Struct Multidisc Optim

DOI 10.1007/s00158-016-1478-2

RESEARCH PAPER

An approximate sequential optimization and reliability

assessment method for reliability-based design optimization
Ping Yi 1 & Zuo Zhu 1 & Jinxin Gong 1

Received: 15 December 2015 / Revised: 17 March 2016 / Accepted: 3 May 2016

# Springer-Verlag Berlin Heidelberg 2016

Abstract Sequential optimization and reliability assessment Keywords Reliability-based design optimization . Sequential
(SORA) is one of the most popular decoupled approaches to optimization and reliability assessment . Minimum
solve reliability-based design optimization (RBDO) problem performance target point . Efficiency
because of its efficiency and robustness. In SORA, the double
loop structure is decoupled through a serial of cycles of deter-
ministic optimization and reliability assessment. In each cycle,
the deterministic optimization and reliability assessment are 1 Introduction
performed sequentially and the boundaries of violated con-
straints are shifted to the feasible direction according to the Compared to traditional deterministic optimization model,
reliability information obtained in the previous cycle. In this reliability-based design optimization (RBDO) is regarded as
paper, based on the concept of SORA, approximate most a more reasonable model design methodology because of the
probable target point (MPTP) and approximate probabilistic rational consideration of uncertainties in material property,
performance measure (PPM) are adopted in reliability assess- manufacturing precision, geometry, and external loads during
ment. In each cycle, the approximate MPTP needs to be re- structures’ lifetime. The RBDO model is generally expressed
served, which will be used to obtain new approximate MPTP as the minimization of objective function under probabilistic
in the next cycle. There is no need to evaluate the performance constraints restricting the failure probability (Frangopol and
function in the deterministic optimization since the approxi- Corotis 1996).
mate PPM and its sensitivity are used to formulate the linear The probabilistic constraints can be evaluated in two ways:
Taylor expansion of the constraint function. One example is reliability index approach (RIA; Nikolaidis and Burdisso
used to illustrate that the approximate MPTP will approach the 1988) and performance measure approach (PMA; Tu et al.
accurate MPTP with the iteration. The design variables and 1999). In RIA, reliability index is employed to replace failure
the approximate MPTP converge simultaneously. Numerical probability in the sub-problem of constraint evaluation.
results of several examples indicate the proposed method is Whereas, in PMA the probabilistic constraints are transformed
robust and more efficient than SORA and other common to check if the probabilistic performance measure (PPM) is
RBDO methods. greater than zero. Youn and Choi (2004) and Yi and Cheng
(2008) found that PMA is more efficient, stable and less de-
pendent on probabilistic distribution types than RIA.
However, because the calculation of either reliability index
in RIA or PPM in PMA is in essence an optimization proce-
dure, the direct approach to RBDO is to form a double loop
* Zuo Zhu optimization by simply connecting algorithms of probabilistic
[email protected] calculation and design optimization. The computational cost
of the double loop structure is prohibitive especially when the
1
Faculty of Infrastructure Engineering, Dalian University of RBDO model contains a large number of probabilistic con-
Technology, Dalian 116024, China straints or the performance functions are highly nonlinear. As
 P. Yi et al.

a result, some approximate methods such as single loop and optimization and reliability assessment. In each cycle, the de-
decoupled approaches have been developed. terministic optimization and reliability assessment are per-
Single loop approaches employ the Karush-Kuhn-Tucker formed sequentially and the boundaries of violated constraints
(KKT) optimality conditions to replace the sub-optimization need to be shifted to the feasible direction according to the
loop, and so the double loop structure is transformed into reliability information obtained in the previous cycle. In this
single loop. Chen et al. (1997) proposed the single loop single paper, based on the concept of SORA, approximate reliability
vector (SLSV) approach, which found the approximate most assessment is adopted to reduce the number of performance
probable target point (MPTP) by using the information of function evaluations. In each cycle, the approximate MPTP
target reliability index and limit state derivatives. Based on needs to be reserved, which will be used to obtain new ap-
the concept of SLSV, Liang et al. (2004) proposed the single proximate MPTP in the next cycle. There is no need to eval-
loop approach (SLA), which was further developed to consid- uate the performance functions in the deterministic optimiza-
er the system reliability requirement (Liang et al. 2007). There tion since the approximate PPM and its sensitivity are used to
are also other single loop approaches proposed in recent years formulate the linear Taylor expansion of the constraint func-
(see e.g. Nguyen et al. 2010; Silva et al. 2010; He et al. 2011; tions. The error analysis shows that in the ε-vicinity of optimal
Zhang et al. 2011). And these methods are well suitable for point, the difference between real constrains and approximate
linear and weakly nonlinear RBDO problems (Chen et al. constraints is of second order of ε in the process of determin-
2013a). istic optimization. One example illustrates that although ap-
Decoupled approaches separate reliability analysis from proximate MPTP are used in this method, as the design vari-
the optimization process and the double structure is ables converge to the optimum design, the approximate MPTP
transformed into serial loops. Wu and Wang (1998) proposed can converge to the accurate MPTP. Numerical results of sev-
the safety factor approach (SFA) based on the concept of ap- eral examples indicate the proposed method is more efficient
proximately equivalent deterministic constraint, where the than SORA and other common RBDO methods.
safety factors are obtained by the reliability analysis during Organization of this paper is as follows: in Section 2, the
the optimization process. Du and Chen(2004) proposed se- basic formulation of RBDO and common RBDO methods
quential optimization and reliability assessment (SORA), such as RIA, PMA, SLSV, SORA are briefly reviewed. The
where the shifting vector is used to decouple the double loop proposed method for solving RBDO problem and error anal-
structure into a serial of loops of deterministic optimization ysis are detailedly presented in Section 3. Illustrative exam-
and reliability assessment. It is noted that the SFA and SORA ples are given in Section 4. And finally, the conclusions are
are identical in concept in spite of different implementation given in Section 5.
(Yang and Gu 2004). Cheng et al. (2006) proposed the con-
cept of sequential approximate programming (SAP) strategy
and Yi and Cheng (2008) further developed this strategy for
PMA-based RBDO problem. In SAP approach, the RBDO
2 Reliability-based design optimization
problem is decomposed into a series of sub-optimization
A typical deterministic optimization can be expressed as follows:
problems and approximate PPM at the current design point
and its sensitivity are adopted, which are obtained by min f ðdÞ

