IOP Conference Series: Earth and Environmental Science

PAPER • OPEN ACCESS You may also like

- Combined Subset Simulation and
Review of Reliability-Based Design Optimization Comprehensive Learning Particle Swarm
Optimization in Reliability-Based Structural
Approach and Its Integration with Bayesian Optimization
Arnut Sutha, Thu Huynh Van and
Sawekchai Tangaramvong
Method
- Effect of surfactants on the lightning
breakdown voltage of palm oil and coconut
To cite this article: Xiangnan Zhang 2018 IOP Conf. Ser.: Earth Environ. Sci. 128 012109 oil based Al2O3 nanofluids
Nur Aqilah Mohamad, Norhafiz Azis,
Jasronita Jasni et al.

- Design optimization under uncertainty and

speed variability for a piezoelectric energy
View the article online for updates and enhancements. harvester powering a tire pressure
monitoring sensor
Amin Toghi Eshghi, Soobum Lee,
Mohammad Kazem Sadoughi et al.

This content was downloaded from IP address 27.34.73.178 on 18/10/2022 at 08:56

ICEESE 2017 IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
128 (2018) 012109 doi:10.1088/1755-1315/128/1/012109

Review of Reliability-Based Design Optimization Approach

and Its Integration with Bayesian Method

Xiangnan Zhang1
Faculty of Vehicle Engineering and Mechanics, Dalian University of Technology
No.2 Linggong Road, Ganjingzi District, Dalian City, Liaoning Province, P.R.C.,
116024
E-mail: 1960946626@qq.com

Abstract. A lot of uncertain factors lie in practical engineering, such as external load
environment, material property, geometrical shape, initial condition, boundary condition, etc.
Reliability method measures the structural safety condition and determine the optimal design
parameter combination based on the probabilistic theory. Reliability-based design optimization
(RBDO) is the most commonly used approach to minimize the structural cost or other
performance under uncertainty variables which combines the reliability theory and optimization.
However, it cannot handle the various incomplete information. The Bayesian approach is utilized
to incorporate this kind of incomplete information in its uncertainty quantification. In this paper,
the RBDO approach and its integration with Bayesian method are introduced.

1. Reliability-based Design Optimization

Various uncertainties lie in the practical engineering process, which have a great influence on the design
optimization results. Some types of constraints such as initial condition and load capacity may be
violated when uncertainties are taken into consideration. Under the circumstance, reliability-based
design optimization (RBDO) is developed to solve the design optimization problem with reliability
constraints. The RBDO process deals with two optimization models, simultaneously. The first one is a
design optimization model, which search the feasible solution in an original random space. The other
one is a reliability analysis model, which transfer the probability distribution into nonlinear mapping
and finds the optimal solution in a standard normal random space. Hence the double-loop strategy is
most often used to deal with RBDO problem [1, 2, 3].
The complexity and large calculating quantity of RBDO model is also a difficult question to be
addressed. A single-loop strategy was proposed to improve computation efficiency. By decoupling the
design optimization model and reliability analysis model, the computational process is greatly simplified
[4, 5]. The RBDO model was initially developed by defining a probabilistic constraint as the reliability,
a so-called reliability index approach (RIA). Then, an inverse approach was proposed by formulating
RBDO with a probabilistic performance measure (PMA). And RIA and PMA have been the mainstream
of RBDO.
In the system parameter design process, the RBDO model [6] can be generally defined as

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
ICEESE 2017 IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
128 (2018) 012109 doi:10.1088/1755-1315/128/1/012109

Minimize Cost(d)
subject to P(Gi ( X)  0) - Φ(-βt )  0, i  1,2,, NP
dL  d  dU , d  R ndv and X  R nrv (1)
Where d  μ(X) is the design vector, X is the random vector, and the probabilistic constraints are
described by the performance function Gi ( X ) with Gi ( X )  0 indicates failure, their probabilistic

