Computers and Structures 80 (2002) 257–269

www.elsevier.com/locate/compstruc

A comparative study on reliability-index

and target-performance-based probabilistic
structural design optimization
Jae-Ohk Lee *, Young-Soon Yang, Won-Sun Ruy
Department of Naval Architecture and Ocean Engineering, College of Engineering, Seoul National University,
Seoul 151-742, Republic of Korea
Received 22 December 2000; accepted 27 November 2001

Abstract
Probabilistic structural design optimization enables designers and engineers to quantitatively take into account the
uncertainties observed in the structural and environmental properties. In this paper, two approaches to determine the
satisfaction of probabilistic constraints are discussed. One is the conventional reliability-index-based approach and
the other is a more recently proposed target-performance-based approach. An algorithm, which detects and eliminates
the excessive zigzagging iterations during the searches for the most probable failure point and the minimum perfor-
mance target point, was incorporated.
The number of iterations required by the two approaches was investigated in three examples: a cantilever beam, a
three-bar truss and a ten-bar truss structure. Based on the results, the target-performance-based approach was found to
be superior to the reliability-index-based one in view of both computational efficiency and numerical stability. Ó 2002
Elsevier Science Ltd. All rights reserved.

Keywords: Probabilistic constraints; Reliability index; The most probable failure point; Target performance; The minimum perfor-
mance target point

1. Introduction ond-moment method (AFOSM) and the second-order

second-moment method calculate the reliability index as
Variations are observed in the mechanical and the a measure of the probabilistic structural safety [1,2].
geometric properties of structures as well as the external Recent studies on the response surface methodology
loads due to uncertainties. Therefore, for more rational techniques have enlarged the availability by enhancing
structural design these variations should be taken into the computational efficiency [3–5].
account, and structural reliability theory provides a In contrast to the deterministic optimization which is
good methodology to do this as it is based on proba- usually expressed by
bility theory. Simulation techniques such as the Monte-
Carlo simulation and importance sampling directly find d; which minimizes f ðdÞ
estimate the failure probability, and what is known as the subjected to gi ðdÞ P 0 ði ¼ 1; 2; . . . ; pÞ ð1Þ
moment methods such as the advanced first-order sec-

probabilistic structural design optimization can be

characterized by the probabilistic constraints [6]. The
*
Corresponding author. Tel.: +82-2-880-7338; fax: +82-2- random variable vector x as well as the design variable
888-9298. vector d is involved in defining the problem of proba-
E-mail address: [email protected] (J.-O. Lee). bilistic optimization:

0045-7949/02/$ - see front matter Ó 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 4 5 - 7 9 4 9 ( 0 2 ) 0 0 0 0 6 - 8
258 J.-O. Lee et al. / Computers and Structures 80 (2002) 257–269

find d; which minimizes f ðdÞ index corresponding to the allowable violation proba-
subjected to P ½gi ðd; xÞ 6 0 6 Pi ði ¼ 1; 2; . . . ; pÞ ð2Þ bility Pi .
The optimization problem shown in Eq. (3) actually
A probabilistic constraint defines the feasible region by has two optimization loops since the MPFP search to
restricting the probability that a deterministic constraint calculate the reliability index is also an optimization
gi is violated within the allowable probability of viola- procedure carried out in random variable space. This
tion Pi . Both simulation techniques and moment meth- nested optimization procedure will be discussed in the
ods are available to determine whether or not the following sections.
probabilistic constraint is satisfied. However, the simu-
lation techniques sometimes require a prohibitively large 2.1.1. Calculation of reliability index
amount of structural analyses in spite of their robust- A normal random variable X can be transformed into
ness. In addition, they do not produce any information a standard normal random variable U by
regarding the sensitivity for more efficient search of the
X l
optimum structural design. U¼ ð4Þ
For this reason, moment methods, especially the r
AFOSM, are frequently used to estimate the probabi- where l and r are the mean and standard deviation of X.
listic constraints with acceptable computations. The re- If all the random variables are transformed into
liability index is calculated by determining the most statistically independent standard normal ones and si-
probable failure point (MPFP) in random variable multaneously the limit state equation g(u) is linear, then
space. It is then compared with target reliability index. the reliability index b is the shortest distance in U space
This approach is known as the reliability-index-based from the origin to the failure surface given by gðuÞ ¼ 0.
one. However, a new method has recently been intro- The failure probability Pf is calculated by
duced, which is based on the same theoretical back-
ground as the moment methods [7]. This new method Pf ¼ UðbÞ ¼ 1  UðbÞ
Z b  
calculates the target performance by determining the 1 1
where UðbÞ ¼ pffiffiffiffiffiffi exp  u2 du ð5Þ
minimum performance target point (MPTP). The sign of 1 2p 2
the target performance indicates the satisfaction of a
probabilistic constraint. This approach can be referred Thus a failure point located closest to the origin should
to as target-performance-based one. be ascertained to calculate the reliability index. There-
In this paper, the basic concepts and formulations of fore, this problem may be described by
these two approaches are discussed in Sections 2.1 and pffiffiffiffiffiffiffiffi
find u; which minimizes b ¼ juj ¼ uT u
2.2, respectively. In Section 2.3, an algorithm to reduce
the number of iterations during the searches for the subjected to gðuÞ ¼ 0 ð6Þ
MPFP and the MPTP is presented, and its effectiveness
By applying the Lagrangian multiplier method with
is illustrated. Three simple structures including a canti-
gðuÞ ¼ 0 approximated into its first order, the following
lever beam, a three-bar truss and a ten-bar truss struc-
update formula can be obtained:
ture are optimized in Section 3 by the two approaches.
 
