49th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference <br>16t AIAA 2008-1801

7 - 10 April 2008, Schaumburg, IL

Reliability Based Design Optimization Formulations for

Component and System Reliability
Mark McDonald1 and Sankaran Mahadevan2
Vanderbilt University, Nashville, TN, 37232

Reliability-based design optimization (RBDO) of mechanical systems is computationally

intensive due to the presence of two types of iterative procedures – design optimization and
reliability estimation. Single-loop RBDO algorithms offer tremendous savings in
computational effort, but they have so far only been able to consider individual component
reliability constraints. This paper presents a single-loop RBDO formulation and an
equivalent formulation that can also include system-level reliability constraints. The
formulations allow the allocation of optimal reliability levels to individual component limit
states in order to satisfy both system-level and component-level reliability requirements.
Four solution algorithms to implement the second, more efficient formulation are developed.
A key feature of these algorithms is to remove the most probable points from the decision
space, thus avoiding the need to calculate Hessians or gradients of limit state gradients. It is
shown that with the proposed methods, systems-level RBDO can be accomplished with
computational expense equivalent to several cycles of computationally inexpensive single-
loop RBDO based on second-moment methods. Examples of this new approach applied to
series, parallel, and combined systems are provided.

Nomenclature

Α = negative normalized limit state function gradient vector

Β = reliability index
µplatethickness = mean plate thickness
ρ = correlation between failure modes
σx = normal stress in x-direction
σy = normal stress in y-direction
σxy = shear stress in x-y plane
σyield = aluminum yield stress
A = truss member cross sectional area
B = vector of component reliability indices
D = vector of design variables
Fy = steel yield stress
g( ) = limit state function
K = set of limit states
Nx = running load (stress per unit thickness) in x-direction
Ny = running load in y-direction
Nxy = running shear load
P = load on truss
R = failure mode correlation matrix
tHC = honeycomb thickness
tplate = plate thickness
u = vector of transformed random variables
u* = most probable point of failure
x = vector of random variables in original space

1
Graduate Student, Department of Civil and Environmental Engineering, VU Station B 356077, Student Member,
AIAA
2
Professor, Department of Civil and Environmental Engineering, VU Station B 356077, Member AIAA

1
American Institute of Aeronautics and Astronautics

Copyright © 2008 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved.
 I. Introduction

T he design of a reliable system at minimum cost, or conversely the maximization of system reliability subject to a
cost constraint, taking into account uncertainties in design and system variables is the general objective of
reliability based design optimization (RBDO). Most RBDO methods in the literature only consider component-level
reliability constraints. This paper investigates efficient methods for including system-level reliability constraints
within RBDO.
RBDO methods fall into three groups depending upon how reliability analysis is incorporated into the
optimization process. Nested algorithms include a full reliability analysis at every step of the design optimization
algorithm. Decoupled methods sequentially perform optimization and either direct or inverse reliability analyses
until convergence is achieved.1-6 Single loop algorithms7-12 perform optimization and reliability analysis
simultaneously. Nested algorithms are usually more expensive than decoupled algorithms; decoupled algorithms are
usually more expensive than single-loop algorithms.
System reliability can be included in the three types of RBDO methods in different ways. In the case of nested
RBDO, systems reliability constraints can be evaluated after each optimization iteration, but this is not
computationally efficient. In decoupled RBDO, systems reliability constraints can be handled in one of two ways.
The probabilistic constraints in the outer optimization can be replaced with a Taylor series expansion, and the failure
probabilities and their gradients are updated after each solution of the outer level optimization probability.6
Alternately, the design can be optimized for a set of component reliability targets while ignoring the system
reliability constraint, and checking to see that the design yielded by the RBDO process meets system level reliability
requirements, and repeating the whole process with heuristically increased component-level reliability targets if
necessary.4
Single-loop algorithms for RBDO have proven to be very computationally efficient13, but currently are only
applicable to component reliability constraints. The purpose of this paper is to develop efficient RBDO methods
capable of handling system-level reliability constraints. A single-loop RBDO formulation and an equivalent
formulation that can include both component-level and system-level reliability constraints are developed. The
formulations allow the allocation of optimal reliability levels to individual component limit states in order to satisfy
both system-level and component-level reliability requirements. Four solution algorithms to implement the second,
more efficient formulation are developed. It is shown that with the proposed methods, systems-level RBDO can be
accomplished with computational expense equivalent to several cycles of computationally inexpensive single-loop
RBDO based on second-moment methods.

II. System Reliability

We will utilize approaches to system reliability based on the first order reliability method (FORM). The first-order
reliability method (FORM) is used extensively in the implementation of single-loop RBDO algorithms. In FORM, a
limit state function g(x) is used to characterize the state of the component as failed or safe, and the probability of
failure of the component is defined as
PF = P{g (x) ≤ 0} = ∫f
g ( x )≤0
x (x)dx (1)

where fx(x) is the joint probability density of variables x1, x2… xn. An analytical evaluation of the integral in Eq. (1)
is possible in only a few special cases, and hence numerical integration is necessary. FORM has been found to be a
computationally efficient and reasonably accurate analytical approximation to the probability integral under
consideration for many problems. See Haldar and Mahadevan14 for details related to the implementation of FORM.
The results of the individual component reliabilities in FORM can also be used to estimate system reliability. For
series and parallel systems the failure probability is defined as

PF , Series = P{U g k (x) ≤ 0} (2.1)

k

PF , Parallel = P{I g k (x) ≤ 0} (2.2)

k
Let B be the vector of reliability indices for each of the limit states and the elements of the matrix R be the dot
products of the corresponding α vectors obtained from the FORM analysis for each distress mode. Then for a series
system, the system failure probability is given by 1 – Φ(B, R), where Φ(B, R) is the standard normal multivariate
CDF with correlation matrix R. Likewise, the failure probability of the parallel system can be calculated from the

2
American Institute of Aeronautics and Astronautics
results of the FORM analysis of its components as Φ(-B, R). More complex systems can be represented as
combinations of series and parallel systems. Details of system reliability evaluation are available in several text
books; see for example Haldar and Mahadevan14 and the references therein. Ambartzumian et al15 has proposed a
very efficient method based on sequentially conditioned importance sampling for evaluation of the multinormal
CDF, and Pandey16 has developed a very accurate moment-based approximation to the multinormal CDF.

