An Effective Approach to Solve

Design Optimization Problems

With Arbitrarily Distributed

Downloaded from http://asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/138/7/071403/6229138/md_138_07_071403.pdf by National Taiwan University of Science and Technology user on 07 September 2023
Uncertainties in the Original
Design Space Using Ensemble of
Gaussian Reliability Analyses
Po Ting Lin1 Reliability-based design optimization (RBDO) algorithms have been developed to solve
Intelligent Unmanned System R&D Group,
design optimization problems with existence of uncertainties. Traditionally, the original
Department of Mechanical Engineering,
random design space is transformed to the standard normal design space, where the reli-
Chung Yuan Christian University,
ability index can be measured in a standardized unit. In the standard normal design
Chungli, Taoyuan 32023, Taiwan
space, the modified reliability index approach (MRIA) measured the minimum distance
e-mail: [email protected]
from the design point to the failure region to represent the reliability index; on the other
hand, the performance measure approach (PMA) performed inverse reliability analysis
Shu-Ping Lin to evaluate the target function performance in a distance of reliability index away from
Intelligent Unmanned System R&D Group,
the design point. MRIA was able to provide stable and accurate reliability analysis while
Department of Mechanical Engineering,
PMA showed greater efficiency and was widely used in various engineering applications.
Chung Yuan Christian University,
However, the existing methods cannot properly perform reliability analysis in the stand-
Chungli, Taoyuan 32023, Taiwan
ard normal design space if the transformation to the standard normal space does not exist
or is difficult to determine. To this end, a new algorithm, ensemble of Gaussian reliability
analyses (EoGRA), was developed to estimate the failure probability using Gaussian-
based kernel density estimation (KDE) in the original design space. The probabilistic
constraints were formulated based on each kernel reliability analysis for the optimization
processes. This paper proposed an efficient way to estimate the constraint gradient and
linearly approximate the probabilistic constraints with fewer function evaluations (FEs).
Some numerical examples with various random distributions are studied to investigate
the numerical performances of the proposed method. The results showed that EoGRA is
capable of finding correct solutions in some problems that cannot be solved by traditional
methods. Furthermore, experiments of image processing with arbitrarily distributed
photo pixels are performed. The lighting of image pixels is maximized subject to the ac-
ceptable limit. Our implementation showed that the accuracy of the estimation of normal
distribution is poor while the proposed method is capable of finding the optimal solution
with acceptable accuracy. [DOI: 10.1115/1.4033548]

1 Introduction design point using an inverse reliability analysis. The performance

measures serve as basis of the probabilistic constraints. Strengths
The existence of uncertainties has driven the need to solve
and weaknesses of both MRIA and PMA have led to the develop-
design problems with considerations of reliability, aside from
ment of a hybrid reliability approach [18,19] by defining a
finding the optimal solutions. This has differentiated RBDO from
selection factor to determine whether PMA or MRIA would be
other design optimization approaches and has been demonstrated
efficient to use.
with modest additional computation costs [1–3]. Many approaches
Most existing RBDO algorithms, such as the aforementioned
have been developed to tackle RBDO problems, which vary in the
methods of MRIA and PMA, require transformation from the
way the problem is perceived. Examples of these approaches are
original design space to the standard normal design space for the
the reliability index approach (RIA) [4–11] and PMA [12–16]. In
processes of reliability analyses. Those methods faced big chal-
RIA, the probabilistic constraints are formulated based on the
lenges when the distribution type of random variable is unknown,
evaluations of reliability indices. This approach, however, had
the information of random distribution is insufficient, or the trans-
been reported to have convergence problems [12] and modifica-
formation to the standard normal design space do not exist or
tions were made by Lin et al. [17] in the MRIA. In MRIA, the def-
cannot be determined easily. A method that does not require trans-
inition of the reliability index had been redefined and had
formation to the standard normal design space is desirable for
resolved convergence problems of the traditional RIA. PMA, on
solving the RBDO problems with arbitrarily distributed uncertain-
the other hand, determines the performance measure of a target
ties in the original design space. Therefore, the design optimiza-
tion problems with arbitrarily distributed uncertainties have drawn
1
Corresponding author. high attentions [20–22] in the past few years. A method called
Contributed by the Design Automation Committee of ASME for publication in
the JOURNAL OF MECHANICAL DESIGN. Manuscript received August 20, 2015; final
EoGRA [21] has been developed to estimate the joint probability
manuscript received May 1, 2016; published online May 18, 2016. Assoc. Editor: density function (JPDF) using KDE [23,24]. EoGRA is capable of
Nam H. Kim. performing the kernel reliability analyses in the original design

Journal of Mechanical Design Copyright V

C 2016 by ASME JULY 2016, Vol. 138 / 071403-1
space and using them to estimate the failure probability without ui  ru gi ðui Þ   r g ðu Þ
x i i
the need of transforming to the standard normal design space. In  þ d  dðkÞ   bf ;i (3)
kru gi ðui Þk kru gi ðui Þk
this paper, the fundamentals of EoGRA are introduced. A new ef-
ficient approach is proposed to enhance the efficiency of EoGRA,
where bf ;i is the ith allowable reliability index.
and a linearly approximated probabilistic constraint is formulated.
Some numerical examples are solved in order to investigate the
numerical performance of the proposed method. 2.2 PMA. In 1999, PMA [12] was developed because the
determination of MPFP in the standard normal design space is

Downloaded from http://asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/138/7/071403/6229138/md_138_07_071403.pdf by National Taiwan University of Science and Technology user on 07 September 2023
costly when the constraint function is far away from the design
2 Fundamentals of RBDO point or the constraint function is highly nonlinear. Especially
when the random variable in the original design space is sup-
A typical RBDO problem is formulated as follows: ported in an interval or a semi-interval, the limit state of the con-
straint function sometimes is mapped to infinity after the
Min f ðdÞ transformation to the standard normal design space. An alternative
(1)
s:t: P½gi ðXÞ > 0  Pf ;i i ¼ 1:::M way was desirable for the situations as described.
In PMA, inverse reliability analysis [28] was utilized for esti-
mation of the target performance of the constraint function at the
where the design variable is the mean d of the random variable X.
tail of random distribution in a distance of bf ;i from the design
This paper considers arbitrarily distributed random variable, that
point. The following subproblem is then solved to determine the
is, X can be mutually dependent and the distribution can be of any
MPTP u#i :
arbitrary shape. In Eq. (1), f is the objective function, and gi is the
ith constraint function. The value of Pf ;i represents the ith allow-
Max gi ðui Þ
able failure probability. The numbers of the constraints and varia- (4)
s:t: kui k ¼ bf ;i
bles are denoted by M and L, respectively.
The most accurate method to evaluate the failure probability of
the arbitrarily distributed random event in Eq. (1) is to use Monte The above equation can be efficiently solved using the hybrid
Carlo simulations (MCS) because the JPDF of X is arbitrary and mean value (HMV) method [29], which selectively chose its
unknown. In this paper, the value of FE represents the number of searching scheme based on the convexity of the constraint func-
the constraint function gðxÞ being evaluated. The evaluation of the tion near the design point. PMA has been widely used because the
ith failure probability in Eq. (1) using MCS equals the number of solution process of HMV is more efficient than the MPFP deter-
sampling points N, which is very costly. mination in MRIA method. Based on the determination of MPTP,
Existing RBDO algorithms often assume that the random varia- the linearly approximated probabilistic constraint is then given by
bles are mutually independent and can be transformed to standard
normal design space. For example, MRIA [17] determines the gi ðu#i Þ þ ðd  dðkÞ Þ  rx gi ðui# Þ  0 (5)
most probable failure point (MPFP) in the standard normal design
space to perform reliability analysis while PMA [12] finds the Both MRIA and PMA methods need the transformation Ti from
most probable target point (MPTP) to conduct inverse reliability the original random design space to the standard normal design
analysis. Sections 2.1 and 2.2 review the reliability analyses of space in order to perform the reliability analyses in Eqs. (2) and
some well-known RBDO algorithms that transform the random (4), respectively. When the distribution type of random variable is
design space to the standard normal design space. The RBDO unknown, the information of random distribution is insufficient,
method that does not require transformation to the standard nor- or the transformation to the standard normal design space do not
mal design space and is capable of performing reliability analysis exist or cannot be determined easily, the existing RBDO methods
in the original design space will be introduced in Sec. 3. can approximate the random variable as some known random dis-
tributions but the accuracy of the determined optimal solution
may be poor. A method that does not require transformation to the
2.1 MRIA. Cornell [25] used the measure of the negative standard normal design space is desirable for design optimization
expected value of the constraint function over its standard devia- problems with arbitrarily distributed uncertainties.
tion as the reliability index. A mean-value first-order second-
moment method [26] evaluated the Cornell reliability index by 3 Introduction to EoGRA
negative constraint function value over the length of constraint
gradient evaluated at the design point in the standard normal The existing RBDO methods that require transformation to the
design space. Hasofer and Lind [27] further evaluated the reliabil- standard normal design space have been reviewed in Sec. 2. These
ity index at the MPFP to provide a better estimation of the tail of reliability analyses in the standard normal design space do not
random distribution. Lin et al. [17] computed the reliability index work for the arbitrarily distributed uncertainties because the trans-
as the vector of MPFP projected on the constraint gradient direc- formation to the normal design space is unknown or difficult to
tion evaluated at the MPFP. A proper measurement of reliability determine. This section introduces the EoGRA method to estimate
index is then determined, i.e., a positive value is obtained when the failure probability in the original design space and solve
the design point is feasible and a negative value for an infeasible RBDO problems with arbitrarily distributed uncertainties. In Sec.
design point. Accordingly, MRIA was developed. 3.1, the estimation of failure probability using the ensemble of
After transforming the design variable to ith standard normal multiple Gaussian kernel reliability analyses is derived. Further-
design variable ui using the transformation of ui ¼ Ti ðxÞ, MRIA more, an efficient gradient approximation is proposed in Sec. 3.2
determines the ith MPFP ui by solving the following subproblem: to greatly reduce the required FEs in EoGRA and utilized to for-
mulate the linearly approximated probabilistic constraint for the
Min kui k optimization processes, which will be shown in Sec. 3.3.
(2)
s:t: gi ðui Þ ¼ 0
3.1 Failure Probability Evaluated by the Ensemble of
The transformations Ti from various types of random variables Multiple Gaussian Kernel Reliability Analyses. Lin and Lin
to the standard normal design space have been introduced in Ref. [21] first used multivariate KDE to approximate the JPDF with a
[13]. The linearly approximated probabilistic constraint of MRIA nonparameterization approach and initiated the concept of
is then formulated based on the evaluation at ui , as shown in the EoGRA. The detailed derivation of EoGRA will be provided in
following equation: this section.