In addition, comparative remarks are made on the basis ðkþ1Þ GTuðkÞ uðkÞ  g uðkÞ
of the results. u ¼ GuðkÞ
GTuðkÞ GuðkÞ
 T
og og og
where GuðkÞ ¼ ; ;...;
2. Two approaches for probabilistic structural design ou1 ou2 oun
 T
optimization og og og
¼ r1 ; r2 ; . . . ; rn ð7Þ
ox1 ox2 oxn
2.1. Reliability-index-based approach
The origin in U space, which corresponds to the mean
When probabilistic constraints are estimated in terms point in X space, is usually assigned to a starting point.
of the reliability index, the probabilistic structural design The limit state equation is approximated into the first
optimization of Eq. (2) may be expressed as order and the standard deviation, which is the second
moment of a random distribution, is considered in this
find d; which minimizes f ðdÞ formulation. This is the reason why this method is called
subjected to bi ðdÞ P bi;target ði ¼ 1; 2; . . . ; pÞ ð3Þ as the first-order second-moment method. Additionally,
the standard normal probabilistic density decreases ex-
where the reliability index bi (d) can be obtained as a ponentially with the distance from the origin in U space.
result of the MPFP search for each limit state equation As the optimum point of Eq. (6), i.e. u , is closer to the
gi ðd; xÞ ¼ 0. The symbol bi;target is the target reliability origin than any other failure point, it is most probable.
 J.-O. Lee et al. / Computers and Structures 80 (2002) 257–269 259

Therefore, u is the most probable failure point (MPFP) where n is the number of random variables and /ð Þ is
and the problem given by Eq. (6) is referred to as the the standard normal probabilistic density function.
MPFP search in this paper. An additional expression for the sensitivity can be
It is worth mentioning that Eq. (6) does not distin- obtained if di ¼ zi , which does not need to calculate the
guish the possible occurrence of a negative reliability derivatives of the limit state equation. Based on the
index. Hence a negative sign should be added when the geometric definition of reliability index, the following
limit state equation is negative at the origin in U space. expression can be derived:
When a non-normal random variable is involved, the ffi 1X
ob o pffiffiffiffiffiffiffi n
ouk
transformation into a standard normal random variable ¼ uT u ¼ uk
by Eq. (4) can be replaced by the Rackwitz–Fiessler ozi ozi b k¼1 ozi
transformation [8]: 1X n
oU1 ðFXk ðxk ; zÞÞ
¼ uk
FX ðxÞ ¼ UðuÞ ð8Þ b k¼1 ozi
1X n
uk oF ðx ; zÞ
where FX ( ) is the non-normal cumulative probability ¼   Xk k ð11Þ
and Uð Þ as defined in Eq. (5) is the standard normal b k¼1 / U1 ðFXk ðxk ; zÞÞ ozi
cumulative probability.