approximate optimal conditions in the vicinity of the MPTP. s:t: h j ðdÞ≤ 0 j ¼ 1; …; N p ð1Þ
In a recent work, Chen et al. (2013b) proposed the adaptive dL ≤ d ≤ dU
decoupling approach (ADA), which employs a new update
angle strategy and a novel feasibility-checking method to im- where d = [d1, d2,…, dn]T is the design variable vector, f(d) is the
prove the efficiency of solving RBDO problem. Among these objective function, hj(d) is the j-th constraint function, and dL
approximate methods, SORA is regarded as one of the most and dU are the lower limit and upper limit of design vector,
popular approaches to solve RBDO problem because of its respectively.
efficiency and robustness (Aoues and Chateauneuf 2010; However, because of the uncertainty in material and geom-
Chen et al. 2013a, b). etry properties as well as external loads during the design and
It should be pointed out that other advanced ap- service period, it is more reasonable to adopt reliability-based
proaches for solving RBDO, such as advanced sampling design optimization (RBDO). RBDO can be generally formu-
methods (Zhuang and Pan 2012) as well as the surro- lated as follows (Tu et al. 1999; Lee et al. 2002; Youn et al.
gate models (Ju and Lee 2008) for RBDO, have been 2003; Yi and Cheng 2008):
reported and showed improved accuracy and efficien-
min f ðdÞ
cies. However, these are not discussed here.
s:t: PðGðd; XÞ≤ 0Þ ≤ Pt ð2Þ
As previously mentioned, in SORA, the double loop struc-
ture is decoupled through a serial of cycles of deterministic dL
≤ d ≤ d U
An approximate sequential optimization and reliability assessment

where X = [X1, X2,…, Xn] T is the random variable vector, and reliability index can be obtained through the following opti-
the design variable vector, d = [d1, d2,…, dn]T repre- mization problem in the U-space:
sents the mean values of the random variables. For clar-
ity, the deterministic constraints are omitted, and only f or any given d find u* ; such that
one probability constraint and bounds on design vari- min β ¼ kUk ð6Þ
ables are considered. However, the problem with more s:t: gðd; UÞ≤ 0
than one probability constraints can be dealt with in a The Hasofer and Lind-Rackwitz and Fiessler (HL-RF)
similar way. G(d, X) is the limit state function or per- method (Rackwitz and Fiessler 1978) is generally employed
formance function and G(d, X) ≤ 0 denotes the failure to calculate the above optimization model.
domain. Pt is the prescribed acceptable failure probabil-
ity. The failure probability can be obtained by
2.2 Performance measure approach(PMA)
P f ðdÞ ¼ PðGðd; XÞ ≤ 0Þ
Z The probabilistic constraint in RBDO model can be also
expressed as follows (Tu et al. 1999):
¼ f ðXÞdx ¼ F G ð0; dÞ ð3Þ
Gðd;XÞ ≤ 0
Gp ðdÞ ¼ F −1
G ðΦð−β t Þ; dÞ≥ 0 ð7Þ
where f(X) is the joint probability density function of X where G p (d) is PPM of the performance function.
and F G (*,d) is the cumulative distribution function Consequently, the RBDO based on PMA can be formulated
(CDF) of G (d,X). as:
The direct formulation of RBDO is a double loop structure
in essence: the outer loop aims at the optimization of design min f ðdÞ
variables while the inner loop handles the reliability assess- s:t: Gp ðdÞ≥ 0 ð8Þ
ment iteratively. RIA and PMA are generally used to perform dL ≤ d≤ dU
reliability assessment under the double loop structure.
PPM can also be obtained through an optimization problem
However, because of the poor efficiency of double loop ap-
in U-space as follows:
proaches, many approximate methods such as SLSV and
SORA have been proposed, which transform the RBDO mod- f or any given d find u* ; such that
el into single loop or serial loop. In this section, these methods min g ðd; UÞ ð9Þ
are briefly reviewed. s:t: kUk ¼ β t

where u* is the MPTP on the sphere surface with the target

2.1 Reliability index approach(RIA)
reliability indexβt and PPM is calculated by,



For RBDO model, the probabilistic constraints can be trans- Gp ðdÞ ¼ g d; u* ¼ G d; x* ð10Þ
formed into reliability index constraints as follows:
The advanced mean-value (AMV) (Wu et al. 1990)

 method is widely used to locate MPTP due to its effi-
βðdÞ ¼ − Φ−1 ð F G ð0; dÞÞ ≥β t ð4Þ
ciency and simplicity. The iterative formula of AMV is
expressed as (11)
whereβ(d) and βt are the reliability index and target reliability


index of the limit state function, respectively, andΦ(·) is the ∇U g u k
standard normal cumulative distribution function. u ¼ −βt
kþ1
ð11Þ
k∇ U g ðu k Þk
If (4) is applied to replace the probabilistic constraint of
RBDO model, we get the RBDO based on RIA: where uk =(u1k, u2k,..., unk) is the k-th iteration point on the target

  ∂g ∂g ∂g

reliability sphere surface. ∇U g uk ¼ ∂u k ; ∂uk ; ⋯; ∂uk is
min f ðdÞ 1 2 n

s:t: βðdÞ ≥ β t ð5Þ the gradient vector of performance function at point uk.
dL ≤ d≤ dU
2.3 Single loop single vector(SLSV)
To calculate reliability indexβ(d), one first needs to trans-
form the original random variable vector X into standard SLSV (Chen et al. 1997) employed approximate MPTP,
Gaussian vector U (zero means, unit variance and independent which is obtained by using the information of target reliability
components), expressed as U = T(X) or X = T−1(U). The index and limit state derivatives, to replace the probabilistic
 P. Yi et al.