III. Novel Formulation for RBDO with Component and System Reliability Constraints
The algorithm presented here for systems level RBDO is FORM-based, as in other single-loop approaches7-13, but
solves the design optimization problem in an augmented decision space which includes the component reliability
targets. The algorithm presented here is similar to the single-loop approaches previously discussed, but uses an
augmented decision space and appropriate multinormal CDF approximations to assure that the system-level
probability of failure is less than a specific threshold.
In this approach the FORM optimality conditions are satisfied for each component in the system. These
optimality conditions become constraints in the design optimization, completely eliminating the inner loop.
However, the component reliability targets are now included in the decision space of the optimization problem,
along with the design vector and each component MPP. The component reliability targets are constrained to be
equal to the norm of the MPP vector in the standard normal space. These component reliability targets, along with
the correlations obtained from the limit state direction cosines α, are then used to calculate the system failure
probability, which must be less than the allowable threshold

min f (d) (3.1)

d ,u k ,B

s.t. Gk (d, u k ) = 0 ∀k ∈ K limit states (3.2)

uk ∇ G (d, u)
=− u k ∀k ∈ K limit states (3.3)
uk ∇ u Gk (d, u)
uk = βk ∀k ∈ K limit states (3.4)
PF , SYS (B, R ) ≤ PF , ALLOWABLE (3.5)

Eq. (3.2) and Eq. (3.3) are the necessary requirements for a valid MPP. Eq. (3.4) requires the reliability indices
to be the value of the norm of the MPP vector in the standard normal space. Eq. (3.5) is the explicit system
reliability requirement. This probability can be evaluated as a deterministic function given any vector of reliability
indices and failure mode correlations, using the methods in Section 2. It should also be noted that bounds can be
placed on any of the component limit state reliability indices should the designer find any arbitrary reliability index
unacceptable for some limit states.
It should be noted that Eq. (3) is formulated in a way that is consistent with All-at-Once approaches to
multidisciplinary design optimization17 and with the Rackwitz-Fiessler algorithm18 for implementation of FORM .
Feasibility, therefore, is required only at convergence, and it is not necessary for the optimization algorithm to
assure feasibility at every step. Feasibility need only be achieved at convergence because the formulation is a typical
nonlinear programming problem, and it can be solved by Lagrange multipliers. Optimal values for the decision
variables are found by solving the first-order optimality conditions as a set of nonlinear equations using, for
instance, a Newton-type algorithm.
In the formulation given in Eq. (3), the system reliability constraint is not simply a feasibility check, but rather it
drives the entire optimization process. This is obvious because if a Newton-step is used to find the stationary points
for the Lagrangian function for this optimization problem, the system reliability constraint will become a term of the
Lagrangian. The inclusion of this constraint allows for the optimal allocation of reliability to the system
components.
If a gradient-based optimization algorithm is used, the bulk of the computational expense is in evaluating the
gradient and/or Hessian of Eq. (3) with respect to the augmented decision space. This involves evaluating the
Hessian of the gradient of the limit state function, and at every iteration, the computational expense required to
perform one optimization iteration measured in terms of limit state evaluations is at least an order of magnitude
more than the limit state evaluations required to perform second-order reliability analysis if a Newton-type of search
algorithm is used. Rackwitz19 has stated that the computational expense involved in reliability-based design
optimization by single loop methods is roughly an order of magnitude greater than that required for reliability

3
American Institute of Aeronautics and Astronautics
analysis alone. This statement is also true for the proposed algorithm. In fact, the expense required for the system-
level RBDO algorithm presented here is on the order of that required for single-loop algorithms which only consider
component reliability constraints. This approach is far less computationally expensive than traditional nested
methods and may be more computationally efficient than decoupled methods. However, in the next section, the
major problems associated with evaluation of gradients and/or Hessians of gradients of limit state functions with
respect to the random variables are addressed and overcome, as equivalent formulations of the single-loop algorithm
derived here are solved over a decision space consisting only of the design variables and reliability targets.
The optimization given in Eq. (3) may be solved directly by an appropriate nonlinear programming technique.
However, this approach is quite inefficient due to the limit state function’s gradients appearing as constraints in the
single loop formulation and the fact that either gradients or Hessians of these gradients must be evaluated at every
step. Therefore, another solution strategy is developed in this section which has the advantage of optimizing with
respect to only the design variables and reliability targets, thereby removing the most probable points from the
decision space. This formulation is given by

min f (d) (4.1)

d,B
s.t.
(4.2)
uk (βk ) = α k * βk ∀k ∈ K limit state functions
g k (d, u k (β k )) = 0 ∀k ∈ K limit state functions (4.3)
Pf ,SYS (B, R ) ≤ Pf , ALLOWABLE (4.4)

The RBDO formulation in Eq. (4) can be solved in four ways, depending on how the vector αk is calculated and
updated: (1) αk can be evaluated corresponding to the current value of the decision vector, without any iterations to
converge to any MPP. (2) αk can be evaluated only between groups of optimization iterations and held constant
within each group. A group is defined as a set of optimizer iterations ending in convergence based on the current
value of αk. Again in this case, αk is calculated as in option 1 above. (3) αk can be calculated based on a full FORM
analysis of the limit state function after each optimization iteration. (This search procedure is similar to a nested
double loop approach, even though it solves the single-loop problem formulation.) (4) αk can be calculated based on
a full FORM analysis of the limit state function between groups of optimization iterations and held constant within
each group. (This search procedure is similar to decoupled or sequential approaches, even though it solves a single-
loop problem formulation). The details of each of the algorithms are given below:

For algorithms to implement options 1 through 4, initialization is as follows:

Step 0a: Initialize the design variables d0. (A good choice for many problems is to choose the design resulting from
optimization at the expecations of the random variables).