071403-2 / Vol. 138, JULY 2016 Transactions of the ASME

 A single-variate KDE [23,24] can generally be defined as   1X N
 
follows: P gi ðXÞ > 0 ¼ Fp bp (12)
1X N N p¼1
qX ð xÞ ffi K p ð xÞ (6)
N p¼1
However, the exact calculation of Eq. (12) requires N kernel reli-
where qX is often utilized as an estimation of the probability den- ability analyses, which is extremely expensive and impossible for
sity function (PDF) of a series of independent and identically dis- practical tests. A new approach is desirable to reduce the required

Downloaded from http://asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/138/7/071403/6229138/md_138_07_071403.pdf by National Taiwan University of Science and Technology user on 07 September 2023
tributed sampling points x1 ; :::; xN . Duong and Hazelton [30] computations in the current form of EoGRA.
showed that the KDE is capable of modeling the JPDF qX of inde-
pendent and identically distributed multivariate sampling points
x1 ; :::; xN with nonzero mutual correlations between each variable, 3.2 An Efficient Approach to Estimate the Gradient
as shown in the following equation: Direction at the MPFP for Each Kernel Reliability Analysis.
The kernel reliability analysis in Eq. (10) and the gradient evalua-
1X N tion at the pth MPFP in Eq. (11) require some evaluations of the
qX ðxÞ ffi Kp ðxÞ (7) constraint function. The total FEs accumulate quickly if the sub-
N p¼1
problem in Eq. (10) were solved for each kernel analysis. There-
fore, an efficient approach to complete the required kernel
where Kp is the pth Gaussian kernel function and is given by reliability analyses is desirable for RBDO in the original design
  space.
1 kx  sp k2 This paper first determines the ith MPFP xi , which has the min-
Kp ðxÞ ¼ L exp  (8) imum distance from the design point d to the ith constraint func-
ð2pÞ2 r 2r2
tion. The following subproblem is solved to determine xi :
It is noted that the definition in Eq. (7) is general for mutually cor-
Min kx  dk
related variables (e.g., the j1 th variable xj1 and the j2 th variable xj2 x (13)
can be correlated) of random sampling points (e.g., the samplings s:t: gi ðxÞ ¼ 0
of the p1 th point xp1 and the p2 th point xp2 are random). Because
Kp in Eq. (8) is a Gaussian function centered at the pth sampling A unit gradient vector vi is then computed at xi and defined as
point sp with a shape parameter r, the pth failure probability of vi  rx gi ðxi Þkrx gi ðxi Þk1 . This paper further assumes that the
the kernel function can be easily computed by the following cu- gradient direction for each kernel reliability analysis is close to
mulative distribution function (CDF): the unit gradient vector vi . This assumption is based on linear tail
ðx approximation [26] at the limit state (e.g., same concept of reli-
ability analysis at the MPFP in the methods of RIA [17] and uni-
Fp ðxÞ ¼ Kp ðxÞdx (9) fied reliability formulation [31]), which can effectively enhance
1
the solution efficiency. It is noted that the tail approximation has
The MPFP for the pth kernel reliability analysis, as shown in lower accuracy when the tail of the random distribution does not
Fig. 1, can be determined by solving the following subproblem in exist (e.g., normal distribution). Therefore, Eq. (11) can be rewrit-
the original design space: ten as follows:
  
xp  sp  vi
bp ffi (14)
Min kx  sp k r
x (10)
s:t: gi ðxÞ ¼ 0
We assume that the gradient direction at xp is very close to the
one at xi . Therefore, the vector xp  sp is very close to
The optimal solution of Eq. (10), xp , is the pth MPFP and the reli-
kxp  sp kvi . The following equation can then be determined:
ability index [17] of the pth reliability analysis can be determined
in terms of the MPFP, as shown below: ðxp  sp Þ  vi ¼ kxp  sp k ¼ ðxi  sp Þ  vi (15)
    
xp  sp  rx gi xp
bp ¼   (11) Finally, the kernel reliability index can be estimated by Eq. (16)
rkrx gi xp k to avoid the N evaluations of the subproblem in Eq. (10). The for-
mulation of the pth reliability analysis is given by
Therefore, the total failure probability in Eq. (1) can be deter-
mined using the ensemble of the failure probabilities evaluated ðxi  sp Þ  vi
bp ffi (16)
from each kernel reliability analysis, as given by the following r
equation:
The estimation in Eq. (16) is exact for linear constraint functions
and is quite effective, based on our implementations in Sec. 4, for
most problems. The accuracy of the estimation reduces as the non-
linearity of the constraint function dramatically increases.
Figure 2 provides a two-dimensional illustration of the estima-
tion of the xp for determining the reliability index in Eq. (16) for
each kernel reliability analysis in Eq. (12). It is noted that the deri-
vations in Sec. 3.1 and the efficient approach in Sec. 3.2 are still
valid for multidimensional problems.

3.3 Formulation of Linearly Approximated Probabilistic

Constraint in EoGRA. Applying the aforementioned reliability
Fig. 1 Determination of MPFP of each kernel analysis in the analyses of EoGRA, the RBDO problem in Eq. (1) can be rewrit-
original design space ten as follows:

Journal of Mechanical Design JULY 2016, Vol. 138 / 071403-3

 most cases) or there is no problem in finding a best-fitted mathe-
matical model (e.g., Weibull distribution [32] and various types of
copula functions [33]) to describe the random distributions. This
paper proposes the EoGRA to approximate the random distribu-
tions that cannot be modeled by any known methods and provide
proper reliability analyses. In this section, the EoGRA is used to
solve three different numerical examples for the investigation of