2.1.2. Sensitivity of reliability index 2.2. Target-performance-based approach

Various techniques such as the feasible direction
method (FDM), the sequential linear programming In the reliability-index-based approach, by compar-
(SLP) and the sequential quadratic programming use the ing the reliability index with target reliability index, it
sensitivity information in order to determine more ef- can be determined whether or not a probabilistic con-
fective search direction for the optimum. In the reliability- straint is satisfied. The reliability index is obtained as a
index-based approach, the sensitivity of reliability index result of the MPFP search and the target reliability in-
with respect to design variables is provided. dex has nothing to do with the MPFP search.
The sensitivity of the reliability index is derived as In contrast, the target-performance-based approach
determines the satisfaction of a probabilistic constraint
db Gd by the sign of the target performance. The procedure of
¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð9Þ
dd GT G calculating the target performance is an inverse of that
u u
for the reliability index. In other words, concerns are
where Gd and Gu denote the derivative vector of the first focused on the points where the distances from the
limit state equation with respect to the design variables origin are equal to the target reliability index. Subse-
and that with respect to the standard normal random quently, one of them, at which the limit state equation
variables at u , respectively [9]. should be the minimum, is selected. Since a negative
A design variable can be divided into two types. One value of a limit state equation indicates that a failure has
is a characteristic value zi that is related to the random occurred, the sign of the minimum limit state equation
distribution, and the other is a deterministic parameter yi can be used as a measure to determine whether a
which is independent of the random distribution. Since probabilistic constraint is satisfied.
zi is implicitly involved in the random variable xi , a limit This is the basic concept of the target-performance-
state equation gð Þ is expressed in terms of xi and yi . based approach, where the minimum performance target
Therefore, considering the transformation of Eq. (8), the point (the MPTP for abbreviation) is searched instead of
derivatives in Eq. (9) are obtained as follows: the MPFP. The MPTP stands for the fact that it is a
target point since its distance from the origin is equal to
ogðx; yÞ the target reliability index. Simultaneously, the limit
Gdi jdi ¼yi ¼
oyi state equation, which is usually considered as perfor-
ogðx; yÞ X n
ogðx; yÞ oxk mance function, becomes the minimum at this point
Gdi jdi ¼zi ¼ ¼ rather than at any other target point. The target per-
ozi k¼1
oxk ozi
formance atarget is the value of the limit state equation
Xn
ogðx; yÞ oFX1 ðUðuk Þ; zÞ calculated at the MPTP, and a probabilistic constraint is
¼ k

k¼1
oxk ozi satisfied if the corresponding target performance is
positive.
ogðx; yÞ ogðx; yÞ oxi ogðx; yÞ oFX1 ðUðui Þ; zÞ
Gui ¼ ¼ ¼ i Thus, target-performance-based approach can be
oui oxi oui oxi oui expressed by
ogðx; yÞ oFX1 ðUðui Þ; zÞ
¼ i
/ðui Þ find d; which minimizes f ðdÞ
oxi oUðui Þ
ð12Þ
ð10Þ subjected to ai;target ðdÞ P 0 ði ¼ 1; 2; . . . ; pÞ
260 J.-O. Lee et al. / Computers and Structures 80 (2002) 257–269

2.2.1. Calculation of target performance compared to that using the MPTP search. This will be
The MPTP search to calculate the target perfor- shown later in Section 3.4.
mance may be defined by
2.2.2. Sensitivity of target performance
find u; which minimizes atarget ¼ gðuÞ As the target performance atarget is the value of the
pffiffiffiffiffiffiffiffi
subjected to juj ¼ uT u ¼ btarget ð13Þ limit state equation at the MPTP, its sensitivity can be
obtained more easily than that of the reliability index.
The sensitivity of the target performance with respect to
and as with the MPFP search this is also an optimization
the design variables is equal to the derivative of the limit
loop nested into Eq. (12).
state equation:
As the MPFP and the MPTP are the same points if
b ¼ btarget , the update formula for the MPTP search is datarget dg
¼ Gd ¼ ð15Þ
given by dd dd

GuðkÞ and the first two formulae in Eq. (10) are also available
uðkþ1Þ ¼ btarget qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð14Þ for this calculation.
GTuðkÞ GuðkÞ Concerning the sensitivities of reliability index and
target performance, it is worth noting that the sensitivity
A comparison of Eq. (13) with Eq. (6) shows that the of reliability index by Eq. (9) is always nonlinear but
objective function and the equality constraint are ex- that of target performance by Eq. (15) can be linear if
changed between each other, and for this reason it is the limit state equation is linear. In view of optimization,
called as the ‘inverse’ AFOSM. The Rackwitz–Fiessler this can be an advantage when using the target-perfor-
transformation of Eq. (8) is also available when a non- mance-based approach over the reliability-index-based
normal random variable is considered. one.
Fig. 1 shows the conceptual iteration history during
the searches for the MPFP and the MPTP. Usually the 2.3. Elimination of the zigzagging iterations
MPFP search requires several iterations for the inter-
mediate points that are updated by Eq. (7) to reach the As both the MPFP search and the MPTP search are
failure surface given by g ¼ 0, while those updated nested sub-optimizations, the iterations required by
by Eq. (14) for the MPTP search immediately lie on them occupy a considerable portion of the total com-
juj ¼ btarget . Furthermore, the position of g ¼ 0 varies putational load and cost. Thus a reduction in the
with the design point d while that of juj ¼ btarget is fixed number of iterations by the two searches is essential for
in U space during the entire optimization: the position of improving computational efficiency.
g ¼ 0 moves away from the origin as the design point d Three representative types of iteration history, which
becomes conservative. This means that iterations by the were observed during the two searches, are shown in
MPFP search should continue until the updated point Fig. 2. The iterations are convergent in (a) and (c), and
by Eq. (7) reaches the failure surface while the region are divergent in (b). Additionally the iteration histories
to be explored by the MPTP search is determined by in (a) and (b) show a typical zigzagging movement. For
the target reliability index. Therefore, during the entire the purpose of a more efficient and stable search an al-
probabilistic optimization, the number of iterations by gorithm is described in this section, which detects and
the MPFP search is expected to show a large fluctuation eliminates the excessive zigzagging iterations.