constraints instead of calculating directly the double-loop op-

timization problem. Wang and Kodiyalam (2002) extended
this approach for non-normal distributions. In SLSVapproach,
KKT optimality conditions is used to replace the sub-
optimization loop so it is well suitable for linear and weakly
nonlinear RBDO problems. SLSV is formulated as:
min f ðd
Þ 
s : t : G dk ; Xk ≤ 0
dL ≤ d≤ dU
 ð12Þ
∇x g uk−1
where Xi ¼ μi −β t α
k k k−1
σ α k−1
¼
k∇x gðuk−1 Þk

2.4 Sequential optimization and reliability

assessment(SORA)

The SORA method (Du and Chen 2004) transforms the double-
loop structure into serial loop. In SORA, reliability assessment
and optimization have been decoupled from each other, which
means that reliability assessment is not required during the de-
terministic optimization process and it is only conducted after the
optimization. The flowchart is shown in Fig. 1.

Fig. 1 Flowchart of the SORA method

3 Approximate sequential optimization

and reliability assessment(ASORA) Consequently, The linear Taylor expansion of the
constraint function G(d − sk) can be further expressed
3.1 Basic formulation as:

As previously mentioned, the SORA method employs a serial




of cycles of reliability assessment and deterministic optimiza- G d−sk d≈dk ¼ G xkMPTP þ ð∇d GÞT d−dk
tion to decouple the double loop structure. The deterministic  2 
optimization model is expressed as (13) þ O d−dk  ð16Þ

min
f ðdÞ ð13Þ

as for xkMPTP, it can be obtained through AMV method:
s:t: G d−sk ≤ 0


where sk is the shifting vector and sk = dk − xkMPTP. dk is the k-th ∇U g u*
u ¼ −β t
*

iteration design variable vector, and xkMPTP is the MPTP in the k ∇ U g ð u* Þ k

ð17Þ
k-th iteration. xMPTP ¼
T −1 ðu* Þ
k

The linear Taylor expansion of the constraint functions Gp ¼ G xkMPTP

G(d − sk) with respect to the design variables at the current
design point dk is adopted, that is, In this paper, approximate MPTP is employed to replace



 real MPTP during the reliability assessment process in order
G d−sk d≈dk ¼ G dk −sk þ ð∇d GÞT d−dk to reduce the number of performance function evaluations.
 2  Approximate MPTP is expressed as:
þ O d−dk  ð14Þ


where ∇dG is the gradient vector of performance function to ∇U g uk−1
u ¼ −β t
k
design variables d at point dk . The value of performance k∇U
g ðuk−1 Þk
function in (14) at the point d = dk can be expressed as: ⌢
xMPTP ¼ T −1 uk
k ð18Þ




 ⌢
k
G dk −sk ¼ G dk − dk −xkMPTP ¼ G xkMPTP ð15Þ Gp ¼ G ⌢ x Þ
MPTP
An approximate sequential optimization and reliability assessment

⌢
where Gp expresses the approximate estimation of PPM.If the
approximate PPM is adopted in (16), then approximate con-
⌢

straint G d−sk can be formulated as:

⌢

k 
G d−s d≈dk

  2 
¼G ⌢ k
xMPTP Þ þ ð∇d GÞT d−dk þ O d−dk  ð19Þ

Because design variables are just the mean values of the

random variables, the sensitivity ∇dG in (19) can be obtained
by the following expression:


 T
∇d G ¼ ∇u G uk ð∇d uÞ ð20Þ

where ∇du can be derived easily according to the formulation

U = T(X), because u are functions of mean values and no
performance function evaluation is needed. The k-th approx-
imate MPTP uk and the performance function’s sensitivity for
the design point dk need to be reserved and then substituted
into (18) during the next cycle to obtain an updated estimation
uk+1, which will be used in the deterministic optimization.
⌢

Because the approximate constraint G d−sk in (19) can be
determined by the approximate PPM and its sensitivity, no
extra performance function evaluation is needed in the deter-
ministic optimization, which can significantly reduce the
number of performance function calls. This proposed method
Fig. 2 Flowchart of the ASORA method
can be named as Approximate Sequential Optimization and
Reliability Assessment (ASORA).
ASORA consists of the following steps. method is suitable for multiple probabilistic constraints. In
RBDO with multiple probabilistic constraints, approximate
Step 1: set k = 0. Implement a deterministic optimization of MPTP corresponding to each probabilistic constraint is ob-
the design problem and the optimal design is chosen tained through (18) and used in the deterministic optimization
as the initial design point, d0 and set ⌢ x
0MPTP ¼μ0x . step.
⌢
Step 2: Transform x MPTP into u . Calculate G ⌢
k k
x kMPTP and
the gradient vector of performance function ∇ug(uk)
at point uk according to finite difference method. 3.2 Error analysis
Then calculate ∇dG from (20).
Step 3: Then sk ¼ dk −⌢ x kMPTP . Constitute and implement the In the above ASORA approach, the approximate PPM is used
following deterministic optimization and obtain dk + 1. to replace the real PPM and the difference between real con-
strains (16) and approximate constrains (19) is studied in the
min f ðdÞ  following.
⌢
 k
T

s:t: G d−sk ≈G ⌢xMPTP þ ð∇d GÞ d−dk ≤ 0 At an arbitrary design point dk − 1, the real PPM Gp(dk − 1)
dL ≤ d≤ dU
and the approximate one Ĝp(dk − 1) are respectively




 ∇u g u*
Step 4: If ‖f(dk + 1) − f(dk)‖/‖(dk) ≤ 10− 3, stop the iteration Gp dk−1 ¼ g u* where u* ¼ −βt ð21Þ
and dk + 1 is the optimal design. Otherwise, set k∇u gðu* Þk


k = k + 1, calculate ⌢ x kMPTP from (18) and return to ⌢
k−1 
k ∇u g uk−1
step 2. Gp d ¼ g u where uk
¼ −β t ð22Þ
k∇u gðuk−1 Þk

The flowchart of ASORA is shown in Fig. 2. For simplicity, in the following Gp(dk − 1) is substituted by
For simplicity, only one probabilistic constraint is involved Gp, Ĝp(dk − 1) by Ĝp . ∇g(u∗) represents ∇ug(u∗), Δg(u∗) for
when deducing the formulation of ASORA. However, this Δuug(u∗).
 P. Yi et al.