Downloaded from http://asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/138/7/071403/6229138/md_138_07_071403.pdf by National Taiwan University of Science and Technology user on 07 September 2023
the numerical performances of proposed method. Several kinds of
random distributions are studied, including a normal distribution,
a uniform distribution, and some artificially generated distribu-
tions, to demonstrate the capability of performing reliability anal-
yses with distributions in multiple shapes and supported in various
ways in the original design space. Section 4.1 introduces the
Fig. 2 Estimation of MPFP of each kernel reliability analysis in procedures of generating four different kinds of distributions,
the original design space followed by three benchmark mathematical examples that are
shown in Sec. 4.2. In each example, it is assumed that certain
Min f ðdÞ amount of sampling points is provided. The methods of selecting
1X N
  (17) the sampling points are not the main focus of this paper.
s:t: Fi;p bi;p  Pf ;i i ¼ 1:::M
N p¼1
4.1 Generations of Arbitrarily Distributed Random
Variables. There exist many real-world design problems that con-
where Fi;p is the CDF of the pth kernel reliability analysis for the tain nonstandard random distributions. Some good examples
ith constraint, and bi;p is the corresponding reliability index. This include image processing [34], signal processing [35], anisotropic
paper linearly expands the probabilistic constraint in Eq. (17) and material properties [36], etc. In this paper, the numerical examples
formulates a linearly approximated probabilistic constraint in consider the mean value of the random variable as the design
Eq. (18) for the solution processes in the proposed method point. Four distributions are artificially generated to demonstrate
concave distributions, distributions that are partially supported in
X
N X
N
Fp ðbðkÞ ðkÞ
rx Fp ðbpðkÞ Þ  NPf ;i the entire domain and partially supported in intervals, and distri-
p Þ þ ðd  d Þ  (18)
p¼1 p¼1
butions with void regions inside. This paper assumes that the cer-
tain amount of sampling points of the arbitrarily distributed
where the superscript ðkÞ represents the information obtained at random variable is known; in fact, the locations of the sampling
the kth iteration. The sensitivity term rx Fp in Eq. (18) is esti- points are assumed to be the only known information about the
mated by finite difference method. The perturbation of finite dif- random distribution. The estimation of the JPDF will be poor for
ference was chosen for acceptable accuracy. The total required the situation of extremely low sampling points; thus, the proposed
FEs of the ith constraint per iteration equal to ðL þ 1Þ [required method cannot perform properly for extremely low sampling
evaluations of determination of the MPFP in Eq. (13)þðL þ 1Þ], points of the random variables.
where the ðL þ 1Þ premultiplier is for the sensitivity analysis in Table 1 lists the procedures of generating four different distri-
Eq. (18), and the summing term of ðL þ 1Þ inside the bracket is butions: A heart-shaped distribution is generated by merging two
for the evaluation of the gradient vi . rotated Gaussian distributions; a “like”-shaped distribution is gen-
erated by merging four Gaussian distributions and one uniform
distribution; a star-shaped distribution is generated by cropping a
4 Numerical Examples uniform distribution into the assigned star shape; and a corona-
For most common engineering problems, the random variables shaped distribution is generated by removing a circular region
are distributed in a known shape (e.g., normally distributed for inside a normal distribution. These four distributions are used to
Table 1 Procedures for distribution generations

Distribution Description/generation procedure

Normal ½r1 ; r2  ¼ ½0:1; 0:3

Uniform X1 2 ½d1  0:25; d1 þ 0:25, X2 2 ½d2  0:5; d2 þ 0:5
Heart-shaped (1) Generate a normal distribution set of N=2 points with ½l1 ; l2  ¼ ½0; 0 and ½r1 ; r2  ¼ ½0:12; 0:2
(2) Generate another normal distribution set of N=2 points with ½l1 ; l2  ¼ ½0:5; 0:5 and ½r1 ; r2  ¼ ½0:2; 0:12
(3) Rotate all sampling points around the origin by 135 deg
Like-shaped (1) Generate a normal distribution set of N=5 points with ½l1 ; l2  ¼ ½0; 0 and ½r1 ; r2  ¼ ½0:15; 0:05
(2) Generate a second normal distribution set of N=5 points with ½l1 ; l2  ¼ ½0; 0:3 and ½r1 ; r2  ¼ ½0:15; 0:05
(3) Generate a third normal distribution set of N=5 points with ½l1 ; l2  ¼ ½0; 0:6 and ½r1 ; r2  ¼ ½0:15; 0:05
(4) Generate a fourth normal distribution set of N=5 points with ½l1 ; l2  ¼ ½0:5; 0:8 and ½r1 ; r2  ¼ ½0:05; 0:15
(5) Generate a uniform distribution set of N=5 points in the x-interval of [0, 0.6] and the y-interval of [0.1, 0.7]
Star-shaped (1) Keep generating uniformly distributed points in the x-interval of [0, 2] and the y-interval of [0, 2](2)
(2) Ignore the generated point if it is located inside the following unit-radius circles:
(a) Center at ½1:6 cos 0:3p þ 1; 1:6 sin 0:3p þ 1
(b) Center at ½1:6 cos 0:7p þ 1; 1:6 sin 0:7p þ 1
(c) Center at ½1:6 cos 1:1p þ 1; 1:6 sin 1:1p þ 1
(d) Center at ½1:6 cos 1:5p þ 1; 1:6 sin 1:5p þ 1
(e) Center at ½1:6 cos 1:9p þ 1; 1:6 sin 1:9p þ 1
(3) Stop the procedure until N points are generated
Corona-shaped (1) Keep generating normally distributed points with ½l1 ; l2  ¼ ½0; 0 and ½r1 ; r2  ¼ ½0:25; 0:25
(2) Ignore the generated point if it is located inside the circle centered at ½0; 0 with a radius of 0.5
(3) Stop the procedure until N points are generated

071403-4 / Vol. 138, JULY 2016 Transactions of the ASME

 Table 2 Problem descriptions of the numerical examples

Ex. Problem descriptions

1 Min f ¼ d1 þ d2
d
s:t: P½g1 ¼ X1  2X2 þ 10 > 0  Pf

Downloaded from http://asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/138/7/071403/6229138/md_138_07_071403.pdf by National Taiwan University of Science and Technology user on 07 September 2023
P½g2 ¼ 2X1  X2 þ 10 > 0  Pf
0:1  d1 ; d2  10

2 Min f ¼d1 þd2

d
   
s:t: P g1 ¼1 X12 X2 =20>0 Pf
2 3
ðX1 þX2 5Þ2 ðX1 X2 12Þ2
P4g2 ¼1  >05 Pf
30 120

0:1d1 ; d2 10

3 ðd1 þ d2  10Þ2 ðd1  d2 þ 10Þ2

Min f ¼ 
d 30 120
  2  
s:t: P g1 ¼ 1  X1 X2 =20 > 0  Pf Fig. 3 Generated random distributions: (a) a heart-shaped dis-
  tribution, (b) a like-shaped distribution, (c) a star-shaped distri-
P g2 ¼ 1 þ Y12 þ Y13  0:6Y14  Y2 > 0  Pf bution, and (d) a corona-shaped distribution. (Small dots
represent the sampling points; the circle at the center repre-
0:1  d1 ; d2  10 sents the mean value of randomly distributed points; and
" # " #" # " #
Y1 0:9063 0:4226 X1 6 dashed ellipses represent the ranges of 1r, 2r, and 3r if treating
where ¼ þ the random points as normal distributions).
Y2 0:4226 0:9063 X2 0
analyses start from the second iteration. The distributions listed in
Table 1 are investigated. Three RBDO algorithms are used to
solve the problems in Table 2: MRIA, PMA, and EoGRA.
Because MRIA and PMA cannot transform the generated distribu-
show a concave distribution supported in the entire design space, tions from the original design space to the standard normal space,
a combined distribution that is partially supported in the entire do- all non-normal distributions are assumed to be mutually independ-
main and partially supported in a semi-interval, a uniform distri- ent and normally distributed with the computed standard devia-
bution in a concave region, and a distribution supported in the tions. EoGRA is used to solve the optimization problems in the
entire design space with a void region at the center, respectively. original design space using two various amounts of sampling
Figure 3 shows the aforementioned distributions with the mean points, N ¼ 500 and 50; 000.
points located at the origin of the coordinate. For RBDO methods In Eq. (12), the estimation of the failure probability of the ith
that require transformation from the original design space to the constraint is computed in terms of N CDF evaluations of
standard normal design space, the generated distributions are to be Fp ðbp Þ. Equation (16) efficiently estimates the bp in terms of the
approximated as mutually independent normal distributions with location of pth sampling point, the gradient vector vi , and the
the measurements of standard deviations. The elliptic contours in location of the MPFP xi . The determination of xi is the only step
Fig. 3 represent the 3r ranges of the approximated normal distri- that requires evaluation of the constraint function. Therefore, the
butions. The computed standard deviations of the heart-shaped required FE of EoGRA can be much fewer than N. Larger N will
distribution are [0.3896, 0.1648]; the ones for the like-shaped dis- improve the estimation of the failure probability but will not
tribution are [0.2493, 0.3045]; the ones for the star-shaped distri- increase the FE.
bution are [0.3596, 0.3606]; and the ones for the corona-shaped
distribution are [0.4329, 0.4337].