Step 0b: Initialize the iteration points for the MPP’s, u0. (A good choice in the absence of other information would
be the zero vector; however, if available, the results of a FORM analysis of the first candidate design would likely
achieve faster convergence).

Algorithm 1:

Step 1: Calculate the α vectors at the current iteration points, ui, as the negative normalized gradient of the limit
state function in u-space with respect to the random variables.

Step 2: Perform one search iteration for the optimization in Eq. (12) to find a new design vector di+1 and new
iteration points for the MPP’s, ui+1. (Recall that u = αβ.)

Step 3: If both di+1 and the ui+1 have converged, then di+1 is a candidate for an optimal design. Otherwise i = i+1,
and go to Step 1.

4
American Institute of Aeronautics and Astronautics
Algorithm 2:

Step 1: Calculate the α vectors at the current iteration points, ui, as the negative normalized gradient of the limit
state function in u-space

Step 2: Solve the optimization in Eq. (12) to find a new design vector di+1 and new iteration points for the MPP,
ui+1. (Recall that u = αβ.)

Step 3: If both di+1 and the ui+1 have converged, then di+1 is a candidate for an optimal design. Otherwise i = i+1,
and go to Step 1

Algorithm 3:

Step 1: Calculate the α vectors from a full FORM analysis of the current design di.

Step 2: Solve the optimization in Eq. (12) to find a new design vector di+1 and new iteration points for the MPP’s,
ui+1. (Recall that u = αβ.)

Step 3: If both di+1 and the ui+1 have converged, then di+1 is a candidate for an optimal design. Otherwise i = i+1,
and go to Step 1.

Algorithm 4:

Step 1: Calculate the α vectors from a full FORM analysis of the current design di.

Step 2: Perform one search iteration for the optimization in Eq. (12) to find a new design vector di+1 and new
iteration points for the MPP’s, ui+1. (Recall that u = αβ.)

Step 3: If both di+1 and the ui+1 have converged, then di+1 is a candidate for an optimal design. Otherwise i = i+1,
and go to Step 1.

All four algorithms are similar in that they all use approximate direction cosines to generate a MPP at which the
design is optimized. All four algorithms update the direction cosines periodically. Upon convergence of the design
variables and the direction cosines the solution to Eq. (4) is equivalent to the solution of Eq. (3). For all four cases
Eq. (4.2) is used as a condition and Eq. (4.3) is a constraint. This is why the MPP’s are dropped from the augmented
decision space for all four solution algorithms for Eq. (4) and results in the removal of many equality constraints
from the formulation in Eq. (3).
For options 1 and 2, each optimization cycle involves approximately the same level of computational expense as
that required for single-loop algorithms based on second moment methods. In most cases convergence should be
obtained after relatively few optimization cycles, with more cycles involved for limit states that are strongly
nonlinear or random variables that are strongly non-normal. If options 1 and 2 have difficulty in convergence,
options 3 and 4 may provide more robustness (and possibly more efficiency for highly nonlinear limit states), with
option 4 usually being more efficient than option 3. All methods will increase in computational expense with the
number of design variables on an order of the square of the number of variables if a Newton step is used. In any
case, it is clear that the formulation in Eq. (4) gives a more efficient optimization procedure than the formulation in
Eq. (3).
Because the approaches of Eq. (4) allow the MPP’s to be dropped from the decision space, it is not
necessary for the optimization algorithm to evaluate gradients of limit state function gradients or Hessians of limit
state function gradients with respect to the random variables. We expect the computational savings of a larger scale
problem to be much greater as more of these calculations are eliminated.
It is desirable that RBDO methods be easy to integrate with any commercial optimization software or
algorithm, without needing to modify or go inside the commercial software. Options 2 and 3 satisfy this
requirement. However, for options 1 and 4, one needs to go inside the optimization algorithm in order to get the

5
American Institute of Aeronautics and Astronautics
design point at the current iteration and then calculate the gradient vectors. This requires that one must control the
process of optimization. This may not be possible with commercial optimization software.
To summarize, single loop methods offer the potential for large computational savings when they succeed. It is
noted that there is no proof of convergence for these or any other single loop or decoupled RBDO methods, and that
these and all other single-loop RBDO formulations are based on FORM, which may not be accurate for all
applications. Optimization using Eq. (3) is relatively inefficient, however, because of the appearance of limit state
function gradients in the constraints. This results in large computational expenditure if a gradient-based solution
approach is directly applied. Also, the decision space is much larger, and the algorithm will likely need more
iterations to converge. Eq. (4) avoids this expense by removing the MPPs from the decision space and the limit state
gradients from the constraints.

IV. Numerical Examples

A. Parallel System Example Problem: RBDO of a Daniels System

Consider the three bar assembly in Fig. 1. All three bars are assumed identical in size, and are made of A36 steel.
They are assumed to each carry a load of P/3 until all three bars have yielded. The steel is assumed to have a
normally distributed yield stress with mean 36 ksi and standard deviation 3 ksi, and the yield stresses of each of the
bars is identically and independently distributed. The load is normally distributed with mean 90 kips and standard
deviation 9 kips.