In the ε-vicinity of the MPTP u∗ For further derivation, the following equations are given

k
* 

* T
k *  
 2  
k−1  


 
2 
⌢  ¼  
Gp ¼ g u ¼ g u þ ∇g u u −u þ O  uk −u*  ∇g u ∇g u* þ Δg u* uk−1 −u* þ O  uk−1 −u*  


T
k *  
2  rhffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 iffi
¼ Gp þ ∇g u* u −u þ O  uk −u*  ¼ ∇T g ðu* Þ∇g ðu* Þ þ 2∇T gðu* ÞΔg ðu* Þðuk−1 −u* Þ þ O kðuk−1 −u* Þk2





  ∇T g u* Δg u* uk−1 −u* 
2 
ð23Þ ¼ ∇g u*  1 þ þ O  uk−1 −u* 
2
k∇g ðu Þk *

In order to study (uk − u∗), (22) is applied ð26Þ


∇g uk−1 1
u −u ¼ −βt
k *
−u* ð24Þ k∇g ðuk−1 Þk
k∇g ðuk−1 Þk
1  ∇T g
u* Δg
u* 
uk−1 −u* 
In the ε-vicinity of the MPTP u∗, the first order Taylor ¼ 1−
expansion of ∇g(uk − 1) can be expressed as k∇g ðu* Þk k∇g ðu* Þk
2




 
2 
∇g uk−1 ¼ ∇g u* þ Δg u* uk−1 −u* þ O  uk−1 −u*  ð27Þ

2 
þ O  uk−1 −u*  ð25Þ By Substituting (25) and (27) into (24),


∇g uk−1
u − u ¼ − βt
k *
− u*
h
 k∇gðu
Þ
k
k−1
 
2 i
∇g u* þ Δg u* uk−1 −u* þ O  uk−1 −u*   ∇T g
u* Δg
u* 
uk−1 −u* 
¼ −βt 1−
k∇gðu* Þk k∇g ðu* Þk
2
ð28Þ

  
þ O  uk−1 −u* 
2
− u*

* 
k−1 * 



 

Δg u u −u ∇g u* ∇T g u* Δg u* uk−1 −u*  
¼ − βt þ β þ O  uk−1 −u* 2
k∇gðu* Þk t 2
k∇gðu* Þk k∇g ðu* Þk

By applying (28) in (23), we have,

* 
k−1 * 
*  T
* 
* 
k−1 *  !

 Δg u u −u ∇g u ∇ g u Δg u u −u
^ p ¼ Gp þ ∇T g u* −β t
G þ βt
k∇gðu* Þk 2
k∇g ðu* Þk k∇gðu* Þk

  
ð29Þ
þ O  uk−1 −u* 
2

2 
¼ Gp þ O  uk−1 −u* 

Equation (29) shows that if uk − 1 is in the ε-vicinity of the ⌢


constraints G d−sk ,which are obtained by real reliability as-
MPTP u∗, Ĝp − Gp is of second order of uk − 1 − u∗. sessment and approximate reliability assessment, respectively,
By applying (29), we further obtain: is of second order of ε in the process of deterministic
optimization.

 ⌢

  k 
⌢
G d−sk −G d−sk ¼ G xkMPTP −G ⌢
x MPTP ¼ Gp −Gp

2 
¼ O  uk−1 −u*  ð30Þ 4 Illustrative examples

It shows that in the ε-vicinity of optimal point, the differ- In this section, several examples are given to demon-
ence between real constrains G(d − sk) and approximate strate the effectiveness, robustness and efficiency of the
An approximate sequential optimization and reliability assessment

Table 1 Iteration history ⌢

of design variables d and the Iter dk xkMPTP xkMPTP
corresponding MPTP x in
Example 1
0 (3.100,2.090) G1 (3.100,2.090) (2.387,1.541)
G2 (3.100,2.090) (3.680,1.402)
G3 (3.100,2.090) (3.717,2.746)
1 (3.138, 3.207) G1 (2.377,1.553) (2.308,2.859)
G2 (3.703,1.422) (3.517,2.391)
G3 (3.651,2.801) (3.758,3.859)
2 (3.476,3.275) G1 (2.318,2.836) (2.662,2.890)
G2 (3.446,2.361) (3.774,2.426)
G3 (3.748,3.868) (4.126,3.898)
3 (3.439,3.287) G1 (2.658,2.901) (2.620,2.913)
G2 (3.791,2.432) (3.758,2.444)
G3 (4.120,3.905) (4.086,3.915)
4 (3.439,3.287) G1 (2.618,2.917) (2.620,2.913)
G2 (3.759,2.446) (3.758,2.444)
G3 (4.082,3.917) (4.086,3.915)

ASORA method. All these examples are tested in Where G1 ðXÞ ¼ X 21 X 2 = 20 − 1

Matlab R2013b and SQP algorithm is used for deter- ðX 1 þ X 2 −5Þ2 ðX 1 −X 2 −12Þ2
ministic optimization. G2 ðXÞ ¼ þ −1
30
2 120 
G3 ðXÞ ¼ 80 = X 1 þ 8X 2 þ 5 − 1

4.1 Example 1

The first example (Yang and Gu 2004) is given as: There are two independent random variables which follow
T normal distribution and three probabilistic constraints and
Find d ¼ ½d 1 ; d 2 
min d1 þ d2 βt ¼ 3:0. The optimal design obtained from deterministic op-
s:t: P G j ðXÞ≥0 ≤ Φ ð−β t Þ ; j ¼ 1; 2; 3 timization are selected as the initial design point, i.e.,
0≤d 1 ≤ 10; 0 ≤d 2 ≤10: X i ∼N ðd i ; 0:3Þ f or i ¼ 1; 2 d0 = [3.100, 2.090]T.