4.2 Descriptions of the Numerical Examples. Table 2 lists

the problem descriptions of the benchmark examples. The first
example is a linear mathematical problem [12,17]. The second
example is a nonlinear mathematical problem, which was origi-
nally a three-constraint problem in the literature [12,17,37–41]. In
this paper, the third constraint is removed because it is inactive.
The third example is a highly nonlinear mathematical problem
[16,19,42]. Similarly, the third constraint that was originally con-
sidered in the third example is removed in this paper because it is
inactive.
In our implementations, two various levels of allowable failure
probabilities, Pf ¼ 1% and 30%, are studied. In all examples, the
initial design variables are [5, 5] and deterministic optimizations
are used in the first iterations for fair comparisons. The initial Fig. 4 Iteration path of solving example 2 with corona-shaped
points have 100% reliability in all examples. The probabilistic distribution and Pf 5 30% using EoGRA with N 5 50; 000

Journal of Mechanical Design JULY 2016, Vol. 138 / 071403-5

 allowable failure probabilities of 1% and 30%, respectively. Figure
10 shows the optimal solutions of example 3 with allowable failure
probability of 1%. Only the results for normal and uniform distribu-
tions are shown because EoGRA cannot find proper solutions of the
highly nonlinear mathematical problem of example 3 when the
allowable failure probability is low and the random variable is of a
complex distribution type. Furthermore, EoGRA is able to find

Downloaded from http://asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/138/7/071403/6229138/md_138_07_071403.pdf by National Taiwan University of Science and Technology user on 07 September 2023
proper solutions of example 3 when the allowable failure probabil-
ity is 30%, as shown in Fig. 11.
Tables 3–5 show the detailed information about the optimiza-
tion performances of MRIA, PMA, and EoGRA. Besides the true
solutions determined by MCS, the optimal solutions were deter-
mined as the convergence criterion satisfied: kdðkÞ  dðk1Þ k
 103 . “Type” represents the distribution type in each test;
“Iter.” stands for the required iteration number for convergence;
Fig. 5 Iteration path of solving example 3 with like-shaped dis- the obtained optimal solution is denoted by d and the optimal
tribution and Pf 5 30% using EoGRA with N 5 50; 000 function value is denoted by f  ; and “MCS” represents the failure
probabilities evaluated at the determined optimal solutions using
4.3 Numerical Results and Discussion. In order to determine MCS with 105 sampling points. For example, [1.001, 0.993] indi-
the true reliability of each design problem, MCS with 105 sampling cates that the failure probabilities of the first and second con-
points are used to directly evaluate the failure probabilities of the straints are 1.001% and 0.993%, respectively, while MRIA was
solutions. Figures 4 and 5 show some typical iteration paths of solv- used to solve example 1 with normally distributed random varia-
ing the examples 2 and 3, respectively, using the proposed method. bles and Pf ¼ 1%. Furthermore, “FE” stands for the required eval-
Figures 6–11 show the optimal solutions (annotated by the circle uation number of the constraint functions for convergence.
surrounded by the sampling points) of the numerical examples To further verify the correctness of the determined optimal sol-
using EoGRA with 50,000 sampling points (annotated by the the utions and the numerical performances of the optimization proc-
small points). In the above figures, the lines/curves that are further esses, the “true” solutions were determined by a large amount of
away from the optimal solutions are the original constraint func- sampling points and shown in the Tables 3–5. An “error” mea-
tions while the lines that are closer to the optimal solutions repre- surement was computed by the following normalized difference in
sent the linearly approximated probabilistic constraints. The black order to represent the accuracy of the determined optimal solu-
points are the design points and the blue points are the sampling tions d
points. Figure 6 shows the optimal solutions of example 1 with vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
!
u L
allowable failure probability of 1%. Figure 7 shows the optimal sol- uX dj  djtrue 2
Error  t 100% (19)
utions of the same example with allowable failure probability of djU  djL
30%. Figures 8 and 9 show the optimal solutions of example 2 with j¼1

Fig. 6 Using EoGRA with N 5 50; 000 to solve example 1 with Pf 5 1% and (a) normal, (b) uniform, (c) heart-shaped, (d) like-
shaped, (e) star-shaped, and (f) corona-shaped distributions

071403-6 / Vol. 138, JULY 2016 Transactions of the ASME

 Downloaded from http://asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/138/7/071403/6229138/md_138_07_071403.pdf by National Taiwan University of Science and Technology user on 07 September 2023
Fig. 7 Using EoGRA with N 5 50; 000 to solve example 1 with Pf 5 30% and (a) normal, (b) uniform, (c) heart-shaped, (d) like-
shaped, (e) star-shaped, and (f) corona-shaped distributions

Fig. 8 Using EoGRA with N 5 50; 000 to solve example 2 with Pf 5 1% and (a) normal, (b) uniform, (c) heart-shaped, (d) like-
shaped, (e) star-shaped, and (f) corona-shaped distributions

Journal of Mechanical Design JULY 2016, Vol. 138 / 071403-7

 Downloaded from http://asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/138/7/071403/6229138/md_138_07_071403.pdf by National Taiwan University of Science and Technology user on 07 September 2023
Fig. 9 Using EoGRA with N 5 50; 000 to solve example 2 with Pf 5 30% and (a) normal, (b) uniform, (c) heart-shaped, (d) like-
shaped, (e) star-shaped, and (f) corona-shaped distributions

where dtrue is the true solution determined by the large amount of 1 for a >¼ 0
HðaÞ ¼ (21)
sampling points; diU and diL are the upper and lower bounds, 0 otherwise
respectively; and L is the number of variables. The sum of viola-
tions (SOV) of the probabilistic constraints is computed by the Table 3 lists the solutions of the linear problem in example 1.
following equation: All methods were able to find proper solutions. Both MRIA and
PMA utilized fewer iterations and FEs than EoGRA. This paper
X
M
uses the MCS results to represent the accuracy of RBDO methods.
SOV ¼ ½HðPMCS;i  Pf ;i Þ ðPMCS;i  Pf ;i Þ (20) The accuracies of MRIA and PMA dropped for non-normal distri-
i¼1
butions due to the poor approximations of mutually independent
normal distributions. It is possible to improve the accuracy of the
where PMCS;i is the failure probability evaluated by MSC, and Pf ;i reliability analyses of uniform distributions because there exists a
is the allowable failure probability. In Eq. (20), H is the Heaviside nonlinear transformation from the uniform design space to the
function and it is given by standard normal design space [13]. However, the transformations

Fig. 10 Using EoGRA with N 5 50; 000 to solve example 3 with Pf 5 1% and (a) normal and (b)
uniform distributions

071403-8 / Vol. 138, JULY 2016 Transactions of the ASME

 Downloaded from http://asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/138/7/071403/6229138/md_138_07_071403.pdf by National Taiwan University of Science and Technology user on 07 September 2023
Fig. 11 Using EoGRA with N 5 50; 000 to solve example 3 with Pf 5 30% and (a) normal, (b) uniform, (c) heart-shaped, (d) like-
shaped, (e) star-shaped, and (f) corona-shaped distributions

from the generated distributions to the standard normal design EoGRA while the solutions of non-normal distributions were too
space are difficult to determine. On the other hand, EoGRA is ca- conservative. The results showed that the proposed method still
pable of finding the solutions for all tested distributions and allow- can determine the optimal solutions with acceptable accuracy
able failure probabilities because it performs the reliability (i.e., errors of less than 2% for normal, heart-shaped, like-shaped,
analyses in Eq. (12) in the original design space directly. Com- and star-shaped distributions) and good reliability (i.e., 0% of
pared with the traditional methods, the proposed method per- SOV for all cases) for conditions of Pf ¼ 30% and N ¼ 50; 000.
formed more accurately (i.e., lower values of error in most cases) Due to errors of linear approximations for the highly nonlinear
with greater reliability (i.e., lower values of SOV in most cases). constraints, the proposed method is slightly conservative for the
As mentioned previously, less accurate performances were found uniform and corona-shaped distributions but is still reliable (i.e.,
for the uniform distributions because of the bigger errors in tail errors of 5–6% with 0% SOV).
approximations in Eq. (14). Because the approximation of kernel reliability analyses in
Table 4 shows the solutions of the nonlinear problem in exam- Sec. 3.2 reduces the accuracy of probability estimation, EoGRA
ple 2. Only one case cannot find the solution properly. In this spe- requires more iterations than MRIA and PMA to find the correct
cific case of allowable failure probability of 1% and normal solutions in simpler problems like examples 1 and 2 with normally
distribution, PMA could not evaluate the failure probability prop- distributed random variables. The advantage of using the approach
erly. At the second iteration during the optimization process, the in Sec. 3.2 is that the required FE will not depend on the number
second MPTP was located outside the feasible design space lead- of sampling points. If a huge amount of sampling points is avail-
ing to improper evaluation of failure probability. Besides this spe- able, the efficiency of the proposed method can be improved.
cial case, all other cases were able to find a solution properly. Another method to reduce the required FEs is to use surrogate
Similarly, both MRIA and PMA utilized fewer iterations and FEs models with 5L uniformly distributed sampling points [43].
than EoGRA but had less accuracy due to poor approximations of The results also show that the developed EoGRA method is ca-
the non-normal distributions. The overall performances of pable of solving design optimization problems in the original
EoGRA were stable as the solutions were properly determined design space without transformation to the standard normal design
with better accuracy and reliability. Similarly, less accurate per- space. The ensemble of reliability analyses in the estimation of
formances were found for the uniform distributions because of the JPDF is suitable for any arbitrarily shaped random distribution.
bigger errors in tail approximations in Eq. (14). However, EoGRA is not perfect. The approximation of gradient
Table 5 shows the solutions of the highly nonlinear example. direction at the MPFP of each kernel reliability analysis and the
Because EoGRA has weaker approximation of the tail of random linear expansion of the probabilistic constraint accumulate errors
distribution when the fewer sampling points are used, it cannot in the overall reliability analysis when the constraint function is
perform proper reliability analyses for the second constraint that nonlinear or the distribution has a complex shape.
is highly nonlinear. The developed EoGRA can only find accepta-
ble solutions when there are more sampling points at the tail of
distribution for the cases with higher allowable failure probabil- 5 Practical Examples in Image Brightening
ities. The estimation of constraint gradient and the linear approxi- Graphics engineers have developed many mathematical meth-
mation of probabilistic constraint also affected the accuracy of ods [44] to enhance the quality of photos or video frames.