models, and their prescribed confidence level t .
FGi (0)
Performance function failure is statistically defined by a cumulative distribution function as
P(G i (X)  0)  FG i (0)   G i (X) 0  f X (x)dx  Φ( β t )
(2)
F ( 0)
In Eq. (2), Gi is a joint probability density function, which needs to be integrated. The first-order
reliability method (FORM) is a widely-used integration method, which combines the approximate
probability and reliability measure approach. The approach has proven to be efficient and accurate. By
inverse transformation, the probabilistic in Eq. (2) can be further expressed in two distinct forms as:
 Si  ( 1 ( FGi (0)))  t
(3)
Gpi  FG1 (( t ))  0
i (4)
 G
Where Si and Pi represent the safety reliability index and the probabilistic performance measure
th
for the i probabilistic constraint, respectively. Using the safety reliability index, Eq. (3) is then
employed to describe the probabilistic constraint in Eq. (1), i.e., the so-called reliability index approach
(RIA). Similarly, Eq. (4) can replace the probabilistic constraint in Eq. (1) with the performance measure,
referred to as the performance measure approach (PMA).

1.1 First-Order Reliability Analysis in RIA


In RIA, the first-order safety reliability index s is obtained by solving the optimization problem
formulated in FORM, with an implicit equality constraint in U-space defined as the limit state function:
minimize U , subject to G( U)  0
(5)
The optimum point on the failure surface is referred to as the most probable failure point (MPFP)
u*G( U ) 0
. Any MPFP search algorithm specifically developed for FORM or general optimizer can be
used to solve Eq. (5).

1.2 First-Order Reliability Analysis in PMA

The first-order reliability analysis in PMA can be formulated as the inverse of the first-order reliability
G
analysis in RIA. The first-order probabilistic performance measure p , FORM is obtained from a nonlinear
optimization problem with an n-dimensional explicit sphere constraint in U-space, defined as
minimize G( U), subject to U  t
(6)
The optimum point on a target reliability surface is identified as the most probable point (MPP)
u*   t
An advanced mean value (AMV) method can be used to numerically solve the inverse PMA problem.

2
ICEESE 2017 IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
128 (2018) 012109 doi:10.1088/1755-1315/128/1/012109

However, it is found that AMV method exhibits poor behavior for concave constraint functions, although
it is effective for convex constraint functions. To overcome difficulties in the AMV method, a conjugate
mean value (CMV) method is proposed for the concave constraint function in PMA [3]. The HMV
method [3] combines both the CMV and AMV methods - the CMV method is used for concave
constraint functions and the AMV method is used for convex constraint functions. The HMV method
has been shown to be very robust and efficient.

1.3 Comparison Between RIA and PMA

As mentioned above, in RBDO model the reliability is represented as probabilistic constraints with RIA.
But problems do exist with this approach. There is defect of slow speed of convergence due to various
reasons. PMA is used to simplify the constraints, which improves the convergence speed by solving an
inverse problem for the FORM. Although the cost function becomes more complicated, the searching
space is largely reduced.
Compared to RIA, PMA has the following advantages: (1) PMA is more robust and effective whether
the probabilistic constrains are feasible or not. (2) PMA can always find the local optimal solution,
nevertheless, the solution infeasibility may arise in RIA due to a variety of reasons such as Gumbel or
uniform distributions. (3) PMA is more effective than RIA when RBDO applies response surface method
(RSM) to conduct reliability analysis and design optimization. However, when combining with Bayesian
approach, RIA is much more convenient and straightforward in formulation.

2. Improved RBDO Approach with Bayesian Method

To accurately infer the probability distributions, we have to get a large number of experimental data
from actual design cases. One of the most common ways is to obtain sample data at some interval from
historical data or from field measurements. However, collecting more samples is not only time-
consuming but also very expensive. In terms of its information content, quantification of uncertainty as
a set of finite samples can be considered as a medium level of information. In this situation, much of
this valuable information is discarded. There's been a lot of research over the past decade to integrate
the incomplete information with uncertainty quantification. Du, et. al., focused on reliability-based
optimization under a mixture of both random and interval variables [7]. This method is quite successful
in accounting for either end of the spectrum although it is not applicable to problems with finite samples.
In some approaches the probability density function of the uncertain quantities is determined based on
the sampled data, which makes the RBDO algorithms more applicable [8] [9].
The maximum entropy principle [10] is also used to derive pdf based on the given data. However,
the process of deriving the pdf will produce an error in the data and it is difficult to propagate the error
in the prediction. Another popular research direction is the use of fuzzy sets to quantify the amount of
uncertainty based on the possibility of the method [11] [12]. This method includes likelihood of evidence
theory [13]. Fuzzy sets can capture different levels of information by controlling the reduction of
membership functions. However, it is difficult to build a membership function because it involves
"expert opinions", and each expert may be different and may even be in conflict. [14] Tried to combine
the possibility of quantification with probability to obtain the advantages of both representations.
In conclusion, RBDO problems with incomplete information are still not well solved in practical
design situations. Our objective in this article is to develop a reliability-based optimization method for
problems with incomplete information. More specifically, we utilize Bayesian statistics to solve
problems for which a mixture of finite samples and probability density functions of the uncertain
quantities is known.