Fig. 3 Iteration history of d and

the corresponding MPTP x for G1
 P. Yi et al.

Table 2 Results of different approaches in Example 1

Objective Optimal design Number of function calls βMCS

1 βMCS
2 βMCS
3

Objective function Performance function

RIA 6.726 (3.439,3.287) 16 882 2.97 3.05 Infinite

PMA 6.728 (3.439,3.289) 12 1431 2.97 3.06 Infinite
SLSV (Yang and Gu 2004) 6.731 (3.434,3.297) 55 484 2.97 3.09 Infinite
SORA 6.726 (3.439,3.287) 15 210 2.97 3.05 Infinite
ASORA 6.725 (3.440,3.285) 20 48 2.97 3.05 Infinite

Table 1 shows the iteration history of design vari- evaluations, which is far less than any other methods.
ables dk, the approximate MPTP ⌢ x kMPTP and accurate The last three rows in Table 2 represent the probabilistic
k constraints evaluated by Monte-Carlo simulation (MCS)
MPTP xMPTP for three performance functions. The ap-
proximate MPTP ⌢ x kMPTP is obtained by (18). The accu- with ten-million samples at the optimum in order to
k
rate MPTP xMPTP is obtained by the AMV method, i.e., check if the probabilistic constraints are satisfied. If
(17). It can be seen that although the error of ⌢ x kMPTP the probabilistic constraint is inactive, like g3, the reli-
k
and xMPTP for each performance function is significant ability will be infinite. As shown in Table 2, the prob-
at the first two iterations, it decreases quickly and ap- abilistic constraint g 1 is violated slightly no matter
proaches zero when the iteration converges. That is, the which method is adopted while the probabilistic con-
design variables d k and approximate MPTP ⌢ x kMPTP con-
straint g2 can satisfy the corresponding target reliability
verge simultaneously. Figure 3 shows the iteration his- for all these methods.
tory of d and the corresponding MPTP x for G1. From
this figure, it can be more easily seen that the design 4.2 Example 2
variables dk and approximate MPTP ⌢ x kMPTP converge
simultaneously. The RBDO model is formulated as (Lee and Lee
The results obtained from five different methods are 2005):
listed in Table 2 and the number of function calls is
used to compare the efficiency of different methods.
All the methods converge to almost the same optimal Find d ¼ ½d 1 ; d 2 ; d 3 ; d 4 ; d 5 ; d 6 ; d 7 ; d 8 ; d 9 ; d 10 T
result and the objective is 6.731 in the reference (Yang min f ðdÞ ¼ d 21 þ d 22 þ d 1 d 2 − 14 d 1 − 16 d 2 þ ðd 3 −10Þ2
and Gu 2004). Although ASORA needs 20 objective þ 4 ðd 4 −5Þ2 þ ðd 5 −3Þ2 þ 2 ðd 6 −1Þ2 þ 5d 27
function calls, which is almost equal to that of SORA þ 7ðd 8 −11
Þ2 þ 2ðd 9 −10Þ2 þ ðd 10 −7Þ2 þ 45
method and is greater than that of other three methods, s:t: P G j ðXÞ > 0 ≤ Φ −β j;t ; β j;t ¼ 3:0; j ¼ 1; …; 8


it can converge after only 48 performance function d i ≥ 0 ; X i e N d i ; 0:022 f o r i ¼ 1 ; … ; 10 ;

Table 3 Results of different

approaches in Example 2 Objective Number of function calls

Objective function Performance function

RIA 27.749 151 121326

PMA 27.749 72 23368
SLSV (Cho and Lee 2011) 27.751 101 9784
SORA 27.750 52 335
ASORA 27.746 54 192
An approximate sequential optimization and reliability assessment

4X 1 þ 5X 2 −3X 7 þ 9X 8
where G1 ðXÞ ¼ − 1
105
G2 ðXÞ ¼ 10 X 1 − 8X 2 − 17 X 7 þ 2 X 8
−8X 1 þ 2X 2 þ 5X 9 −2X 10
G 3 ðX Þ ¼ − 1
12
2 2
3ðX 1 −2Þ þ 4ðX 2 −3Þ þ 2X 3 −7X 4
2
G 4 ðX Þ ¼ −1
120
2
5X 1 þ 8X 2 þ ðX 3 −6Þ −2X 4
2
G 5 ðX Þ ¼ − 1
40
2 2
0:5ðX 1 −8Þ þ 2ðX 2 −4Þ þ 3X 5 −X 6
2
G 6 ðX Þ ¼ −1
30
2 Fig. 4 A welded beam
G7 ðXÞ ¼ X 1 þ 2 ðX 2 −2Þ −2X 1 X 2 þ 14X 5 −6X 6
2