Fig. 1. Conceptual iteration history during the searches for the MPFP and the MPTP.
 J.-O. Lee et al. / Computers and Structures 80 (2002) 257–269 261

Fig. 2. Three representative types of iteration history during the searches for the MPFP and the MPTP.

Suppose uðk1Þ , uðkÞ and uðkþ1Þ are the intermediate tion as shown in Fig. 5. The inner loop is either the
points successively updated during the searches. Exces- MPFP search or the MPTP search, which estimates the
sive zigzagging iterations are made when the direction of probabilistic constraints in terms of the reliability index
the third intermediate point, i.e. uðkþ1Þ , lies closer to that or the target performance.
of uðk1Þ than to that of uðkÞ . Hence, the zigzagging iter- A cantilever beam, a three-bar truss and a ten-bar
ation can be detected by satisfying truss structure are optimized using the two approaches,
and a comparison will be made on the basis of the re-
uðk1ÞT uðkþ1Þ uðkÞT uðkþ1Þ
> ðkÞ ðkþ1Þ ð16Þ sults.
juðk1Þ jjuðkþ1Þ j ju jju j

and in such a situation uðkþ1Þ should be replaced by unew : 3.1. Cantilever beam problem

uðk1Þ þ uðkÞ A cantilever beam of breadth b and depth d as shown

unew ¼ bnew ð17Þ
juðk1Þ þ uðkÞ j in Fig. 6 [6] was optimized to minimize the weight. The
cross-section area was the objective function, and the
where bnew is the magnitude of unew and it can be de- following four limit state equations were used for
termined by the probabilistic constraints:
8    
>
>
> bðkÞ g uðk1Þ  bðk1Þ g uðkÞ 0:3Eb3 d 6Q1 L
< g1 ¼  Q2 6 0 g2 ¼ A  N0 60
bnew ¼ gðuðk1Þ Þ  gðuðkÞ Þ ð18Þ L2 bd 2
>
> for the MPFP search 4Q2 L3 6Q2 L
>
:b g3 ¼ D0  60 g4 ¼ R  60
for the MPTP search 3 ð19Þ
target Ebd bd 2
where L ¼ 30:0 ½in:; N0 ¼ 2:0 106 ½cycles;
in order to satisfy the equality constraint in Eq. (6) and
Eq. (13). D0 ¼ 0:15 ½in:
Fig. 3 shows four illustrative examples of excessive
where Q1 and Q2 are the fatigue load and design load
zigzagging iterations, where the upper two are the iter-
respectively. Each of Eq. (19) defines in turn, the failure
ation histories during the MPFP search and the lower
by buckling, fatigue, excessive deflection and stress. The
two are those during the MPTP search. The solid line is
SLP [10] with the starting point ðb; dÞ0 at ð2:0; 2:0Þ was
the failure surface on which the MPFP should be located
used as the optimization algorithm. The statistical
and the dotted circle corresponds to juj ¼ btarget where
properties of random variables are summarized in Table
the MPTP should lie. With the exception of (b), the it-
1. As for the target reliability index, a value of 3.0 was
erations are unfortunately divergent. By applying the
set for g1 and g2 , and 2.0 was set for g3 and g4 , respec-
proposed algorithm, each was improved as shown in
tively.
Fig. 4: (a)–(d) in Fig. 4 are the improved histories of (a)–
Fig. 7 shows the constraint plots in addition to the
(d) in Fig. 3, respectively. It can be seen from Fig. 4 that
optimum points in the cases of the deterministic opti-
all iterations are convergent with a much smaller num-
mization and the probabilistic optimization, respec-
ber of iterations.
tively. In the case of probabilistic one, two constraint
plots have been drawn: The solid lines were obtained by
3. Application examples assuming that all the target reliability indexes were zeros
while the dotted lines correspond to those predefined in
Probabilistic structural design optimization by the this problem. The solid lines approximately coincide
reliability-index-based approach or by the target-per- with the deterministic constraints but not exactly due to
formance-based approach has a two-looped configura- some non-symmetrically distributed random variables.
262 J.-O. Lee et al. / Computers and Structures 80 (2002) 257–269

Fig. 3. Illustrative examples of excessive zigzagging iterations.