P
Figure 1. Three-Bar Daniels System

We wish to minimize the bar size (cross section area, assumed deterministic), subject to the probability of all
three bars yielding is less than 1/1000. (Note that while A is assumed deterministic and is a design variable, the bar
cross sectional areas could also be modeled as random. In this case, the designer might be able to control the
statistical moments of the distribution of A.) The limit state functions for the system are g i = AF yi − P / 3 for i =
1,2,3. The formal statement of the problem is:

min A (5.1)
A,B
s.t.
(5.2)
u i ( β i ) = α i * β i ∀i ∈ 1,2,3
g i (d, u i ( β i )) = 0 ∀i ∈ 1,2,3 (5.3)
P{( g1 ≤ 0) I(g 2 ≤ 0)I(g 3 ≤ 0)} ≤ 0.001 (5.4)

6
American Institute of Aeronautics and Astronautics
 A≥0 (5.5)

Thus the design variable is the cross sectional area of the bars. The random variables are the load and the yield
stress. The design bound is that the area must be greater than zero.
Pandey’s 16 method is used for evaluating the system reliability in Eq. (13.4). Optimization according to Eq. (3)
yielded the results shown in Table 1. This solution started from Fy = 33 ksi for all three bars, P = 99 kips for all three
limit states, and from initial reliability targets of 2.0 for all three component reliability targets, the final values of
which are all to be decided by the optimization algorithm. This problem was hot started instead of starting at the
mean values due to the numerical instability of Eq. (3), and the same starting point was used for the sake of
comparison when the algorithms for Eq. (4) were applied. The original design was the optimal design taken at the
means of the random variables (A=0.866 in2). Convergence was based on a tolerance of 0.01 percent for the design
variables and 0.00001 for the equality constraints. Solution of the problem was done using a Newton search with a
forward difference gradient evaluation method. The optimal solution was found after 74 iterations and required
1,110 evaluations of each limit state function.
It can be seen from the results that at the end of the optimization algorithm the system reliability is exactly equal
to the required system reliability of 0.001. The algorithm used here is able to satisfy the system reliability constraint
exactly as it is evaluated in each step of the optimization algorithm. The results show that at optimality, the unit
vector pointing in the direction of the MPP is also equal to the negative normalized gradient vector of the failure
surface evaluated at the MPP, shown as alpha in Table 1, Furthermore, each limit state function is equal to zero at
the MPP corresponding to it. Hence, we see that the solution satisfies all the optimality conditions for FORM. It
should also be noted that restrictions could be placed on component reliability targets if the designer wished to
assure certain component reliabilities in addition to the system reliability.
This problem was also solved as formulated in Eq. (4) with all four solution strategies described in the previous
section. The original iteration point for the random variables was taken to be the means of the random variables. The
original design was taken as A=0.866 in2 and the starting reliability indices were taken to be 2.0. All four options
yielded identical results to those shown in Table 1. To further illustrate the value of this single-loop approach to
system-level RBDO, this same problem is solved by using a decoupled performance measure approach, which
utilizes inverse FORM, and specifying component reliability targets. The system reliability constraint is enforced by
evaluating the system reliability after the decoupled optimization is complete. If the system reliability constraint is
not met, then the active component reliability targets are arbitrarily increased without any quantitative search
procedure. The results for optimization using several component reliability indices are summarized in Table 2. It
should be noted that FORM is exact in this case because of the linearity of the limit state functions in normal
random variables. Included in the table is the component reliability target, the cross-sectional area for the optimal
design, the yield stress and loading MPP in the original space for each component, and the system probability of
failure.
For the case where the target reliability indices were equal to 2.0, the optimization started from the same values
as those used for optimization in accordance with Eq. (4). The design optimization using the decoupled PMA
method required three complete cycles of optimization and inverse reliability analysis to converge to the solution in
Table 2. It required 11 iterations of an optimization algorithm and 12 iterations of an inverse reliability analysis
algorithm that utilized a Newton search procedure. The analysis required 69 function calls per limit state in all.
Comparison of the computational effort required for the three methods are shown in Table 3.
Inspection of the results in Table 2 for a component reliability index of 2.0 shows that the decoupled PMA
method produced a design very close to those obtained by the single-loop methods. Also, the component reliability
targets chosen have a major effect on the design, the design points, and the system reliability. In order to achieve the
optimal system design, multiple attempts of the decoupled PMA approach may be necessary to achieve the proper
component reliability targets, as opposed to the methods presented in this paper where the optimization algorithm
selects the component reliability targets. The arbitrary choice of component reliability targets for individual limit
states in a decoupled PMA approach make it inefficient to use for optimal allocation of reliability to components
corresponding to an overall system reliability target.
Table 3 contains a comparison of the computational expense required for all six approaches under consideration.
The column titled “additional function calls” includes the function evaluations required for evaluation of the
gradients of the limit state functions with respect to the random variables, or the additional calls necessary to
implement either direct or inverse FORM. The most striking observation is the computational expense required by
the single loop formulation of Eq. (3). The required computational effort is much higher for several reasons. Most
important is the fact that the formulation requires Hessians of gradients of the limit state functions, as discussed in
the previous sections. Also of considerable importance is that the decision space was much larger and that the

7
American Institute of Aeronautics and Astronautics
problem had many more constraints than the other two formulations had. These factors resulted in a much slower
convergence to the optimal design by the formulation in Eq. (3).