G8 ðXÞ ¼ − 3 X 1 þ 6 X 2 þ 12 ðX 9 −8Þ2 − 7 X 10
shear stress, bending stress, buckling, and displacement are
used. The objective of this problem is to minimize the welding
cost and the RBDO model is expressed as:
There are ten independent random variables, all of Find d ¼ ½d 1 ; d 2 ; d 3 ; d 4 T
which follow normal distribution, and 8 probabilistic min f ðd; zÞ ¼ c1 d 21 d 2 þ c2 d 3 d 4 ðz2 þ d 2 Þ



constraints. The optimal results obtained from determin- s : t : P G j ðXÞ > 0 ≤ Φ −β j;t ; β j;t ¼ 3:0 ; j ¼ 1 ; … ; 5 :
istic optimization are selected as the initial design point, 3:175 ≤ d 1 ≤ 50:8 ; 0 ≤ d 2 ≤ 254 ; 0 ≤ d 3 ≤ 254 ; 0 ≤ d 4 ≤ 50:8
i.e., d0 = [2.17, 2.36, 8.77, 5.10, 0.99, 1.43, 1.32, 9.83, 8.28, X i ∼N ðd i ; 0:1693Þ f or i ¼ 1; 2; X i ∼N ðd i ; 0:0107Þ f or i ¼ 3; 4:
8.38] T.The optimization results obtained by the five
where G1 ðX; zÞ ¼ τ ðX ; zÞ=z6 −1; G2 ðX; zÞ ¼ σ ðX ; zÞ=z7 −1
methods are listed in Table 3. The optimal results are
G3 ðXÞ ¼ X 1 = X 4 − 1 ; G4 ðX; zÞ ¼ δ ðX ; zÞ = z5 − 1
almost equal to that in the reference (Lee and Lee G5 ðX; zÞ ¼ 1 − Pc ðX ; zÞ z1
2005), where the objective is 27.80. The number of n o1=2
τ ðX; zÞ ¼ t ðX; zÞ2 þ 2t ðX; zÞ2 ttðX; zÞX 2 =2RðXÞ þ ttðX; zÞ2
objective function for ASORA is almost equal to that z1
for SORA but is less than other three methods. And t ðX; zÞ ¼ pffiffiffi ; tt ðX; zÞ ¼ M ðX; zÞ R ðXÞ = J ðXÞ
2X 1 X 2 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
performance function evaluations are much less than
X 22 þ ðX 1 þ X 3 Þ2
other methods when ASORA method is employed. The M ðX; zÞ ¼ z1 ðz2 þ X 2 =2Þ ; R ðXÞ ¼
2
probabilistic constraints at the optimum are evaluated by pffiffiffi n o
2
MCS with ten-million samples, shown in Table 4. It can J ðXÞ ¼ 2X 1 X 2 X 2 =12 þ ðX 1 þ X 3 Þ =4
2

be seen that six probabilistic constraints are active and 6z1 z2 4z1 z32
σðX; zÞ ¼ ; δðX; zÞ ¼
they all satisfy the corresponding target reliability. X 23 X 4 z3 X 33 X 4
pffiffiffiffiffiffiffiffi  rffiffiffiffiffi
4:013X 3 X 34 z3 z4 X 3 z3
Pc ðX; zÞ ¼ 1−
6z22 4z2 z4
4.3 Example 3

A welded beam (Cho and Lee 2011) is shown in Fig. 4 and There are some system parameters in the above expression
there are four independent random variables, all of which have and they are listed in Table 5.
normal distribution. Five probabilistic constraints related to

Table 5 System
Table 4 Evaluation of probabilistic constraints at the optimum in parameters for the z1 2.6688 × 104(N)
Example 2 by MCS welded beam z2 3.556 × 102(mm)
z3 2.0685 × 102(MPa)
βMCS
1 βMCS
2 βMCS
3 βMCS
4 βMCS
5 βMCS
6 βMCS
7 βMCS
8
z4 8.274 × 104(MPa)
RIA 3.00 3.00 3.00 3.00 3.00 Infinite 3.00 Infinite z5 6.35(mm)
PMA 3.00 3.00 3.00 3.00 3.00 Infinite 3.01 Infinite z6 9.377 × 101(MPa)
SLSV 3.00 3.00 3.01 3.04 3.00 Infinite 3.00 Infinite z7 2.0685 × 102(MPa)
SORA 3.00 3.00 3.00 3.00 3.00 Infinite 3.01 Infinite c1 6.74135 × 10−5($/mm3)
ASORA 3.00 3.00 3.00 3.00 3.00 Infinite 3.01 Infinite c2 6.74135 × 10−6($/mm3)
 P. Yi et al.

Table 6 Results of different

approaches in Example 3 Objective Optimal design Number of function calls

Objective Performance
function function

RIA 2.591 (5.730,200.91,210.60,6.239) 23 9590

PMA 2.592 (5.728,200.97,210.71,6.238) 17 1105
SLSV (Cho and Lee 2.592 (5.728,200.99,210.72,6.238) 17 535
2011)
SORA 2.592 (5.731,200.93,210.64,6.242) 73 125
ASORA 2.591 (5.730,200.90,210.60,6.239) 77 90

The optimal result obtained from deterministic opti- 4.4 Example 4

mization is selected as the initial design point, i.e.,
d0 = [6.208, 157.82, 210.62, 6.208]T. The optimization re- A speed reducer (Cho and Lee 2011) is shown in Fig. 5
sults obtained by the five methods are listed in Table 6 and there are seven independent random variables, in-
and the probabilistic constraints at the optimum are cluding gear width (X1), gear module (X2), the number
evaluated by MCS with ten-million samples, presented of pinion teeth (X3), distance between bearings (X4, X5)
in Table 7. As shown in Table 6, the optimal points and diameter of each shaft (X6, X7), and 11 probabilistic
obtained from different methods are almost equal to constraints related to bending stress, contact stress, lon-
those from the reference (Cho and Lee 2011), where gitudinal displacement, stress of the shaft, and geometry
the objective is 2.593 and the optimal design is constraints. The objective is the minimization of weight
(5.731, 200.93, 210.64, 6.242). However, the number and all random variables follow normal distribution. The
of function calls is different for different methods. As RBDO model is expressed as:
double loop methods, RIA and PMA are not efficient
while the efficiency of SLSV and SORA is improved Find d ¼ ½d 1 ; d 2 ; d 3 ; d 4 ; d 5 ; d 6 ; d 7 T


significantly compared to the double methods. Cho and min f ðdÞ ¼ 0:7854 d 1 d 22 3:3333d 23 þ 14:9334d 3 −43:0934
Lee (2011) proposed another improved SORA method,



