, µ10 , µ11 = 0.0 (barrier height and position)
the vehicle weight is simultaneously increased, which is where Ri (i = 1 ~ 10) are the desired reliabilities for each
undesirable. For this reason, an optimization problem is
formed by minimizing the vehicle weight subject to a number constraint, FAL is the dummy abdomen load, Dup , Dmid and Dlow
of safety constraints on the dummy according to the EEVC are the dummy upper rib, middle rib and lower rib deflections
procedure. They include HIC, abdomen load, rib deflection or and VCup , VC mid and VClow are the dummy upper chest, middle
V*C, and pubic symphysis force. chest and lower chest viscous criterion values, respectively.
Also, Fps is the dummy pubic symphysis force, V B − Pillar is the
Table 3. Description and range of design variables for
vehicle side impact example. velocity at the middle B-pillar position and V door is the B-
pillar velocity at door belt line. A 99.87% reliability ( β = 3 )
Random Design Variables &
Range is used for all ten constraints. All random variables and
Parameters
Description Minimum Nominal Maximum parameters are assumed normally distributed with standard
B-Pillar inner 0.5 1.0 1.5 deviations σ 1− 4,6,7 = 0.03 , σ 5 = 0.05 , σ 8,9 = 0.006 and σ 10,11 = 10.0 .
The deterministic optimization reduced the vehicle weight to
Thickness (mm)

B-Pillar reinforcement 0.45 0.9 1.35

Floor side inner 0.5 1.0 1.5 23.5 (see Table 4). However, the reliability of the front-door
Cross members 1 & 2 0.5 1.0 1.5 velocity ( G2 ), lower rib deflection ( G9 ) and pubic force ( G10 )
door beam 0.875 1.75 2.625
Door belt line constraints were found to be unacceptably low. To the
0.4 0.8 1.2 contrary, the reliability-based design optimization increases
reinforcement
Roof rail 0.4 0.8 1.2 slightly the vehicle weight to 28.6 but satisfies all probabilistic
Material of B-Pillar inner mild steel mild steel HS steel constraints with at least a 99.87% reliability. The RBDO
Material of floor side inner mild steel mild steel HS steel problem was solved with the DLP/PMA, SORA and single-
Barrier height (mm) -30 0 +30 loop methods. Due to the excessive computational cost
Barrier hitting position (mm) -30 0 +30 associated with vehicle crashworthiness, a set of response
surfaces was used.

9 Copyright © 2004 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/30/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 The traditional, double-loop RBDO algorithm requires nested
34 optimization loops, where the design optimization (outer)
loop, repeatedly calls a series of reliability (inner) loops. Due
Obj. Function 29 to the nested optimization loops, the computational effort can
be prohibitive for practical problems. For this reason, efforts
24
Det. Opt.
to improve the RBDO efficiency have been pursued in 1)
19 improving the efficiency of the reliability analysis and 2) in
Single-Loop
(Beta=3) developing new RBDO formulations that either decouple the
14 design optimization and reliability loops or combine them into
1 2 3 4
an equivalent single-loop approach. The reliability index
Iteration No. approach (RIA) and the performance measure approach
Figure 12. Comparison of single-loop method with (PMA) have been extensively used for the inner reliability
deterministic optimization for the side impact example. loop. It has been found that the PMA can be computationally
more efficient than RIA and numerically more stable.
Among all decoupling RBDO methods reported in the
Fig. 12 and Table 4 compare the single-loop method literature, the sequential optimization and reliability
with deterministic optimization for 99.87% reliability assessment (SORA) method is the most promising. It consists
( β = 3 ) for all constraints. Both single-loop and of a sequence of deterministic optimization followed by
reliability analyses. Although SORA is more efficient than the
deterministic optimization calculate their respective optimum double-loop RBDO methods, it can still be costly due to the
design in three iterations and 35 function evaluations. This is required repeated reliability analyses which follow each
another indication of the excellent efficiency of the single- deterministic optimization.
loop method. The deterministic and single-loop optima are A robust single-loop RBDO algorithm for both normal
23.58 and 28.69, respectively. As shown in Table 4, the and non-normal random variables has been presented in this
number of function evaluations for single-loop, DLP/PMA paper which greatly improves on the existing single-loop,
and SORA are 35, 13846 and 808, respectively. All three single vector (SLSV) algorithm. Based on the percentile
RBDO methods have given comparable optimum solutions. formulation of the reliability assessment, it collapses the
nested optimization loops into an equivalent single-loop
Table 4. Comparison of RBDO methods for the side impact optimization process by imposing the Karush-Kuhn-Tucker
example. optimality conditions of the reliability loops as equivalent
Design SINGLE deterministic equality constraints of the design optimization
Det. Opt DLP/PMA SORA
Variables LOOP loop. It therefore, converts the probabilistic optimization
x1 0.5 0.9436 0.7916 0.81 problem into an equivalent deterministic optimization
x2 1.2257 1.35 1.35 1.35 problem, eliminating the need for calculating the Most
x3 0.5 0.9127 0.6919 0.7277 Probable Point (MPP) in repeated reliability assessments. Its
x
accuracy is the same with the double-loop approach and its
1.2071 0.9913 1.5 1.5
4
efficiency is almost equivalent to deterministic optimization.
x5 0.875 0.9026 0.9466 0.875
Several numerical applications including an automotive
x6 0.9352 1.2 1.2 1.2 vehicle side impact example, demonstrated the accuracy and
x 7 0.4 0.4 0.4 0.4 superior efficiency of the proposed single-loop RBDO
Objective f(X) 23.5856 28.6528 28.4851 28.6977 algorithm.
G1 (X ) 0.4364 0.2911 0.5182 0.5222 As a final remark, it should be noted that the proposed
G 2 (X ) 0 0 0 0 single-loop method is practically an equivalent gradient-based
G 3 (X )
deterministic optimization method. Therefore, it has all
4.3359 2.0127 0.0514 0.0512
practical limitations of gradient-based optimization methods.
G 4 (X ) 3.5668 1.8895 0.0277 0.0582
It may converge for example, to a local optimum which may
G 5 (X ) 0.0897 0.1268 0.2280 0.239 not be the true MPP or it may be unable to converge at all,
G 5 (X ) 0.1163 0.1284 0.3134 0.3179 depending on the initial point, for highly nonlinear nonconvex
G 7 (X ) 0.0275 0.0455 0.0483 0.0512 problems. The reader must be aware that gradient-based
G 8 (X ) 0 0.0001 0 0 probabilistic optimization methods carry all practical
G 9 (X ) 0.5577 0.0881 0.03788 0.0378
limitations of gradient-based deterministic optimization.
G 10 ( X ) 0.0276 0.0003 0 0.0371
ACKNOWLEDGMENTS
No. of Iteration 3 8(outer) 3(cycles) 3
No. of F. E. 35 13846 808 35
The present study was performed with funding from the
General Motors Research and Development Center. The
support is gratefully acknowledged.
The authors would like to thank Dr. Ren-Jye Yang from
5. SUMMARY AND CONCLUSIONS Ford Motor Company for providing the definition, description
Reliability-Based Design Optimization (RBDO) provides and response surfaces for the vehicle side impact example.
optimum designs in the presence of uncertainty in design Special thanks are also due to Prof. Wei Chen from
variables/parameters and simulation models. It is therefore, a Northwestern University for providing the SORA results for
powerful tool which can assist in design under uncertainty.