Concerning the probabilistic optimum point Points B and D explain the opposite situation, and point
ð0:28; 3:95Þ, it can be seen from Fig. 7 that it is located O is the optimum condition. Namely, a probabilistic
away from the deterministic constraints or the proba- constraint is determined along the line g ¼ 0 by the re-
bilistic constraints drawn by the solid line: The proba- liability-index-based approach while it is done along the
bilistic optimum point is located inside the feasible line b ¼ btarget by the target-performance-based one.
region to deserve the appropriate safety margin given by The iteration history including the number of itera-
the target reliability indexes. tions by the MPFP search and the MPTP search are
Fig. 8 displays the difference between the two ap- summarized in Tables 2 and 3, respectively for the two
proaches, where the relationship between the reliability approaches. The numbers in the parentheses are those
index and the minimum value of the limit state equation when the zigzagging iterations are not eliminated.
g3 has been plotted for three design points. At the For both approaches the same optimums was ob-
starting point, i.e. ðb; dÞ ¼ ð2:0; 2:0Þ, points A and C are tained after eleven iterations by the SLP. The reduction
obtained by the MPFP search and the MPTP search, in computational cost by eliminating the zigzagging it-
respectively. Since point A stands for b < 2:0 and point erations is remarkable in the case of the reliability-index-
C represents a negative target performance, it can be based approach. Moreover, considerable differences can
said that the probabilistic constraint for the target reli- be found in the total number of iterations by the MPTP
ability index of 2.0 is not satisfied at the starting point. search compared with that by the MPFP search: a much
 J.-O. Lee et al. / Computers and Structures 80 (2002) 257–269 263

Fig. 4. Improved iteration history by eliminating excessive zigzagging iterations.

smaller number of iterations were undertaken when the g7;8 : 0:5 6 A1 6 1:2; g9;10 : 0:0 6 A2 6 4:0;
target performance was used instead of the reliability g11;12 : 0:5 6 A3 6 1:2 ð21Þ
index. This is especially so in the case of g2 , where the
ratio between the total numbers of iterations by the two
The FDM [10] with the starting point ðA1 ; A2 ; A3 Þ0 at
searches is 187:33.
ð0:85; 2:0; 0:85Þ was applied as the optimization algo-
rithm. Six random variables are taken into account and
3.2. Three-bar truss problem the statistical properties are summarized in Table 4.
Since the problem is strictly symmetric, A1 and A3 were
The section areas of the three-bar truss structure in treated as one design variable. In addition, based on
Fig. 9 [6] were optimized to minimize the weight. The engineering judgment, only g2 and g4 were considered as
limit state equations define the failures by excessive the limit state equations.
stress in each element: The target reliability index was assumed to be 2.0 and
3.0 in turn, and the corresponding results are summa-
g1;2 : r þ
a 6 r1 6 ra ; g3;4 : r þ
a 6 r2 6 ra ; rized in Tables 5 and 6. The numbers in parentheses are
g5;6 : r þ
a 6 r3 6 ra ð20Þ the total number of iterations during either the MPFP
search or the MPTP search when the zigzagging itera-
and the following side constraints were considered: tions were not eliminated.
264 J.-O. Lee et al. / Computers and Structures 80 (2002) 257–269

Fig. 5. Two-looped configuration of probabilistic structural design optimization.

Fig. 6. The cantilever beam problem [6].

Table 1
Statistical properties for the cantilever beam problem (* for median)
Description Distribution Mean CoV
Q1 Fatigue load [klb] Log-normal 0.4 0.15
Q2 Design load [klb] Log-normal 0.5 0.15
E Young’s modulus [ksi] Normal 30 000.0 0.10
A Fatigue strength coefficient [ksi] Log-normal 1:46E þ 10 0.50
R Yield strength [ksi] Weibull 50.0 0.12

The optimum points obtained by the two approaches As for the number of iterations by the MPFP search
were almost the same, and as expected, the optimum and the MPTP search, more were undertaken by the
points for btarget ¼ 3:0 are more conservative than those atarget -based approach than by the b-based one when
for btarget ¼ 2:0. The larger number of iterations by the btarget ¼ 2:0. However, considering the number of itera-
FDM using the target performance is required than tions by the FDM, the average number of iterations by
using the reliability index when btarget ¼ 2:0, but they are the MPTP search is even smaller than that by the MPFP
similar when btarget ¼ 3:0. search. This is apparently true when btarget ¼ 3:0. Thus,
 J.-O. Lee et al. / Computers and Structures 80 (2002) 257–269 265

Fig. 7. Deterministic and probabilistic constraint plots for the cantilever beam problem.