þ1:508d 1 d 26 þ d 27 þ 7:477 d 36 þ d 37 þ 0:7854 d 4 d 26 þ d 5 d 27
which needed 49 objective function evaluations and 260


s : t : P G j ðXÞ > 0 ≤ Φ −β j;t ; β j;t ¼ 3:0 ; j ¼ 1 ; … ; 11
performance function evaluations. Though the number
of objective calls is greater than other methods, it is 2:6 ≤ d 1 ≤ 3:6; 0:7 ≤ d 2 ≤ 0:8; 17 ≤ d 3 ≤ 28; 7:3 ≤ d 4 ≤ 8:3;
seen that ASORA is the most efficient method, which 7:3 ≤ d 5 ≤ 8:3 ; 2:9 ≤ d 6 ≤ 3:9 ; 5:0 ≤ d 7 ≤ 5:5
can converge after only 90 performance function evalu- X i∼ N ðd i ; 0:005Þ f o r i ¼ 1 ; ⋯ ; 7
ations. From Table 7, all the probabilistic constraints
except g4 are active and the reliability index of each
probabilistic constraint satisfies the target reliability
index.

Table 7 Evaluation of probabilistic constraints at the optimum in

Example 3 by MCS

βMCS
1 βMCS
2 βMCS
3 βMCS
4 βMCS
5

RIA 3.01 3.01 3.00 Infinite 3.01

PMA 3.01 3.53 3.01 Infinite 2.98
SLSV 3.01 3.58 3.01 Infinite 2.98
SORA 3.01 3.58 3.01 Infinite 2.98
ASORA 3.01 3.58 3.01 Infinite 2.98
Fig. 5 A speed reducer
An approximate sequential optimization and reliability assessment

Table 8 Results of different

approaches in Example 4 Objective optimal design Number of function calls

Objective Performance
function function

RIA 3038.58 (3.577,0.700,17.000,7.300,7.754,3.365,5.302) 17 29096

PMA 3040.02 (3.578,0.700,17.000,7.300,7.764,3.366,5.302) 13 19435
SLSV (Cho and 3048.45 (3.589,0.700,17.000,7.300,7.783,3.369,5.307) 17 1771
Lee 2011)
SORA 3040.02 (3.580,0.700,17.000,7.300,7.764,3.366,5.302) 21 500
ASORA 3038.61 (3.577,0.700,17.000,7.300,7.754,3.365,5.302) 22 396

27 397:5 1:93X 34 shown in Table 9, there are only four active probabilistic con-
where G1 ðXÞ ¼ − 1 ; G 2 ðXÞ ¼ − 1 ; G 3 ðXÞ ¼ −1;
X 1 X 22 X 3 X 1 X 22 X 23 X 2 X 3 X 46
r ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
straints at the optimal points and all the reliability indexes of
1:93X 35
745X 4
X 2X 3 þ 16:9  106 probabilistic constraints at the optimal points are greater than
G 4 ðXÞ ¼ 4
− 1 ; G 5 ðXÞ ¼ − 1100 ; the target reliability index.
X 2X 3X 7 0:1X 36
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2 ffi
745X 5
þ 157:5  106
X 2X 3 X1
G 6 ðXÞ ¼ −850; G7 ðXÞ ¼ X 1 X 3 −40; G8 ðXÞ ¼ 5 − ;
0:1X 37 X2
5 Conclusions
X1 1:5X 6 þ 1:9 1:1X 7 þ 1:9
G9 ðXÞ ¼ − 12; G10 ðXÞ ¼ − 1; G11 ðXÞ ¼ −1
X2 X4 X5
In this paper, an efficient approach based on SORA algorithm
has been proposed to solve RBDO problem. In order to reduce
The optimal result obtained from deterministic optimiza- the number of performance function analysis, approximate re-
tion is selected as the initial design point, i.e., d0 = [3.5, 0.7, liability assessment is adopted. In each cycle, the approximate
17.0, 7.3, 7.72, 3.35, 5.29]T. The optimization results obtained MPTP is reserved, which will be used to obtain new approxi-
by the five methods are listed in Table 8 and the probabilistic mate MPTP in the next cycle. There is no need to evaluate the
constraints at the optimum are evaluated by MCS with ten- performance functions in the deterministic optimization since
million samples, presented in Table 9. From Table 8, it can be the approximate PPM and its sensitivity are used to formulate
seen that all these methods can converge to almost the same the linear Taylor expansion of the constraint functions. The
optimal points, which are also equal to those from the refer- error analysis shows that in the ε-vicinity of optimal point,
ence (Cho and Lee 2011), where the objective is 3038.61 and the difference between real constrains and approximate con-
the optimal design is (3.580,0.700,17.000,7.300, 7.764,3.366, straints, which are obtained by real reliability assessment and
5.302). However, the number of function calls varies consid- approximate reliability assessment, respectively, is of second
erably. Cho and Lee (2011) proposed another improved order of ε in the process of deterministic optimization. The
SORA method, which needed 36 objective functions evalua- results of numerical examples also verify the efficiency and
tions and 627 performance functions evaluations. ASORA robustness of ASORA. Although sometimes the number of
demonstrates good convergence behavior and converges after objective function evaluation is greater than other methods,
only 22 objective function evaluations and 396 performance ASORA is the most efficient method because it significantly
function evaluations, which is the most efficient method. As reduces the performance function evaluations.