Table 1. RBDO Results for Three Bar Test Problem

DESIGN VARIABLE
Bar Cross-Sectional Area (in2) 1.08
LIMIT STATE FUNCTION VALUES
Bar 1 0
Bar 2 0
Bar 3 0
RELIABILITY TARGETS
Bar 1 Reliability Index 2.02
Bar 2 Reliability Index 2.02
Bar 3 Reliability Index 2.02
BAR 1 X* u* u*/lu*l α
Fy 31.55 -1.483 -0.734 -0.734
P 102.35 1.372 0.679 0.679
BAR 2 X* u* u*/lu*l α
Fy 31.55 -1.483 -0.734 -0.734
P 102.35 1.372 0.679 0.679
BAR 3 X* u* u*/lu*l α
Fy 31.55 -1.483 -0.734 -0.734
P 102.35 1.372 0.679 0.679
SYSTEM FAILURE PROBABILITY
PF,SYSTEM 0.001

Table 2. RBDO Results Using Decoupled PMA Approach

Component β A Fy P PF, System
1.0 0.948 33.94 96.53 3.67E-02
1.3 0.985 33.26 98.33 1.52E-02
1.7 1.037 32.32 100.62 3.75E-03
2.0 1.078 31.60 102.24 1.09E-03
2.3 1.121 30.85 103.78 2.77E-04
2.7 1.182 29.82 105.70 3.14E-05
3.0 1.229 29.02 107.04 4.98E-06
3.3 1.280 28.20 108.29 6.90E-07
3.7 1.352 27.08 109.81 2.91E-08
4.0 1.409 26.21 110.84 2.17E-09

Table 3. Comparison of Computational Effort for Three Bar Test Problem

8
American Institute of Aeronautics and Astronautics
 Optimizer Additional Total Function
Method Function Calls Function Calls Calls
Eq. (3) 1110 0 1110
Eq. (4) Option 1 42 21 63
Eq. (4) Option 2 30 6 36
Eq. (4) Option 3 24 24 48
Eq. (4) Option 4 30 12 42
Decoupled PMA 33 36 69

It is also important to notice that while the decoupled PMA procedure only required 69 limit state function
evaluations; this algorithm was hot started at the optimal values for the component reliability targets. The optimal
allocation of reliability to components under decoupled PMA requires a heuristic approach and would require
several iterations of the sequential process. The selection of an optimal set of reliability indices was simple in this
case due to the symmetry of the example problem. For practical problems selection of the optimal component target
reliability indices to satisfy a system reliability target is much more difficult, and the computational expense
required to achieve performance on the level of the formulation in Eq. (4) is likely to be much greater.
The advantage of the strategies used to solve Eq. (4) is that the optimization algorithm selected the component
reliability indices to optimize system reliability, and that the algorithm optimized the design without needing to
make a large number of function calls. These approaches were much faster than the other two because it made
intelligent use of the FORM KKT optimality conditions and used limited numbers of gradient evaluations in the
search procedure.
Among previous studies, the only viable decoupled approach to RBDO problems with system reliability
constraints involves representing probabilistic constraints with a Taylor series approximation, using the gradients of
the failure probability as calculated from a separate reliability analysis6. Perhaps one of the most attractive features
of the novel approach presented here is that component reliability targets are selected by the optimization algorithm
in a manner which provides allocation of reliability to the components of the system without having to represent the
probabilistic system-reliability constraint as a Taylor series as in the reliability index approach, thus not requiring
computation of the gradients of the system failure probability in a decoupled reliability analysis.

B. Series System Example: RBDO of a Liquid Hydrogen Tank

In this example problem the methods developed in this paper will be applied to a problem adapted from Smith
and Mahadevan 20. Under consideration is the design of a liquid hydrogen fuel tank to be used on a space launch
vehicle. The tank is to have a honeycomb sandwich design with top and bottom plates of aluminum alloy AL2024
and Hexcell 1/8”-5052.0015 as the sandwich material. The tank is subjected to stresses caused by ullage pressure,
head pressure, axial forces due to accelerations, and bending and shear stresses caused by fuel weight. The tank can
fail in three modes: von Mises strength, isotropic strength, or honeycomb buckling. For each failure mode an
appropriate limit state function can be written.
For the case of von Mises stresses in two dimensions the limit state function can be written as
gVM = σ yield − σ x 2 + σ y 2 − σ xσ y + 3σ xy 2 (6)
and the limit state function for isotropic strength can be written as
g ISO = σ yield − σ y (7)
However, for convenience, the design can be based on running loads, defined as the stress per unit of thickness
under plane stress conditions. Taking the yield stress as 84,000 psi times the plate thickness and using running loads
gives the following limit state function for the von Mises stress criterion:
84000 * t plate
gVM = −1 (8)
N x 2 + N y 2 − N x N y + 3N xy 2
The isotropic strength criterion is defined in terms of running loads to be

9
American Institute of Aeronautics and Astronautics
 84000 * t plate
g ISO = −1 (9)
Ny
The honeycomb buckling limit state function is more complex and is evaluated using a response surface generated
from finite element analysis performed using the structural sizing program Hypersizer 21, given by Smith and
Mahadevan20 as

g HCB = .847 + .96x1 + .986x 2 − .216x3 + .077x12 + .11x 2 2 + .007x3 2 + .378x1 x2 − .106x1 x3 − .11x 2 x3 (10)

where

x1 = 4(tplate - .075), x2 = 20(thc – 0.1), and x3 = -6000(1/Nxy + .003). (11)

The tank is divided longitudinally into ten sections and each section is divided into four panels as shown in
Figure 2. Each panel of the tank can fail in any one of the three failure modes at any location along the panel. This
is obviously a series system with an infinite number of limit states. However, the designs for each panel will be
based on a critical location, resulting in a finite number of limit state functions. Since the side panels are identical,
there are 30 limit states to consider. From this point forward the design of the number 10 side panel is under
consideration.

y
1
10
x

Top
Aluminum Plate

Side
Hexcell

Bottom

Figure 2. Schematic of Liquid Hydrogen Tank

The loads are considered random variables, as is the honeycomb thickness, which is an additional resistance
variable. The plate thickness is considered to be a random design variable, for which the designer selects a mean, but
not the standard deviation. Smith and Mahadevan (2005) characterized the distributions of the load variables as a
result of Monte Carlo simulation using Hypersizer. The assumptions for the random variables under consideration
are summarized in Table 4.
The tank designer should choose the minimum plate thickness subject to the condition that the series system
failure probability is less than 0.001. The formal statement of the problem is:

Table 4. Probability Distributions for Tank Random Variables

Variable Distribution Type Mean Standard Deviation
tplate Normal Des. Var. 0.005
tnc Normal 0.1 0.01
Nx Normal 13 60

10
American Institute of Aeronautics and Astronautics
 Ny Normal 4751 48
Nxy Normal -684 11

min µ plate thickness (12.1)

µ plate thickness ,B

u i ( β i ) = α i * β i ∀i ∈ I limit state functions (12.2)

g i (d, u i ( β i )) = 0 ∀i ∈ I limit state functions (12.3)
P{( gVM ≤ 0) U(g ISO ≤ 0) U(g HCB ≤ 0)} ≤ 0.001 (12.4)
µ plate thickness ≥ 0 (12.5)

Thus the design variable is the mean value of the plate thickness. The random variables are the plate thickness,
honeycomb thickness, and the running loads Nx, Ny, and Nxy,. The design bound is that the mean of the plate
thickness must be greater than zero.
Pandey’s16 method is used for multinormal integration. When the formulation of Eq. (3) was attempted starting
from the mean values plus one standard deviation for the load variables, the mean values minus one standard
deviation for the resistance variables, target reliability indices of 3, and the starting mean plate thickness of 0.058
inches, as found by optimization at mean values of the random variables, a Newton-type algorithm was not able to
find a feasible solution after 100 iterations. When the optimization approach described in Eq. (4), options 1 and 2,
was applied, starting from the same design, same target reliability indices, and the mean values of the random
variables, the elliptical nature of the von Mises failure surface caused the von Mises projected MPP to oscillate
about two regions of the limit state surface and prevented convergence. Convergence was obtained by using the
approach of Eq. (12) with options 3 and 4, obtaining identical results with either solution strategy. The results of the
optimization are shown in Table 5. Table 6 shows the final values of the iteration points for each limit state. Table 7
gives a summary of the limit state function evaluations required for the optimization and reliability analysis cycles
using the approach of Eq. (12), option 3. Table 8 gives a summary of the limit state function evaluations required
for the optimization and reliability analyses for the approach of Eq. (12), option 4.

Table 5. Results of Tank System RBDO

Variable Final Value
Mean Plate Thickness 0.07433
β, von Mises 3.242
β, Isotropic Strength 3.531
β, Honeycomb Buckling 9.561
System Failure Probability 0.001

Table 6. Final Iteration Points

Variable von Mises Isotropic Strength Honeycomb Buckling
Plate Thickness 0.0582 0.567 0.0696
Honeycomb Thickness 0.1 0.1 0.00518
Nx 13 13 13
Ny 4751 4770 4751
Nxy -685.7 -684 -677.5

Table 7. Number of Function Evaluations - Eq. (12) Option 3

von Mises Isotropic Strength Honeycomb Buckling
Optimizer Calls 60 60 60
FORM Calls 60 39 48

11
American Institute of Aeronautics and Astronautics
 Total Calls 120 99 108

Table 8. Number of Function Evaluations - Eq. (12) Option 4

von Mises Isotropic Strength Honeycomb Buckling
Optimizer Calls 72 72 72
FORM Calls 20 12 16
Total 92 84 88

This example shows that the solution strategies developed in Section 6 to solve the systems-level RBDO
problem offer a large advantage because each optimization cycle requires only a few limit state function evaluations,
but rapidly converges to the optimal design for many practical cases. Solving the single loop problem as a series of
successive optimizations also allows for appropriate component-level reliability and sensitivity analysis to be
performed between cycles of optimization to provide more accurate starting points for each successive optimization.
This improves the efficiency and robustness of the solution process. The ability to use the results of direct reliability
analysis to find the MPP is very useful in the case of nonlinear limit states, where the gradients change throughout
the space, because gradients are evaluated at a point much closer to the MPP associated with the final optimal
design. The ability of the solution process to harness such information may save multiple optimization cycles in
practice. Furthermore, each successive optimization cycle and FORM analysis will converge more quickly as the
starting points are closer to the optimal values.
An additional result is that option 4 required 12 optimization iterations as compared with 10 for option 3. For
highly nonlinear cases it is possible that while option 3 requires a direct FORM analysis of each limit state between
analyses, the reduction in the number of required optimization iterations may allow this to be less expensive that
option 4. The relative expense of the two options, however, is problem specific.
However, it should be noted that FORM is not always accurate for nonlinear limit states. At the optimum
solution for this problem, while FORM gave a system failure probability estimate of 0.0010, Monte Carlo
Simulation gave a system failure probability estimate of 0.0005. This affirms that the design is feasible, although the
first-order probability integration error led to a conservative design, and thus needs to be addressed. This can be
done heuristically by solving the optimization with higher values for the FORM-estimated system failure probability
constraint, or with lower values for system level reliability constraints4. Thus the problem was solved again by
modifying the FORM-estimated system failure probability constraint. The appropriate value for the constraint was
selected based on a simple bisection approach. For the resulting optimum solution, the FORM estimate of the failure
probability was 0.0021 while the Monte Carlo simulation estimate was 0.001, thus providing a less conservative, yet
still feasible design. The mean plate thickness (design variable), reliability targets (decision variables), and failure
probabilities are shown in Table 9. From the results in Table 9 it is clear that the von Mises failure criterion still
controls the design, and that accounting for the first-order probability integration error allowed a slightly less
conservative design.

Table 9. Updated Results of Tank System RBDO

Variable Solution 1 Solution 2
Mean Plate Thickness 0.0743 0.0733
β, von Mises 3.242 3.035
β, Isotropic Strength 3.531 3.326
β, Honeycomb Buckling 9.543 9.527
System Failure Probability (FORM) 0.001 0.0021
System Failure Probability (MCS) 0.0005 0.001
Note: Solution 1 was based on FORM alone. Solution 2 was obtained after the Monte Carlo check of Solution 1.