Journal of Mechanical Design JULY 2016, Vol. 138 / 071403-9

One effective method is to (1) transform the image pixels in the Max d1
red, green, and blue (RGB) color space to the CIELAB [44] color s:t: PðX1 > LU Þ  Pf (22)
space and (2) enhance the L components (e.g., lighting) of the where X ¼ TðYÞ
pixel points within some desired design rules. Therefore, an opti-
mization process of lighting enhancement can be defined as In the above equation, Y is the random pixel points in the RGB
follows: color space; T is the transformation to the CIELAB color space

Downloaded from http://asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/138/7/071403/6229138/md_138_07_071403.pdf by National Taiwan University of Science and Technology user on 07 September 2023
Table 3 Solutions of example 1

Type Pf (%) Method N Iter. d f MCS (%) FE Error (%) SOV (%)

Normal 1 True — — [3.4272, 3.9924] 7.420 — — — —

MRIA — 3 [3.4208, 3.9971] 7.418 [1.001, 0.993] 60 <1 <1
PMA — 3 [3.4209, 3.9970] 7.418 [0.935, 0.989] 36 <1 0
EoGRA 500 10 [3.4291, 4.0610] 7.490 [0.512, 0.580] 162 1 0
EoGRA 50,000 10 [3.4408, 3.9963] 7.437 [0.920, 0.742] 162 <1 0
Normal 30 True — — [3.3536, 3.4823] 6.836 — — — —
MRIA — 3 [3.3530, 3.4829] 6.836 [30.046, 30.052] 60 <1 <1
PMA — 3 [3.3529, 3.4831] 6.836 [30.062, 30.031] 36 <1 <1
EoGRA 500 6 [3.3647, 3.4582] 6.823 [32.187, 30.246] 90 <1 2
EoGRA 50,000 5 [3.3569, 3.4856] 6.843 [29.618, 29.129] 72 <1 0
Uniform 1 True — — [4.2396, 3.7402] 7.980 — — — —
MRIA — 3 [3.5049, 3.9397] 7.445 [0, 0.118] 54 8 0
PMA — 3 [3.5049, 3.9398] 7.445 [0, 0.122] 36 8 0
EoGRA 500 9 [3.5455, 3.8015] 7.347 [0.487, 0.545] 144 7 0
EoGRA 50,000 9 [3.5527, 3.8025] 7.355 [0.444, 0.436] 144 7 0
Uniform 30 True — — [3.7203, 3.3649] 7.085 — — — —
MRIA — 3 [3.3720, 3.4700] 6.842 [34.365, 30.864] 54 4 5
PMA — 3 [3.3720, 3.4701] 6.842 [34.341, 30.947] 36 4 5
EoGRA 500 5 [3.3638, 3.5049] 6.869 [31.378, 29.483] 72 4 1
EoGRA 50,000 5 [3.3557, 3.5220] 6.878 [30.037, 29.341] 72 4 <1
Heart-shaped 1 True — — [3.8838, 3.5427] 7.427 — — — —
MRIA — 3 [4.1728, 3.5071] 7.680 [0.102, 0] 60 3 0
PMA — 3 [4.1728, 3.5072] 7.680 [0.139, 0.002] 36 3 0
EoGRA 500 9 [3.9224, 3.5301] 7.453 [0.822, 0.540] 144 <1 0
EoGRA 50,000 9 [3.8992, 3.5500] 7.449 [0.735, 0.678] 144 <1 0
Heart-shaped 30 True — — [3.6470, 3.3373] 6.984 — — — —
MRIA — 3 [3.5225, 3.3725] 6.895 [34.562, 42.209] 60 1 17
PMA — 3 [3.5225, 3.3725] 6.895 [34.733, 42.160] 36 1 17
EoGRA 500 5 [3.6441, 3.3514] 6.996 [28.344, 29.703] 72 <1 0
EoGRA 50,000 5 [3.6444, 3.3402] 6.985 [30.084, 30.484] 72 <1 <1
Like-shaped 1 True — — [3.7268, 3.7666] 7.493 — — — —
MRIA — 3 [3.7291, 3.9008] 7.630 [0.014, 0.365] 60 1 0
PMA — 3 [3.7291, 3.9009] 7.630 [0.011, 0.365] 36 1 0
EoGRA 500 9 [3.7259, 3.7724] 7.498 [0.820, 1.059] 144 <1 <1
EoGRA 50,000 9 [3.7411, 3.7863] 7.527 [0.515, 0.789] 144 <1 0
Like-shaped 30 True — — [3.4705, 3.5010] 6.972 — — — —
MRIA — 3 [3.4225, 3.4612] 6.884 [37.127, 39.419] 60 1 16
PMA — 3 [3.4226, 3.4613] 6.884 [37.101, 39.232] 36 1 16
EoGRA 500 6 [3.4372, 3.5393] 6.977 [28.148, 32.058] 90 1 2
EoGRA 50,000 5 [3.4688, 3.5064] 6.975 [29.687, 30.036] 72 <1 <1
Star-shaped 1 True — — [3.8046, 3.9501] 7.755 — — — —
MRIA — 3 [3.9561, 3.9592] 7.915 [0.309, 0.001] 60 2 0
PMA — 3 [3.9561, 3.9593] 7.915 [0.297, 0.005] 36 2 0
EoGRA 500 9 [3.7984, 3.9038] 7.702 [1.578, 1.313] 144 <1 <1
EoGRA 50,000 9 [3.8094, 3.9571] 7.767 [0.893, 0.813] 144 <1 0
Star-shaped 30 True — — [3.5035, 3.4998] 7.003 — — — —
MRIA — 3 [3.4737, 3.4744] 6.948 [33.078, 32.805] 60 <1 6
PMA — 3 [3.4738, 3.4744] 6.948 [33.181, 32.871] 36 <1 6
EoGRA 500 6 [3.5526, 3.4783] 7.031 [29.370, 27.298] 90 1 0
EoGRA 50,000 5 [3.5015, 3.4993] 7.001 [29.921, 30.063] 72 <1 <1
Corona-shaped 1 True — — [3.8954, 3.8892] 7.785 — — — —
MRIA — 3 [4.0833, 4.0859] 8.169 [0.025, 0.019] 60 3 0
PMA — 3 [4.0833, 4.0859] 8.169 [0.018, 0.028] 36 3 0
EoGRA 500 9 [3.9332, 3.8289] 7.762 [1.649, 0.930] 144 1 <1
EoGRA 50,000 9 [3.8972, 3.8978] 7.795 [0.818, 0.871] 144 <1 0
Corona-shaped 30 True — — [3.5929, 3.5898] 7.183 — — — —
MRIA — 3 [3.5024, 3.5029] 7.005 [37.464, 37.286] 60 1 15
PMA — 3 [3.5025, 3.5029] 7.005 [37.654, 37.550] 36 1 15
EoGRA 500 5 [3.6021, 3.5857] 7.188 [30.105, 29.552] 72 <1 <1
EoGRA 50,000 5 [3.5960, 3.5881] 7.184 [30.043, 30.087] 72 <1 <1

071403-10 / Vol. 138, JULY 2016 Transactions of the ASME

[44]; X is the random pixel points in the CIELAB color space; d pixel distribution is unknown and cannot be transformed to stand-
is the mean point of X; and d1 is the L component of d as X1 is the ard normal space. Therefore, treating X as a normal distribution
L component of X. In this example, the acceptable upper limit of will introduce large errors in accuracy.
lighting LU is 90 and the allowable failure probability Pf is 5%. Figure 12 shows the result of photo enhancement by solving
The distribution of the image pixels in one photo or video frame is Eq. (22). The value of Pnorm represents the failure probability of
completely random and arbitrarily shaped; in other words, the the results as X is treated as normally distributed. The value of