Ai of each element was treated as a normal random

variable and its mean value was the design variable.
Young’s modulus E and the external force P were also
considered as random variables. The statistical proper-
ties are summarized in Table 7.
A value of 0:1 ½lb=in:3  was assumed for the material
density. The mean of each section area, lAi , should not
be smaller than 0:10 ½in:2 , and as a probabilistic con-
straint the probability that the deflection at node 2 in
Fig. 10 exceeds 2.0 [in.] was restricted by a target reli-
ability index of 3.0. The SLP with the starting point at
lAi ¼ 10:0 was used as the optimization algorithm.
In Table 8, the results by the two approaches as well
as those from [11,12] are summarized. The optimum
points obtained by the two approaches are almost the
same and the target-performance-based approach shows
the better computational efficiency compared with the
reliability-index-based approach, when considering the
number of iterations by both the SLP and the MPFP/
Fig. 8. Difference between reliability-index and target-perfor- MPTP search. In particular, the reliability-index-based
mance-based approaches. approach failed without eliminating the zigzagging it-
erations.

it can be said that a smaller number of iterations are

3.4. Comparative remarks on the two approaches
usually required by the MPTP search than the MPFP
search irrespectively of the target reliability index.
In structural optimization most of the computational
The same optimum was obtained when A1 and A3
cost is usually ascribed to structural analyses to estimate
were treated as different design variables and when g6 in
the constraints. In the case of probabilistic optimization,
addition to g2 and g4 was also involved as a limit state
the computational cost depends on the number of iter-
equation on purpose. However, it should be noted that
ations by the MPFP search and the MPTP search since
optimization by the reliability-index-based approach
they use the structural analysis directly to estimate limit
failed when either of g1 , g3 and g5 was involved, but it
state equations and their derivatives. This is the reason
did not matter with the target-performance-based ap-
why more attention is paid to the situation where a
proach. This will be discussed again in Section 3.4.
significant difference is observed between the numbers of
iterations by the two searches.
3.3. Ten-bar truss problem Such a situation is found in the cantilever beam
problem, especially in the number of iterations for g2 .
A ten-bar truss structure, shown in Fig. 10 [11,12], From Tables 2 and 3, it can be seen that the total
was optimized to minimize the weight. The section area number of iterations by the MPFP search is 187 while
266 J.-O. Lee et al. / Computers and Structures 80 (2002) 257–269

Table 2
Iteration history during reliability-index-based approach for the cantilever beam problem
k ðb; dÞk fk Number of iterations by the MPFP search
g1 g2 g3 g4
0 (2.00,2.00) 4.0000 3 (3) 20 (21) 7 (15) 9 (8)
1 (1.50,2.22) 3.3284 4 (4) 17 (23) 7 (8) 9 (12)
2 (1.18,2.42) 2.8522 4 (4) 22 (45) 6 (6) 9 (14)
3 (0.92,2.62) 2.4233 5 (5) 20 (25) 8 (8) 9 (14)
4 (0.74,2.83) 2.0994 6 (6) 20 (21) 8 (12) 9 (12)
5 (0.60,3.03) 1.8320 11 (0) 13 (13) 8 (8) 9 (8)
6 (0.50,3.24) 1.6130 13 (20) 16 (15) 8 (9) 10 (7)
7 (0.42,3.44) 1.4307 16 (36) 12 (12) 8 (8) 10 (14)
8 (0.35,3.65) 1.2774 14 (18) 13 (35) 8 (10) 9 (8)
9 (0.30,3.85) 1.1474 10 (15) 20 (23) 8 (10) 12 (8)
10 (0.28,3.95) 1.1054 8 (8) 14 (23) 9 (9) 9 (10)
11 (0.28,3.95) 1.1087
Total number of iterations by the MPFP search 94 (129) 187 (256) 85 (103) 104 (115)

ðb; dÞ ¼ ð0:2806; 3:9516Þ, f ¼ 1:1087

Table 3
Iteration history during target-performance-based approach for the cantilever beam problem
k ðb; dÞk fk Number of iterations by the MPTP search
g1 g2 g3 g4
0 (2.00,2.00) 4.0000 4 (4) 3 (3) 7 (8) 6 (6)
1 (1.50,2.22) 3.3258 4 (4) 3 (3) 7 (8) 6 (6)
2 (1.18,2.42) 2.8508 4 (4) 3 (6) 7 (8) 6 (6)
3 (0.92,2.62) 2.4227 4 (4) 3 (3) 7 (8) 6 (6)
4 (0.74,2.83) 2.0990 5 (5) 3 (3) 7 (8) 6 (6)
5 (0.60,3.03) 1.8320 6 (6) 3 (3) 7 (8) 7 (7)
6 (0.50,3.24) 1.6131 6 (6) 3 (3) 7 (8) 7 (7)
7 (0.42,3.44) 1.4309 7 (7) 3 (3) 7 (8) 8 (8)
8 (0.35,3.65) 1.2776 8 (8) 3 (3) 7 (8) 8 (8)
9 (0.30,3.85) 1.1476 4 (4) 3 (3) 7 (8) 7 (7)
10 (0.28,3.95) 1.1092 9 (12) 3 (3) 7 (8) 7 (7)
11 (0.28,3.95) 1.1087
Total number of iterations by the MPTP search 61 (64) 33 (36) 77 (88) 74 (74)
ðb; dÞ ¼ ð0:2806; 3:9516Þ, f ¼ 1:1087