C. Combined Systems Example: Indeterminate Truss RBDO

Design of combined systems, which are combinations of series and parallel systems as described in Section 3,
can also be accomplished using the methodology developed in this paper. Consider the example of reliability-based

12
American Institute of Aeronautics and Astronautics
design optimization of a six-bar statically indeterminate truss as shown in Fig. 7. There are six limit states
(neglecting the buckling failure mode) for yielding in each individual bar. Failure of the entire truss would result
after failure of any two bars, resulting in fifteen minimal cut sets. These sets are {1,2}, {1,3}, {1,4}, {1,5}, {1,6},
{2,3}, {2,4}, {2,5}, {2,6}, {3,4}, {3,5}, {3,6}, {4,5}, {4,6}, and {5,6}. Assuming that the two diagonal members
carry equal load the six limit state functions can be written as follows:

g1 = A1 F1 − 0.707 P (13.1)
g 2 = A2 F2 − 0.707 P (13.2)
g 3 = A3 F3 − 0.500 P (13.3)
g 4 = A4 F4 − 0.500 P (13.4)
g 5 = A5 F5 − 0.500 P (13.5)
g 6 = A6 F6 − 0.500 P (13.6)

where P is the applied load, Ai is the cross-sectional area of the ith bar, and Fi is the yield stress for the material of the
ith bar. The designer wishes to minimize the total structural weight, which is proportional to 1.414(A1 + A2) + A3 + A4
+ A5 + A6, subject to the constraint that the system failure probability be less than 0.001. The load is assumed
normally distributed with a mean of 1000 kips and a standard deviation of 100 kips, and the yield stress of each bar
is assumed normally distributed with mean 36 and standard deviation 3.
6

P
1 2
L 3 5

L
Figure 3. Six Bar Indeterminate Truss

The cross-sectional areas of the bars are assumed deterministic, although the optimization could be performed
with respect to the mean of the cross sectional area, and treating the cross sectional areas as random variables. The
formal statement of the problem is:

min 1.414( A1 + A2 ) + A3 + A4 + A5 + A6 (14.1)

A ,B
s.t.
(14.2)
u i ( β i ) = α i * β i ∀i ∈ 1,2,3,4,56
g i (d, u i ( β i )) = 0 ∀i ∈ 1,2,3,4,5,6 (14.3)
P( U{( g i ≤ 0)I(g j ≤ 0)}) ≤ 0.001 ∀ i ∈ 1,...,5 j ∈ i + 1,...,6 (14.4)
A1 , A2 , A3 , A4 , A5 , A6 ≥ 0 (14.5)

Thus the design variables are the cross sectional areas of the bars. The random variables are the load and the yield
stresses. The design bound is that the areas must be greater than zero.
The failure probability for each cut set was evaluated as a parallel system using Pandey’s 16 method for
multinormal integration, and the combined system failure probability was estimated using first-order bounding. It

13
American Institute of Aeronautics and Astronautics
should be noted that while the overall system failure probability is approximated by a bounding formula, the
component reliabilities are exact due to the linearity of the limit state functions in the random variables. The
problem was attempted using a direct application of the approach in Eq. (3) and all four options for implementing
Eq. (4). The starting values for the design variables were obtained by optimization at the mean values. They are A1 =
A2 = 19.63 and A3 = A4 = A5 = A6 = 13.88. The starting values for all random variables were the means. When the
approach of Eq. (3) was attempted, the algorithm did not converge from the given starting point. However, all four
options for implementation of Eq. (4) converged to the results given in Table 10. A summary of the computational
expense required for each method is given in Table 11.

Table 10. Results of Indeterminate Truss System RBDO

Bar Area Reliability Index
1 28.57 2.89
2 28.32 2.83
3 20.94 3.16
4 20.83 3.12
5 20.66 3.06
6 20.29 2.92

Table 11. Number of Function Calls for

Indeterminate Truss Problem
Optimizer Additional Total Function
Method Function Calls Function Calls Calls
Eq. (3) No convergence
Eq. (4) Option 1 150 78 228
Eq. (4) Option 2 156 6 162
Eq. (4) Option 3 144 144 288
Eq. (4) Option 4 150 12 162

In Table 10, each execution of the optimizer requires six limit state functions to be evaluated. Thus 150 function
calls translates to 25 optimizer iterations. The additional function calls (i.e., limit state evaluations) in options 1 and
2 are to evaluate direction cosines at the iteration points. Whereas in options 3 and 4, the additional function calls
are to implement a full FORM analysis.
An observation of the results in Table 10 shows that options 1 and 3 were more expensive than options 2 and 4,
much more than in the first example. This can be partially attributed to the larger decision space in this example,
which requires more optimizer iterations to obtain convergence. In this case, since the limit state functions were
linear in the random variables and the random variables are normally distributed, there is no difference in the α
vectors when calculated at the means of the random variables as opposed to the MPP, and the function calls required
to obtain true FORM MPPs are not necessary to obtain correct direction cosines. As a consequence, options 2 and 4
were the best choices because they did not require limit state function evaluations (for calculating direction cosines)
at every optimization iteration. Option 2 required one more iteration than option 4 to obtain convergence, although
option 4 required more limit state function calls to perform direct FORM analysis.