2.1 Bayesian Inference Method

Consider a Bernoulli process, such as a coin toss, whose probabilities of “success” and “failure” are p
and (1-p) respectively. Given N independent trials, the probability of having r successes out of these
trials follows a Binomial distribution: r~bin(N,p) In a Binomial distribution, the probability of success
p of each trial is known, and we want to predict the outcome of a trial from it. An inference of this

3
ICEESE 2017 IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
128 (2018) 012109 doi:10.1088/1755-1315/128/1/012109

process seeks to calculate p based on the outcomes of the trials. Given r successes out of N trials, the
probability distribution of p can be calculated using Bayes’ theorem shown in Eq. (7)
f ( p)  f (r | p)
f ( p | r) 
1
0 f ( p)  f (r | p)dp (7)
In this equation f ( p ) is the prior distribution of p, f( p | r ) is the posterior distribution of p, and
f ( r | p ) is the likelihood of r given p. The integral in the denominator is a normalizing factor to make
the probability distribution proper.
The posterior distribution is the distribution of interest. It is the estimate of p based on the outcome
of the trials. The prior distribution is our knowledge about p before the information from the trials. Many
types of prior distribution have been proposed, but it has been shown that the analysis is insensitive to
the choice of prior if the number of trials is large enough. In this article we assume a uniform prior for
p. This assumption implies that any values of p between 0 and 1 are equally probable. It is a conservative
choice based on the maximum ignorance principle. The only information we have a priori is that p must
be within [0, 1], and a uniform distribution can be fully defined based solely on this information. The
likelihood function is the probability that there will be r successes out of N trials if the probability of
success is p, which is the binomial distribution of r. Using a uniform prior and a Binomial likelihood
function in Eq. (7) results in a Beta posterior distribution, Eq. (8), where α=r+1 and β=(N-r) +1
(   )  1
f ( p | r)  p (1  p)  1
( )(  ) (8)
In other words, p is distributed according to a Beta distribution whose two parameters depend on the
outcome of the trials: p ~ Beta (r  1, ( N  r )  1) .
One very important feature of Bayes’ theorem is that it facilitates an updating scheme to account for
additional information. Suppose that after the N1 initial trials, we conduct N 2 additional trials and
observe r2 more successes. In Bayes’ theorem, the posterior distribution from the N1 trials can be used
as the prior distribution for the N 2 trials, thus creating a chain of analysis based on additional
information. Using beta(1, 1 ) for the prior distribution in Eq.(3), the posterior distribution of p after
the additional N 2 trials is as shown in Eq.(6) Here  2  1  r2 and  2  1  ( N 2  r2 ) .
It can be immediately seen that the new posterior is also a Beta distribution with the two parameters
equal to the prior parameters updated by the new information. This is a conjugacy feature of a Beta
distribution when used in Bayes’ theorem. This conjugacy is not limited to just two successive trials. If
( N , N ,, N )
1 2 k trials, then the posterior of the (k-1) th trial becomes the prior
there is a sequence of
of the kth trial, except for k=1 whose prior is Beta (1,1) (mathematically, a uniform distribution is
equivalent to a Beta(1,1) distribution). In general, the posterior distribution of p at the kth trial is Beta
   k 1  rk and  k   k 1  ( N k  rk ), k  1 .
(αk, βk) where k