Downloaded from http://asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/138/7/071403/6229138/md_138_07_071403.pdf by National Taiwan University of Science and Technology user on 07 September 2023
Table 4 Solutions of example 2

Type Pf (%) Method N Iter. d f MCS (%) FE Error (%) SOV (%)

Normal 1 True — — [3.1442, 2.8034] 5.948 — — — —

MRIA — 3 [3.1336, 2.8034] 5.937 [1.077, 0.979] 72 <1 <1
PMA — Unable to find a solution near the true solution
EoGRA 500 10 [3.1781, 2.8345] 6.013 [0.525, 1.008] 270 <1 <1
EoGRA 50,000 10 [3.1669, 2.8085] 5.975 [0.661, 1.103] 270 <1 <1
Normal 30 True — — [3.1237, 2.2297] 5.353 — — — —
MRIA — 3 [3.1176, 2.2298] 5.347 [30.676, 29.439] 60 <1 <1
PMA — 3 [3.1150, 2.2334] 5.349 [30.579, 28.652] 48 <1 <1
EoGRA 500 5 [3.1146, 2.2198] 5.311 [32.273, 30.296] 114 <1 2
EoGRA 50,000 5 [3.1198, 2.2312] 5.351 [30.106, 29.317] 114 <1 <1
Uniform 1 True — — [3.6540, 3.0540] 6.708 — — — —
MRIA — 4 [3.1605, 2.8213] 5.9818 [0.105, 0] 96 6 0
PMA — 4 [3.1594, 2.8211] 5.9805 [0.103, 0] 114 6 0
EoGRA 500 9 [3.2095, 2.7069] 5.916 [0.393, 0.981] 240 6 0
EoGRA 50,000 9 [3.2264, 2.7011] 5.928 [0.187, 2.201] 240 6 1
Uniform 30 True — — [3.3038, 2.3999] 5.704 — — — —
MRIA — 3 [3.1212, 2.2345] 5.356 [33.209, 32.288] 60 2 5
PMA — 4 [3.1220, 2.2343] 5.356 [33.032, 32.125] 72 2 5
EoGRA 500 5 [3.1310, 2.2612] 5.392 [29.276, 30.363] 114 2 <1
EoGRA 50,000 5 [3.1170, 2.2660] 5.383 [30.578, 28.687] 114 2 <1
Heart-shaped 1 True — — [3.3663, 2.7208] 6.087 — — — —
MRIA — 5 [3.5685, 2.8604] 6.429 [0.004, 0.363] 137 2 0
PMA — 5 [3.5685, 2.8604] 6.429 [0.003, 0.389] 174 2 0
EoGRA 500 9 [3.3843, 2.6686] 6.053 [1.132, 2.755] 240 1 2
EoGRA 50,000 9 [3.4049, 2.6771] 6.082 [0.787, 2.841] 240 1 2
Heart-shaped 30 True — — [3.2428, 2.3436] 5.586 — — — —
MRIA — 4 [3.1716, 2.2832] 5.455 [42.485, 33.360] 87 1 16
PMA — 4 [3.1717, 2.2832] 5.455 [42.621, 33.211] 81 1 16
EoGRA 500 5 [3.2268, 2.3447] 5.572 [31.971, 28.703] 120 <1 2
EoGRA 50,000 5 [3.2326, 2.3450] 5.578 [31.356, 28.971] 120 <1 1
Like-shaped 1 True — — [3.3056, 2.7957] 6.101 — — — —
MRIA — 4 [3.2644, 2.9764] 6.241 [0.417, 0] 102 2 0
PMA — 4 [3.2650, 2.9761] 6.241 [0.340, 0] 108 2 0
EoGRA 500 10 [3.3954, 2.7237] 6.119 [0.545, 4.392] 270 1 3
EoGRA 50,000 10 [3.3766, 2.7442] 6.121 [0.560, 2.930] 270 1 2
Like-shaped 30 True — — [3.2223, 2.2782] 5.501 — — — —
MRIA — 4 [3.1356, 2.2733] 5.409 [39.892, 23.543] 84 1 10
PMA — 4 [3.1356, 2.2733] 5.409 [39.970, 23.570] 69 1 10
EoGRA 500 6 [3.2124, 2.2932] 5.506 [29.721, 27.988] 150 <1 0
EoGRA 50,000 6 [3.2085, 2.2892] 5.498 [30.591, 27.947] 150 <1 <1
Star-shaped 1 True — — [3.3167, 3.0686] 6.385 — — — —
MRIA — 4 [3.4058, 3.2076] 6.613 [0.011, 0.187] 108 2 0
PMA — 5 [3.4058, 3.2076] 6.613 [0.010, 0.163] 165 2 0
EoGRA 500 10 [3.4793, 2.9852] 6.465 [0.150, 3.743] 294 2 3
EoGRA 50,000 10 [3.4225, 2.9665 6.389 [0.569, 3.664] 294 1 3
Star-shaped 30 True — — [3.2030, 2.3459] 5.549 — — — —
MRIA — 4 [3.1521, 2.3343] 5.486 [34.071, 28.184] 84 1 4
PMA — 4 [3.1521, 2.3342] 5.486 [34.113, 28.072] 60 1 4
EoGRA 500 6 [3.2242, 2.3796] 5.604 [27.259, 28.513] 150 <1 0
EoGRA 50,000 6 [3.1734, 2.3765] 5.550 [30.601, 25.974] 150 <1 <1
Corona-shaped 1 True — — [3.3809, 3.0823] 6.463 — — — —
MRIA — 5 [3.4966, 3.4230] 6.920 [0.017, 0.016] 132 4 0
PMA — 5 [3.4966, 3.4231] 6.920 [0.023, 0.024] 186 4 0
EoGRA 500 9 [3.4694, 2.9846] 6.454 [0.580, 4.795] 264 1 4
EoGRA 50,000 9 [3.4643, 2.9811] 6.445 [0.682, 4.977] 264 1 4
Corona-shaped 30 True — — [3.2497, 2.5239] 5.774 — — — —
MRIA — 4 [3.1621, 2.3884] 5.551 [39.894, 33.328] 84 2 13
PMA — 4 [3.1622, 2.3884] 5.551 [39.797, 33.656] 69 2 13
EoGRA 500 6 [3.1706, 2.5337] 5.704 [34.302, 26.862] 150 1 4
EoGRA 50,000 6 [3.2227, 2.5274] 5.750 [31.587, 28.878] 150 <1 2

Journal of Mechanical Design JULY 2016, Vol. 138 / 071403-11

PEoGRA represents the failure probability of the results as EoGRA distribution, the accuracy of the determined result is poor (i.e.,
is used to solve Eq. (22). It is noted that the pixel distribution of failure probability not closed to 5%). Our implementation shows
the tested photo is completely random and cannot be described by that the proposed method is capable of providing accurate estima-
any known mathematical function. The black circle represents the tion of the JPDF of the random pixel distribution and finding the
mean point of the distribution. Because the estimation of normal optimal solution with acceptable accuracy (i.e., failure probability
distribution cannot properly approximate the tail of the random closed to 5%).

Downloaded from http://asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/138/7/071403/6229138/md_138_07_071403.pdf by National Taiwan University of Science and Technology user on 07 September 2023
Table 5 Solutions of example 3

Type Pf (%) Method N Iter. d f MCS (%) FE Error (%) SOV (%)