V. Conclusion
A novel approach to the reliability based design optimization of series, parallel, and general systems has been
developed using the optimality conditions of the first order reliability method and a single-loop optimization
framework. The proposed approach can optimally allocate component reliability targets, but it can also incorporate
user-specified component reliability targets if required; thus it has the flexibility to accommodate both types of

14
American Institute of Aeronautics and Astronautics
problems. A single loop formulation (Eq. 11) and an equivalent formulation (Eq. 12) have been developed and
demonstrated with three example problems. The formulation in Eq. (12) is much more efficient and robust, and four
algorithms for implementing this formulation are developed. These algorithms allow for the optimal allocation of
reliability across components in a system, and allow for computationally inexpensive solutions for systems-level
RBDO.
The proposed solution algorithms remove the most probable points from the decision space, thus reducing and
even eliminating the need to compute gradients and/or Hessian of limit state function gradients with respect to the
random variables. This feature offers tremendous improvement in computational efficiency and strongly enables
systems-level RBDO. The most efficient of these algorithms has approximately the same order of computational
effort as single-loop component-level RBDO using second-moment methods, by far the least computationally
expensive RBDO strategy presently in use. Furthermore, this solution process allows the designer to take advantage
of information gained from direct reliability analysis done on a particular design. The ability to harness such
knowledge has been shown to be extremely valuable in improving the robustness and efficiency of single-loop
RBDO methods.

Acknowledgments
This study was partially supported by funds from the Dwight D. Eisenhower Fellowship in Transportation
Engineering and the National Science Foundation's IGERT graduate program in Reliability and Risk Engineering
and Management at Vanderbilt University (Grant No. 0114329). The support is gratefully acknowledged.

References
1
Torng, T., and Yang, R., 1993, “Robust Structural System Design Using a System Reliability-Based Design Optimization
Method,” in Probabilistic Structural Mechanics: Advances in structural reliability method, P. Spanos and Y. Wu, eds.Springer-
Verlag, Berlin, pp. 534–549.
2
Wu, Y., Shin, Y., Sues, R., and Cesare, M., 2001, “Safety Factor Based Approach for Probabilistic-Based Design
Optimization,” in Proceedings of 42nd AIAA Structural Dynamics and Materials Conference. AIAA-2001–1522.
3
Wu, Y., and Wang, W., 1998, “Efficient Probabilistic Design by Converting Reliability Constraints to Approximately
Equivalent Deterministic Constraints,” J. Integr. Des. Process Sci., 2, pp. 13–21.
4
Royset, J.O., Der Kiureghian, A,. and Polak, E., 2001, “Reliability-Based Optimal Structural Design by the Decoupled
Approach,” Reliability and Structural Safety, Vol. 73, pp. 213-221.
5
Du, X., and Chen, W., 2004, “Sequential Optimization and Reliability Assessment Method for Efficient Probabilistic
Design,” J. Mech. Des., 126, pp. 225–233.
6
Zou, T., and Mahadevan, S., 2006, “Versatile Formulation for Multiobjective Reliability-Based Design Optimization,”
ASME Journal of Mechanical Design, Vol. 128, pp. 1217-1226.
7
Madsen, H., and Hansen, P., 1992, “A Comparison of Some Algorithms for Reliability Based Structural Optimization and
Sensitivity Analysis,” Proceedings of the 4th IFIP WG 7.5 Conference, Munich, Germany, R. Rackwitz and P. Thoft-
Christensen, Eds., Springer-Verlag, Berlin, pp. 443–451.
8
Chen, X., Hasselman, T. K., and Neill, D. J., 1997, “Reliability Based Structural Design Optimization for Practical
Applications,” Proceedings of the 38th IAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Material Conference,
Kissimmee, Florida, AIAA-97-1403.
9
Kuschel, N., and Rackwitz, R., 1997. “Two Basic Problems in Reliability-Based Structural Optimization,” Math. Methods
Oper. Res., 46, pp. 309–333.
10
Kuschel, N., and Rackwitz, R., 2000, “Time-Variant Reliability-Based Structural Optimization Using SORM,”
Optimization, 47, pp. 349–368.
11
Agarwal, H., and Renaud, J., 2004, “A Unilevel Method for Reliability Based Design Optimization,” in Proceedings of 45th
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Material Conference, AIAA-2004–2029.
12
Liang, J., Mourelatos, Z., and Tu, J., 2004, “A Single-Loop Method for Reliability-Based Design Optimization,” in
Proceedings of the ASME Design Engineering Technical Conferences.
13
Yang, R., and Gu, L., 2004, “Experience with Approximate Reliability-Based Optimization Methods,” Struct. Multidiscip.
Optim., 26, pp. 152–159.
14
Haldar, A., Mahadevan, S., 2002,. Probability, Reliability and Statistical Methods in Engineering Design, J. Wiley & Sons,
New York, NY, 304 pp., ISBN-10: 0-471-33119-8.
15
Ambartzumian R, Der Kiureghian A, Ohanian V, Sukiasian H., 1997, “Multinormal Probability by Sequential Conditioned
Importance Sampling,” In: Advances in Safety and Reliability, Proc. ESREL ’97, 17–20 June 1997, Lisbon, pp. 1261–8 [Vol. 2].
16
Pandey MD., 1998. “An Effective Approximation to Evaluate Multinormal Integrals,” Structural Safety, Vol. 20, pp. 51–
67.

15
American Institute of Aeronautics and Astronautics
 17
Alexandrov, N. M.; Lewis, R. M.: “Algorithmic Perspectives on Problem Formulation in MDO,” 8th
AIAA/USAF/NASA/ISSMO Symposium on MA&O, Long Beach, CA, AIAA Paper 2000-4718.
18
Rackwitz, R. and B. Fiessler, 1978, “Structural Reliability under Combined Load Sequences,” Comput. Struct., Vol. 9, pp
489 - 494.
19
Rackwitz, R., 2001. “Reliability Analysis, a Review and Some Perspectives,” Structural Safety, Vol. 23, pp. 365-395.
20
Smith, N. and Mahadevan, S., 2005, “Integrating System-Level and Component-Level Designs Under Uncertainty.” Journal
of Spacecraft and Rockets, Vol. 42, No.4, pp. 752-760.
21
Collier, C., Yarrington, P., and Pickenheim, M., 1999, “The Hypersizing Method for Structures,” presented at NAFEMS
World Congress ’99, Newport, Rhode Island, April 25-28, 1999.

16
American Institute of Aeronautics and Astronautics