2.2 Bayesian Reliability Estimate

gj
In an optimization problem, an inequality constraint iiiiies tee iesinn space into too iomains
t t [X , P ] 
feasible ani infeasible. Suppose for noo teat tee pif’s of all X ani P are unknoon ∅, ani
teat teere are only N samples aiailable. Eace of teese samples is a realization of (X, P) in tee iesinn
space, so eace sample corresponis to eiteer a feasible or an infeasible realization. Let r ani (N-r) be tee
Fg j (0)
number of feasible ani infeasible realizations, respectiiely. Tee noal is to estimate niien teese

4
ICEESE 2017 IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
128 (2018) 012109 doi:10.1088/1755-1315/128/1/012109

Fg j (0)
r ani (N-r) realizations. Teis estimation problem is analonous to tee aayesian inference
gj
problem iiscussei preiiously. If oe use a coin toss as an analony, teen is tee “coin” oite only too
Fg j (0)
mutually exclusiie outcomes, tee N samples are tee “coin tosses” or trials, ani is tee
“probability of eeai,” oeere oe eaie iesinnatei a feasible realization as a “eeai.” Tee problem of
Fg j (0)
inferrinn niien r ani (N-r) realizations is teen equiialent to tee problem of inferrinn p, tee
probability of eeais, niien teat teere are r eeais out of N coin tosses. From tee aboie iiscussion oe
Fg j (0) ~ beta( ,  )
obtain , oeere α=r+1 ani β=(N-r)+1.

j
Fg (0)
A aeta iistribution is strictly bouniei oitein [0,1]. Teis makes tee estimate proper because
is a cif. Tee seape of a aeta iistribution iepenis on tee number of feasible-infeasible realizations. If
teere are many feasible realizations, it oill be skeoei to tee rinet tooaris one; if teere are many
infeasible realizations, it oill be skeoei to tee left tooaris zero (Fin.2). Teis is also a proper beeaiior
Fg (0)
j
for a estimate. If a iesinn eas many feasible realizations, oe expect its reliability to be eine,
ani iice iersa.
Fg (0)
j
Teis estimate of assumes teat tee pif’s of all X ani P are unknoon. Ween some of tee pif’s
are knoon ([Xt, Pt]≠∅)， tee coin toss analony still eolis proiiiei teat teis pif information is accountei
for. Tee iifference betoeen tee too cases lies in tee calculation of tee feasible-infeasible realization of
a iesinn. In tee case oeere all pif’s are unknoon, one (Xs,Ps) sample niies a iistinct feasible or
infeasible realization, eiteer to tee left or rinet of nj =0 in tee nj axis. In contrast, oeen some of tee pif’s
are knoon, one sample results in a iistribution of nj ialues (Fin.3).
Since eace (Xs, Ps) sample noo eas a iistribution of nj ialues, it is not possible to simply count tee
number of feasible or infeasible realizations. Weat oe can calculate insteai is tee probability teat nj is
Pr[ g ( X , P )  0 | ( X , P ) ]
j t t s s k
feasible niien tee kte sample of (Xs,Ps) Teis probability is tee
expectei feasible realization of one sample. Tee sum of tee probabilities of all samples is teen tee
expectei total number of feasible realizations of tee iesinn
N
E (r )   Pr[g j (Xt , Pt )  0 | (X s , Ps )k ]
k 1 (9)

5
ICEESE 2017 IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
128 (2018) 012109 doi:10.1088/1755-1315/128/1/012109

Usinn E(r) as tee expectei number of eeais, oe can use tee coin toss analony anain to estimate
Fg j (0)
. Teis neo estimate is niien by a aeta iistribution oite α=E(r)+1 ani β=(N-E(r)+1. Notice teat
Eq. (9) is ialii for bote oeen only (Xs,Ps) samples are aiailable ani oeen teere is a mix of (Xs,Ps)
samples ani (Xt,Pt) pif’s. In tee case oeere teere is no knoon pif, eace probability in Eq.(9) becomes
g j ((X s , Ps ) k )  0
an iniicator function oeere Ik=1 if ani Ik=0 oteeroise.
Fg (0)
j
Unier incomplete information, can be estimatei oite a aeta iistribution. Tee precision of
teis estimate (nrapeically iepictei as tee spreai of tee iistribution) iepenis on tee number of samples.
As N increases, tee estimate becomes more precise. In tee extreme teat N→∞, tee iistribution coniernes
Fg j (0)
to a Dirac ielta function meaninn teat is knoon exactly, oeice is tee same as knooinn tee pif’s
of all X ani P. Teis link betoeen N ani precision of tee estimate is capturei by tee aayesian upiate.
Follooinn preiious iiscussion, aiiinn N2 samples to an existinn N1