that by the MPTP is only 33. As can be seen in Fig. 7, the MPFP and the MPTP search for several values of d
the infeasible region with respect to g2 is very narrow with b fixed at 2.0. With increasing of d, both the reli-
and is restricted around the origin of the design space. ability index and the target performance are expected to
For this reason, the reliability index for g2 may be rel- increase. What is important is the number of iterations
atively large compared with that for the other limit state needed for convergence: the number of iterations by the
equations throughout the optimization. Considering the MPFP search increases with reliability index. However,
conceptual difference between the two searches as was the number of iterations by the MPTP search is inde-
illustrated in Fig. 1, it is necessary to survey the rela- pendent of the target performance. These results agree
tionship between the number of iterations by the MPFP fairly well with the expectation from Fig. 1 and the
search and the magnitude of the reliability index. discussion in Section 2.2.1.
Fig. 11 shows the change of the reliability index and Therefore, it can be concluded that the reliability-
the target performance as the iteration proceeds during index-based approach may be inferior to the target-
 J.-O. Lee et al. / Computers and Structures 80 (2002) 257–269 267

normal tail transformation by Eq. (8). However, it

cannot be explicitly expressed in an exact closed form.
Although several approximated expressions are avail-
able, their accuracy in the range of a relatively large u
has not been proven since the area at each tail of the
standard normal probabilistic density function is almost
negligible.
Though the value of UðuÞ itself may be meaningless
at an extremely large u, it can cause problems when
combined with the significant number of digits specific
to the computer programs. This is particularly so in the
normal tail transformation or in calculating the sensi-
tivity of the reliability index by Eqs. (9) and (10), con-
sequently providing for an unexpected abnormal
termination of the optimization procedure. Since the
Fig. 9. The three-bar truss structure [6]. target-performance-based approach does not calculate
the reliability index, this problem matters only with the
reliability-index-based one. The failure of the reliability-
performance-based one when many constraints remain index-based approach for the three-bar truss problem
inactive and thus they have a large reliability index when either of g1 , g3 and g5 was involved is evidence for
throughout the optimization. this.
A large reliability index can cause another problem This kind of numerical trouble caused by a large re-
combined with the standard normal cumulative proba- liability index restricts the available optimization algo-
bility function. The standard normal cumulative prob- rithms and the starting points. For example, both the
ability function UðuÞ plays an important role in MPFP search and the MPTP search were performed for
calculating the failure probability by Eq. (5) and in the g3 in the cantilever beam problem at the twenty-four

Table 4
Statistical properties for the three-bar truss problem (* for median)
Description Distribution Mean CoV
P1 External load [lbs] Gumbel 20000.0 0.20
P2 External load [lbs] Gumbel 20000.0 0.20
h1 Loading angle Normal p=4 0.10
h2 Loading angle Normal p=4 0.10
rþ
a Tensile allowable stress [lbs] Log-normal 20000.0 0.05
r
a Compressible allowable stress [lbs] Log-normal 15000.0 0.05

Table 5
Summary of results when btarget ¼ 2:0 for the three-bar truss problem
Iteration by FDM ðA1 ¼ A3 ; A2 Þ f Total number of iterations by the
MPFP/MPTP search
g2 g4
b-based 6 (0.5613,1.5005) 3.0882 49 (49) 42 (44)
atarget -based 10 (0.5618,1.5002) 3.0892 65 (65) 50 (50)

Table 6
Summary of results when btarget ¼ 3:0 for the three-bar truss problem
Iteration by FDM ðA1 ¼ A3 ; A2 Þ f Total number of iterations by the
MPFP/MPTP search
g2 g4
b-based 7 (0.9694,1.5952) 4.3371 63 (63) 47 (49)
atarget -based 8 (0.9706,1.5942) 4.3403 48 (48) 40 (40)
268 J.-O. Lee et al. / Computers and Structures 80 (2002) 257–269

at this stage. When a gradient-based optimization tech-

nique is adopted, one should be careful in selecting the
starting point and also in screening the inactive con-
straints so as to prevent a large reliability index.