10 Copyright © 2004 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/30/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 all three examples as well as many useful discussions. The 16. Gu, L. and Yang, R. J., “Optimization and Robustness for
continuous support and enthusiasm of Bob Lust, Mary Fortier, Crashworthiness of Side Impact,” International Journal of
John Cafeo and Xuru Ding of General Motors Corporation and Vehicle Design, 26(4), 348-360, 2001.
Prof. S. Mahadevan from Vanderbilt University during this 17. Rackwitz, R. and Fiessler, B., “Reliability under
study, are greatly appreciated. Combined Random Load Sequences,” Computers and
Structures, 9(5), 489-494, 1978.
REFERENCES 18. Breitung, K., “Asymptotic Approximations for Multi-
1. Reddy, M. V., Granhdi, R. V. and Hopkins, D. A., normal Integrals”, Journal of Engineering Mechanics, ASCE,
“Reliability Based Structural Optimization: A Simplified 110(3), 357-366, 1984.
Safety Index Approach,” Computers and Structures, 53(6), 19. Tvedt, L., “Distribution of Quadratic Forms in Normal
1407-1418, 1994. Space – Application to Structural Reliability”, Journal of
2. Yu, X., Chang, K. H. and Choi, K. K., “Probabilistic Engineering Mechanics, 116(6), 1183-1197, 1990.
Structural Durability Prediction,” AIAA Journal, 36(4), 628- 20. Köylüoğlu, H. U. and Nielsen, S. R. K., “New
637, 1998. Approximations for SORM Integrals”, Structural Safety, 5,
3. Lee, J. O., Yang, Y. O. and Ruy, W. S., “A Comparative 119-276, 1994.
Study on Reliability Index and Target Performance Based 21. Cai, G. Q. and Elishakoff, I., “Refined Second-Order
Probabilistic Structural Design Optimization,” Computers and Reliability Analysis,” Structural Safety, 14, 267-276, 1994.
Structures, 80, 257-269, 2002. 22. Du, X. and Chen, W., “Sequential Optimization and
4. Mahadevan, S., “Probabilistic Optimum Design of Framed Reliability Assessment Method for Efficient Probabilistic
Structures,” Computers and Structures, 42(3), 365-374, 1992. Design,” Proceedings of ASME Design Engineering Technical
5. Pu, Y., Das, P. K. and Faulkner, D., “A Strategy for Conferences (DETC), Paper # DETC2002/DAC-34127, 2002.
Reliability-Based Optimization,” Engineering Structures, 19, 23. Royset, J. O., Der Kiureghian, A. and Polak, E.,
276-282, 1997. “Reliability-based Optimal Design of Series Structural
6. Wu, Y.-T., “Computational Methods for Efficient Structural Systems,” Journal of Engineering Mechanics, 607-614, 2001.
Reliability and Reliability Sensitivity Analysis,” AIAA 24. Royset, J. O., Der Kiureghian, A. and Polak, E.,
Journal, 32(8), 1717-1723, 1994. “Reliability-based optimal structural design by the decoupling
7. Wu, Y. –T. and Wang, W., “A New Method for Efficient approach,” Reliability Engineering & System Safety, 73, 213-
Reliability-Based Design Optimization,” Proceedings of 7th 221, 2001.
Special Conference on Probabilistic Mechanics & Structural 25. Thanedar, P. B. and Kodiyalam, S., “Structural
Reliability, 274-277, 1996. Optimization Using Probabilistic Constraints”, Structural
8. Padmanabhan, D. and Batill, S. M., “Decomposition Optimization, 4, 236-240, 1992.
Strategies for Reliability Based Optimization in 26. Kwak, B. M. and Lee, T.W., “Sensitivity Analysis for
Multidisciplinary System Design,” Proceedings of 9th Reliability-Based Optimization using an AFOSM Method,”