estimate from a aeta(1,1) to a aeta(2,2), oeere  2  1  E (r2 ) ani

Fg j (0)
samples reiises tee
 2  1  ( N  E (r2 )) .

2.3 Bayesian Approach Integrated with Reliability-Based Optimization

Due to tee incompleteness of tee raniom iariable information, oe can only obtain tee iistribution of
tee iesinn reliability rateer tean tee exact ialue. Teerefore, in iesinn optimization, oe cannot ietermine
oeeteer tee reliability of tee iesinn is less tean or nreater tean tee tarnet ialue. Teere is not enoune
information to make suce an exact statement. Teen, tee optimization requires aiiitional measures to
ietermine oeeteer tee iesinn can be consiierei to meet tee reliability requirements. We iefine a
quantity oe call confiience to be suce as a measure. For a iesinn, tee confiience of teat iesinn oite
respect to tee jte inequality constraint is iefinei to be tee probability teat it oill meet or exceei tee
reliability tarnet.
 j ( μ X )  Pr[ Fg j (0) |   Rj]
X j=1, 2, ..., J (10)
 j 0  j 1
means teat tee iesinn is certainly not reliable, oeile means teat tee iesinn certainly
Fg j (0)
meets or exceeis tee tarnet. Since tee estimate is a aeta iistribution, tee confiience can also
 j ( μX )  1   B (R j )
j j
 B ()
be oritten as oeere is tee cif of tee jte aeta iistribution, for all
j=1, ... , J.
Ween oe eaie complete information renariinn tee raniom iariables/parameters, optimization seeks
a iesinn oite tee best objectiie ialue ani oite reliability nreater tean or equal to Rj. Teere is a traie-
off betoeen minimizinn f ani satisfyinn tee reliability constraints. To account for teis traie-off, tee
problem must be reformulatei into a multi-objectiie optimization problem oeere tee objectiies are to
 
minimize f ani to maximize tee j ’s. Teere is a J number of j ; eooeier, in reliability-basei
optimization a iesinn is consiierei feasible if all probabilistic constraints are satisfiei. Follooinn teis
 s ( μ X )  min  j ( μ X )
j 1,, J
coniention, tee j’s can be lumpei into , a quantity callei tee oierall
confiience of a iesinn. Usinn teis measure tee multi-objectiie problem teen becomes as seoon in Eq.
(11).

6
ICEESE 2017 IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
128 (2018) 012109 doi:10.1088/1755-1315/128/1/012109