Table 9 Evaluation of probabilistic constraints at the optimum in Example 4 by MCS

βMCS
1 βMCS
2 βMCS
3 βMCS
4 βMCS
5 βMCS
6 βMCS
7 βMCS
8 βMCS
9 βMCS
10 βMCS
11

RIA Infinite Infinite Infinite Infinite 3.02 3.00 Infinite 3.00 Infinite Infinite 3.02
PMA Infinite Infinite Infinite Infinite 3.15 3.08 Infinite 3.07 Infinite Infinite 4.22
SLSV Infinite Infinite Infinite Infinite 3.65 4.07 Infinite 3.50 Infinite Infinite Infinite
SORA Infinite Infinite Infinite Infinite 3.15 3.08 Infinite 3.07 Infinite Infinite 4.21
ASORA Infinite Infinite Infinite Infinite 3.18 3.08 Infinite 3.07 Infinite Infinite 4.21
 P. Yi et al.

Acknowledgments The authors would like to deeply thank the anony- Liang J, Mourelatos ZP, Tu J (2004). A single-loop method for reliability-
mous reviewers for their useful and helpful comments and gratefully based design optimization. In: Proceedings of ASME design engi-
acknowledge the financial support provided by the National Natural neering technical conferences, paper N0DETC2004/DAC-57255
Scientific Foundation of China (51479027) and the Key Project of Liang J, Mourelatos ZP, Nikolaidis E (2007) A single-loop approach for
Chinese National Programs for Fundamental Research and system reliability-based design optimization. J Mech Des 129(12):
Development (2015CB057703). 1215–1224
Nguyen TH, Song J, Paulino GH (2010) Single-loop system reliability-
based design optimization using matrix-based system reliability
method: theory and applications. J Mech Des 132(1):0110051–
References 01100511
Nikolaidis E, Burdisso R (1988) Reliability based optimization: a safety
Aoues Y, Chateauneuf A (2010) Benchmark study of numerical methods index approach. Comput Struct 28(6):781–788
for reliability-based design optimization. Struct Multidiscip Optim Rackwitz R, Fiessler B (1978) Structural reliability under combined ran-
41:277–294 dom load sequence. Comput Struct 9(5):489–494
Chen D, Hasselman T, Neill D (1997) Reliability-based structural design Silva M, Tortorelli DA, Norato JA, Ha C, Bae HR (2010) Component and
optimization for practical applications. In: Proceeding of the system reliability-based topology optimization using a single-loop
38thAIAA/ASME/ASCE/AHS/ASC structures, structural dynam- method. Struct Multidiscip Optim 41(1):87–106
ics, and material conference, Kissimmee, AIAA-97-1403 Tu J, Choi KK, Park YH (1999) A new study on reliability based design
Chen Z, Qiu H, Gao L, Su L, Li P (2013a) An adaptive decoupling optimization. J Mech Des 121(4):557–564
approach for reliability-based structural optimization. Comput Wang L, Kodiyalam S (2002) An efficient method for probabilistic and
Struct 84(21):1353–1367 robust design with non-normal distributions. In: Proceeding of the
Chen Z, Qiu H, Gao L, Li P (2013b) An optimal shifting vector approach 43th AIAA/ASME/ASCE/AHS/ASC structures, structural dynam-
for efficient probabilistic design. Struct Multidiscip Optim 47(6): ics, and material conference, Denver, AIAA-2002-1754
905–920
Wu YT, Wang W (1998) Efficient probabilistic design by converting
Cheng G, Xu L, Jiang L (2006) A sequential approximate programming
reliability constraints to approximately equivalent deterministic con-
stragety for reliability-based structural optimization and reliability
straints. J Integr Des Process Sci 2(4):13–21
assessment method. Struct Saf 33(1):42–50
Wu YT, Millwater HR, Cruse TA (1990) Advanced probabilistic struc-
Cho TM, Lee BC (2011) Reliability-based design optimization using
tural analysis method for implicit performance functions. AIAA J
convex linearization and sequential optimization and reliability as-
28(9):1663–1669
sessment method. Struct Saf 33(2011):42–50
Du X, Chen W (2004) Sequential optimization and reliability assessment Yang RJ, Gu L (2004) Experience with approximate reliability-based
method for efficient probabilistic design. J Mech Des 126(2):225–233 optimization methods. Struct Multidiscip Optim 26(2):152–159
Frangopol DM, Corotis RB (1996) Reliability-based structural system Yi P, Cheng GD (2008) Futher study on efficiency of sequential approx-
optimization: state-of-the-art versus state-of-the-practice. Proc. imate programming for probabilistic structural design optimization.
12th conference held in conjunction with structures congress XIV, Struct Multidiscip Optim 35(6):509–522
67–78 Youn BD, Choi KK (2004) An investigation of nonlinearity of
He Q, Li Y-S, Ao L-B, Wen Z-X, Yue Z-F (2011) Reliability and multi- reliability-based design optimization approaches. J Mech Des
disciplinary design optimization for turbine blade based on single- 126(5):403–411
loop method. J Propulsion Technol 32(5):658–663 Youn BD, Choi KK, Park YH (2003) Hybrid analysis method for reli-
Ju BH, Lee BC (2008) Reliability-based design optimization using a mo- ability based design optimization. J Mech Des 125(3):221–232
ment method and a kriging metamodel. Eng Optim 40(5):421–438 Zhang X, Huang H, Xie M (2011) Single loop methods for reliability
Lee JJ, Lee BC (2005) Efficient evaluation of probabilistic constraints based design optimization. J Univ Electron Sci Technol China 40(1):
using an envelope function. Eng Optim 37(2):185–200 157–160
Lee JO, Yang YS, Ruy WS (2002) A comparative study on reliability- Zhuang XT, Pan R (2012) A sequential sampling strategy to improve
index and target-performance-based probabilistic structural design reliability-based design optimization with implicit constraint func-
optimization. Comput Struct 80:257–269 tions. J Mech Des 134:0210021–02100210