AIAA/SSMO Symposium on Multidisciplinary Analysis and Computers and Structures, 27(3), 399-406, 1987.
Optimization, 2002. 27. Chen, X., Hasselman, T. K. and Neill, D. J., “Reliability
9. Zou, T., Mourelatos, Z. P., Mahadevan, S. and J. Tu, Based Structural Design Optimization for Practical
“Component and System Reliability Analysis Using an Applications,” Proceedings of the 38th
Indicator Response Surface Monte Carlo Approach,” AIAA/ASME/ASCE/AHS/ASC Structures, Structural
Proceedings of ASME Design Engineering Technical Dynamics, and Materials Conference, 1997.
Conferences, DETC2003/DAC-48708, 2003. 28. Wang, L. and Kodiyalam, S., “An Efficient Method for
10. Tu, J., “Design Potential Concept for Reliability-Based Probabilistic and Robust Design with Non-Normal
Design Optimization,” Ph.D. Thesis, The University of Iowa, Distributions,” Proceedings of 43rd AIAA SDM Conference,
1999. Denver, Colorado, 2002.
11. Tu, J., Choi, K. K. and Park, Y. H., “A New Study on 29. Yang, R. J. and Gu, L., “Experience with Approximate
Reliability-Based Design Optimization”, ASME Journal of Reliability-Based Optimization Methods”, AIAA-2003-1781,
Mechanical Design, 121, 557-564, 1999. 2003.
12. Tu, J., Choi, K. K. and Park, Y. H., “Design Potential 30. Gu, L. and Yang, R. J., “Recent Applications on
Method for Robust System Parameter Design”, AIAA Journal, Reliability Based Optimization of Automotive Structures,”
39(4), 667-677, 2001. Proceedings of SAE World Congress, Paper No. 2003-01-
13. Choi, K. K. and Youn, B. D., “Hybrid Analysis Method 0152, 2003.
for Reliability-Based Design Optimization,” Proceedings of 31. Wu, Y.-T., Shin, Y., Sues, R. and Cesare, M., “Safety –
ASME Design Automation Conference, 2001. Factor Based Approach for Probabilistic – Based Design
14. Wu, Y.-T. and Wang, W., “Efficient Probabilistic Design Optimization,” 42nd AIAA/ASME/ASCE/AHS/ASC Structures,
by Converting Reliability Constraints to Approximately Structural Dynamics and Materials Conference, Seattle, WA,
Equivalent Deterministic Constraints, ” Journal of Integrated 2001.
Design and Process Sciences, 2(4), 13-21, 1998. 32. Haldar, A. and Mahadevan, S., Probability, Reliability and
15. Yang, R. J., Gu, L., Tho, C. H., Choi, K. K. and Youn, B. Statistical Methods in Engineering Design, John Wiley &
D., “Reliability-Based Multidisciplinary Design Optimization Sons, Inc., 2000.
of a Full Vehicle System,” Proceedings of the 43rd 33. Ang, A. H-S. and Tang, W. H., Probability Concepts in
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Engineering Planning and Design; Vol. II – Decision, Risk
Dynamics, and Materials Conference, 2002. and Reliability, ISBN 0-471-03201-8, 2000.

11 Copyright © 2004 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/30/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 34. Hohenbichler, M. and Rackwitz, R., “Non-Normal
Dependent Vectors in Structural Reliability,” Journal of the
Engineering Mechanics, ASCE, 107(6), 1227-1238, 1981.
35. Papalambros, P. Y. and Wilde, D. J., Principles of Optimal
Design; Modeling and Computation, 2nd Edition, Cambridge
University Press, 2000.
36. Vanderplaats, G. N., Numerical Optimization Techniques
for Engineering Design, McGraw-Hill, 1984.

12 Copyright © 2004 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 01/30/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use