minimize f ( μ X , μ P )
X
maximize  s ( μ X )
X
subject to：0   s ( μ X )  1
(11)
Soliinn equation (8) usually proiuces a set of Pareto optimal solution sets, rateer tean tee optimal
iesinn. Teis is tee result of incomplete information basei on X ani P it is not possible to ietermine
exactly oeice iesinn is optimal ani reliable. On tee contrary, oe can only net a set of iesinns, eace
iesinn eas a certain ienree of probability to become a true optimal ani reliable iesinn (corresponiinn
to tee complete information).
Tee multi-objectiie formula in tee formula (11) alloos tee Pareto solution to be increasei or reiucei
accoriinn to tee tarnet f to tee true best reliable iesinn. Teis is because tee past is not feasible for tee
iesinn of probability constraints. If tee iistance probability bouniary is not too far, teere may be a non-
zero. If tee problem is finite ani local monotonic, tee relaxation of tee actiie probability constraint
improies tee optimal f-ialue. As a result, tee iesinn farteer aoay from tee actiie bouniary to tee
infeasible iomain, tee better. Teis traie-off iesinn is part of tee forefront of Pareto, oeose ialue is
beyoni tee true best ani reliable iesinn, oeice in turn is establiseei. Tee farteer aoay from actiie
boriers to iiable areas, tee better tee iesinn is, ani tee f becomes farteer. Some of tee Pareto bouniaries
formei by teese iesinns impact tee true ani reliable iesinn of tee f-ialue.
If tee information of tee raniom quantities is only aiailable as samples, i.e., [Xt,Pt]=∅, tee Pareto
frontier of Eq.(11) oill be necessarily iiscrete. Teis property can be explainei as folloos. Tee aeta
Fg (0)
j
iistribution of tee estimate is createi from upiatinn a uniform prior iistribution oite tee
number of feasible-infeasible realizations. For N samples, tee possible r ialues are iiscrete from a
minimum of 0 to a maximum of N. So teere are only a total of (N+1) iifferent aeta iistributions possible,
ani teis translates to only (N+1) iiscrete possible ialues of s. Since s is one of tee too objectiies in Eq.
(11). tee Pareto frontier can only contain teese particular ialues, ani eence is iiscrete. If tee pif’s of
some of X ani P are knoon, tee iiscreteness of s iisappears. Hooeier, tee Pareto frontier may be
iiscrete or continuous iepeniinn on tee continuity of f.

3. Multi-objective Algorithm for RBDO

In single-objective optimization, the best possible design or decision is usually the global minimum or
global maximum, depending on whether the optimization problem is minimized or maximized. PSO and
GA algorithms have been widely used in single target RBDO problems. In recent years, some scholars
have developed a hybrid GA-PSO method for solving the RBDO redundancy assignment problem. On
the other hand, there may not be a solution to the best (global minimum or maximum) of all targets for
multiple targets. In multi-objective optimization, there is a set of solutions that, when considering all
targets, the remaining solutions in the search space outperform other solutions in the space in one or
more goals (not all). These solutions are called Pareto optimal or non-dominant solutions (Srinivas and
Deb (1994)), and the rest of the solution is called the dominant solution. Because any solution in a non-
dominant collection cannot be considered to be superior to the other, any of them is an acceptable
solution. Since the reliability of each component is an interval value, the reliability of the system will
be differentiated. MOMS-GA-based methods have been proposed to solve multi-objective and multi-
state reliability optimization with interval targets [15]. In addition, the multi-objective optimization
problem based on reliability has been transformed into a single target unconstrained optimization
problem by using penalty function techniques [16].

4. Multi-state Algorithm for RBDO

Reliability is iefinei as tee probability teat tee ieiice or system can satisfactorily perform its inteniei
function oitein a specifiei time unier specifiei coniitions. Hooeier, tee traiitional reliability meteoi

7
ICEESE 2017 IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
128 (2018) 012109 doi:10.1088/1755-1315/128/1/012109

assumes teat tee system ani its components can only be in a fully operational state or a complete failure
state, teat is, tee intermeiiate state is not allooei. Teis assumption eelps to ieielop a pooerful ani
broai teeory to analyze tee performance of tee system. Hooeier, in some cases, tee traiitional teeory
of reliability cannot represent tee real beeaiior of tee system. Teis situation may be a major iraoback
oeen tee system eas a series of intermeiiate states teat tee traiitional reliability estimates io not take
into account. In orier to iescribe tee satisfactory performance of tee equipment or system, oe may neei
to use multi-leiel satisfaction, for example, excellent, aierane ani poor. Some stuiies eaie proposei
multi-state reliability as a supplemental teeory to ieal oite traiitional reliability teeory ani moiel
problems of systematic analysis. Teen, in a multi-state system, tee system ani its components are
allooei to experience too or more possible states, suce as full oork, partial or partial failure, ani
complete failure [15].

5. Conclusion
In this paper, the reliability-based design optimization model is introduced in detail. With the integration
of reliability theory and optimization methods, RBDO can deal with engineering problems under
uncertainties. As the most commonly used approach, RIA represents the reliability as probabilistic
constraints, while PMA simplifies the constraints. PMA is more effective and robustness than RIA.
RBDO Approach can also be improved by combining Bayesian Method when information is incomplete.
This kind of problem can be converted into multi-objective and multi-state reliability optimization with
interval targets, which meets the practical requirements more.