Normal 1 True — — [5.0031, 1.4935] 1.112 — — — —

MRIA — Unable to find a solution near the true solution
PMA — Unable to find a solution near the true solution
EoGRA 500 15 [5.1453, 1.5698] 1.176 [0.335, 8.762] 519 2 8
EoGRA 50,000 14 [5.0967, 1.5217] 1.154 [0.617, 5.423] 447 1 4
Normal 30 True — — [5.1474, 0.9123] 1.171 — — — —
MRIA — Unable to find a solution near the true solution
PMA — Unable to find a solution near the true solution
EoGRA 500 6 [5.1177, 0.9265] 1.157 [29.669, 20.729] 153 <1 0
EoGRA 50,000 6 [5.1106, 0.9330] 1.153 [29.162, 18.732] 153 <1 0
Uniform 1 True — — [4.8392, 1.0896] 1.023 — — — —
MRIA — Unable to find a solution near the true solution
PMA — Unable to find a solution near the true solution
EoGRA 500 16 [4.7575, 1.6723] 1.002 [0, 0] 525 6 0
EoGRA 50,000 16 [4.7506, 1.6735] 0.999 [0, 0] 525 6 0
Uniform 30 True — — [4.5212, 1.9202] 0.901 — — — —
MRIA — 3 [3.5048, 1.7943] 0.406 [34.135, 0] 96 10 4
PMA — Unable to find a solution near the true solution
EoGRA 500 10 [4.6867, 1.2393] 0.954 [23.066, 19.008] 321 7 0
EoGRA 50,000 11 [4.6316, 1.2991] 0.929 [20.67, 15.266] 357 6 0
Heart-shaped 1 True — — [4.6345, 1.4007] 0.935 — — — —
MRIA — 5 [4.5065, 1.6357] 0.884 [0.035, 0] 155 3 0
PMA — Unable to find a solution near the true solution
EoGRA 500 Unable to find a solution near the true solution
EoGRA 50,000 Unable to find a solution near the true solution
Heart-shaped 30 True — — [4.8718, 0.9818] 1.035 — — — —
MRIA — 4 [4.9750, 0.9187] 1.084 [34.679, 39.998] 87 1 15
PMA — 4 [4.9731, 0.9195] 1.084 [34.522, 39.958] 99 1 15
EoGRA 500 9 [4.8327, 1.0514] 1.018 [21.179, 25.053] 285 1 0
EoGRA 50,000 8 [4.8012, 1.0636] 1.003 [21.394, 21.006] 249 1 0
Like-shaped 1 True — — [4.8041, 1.4617] 1.019 — — — —
MRIA — 5 [4.7051, 1.6504] 0.977 [0.018, 0] 144 2 0
PMA — Unable to find a solution near the true solution
EoGRA 500 Unable to find a solution near the true solution
EoGRA 50,000 Unable to find a solution near the true solution
Like-shaped 30 True — — [5.0926, 0.9855] 1.146 — — — —
MRIA — 4 [5.0468, 0.9501] 1.122 [36.061, 27.797] 90 1 6
PMA — 4 [5.0467, 0.9501] 1.122 [36.115, 27.629] 126 1 6
EoGRA 500 7 [5.0034, 1.0488] 1.103 [27.208, 20.230] 183 1 0
EoGRA 50,000 7 [4.9795, 1.0681] 1.092 [26.444, 17.411] 183 1 0
Star-shaped 1 True — — [4.7011, 1.7237] 0.977 — — — —
MRIA — 5 [4.5762, 1.8835] 0.925 [0.343, 0] 157 2
PMA — Unable to find a solution near the true solution
EoGRA 500 Unable to find a solution near the true solution
EoGRA 50,000 Unable to find a solution near the true solution
Star-shaped 30 True — — [5.0004, 1.0469] 1.102 — — — —
MRIA — 4 [4.9873, 1.0030] 1.094 [33.926, 29.460] 96 <1 4
PMA — 4 [4.9868, 1.0033] 1.094 [33.776, 29.206] 132 <1 4
EoGRA 500 8 [4.8465, 1.1309] 1.028 [27.771, 18.286] 237 2 0
EoGRA 50,000 7 [4.8762, 1.1350] 1.043 [26.616, 20.471] 207 2 0
Corona-shaped 1 True — — [4.6406, 1.7535] 0.951 — — — —
MRIA — 5 [4.5347, 2.1049] 0.911 [0.026, 0.011] 161 4 0
PMA — Unable to find a solution near the true solution
EoGRA 500 Unable to find a solution near the true solution
EoGRA 50,000 Unable to find a solution near the true solution
Corona-shaped 30 True — — [5.0190, 1.1805] 1.114 — — — —
MRIA — 4 [4.9489, 1.0566] 1.076 [39.118, 29.783] 96 1 9
PMA — 4 [4.9477, 1.0572] 1.076 [39.120, 29.818] 144 1 9
EoGRA 500 10 [4.6033, 1.5526] 0.927 [10.297, 1.931] 321 6 0
EoGRA 50,000 9 [4.6420, 1.5075] 0.943 [13.247, 3.609] 285 5 0

071403-12 / Vol. 138, JULY 2016 Transactions of the ASME

 Downloaded from http://asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/138/7/071403/6229138/md_138_07_071403.pdf by National Taiwan University of Science and Technology user on 07 September 2023
Fig. 12 Results of photo enhancements: (a) original photo, (b) pixel distribution in CIELAB, (c) enhanced photo by treating
image pixels as normally distributed (Pnorm 5 0.2811%), and (d) enhanced photo by EoGRA (PEoGRA 5 4.9485%)

6 Conclusions proposed method also showed acceptable performances and accu-

racies in the second example, which has nonlinear constraints.
This paper has reviewed two well-known RBDO methods that
However, EoGRA was known to have weaker estimations of the
transform the original random design space to the standard normal
tails of random distributions when fewer sampling points were
space to perform reliability analyses: MRIA and PMA. Our imple-
used or the allowable reliability level is very high. In the end,
mentations have shown that the existing methods cannot trans-
EoGRA was not able to solve the third highly nonlinear example
form arbitrary distributions to the standard normal space and can
with some complex distributions unless a large amount of
only approximate it as normal distributions. In the end, poor
sampling points were used. Using the proposed estimations of
approximations led to inaccurate estimations of failure probabil-
constraint gradients, no additional FEs are necessary for the reli-
ities. On the other hand, the method of EoGRA was introduced
ability analyses with greater amounts of sampling points.
with an efficient approach of estimating the constraint gradient.
A series of real photos were also tested in this paper. The pixel
EoGRA performs multiple kernel reliability analyses in the origi-
distribution of each photo was completely random and arbitrarily
nal design space based on the approximated gradient information
shaped. The estimation of normal distribution could not provide
for the estimations of failure probabilities. Linearly approximated
accurate approximation of the tail of the arbitrarily shaped pixel
probabilistic constraints were then formulated based on the en-
distributions. The accuracy of the results based on normal estima-
semble of kernel analyses and utilized for the optimization
tion was poor. On the other hand, the proposed method was able
processes.
to provide proper estimation of the pixel distribution using
Three benchmark examples were studied with various types of
Gaussian-based kernel density functions. The lighting of the
random variables. Besides a normal distribution and a uniform
image pixels was then maximized subject to the allowable proba-
distribution, some complex distributions were artificially gener-
bility of over enhancement. The experimental results showed that
ated to represent concavely shaped random variables and distribu-
EoGRA is capable of finding the optimal solution with acceptable
tions supported in various ways. The results showed that both
reliability.
MRIA and PMA can only perform proper reliability analyses for
normally distributed random variables and have poor performan-
ces when approximating the complex random variables as normal Acknowledgment
distributions. EoGRA provided accurate reliability analyses for The support from Ministry of Science and Technology, Taiwan
any kinds of random distributions with different amounts of sam- (Grant Nos. MOST 104-2218-E-033-013 and MOST 103-2221-E-
pling points when the constraint functions were linear. The 033-015) was greatly appreciated. Furthermore, the supports from

Journal of Mechanical Design JULY 2016, Vol. 138 / 071403-13

Center for Robotics Research, Research Center for Microsystem Distributed Design Uncertainties,” 11th World Congress of Structural and
Engineering, and Center for Biomedical Technology at Chung Multidisciplinary Optimization, WCSMO-11, Sydney, Australia, June 7–12,
Paper No. 1211.
Yuan Christian University (CYCU), Taiwan were also appreciated. [23] Rosenblatt, M., 1956, “Remarks on Some Nonparametric Estimates of a Den-
sity Function,” Ann. Math. Stat., 27(3), pp. 832–837.
[24] Parzen, E., 1962, “On Estimation of a Probability Density Function and Mode,”
References Ann. Math. Stat., 33(3), pp. 1065–1076.
[1] Chen, X., Hasselman, T. K., and Neill, D. J., 1997, “Reliability Based Structural [25] Cornell, C. A., 1969, “A Probability-Based Structural Code,” J. Am. Concr.
Design Optimization for Practical Applications,” AIAA Paper No. 97-1403. Inst., 66(12), pp. 974–985.