4. Conclusions

Both the MPFP and MPTP searches, which calculate

the reliability index and the target performance respec-
tively to estimate the probabilistic constraints, were
discussed in this paper. A cantilever beam, a three-bar
truss and a ten-bar truss structure were optimized using
Fig. 10. The ten-bar truss structure [11,12]. the reliability-index-based and the target-performance-
based approach. A comparison based on the results of
the two approaches was performed.
Though the same optimum results were obtained by
points shown in Fig. 12. In that figure the number of both approaches, there has been observed a considerable
iterations by the two searches are written on the selected difference in the number of iterations between them. The
points and the indicates the occurrence of the abnor- number of iterations by the MPFP search increases with
mal termination of the search. In the case of the MPFP reliability index while that by the MPTP search is in-
search, the number of iterations shows a large fluctua- dependent of the target performance. In addition, the
tion ranged from 1 to 28. Furthermore, abnormal ter- abnormal termination of the MPFP search is also at-
minations were reported in the region where a large tributed to the numerical instability caused by a large
reliability index was expected. reliability index.
Therefore, more care should be taken in selecting the Therefore, the conventional reliability-index-based
optimization algorithm and in determining the starting approach is not computationally attractive, compared
point. For instance, the genetic algorithm generates a with the recently introduced target-performance-based
large population of design points over the entire design approach, when many probabilistic constraints remain
space and all the constraints are estimated from these inactive throughout the optimization and thus a large
points. The numerical instability can possibly be induced reliability index is expected.

Table 7
Statistical properties for the ten-bar truss problem
Description Distribution Mean CoV
7
E Young’s modulus [psi] Normal 10 0.05
P External load [lb] Normal 105 0.05
Ai Section area ði ¼ 1; 2; . . . ; 10Þ [in:2 ] Normal lAi (design var.) 0.05

Table 8
Summary of results for the ten-bar truss problem
Number of Total number of iterations f A1 A2 A3 A4 A5
iterations for by the MPFP/MPTP A6 A7 A8 A9 A10
the optimum search
b-based 25 244 (failed) 6144.6 38.51 0.10 26.88 19.07 0.10
0.10 6.18 27.35 27.13 0.10
atarget -based 13 93 (198) 6072.1 39.16 0.10 27.67 19.16 0.10
0.10 2.64 28.03 27.49 0.10
[11] 10 Not given 6242.7 Not given
[12] 19 Not given 6303.5 39.70 0.55 27.49 19.51 0.10
0.66 4.78 28.35 27.86 0.60
 J.-O. Lee et al. / Computers and Structures 80 (2002) 257–269 269

Fig. 11. Change of reliability index and target performance during the MPFP search and the MPTP search.

Fig. 12. Number of iterations for g3 at various design points of the cantilever beam problem.

References [7] Tu J, Choi KK, Park YH. A new study on reliability-based

design optimization. J Mech Des 1999;121:557–64.
[1] Ditlevsen O, Madsen HO. Structural reliability method. [8] Rackwitz R, Fiessler B. Structural reliability under combined
New York: John Wiley & Sons; 1996. random load sequences. Comput Struct 1978;9:489–94.
[2] Yang, Young-Soon, Suh Yong-Suck, Lee Jae-Ohk. Struc- [9] Bjerager P, Krenk S. Sensitivity measures in structural
tural reliability engineering. Seoul: Seoul National Univer- reliability analysis. In: Thoft-Christensen P, editor. Pro-
sity Press; 1999 [in Korean]. ceedings of the First IFIP Working Conference on Reli-
[3] Rajashekhar MR, Ellingwood BR. A new look at the ability and Optimization of Structural Systems. Berlin:
response surface approach for reliability analysis. Struct Springer-Verlag; 1987. p. 459–70.
Safety 1993;12:205–20. [10] Vanderplaats GN. Numerical optimization techniques for
[4] Yang, Young-Soon, Lee Jae-Ohk, Kim Bong-Jae. Struc- engineering design with applications. New York: McGraw-
tural reliability analysis using commercial FEM package. Hill, Inc; 1984.
Proceedings of the 6th International Offshore & Polar [11] Luo X, Grandhi RV. ASTROS for reliability-based
Engineering Conference, vol. 4. 1996. p. 387–94. multidisciplinary structural analysis and optimization.
[5] Zheng Y, Das PK. Improved response surface method and Comput Struct 1997;62:717–45.
its application to stiffened plate reliability analysis. Eng [12] Pettit CL, Grandhi RV. Multidisciplinary optimization of
Struct 2000;22:544–51. aerospace structures with high reliability. 8th ASCE
[6] Thanedar PB, Kodiyalam S. Structural optimization using Specialty Conference on Probabilistic Mechanics and
probabilistic constraints. Struct Optimiz 1992;4:236–40. Structural Reliability, PMC 2000-132.