References
[1] Thanedar, P.B., and Kodiyalam, S., “Structural Optimization Using Probabilistic Constraints,”
Journal of Structural Optimization, Vol. 4, pp. 236-240, 1992.
[2] Chen, X., Hasselman, T.K., and Neill, D. J., “ReliabilityBased Structural Design Optimization for
Practical Applications,” 38th AIAA SDM Conference, AIAA-97-1403, 1997.
[3] Wang, L, Kodiyalam, S., “An Efficient Method for Probabilistic and Robust Design with Non-
normal Distributions,” 43rd AIAA SDM Conference, Denver, Colorado, 2003
[4] Wu,Y-T.,Shin,Y.,Sues,R.H.,and Cesare, M. A.,“SafetyFactor Based Approach for Probability-
Based Design Optimization,” AIAA-2001-1522, 42nd AIAA/ASME/ASC/AHS/ASC SDM
Conference & Exhibit, Seattle, Washington, April, 2001
[5] Du, X., and Chen, W., “Sequential Optimization and Reliability Assessment Method for Efficient
Probabilistic Design,” ASME Design Engineering Technical Conferences, Montreal, Canada,
September 2002.
[6] Youn, B.D., Choi, K.K., and Park, Y.H., Hybrid Analysis Method for Reliability-Based Design
Optimization, Journal of Mechanical Design, ASME, Vol. 125, No. 2, pp. 221-232, 2003 and
Proceedings of 2001 ASME Design Engineering Technical Conferences: 27th Design
Automation Conference, Pittsburgh, PA, 2001.
[7] Du, X., and Sudjianto, A., 2005, “Reliability-Based Design With the Mixture of Random and
Interval Variables,” ASME J. Mech. Des., 127(6) pp. 1068–1076.
[8] Zong, Z., and Lam, K. Y., 2000, “Bayesian Estimation of Complicated Distributions,” Struct. Safety,
221, pp. 81–95.
[9] Zong, Z., and Lam, K. Y., 2001, “Bayesian Estimation of 2-Dimensional Complicated
Distributions,” Struct. Safety, 23(2) pp. 105–121.
[10] Theil, H., and Fiebig, D. G., 1984, Exploiting Continuity: Maximum Entropy Estimation of
Continuous Distributions, Ballinger, Cambridge, MA.
[11] Choi, K. K., Diu, L., and Youn, B. D., 2004, “A New Fuzzy Analysis Method for Possibility-Based
Design Optimization,” In Proc. of the 10th AIAA/ISSMO Multidisciplinary Analysis and
Optimization Conf., Albany, NY, Aug. 30–Sept. 1, Paper No. AIAA2004–4585.

8
ICEESE 2017 IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
128 (2018) 012109 doi:10.1088/1755-1315/128/1/012109

[12] Mourelatos, Z. P., and Zhou, J., 2004, “Reliability Estimation and Design With Insufficient Data
Based on Possibility Theory,” In Proc. of the 10th AIAA/ ISSMO Multidisciplinary Analysis and
Optimization Conf., Albany, NY, Aug. 30–Sept. 1, Paper No. AIAA2004–4586.
[13] Agarwal, H., 2004, “Reliability Based Design Optimization: Formulations and Methodologies,”
Ph.D dissertation, University of Notre Dame, South Bend, IN.
[14] Langley, R. S., 2000, “Unified Approach to Probabilistic and Possibilistic Analysis of Uncertain
Systems,” J. Eng. Mech. Div., Am. Soc. Civ. Eng., 126(11) pp. 1163–1172.
[15] Heidi A. Taboada, Jose F. Espiritu, and David W. Coit MOMS-GA: A Multi-Objective Multi-State
Genetic Algorithm for System Reliability Optimization Design Problems EEE
TRANSACTIONS ON RELIABILITY, VOL. 57, NO. 1, MARCH 2008
[16] Laxminarayan Sahoo a,, Asoke Kumar Bhunia b, Parmad Kumar Kapur c Genetic algorithm based
multi-objective reliability optimization in interval environment Computers & Industrial
Engineering 62 (2012) 152–160