Downloaded from http://asmedigitalcollection.asme.org/mechanicaldesign/article-pdf/138/7/071403/6229138/md_138_07_071403.pdf by National Taiwan University of Science and Technology user on 07 September 2023
[2] Youn, B. D., Choi, K., Yang, R.-J., and Gu, L., 2004, “Reliability-Based Design [26] Madsen, H. O., Krenk, S., and Lind, N. C., 1986, Methods of Structural Safety,
Optimization for Crashworthiness of Vehicle Side Impact,” Struct. Multidiscip. Prentice-Hall, Englewood Cliffs, NJ.
Optim., 26(3–4), pp. 272–283. [27] Hasofer, A. M., and Lind, N. C., 1974, “Exact and Invariant Second Moment
[3] Gu, L., and Yang, R., 2003, “Recent Applications on Reliability-Based Optimi- Code Format,” J. Eng. Mech. Div., 100(1), pp. 111–121.
zation of Automotive Structures,” SAE Technical Paper No. 2003-01-0152. [28] Rubinstein, R. Y., and Kroese, D. P., 2007, Simulation and the Monte Carlo
[4] Nikolaidis, E., and Burdisso, R., 1988, “Reliability Based Optimization: A Method, Wiley, 2007, Wiley, Hoboken, NJ.
Safety Index Approach,” Comput. Struct., 28(6), pp. 781–788. [29] Youn, B. D., Choi, K. K., and Park, Y. H., 2003, “Hybrid Analysis Method for
[5] Enevoldsen, I., 1994, “Reliability-Based Optimization as an Information Tool,” Reliability-Based Design Optimization,” ASME J. Mech. Des., 125(2),
Mech. Struct. Mach., 22(1), pp. 117–135. pp. 221–232.
[6] Enevoldsen, I., and Sorensen, J. D., 1994, “Reliability-Based Optimization in [30] Duong, T., and Hazelton, M. L., 2005, “Convergence Rates for Unconstrained
Structural Engineering,” Struct. Saf., 15(3), pp. 169–196. Bandwidth Matrix Selectors in Multivariate Kernel Density Estimation,”
[7] Chandu, S. V. L., and Grandhi, R. V., 1995, “General Purpose Procedure for J. Multivar. Anal., 93(2), pp. 417–433.
Reliability Based Structural Optimization Under Parametric Uncertainties,” [31] Lin, P. T., 2013, “Ensemble of Unified Reliability Formulations (EURF),” 10th
Adv. Eng. Software, 23(1), pp. 7–14. World Congress on Structural and Multidisciplinary Optimization, WCSMO10,
[8] Frangopol, D. M., and Corotis, R. B., 1996, “Reliability-Based Structural Sys- Orlando, FL, May 20–24, Paper No. 5476.
tem Optimization: State-of-the-Art Versus State-of-Practice,”12th Conference [32] Chan, K.-Y., Papalambros, P. Y., and Skerlos, S. J., 2010, “A Method for
Held in Conjunction With Structures Congress XIV: Analysis and Computation, Reliability-Based Optimization With Multiple Non-Normal Stochastic Parame-
Chicago, IL, April 15–18, pp. 67–78. ters: A Simplified Airshed Management Study,” Stochastic Environ. Res. Risk
[9] Wu, Y. T., and Wang, W., 1996, “A New Method for Efficient Reliability- Assess., 24(1), pp. 101–116.
Based Design Optimization,” 7th Special Conference on Probabilistic Mechan- [33] Noh, Y., Choi, K. K., and Lee, I., 2010, “Identification of Marginal and Joint
ics and Structural Reliability, pp. 274–277. CDFs Using Bayesian Method for RBDO,” Struct. Multidiscip. Optim., 40,
[10] Carter, A. D. S., 1997, Mechanical Reliability and Design, Wiley, New York. pp. 35–51.
[11] Grandhi, R. V., and Wang, L. P., 1998, “Reliability-Based Structural Optimiza- [34] Zhu, X., and Milanfar, P., 2010, “Automatic Parameter Selection for Denoising
tion Using Improved Two-Point Adaptive Nonlinear Approximations,” Finite Algorithms Using a No-Reference Measure of Image Content,” IEEE Trans.
Elem. Anal. Des., 29(1), pp. 35–48. Image Process., 19(12), pp. 3116–3132.
[12] Tu, J., Choi, K. K., and Park, Y. H., 1999, “A New Study on Reliability Based [35] De La Torre, A., Peinado, A. M., Segura, J. C., Perez-C ordoba, J. L., Benıtez,
Design Optimization,” ASME J. Mech. Des., 121(4), pp. 557–564. M. C., and Rubio, A. J., 2005, “Histogram Equalization of Speech Representa-
[13] Youn, B. D., and Choi, K. K., 2004, “An Investigation of Nonlinearity of tion for Robust Speech Recognition,” IEEE Trans. Speech Audio Process.,
Reliability-Based Design Optimization Approaches,” ASME J. Mech. Des., 13(3), pp. 355–366.
126(3), pp. 403–411. [36] Gillman, A., Matous, K., and Atkinson, S., 2013, “Microstructure-Statistics-
[14] Youn, B. D., and Choi, K. K., 2004, “Selecting Probabilistic Approaches for Property Relations of Anisotropic Polydisperse Particulate Composites Using
Reliability-Based Design Optimization,” AIAA J., 42(1), pp. 124–131. Tomography,” Phys. Rev. E, 87(2), p. 022208.
[15] Youn, B. D., Choi, K. K., and Du, L., 2005, “Adaptive Probability Analysis [37] Yang, R. J., and Gu, L., 2004, “Experience With Approximate Reliability-
Using an Enhanced Hybrid Mean Value Method,” Struct. Multidiscip. Optim., Based Optimization Methods,” Struct. Multidiscip. Optim., 26(2), pp. 152–159.
29(2), pp. 134–148. [38] Cheng, G., Xu, L., and Jiang, L., 2006, “A Sequential Approximate Program-
[16] Youn, B. D., Choi, K. K., and Du, L., 2005, “Enriched Performance Measure ming Strategy for Reliability-Based Structural Optimization,” Comput. Struct.,
Approach for Reliability-Based Design Optimization,” AIAA J., 43(4), 84(21), pp. 1353–1367.
pp. 874–884. [39] Gea, H. C., and Oza, K., 2006, “Two-Level Approximation Method for
[17] Lin, P. T., Gea, H. C., and Jaluria, Y., 2011, “A Modified Reliability Index Reliability-Based Design Optimisation,” Int. J. Mater. Prod. Technol., 25(1–3),
Approach for Reliability-Based Design Optimization,” ASME J. Mech. Des., pp. 99–111.
133(4), p. 044501. [40] Yi, P., Cheng, G., and Jiang, L., 2008, “A Sequential Approximate Program-
[18] Lin, P. T., Jaluria, Y., and Gea, H. C., 2010, “A Hybrid Reliability Approach ming Strategy for Performance-Measure-Based Probabilistic Structural Design
for Reliability-Based Design Optimization,” ASME Paper No. DETC2010- Optimization,” Struct. Saf., 30(2), pp. 91–109.
28871. [41] Yi, P., and Cheng, G., 2008, “Further Study on Efficiency of Sequential Ap-
[19] Lin, P. T., Manuel, M. C. E., Liu, Y.-H., Chou, Y.-C., Ting, Y., Shyu, S.-S., proximate Programming for Probabilistic Structural Design Optimization,”
Chen, C.-K., and Lee, C.-L., 2014, “A Multifaceted Approach for Safety Design Struct. Multidiscip. Optim., 35(6), pp. 509–522.
and Probabilistic Optimization,” Math. Probl. Eng., 2015, p. 817820. [42] Lee, I., Choi, K. K., Du, L., and Gorsich, D., 2008, “Inverse Analysis Method
[20] Lai, X., 2014, “A Modified Method to Compute the Reliability With Arbitrary Using MPP-Based Dimension Reduction for Reliability-Based Design Optimi-
Independent Random Variables,” Proc. Inst. Mech. Eng., Part E, 228(4), pp. zation of Nonlinear and Multi-Dimensional Systems,” Comput. Methods Appl.
320–331. Mech. Eng., 198(1), pp. 14–27.
[21] Lin, P. T., and Lin, S.-P., 2014, “Ensemble of Gaussian-Based Reliability Anal- [43] Lin, P. T., Jaluria, Y., and Gea, H. C., 2009, “Parametric Modeling and Optimi-
yses (EoGRA) in the Original Design Space for Design Optimization Under zation of Chemical Vapor Deposition Process,” ASME J. Manuf. Sci. Eng.,
Arbitrarily Distributed Design Uncertainties,” 8th China–Japan–Korea Joint 131(1), p. 011011.
Symposium on Optimization of Structural and Mechanical Systems, CJK- [44] Lin, B. R., 2015, “Development of Fuzzy Automatic Contrast Enhancement
OSM8, Gyeongju, Korea, May 25–29, Paper No. 0187. Based on Fuzzy C-Means Clustering in CIELAB Color Space,” M.S. thesis,
[22] Lin, P. T., 2015, “Utilization of Gaussian Kernel Reliability Analyses in the Department of Mechanical Engineering, Chung Yuan Christian University,
Gradient-Based Transformed Space for Design Optimization With Arbitrarily Taoyuan City, Taiwan.

071403-14 / Vol. 138, JULY 2016 Transactions of the ASME