See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/245426180

Robust Design: An Overview

Article  in  AIAA Journal · January 2006

DOI: 10.2514/1.13639

CITATIONS READS

447 10,771

4 authors, including:

Kwang-Hyeon Hwang Tae Hee Lee

Hyundai Motor Company Hanyang University
11 PUBLICATIONS   587 CITATIONS    252 PUBLICATIONS   2,699 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Modeling and optimization of wheeled ground/unmanned combat system View project

(Hyundai Motor Company) Development and comparison study of measures for wire and pipe arrangement of commercial vehicle View project

All content following this page was uploaded by Tae Hee Lee on 17 March 2016.

The user has requested enhancement of the downloaded file.

AIAA JOURNAL
Vol. 44, No. 1, January 2006

Robust Design: An Overview

Gyung-Jin Park∗
Hanyang University, Gyeonggi-do 426-791, Republic of Korea
Tae-Hee Lee†
Hanyang University, Seoul 133-791, Republic of Korea
Kwon Hee Lee‡
Dong-A University, Busan 604-714, Republic of Korea
and
Kwang-Hyeon Hwang§
Hyundai and Kia Motors, Gyeonggi-do 445-706, Republic of Korea

Robust design has been developed with the expectation that an insensitive design can be obtained. That is, a
product designed by robust design should be insensitive to external noises or tolerances. An insensitive design has
more probability to obtain a target value, although there are uncertain noises. Theories of robust design have been
developed by adopting the theories of other fields. Based on the theories, robust design can be classified into three
methods: 1) the Taguchi method, 2) robust optimization, and 3) robust design with the axiomatic approach. Each
method is reviewed and investigated. The methods are examined from a theoretical viewpoint and are discussed
from an application viewpoint. The advantages and drawbacks of each method are discussed, and future directions
for development are proposed.

Nomenclature p = probability of success

Acr = area of common range p = design parameter vector
Asr = area of system range q = number of equality constraints
b = design variable vector r = number of inequality constraints
bL = lower bound of b s = scale factor
bU = upper bound of b u i (z ib ) = probability density function of z ib
p p
DPi = ith design parameter vi (z i ) = probability density function of z i
di− = shortage of the ith goal wi = weighting factor for the ith objective
d +j = overachievement of the ith goal function
zb = noise vector of b
d(zb , z p ) = joint probability density function
zp = noise vector of p
E[·] = expectation value of ·
η = signal-to-noise ratio
F(b, p, z) = objective function with noises
µ = mean of characteristic function f
FRi = ith functional requirement
µb = mean vector for b when the noises are
f (b) = objective function
considered
f new = normalized multiobjective function
µbi = mean of bi
G j (b, p, z), g j,new = jth constraint with noises
µ f (b, p), µ f = mean of characteristic function f
g j (b) = jth inequality constraint
f µ∗f = base value for µ
gj = goal of the jth objective function µp = mean vector for p when the noises are
h i (b) = ith equality constraint considered
h i (di− , di+ ) = function made of di− and d +j σ = standard deviation of characteristic
I = information content function f
k = constant for the loss function, σbi = standard deviation of bi
constant for g j,new σ f (b, p), σ f = standard deviation of characteristic
L( f ) = loss function function f
m = target value of the characteristic σ ∗f = base value for σ
function f σgi = standard deviation of gi
P = probability for satisfaction of a constraint σ pi = standard deviation of pi
P(b) = penalty function σy = standard deviation of characteristic
P j,0 = lower limit of P function y
φj(pj) = variable for probability
Received 25 September 2004; revision received 4 July 2005; accepted for ψ(b) = characteristic function of inner array
publication 5 July 2005. Copyright  c 2005 by the American Institute of
Aeronautics and Astronautics, Inc. All rights reserved. Copies of this paper
may be made for personal or internal use, on condition that the copier pay I. Introduction
the $10.00 per-copy fee to the Copyright Clearance Center, Inc., 222 Rose-
wood Drive, Danvers, MA 01923; include the code 0001-1452/06 $10.00 in
correspondence with the CCC.
∗ Professor, Department of Mechanical Engineering, College of Engineer-
A S the engineering environment becomes extremely competi-
tive, quality products are required in industries. Unexpected
deviations from the function that a designer initially intended for
ing Science, 1271 Sa-Dong, Sangrok-gu, Ansan City; gjpark@hanyang. are caused by variations in various engineering processes. Robust
ac.kr. Senior Member AIAA. design is to prevent such phenomena. Robust design has been de-
† Professor, School of Mechanical Engineering. veloped to improve product quality and reliability in industrial en-
‡ Assistant Professor, Department of Mechanical Engineering. gineering. Recently, this technology has been expanded to various
§ Senior Engineer.
design areas.
181
182 PARK ET AL.

First, we survey the definition of robust design. Many researchers g j (b) ≤ 0, j = 1, . . . , r (4)
have defined it. Taguchi, who is the pioneer of robust design, said
“robustness is the state where the technology, product, or process b L ≤ b ≤ bU (5)
performance is minimally sensitive to factors causing variability
(either in the manufacturing or user’s environment) and aging at The equality constraints may not be included in the formulation
the lowest unit manufacturing cost.”1 Suh said “robust design is de- because they are usually eliminated before the optimization process.
fined as the design that satisfies the functional requirements even The preceding functions can be evaluated by two methods. One
though the design parameters and the process variables have large is to use experiments when the function is not mathematically de-
tolerances for ease of manufacturing and assembly. This definition fined. A number of case studies using the traditional Taguchi method
of robust design states that the information content is minimized.”2 are attributed to this case. The other is to use mathematical expres-
Box said that “Robustifying a product is the process of defining its sions that are directly calculated or approximated by numerical ap-
specifications to minimize the product’s sensitivity to variation.”3 proaches. The present review is mainly focused on the latter method.
Although different expressions are used, their meanings are simi- When noises or perturbations exist, the objective and the constraint
lar. A common aspect of the definitions is that robust design is a functions are modified as follows:
design insensitive to variations, and this concept is accepted in the
engineering community. f (b) → f (b + zb , p + z p ) (6)
Conventional design methodologies have been developed to im-
g j (b) → g j (b + zb , p + z p ) (7)
prove the performances of products.4−7 If consideration of robust-
ness is added, a new direction should be taken in addition to the
where p ∈ R 0 is the design parameter vector, and it is usually re-
conventional concept. Theories of robust design have been devel-
garded as a constant vector in the design process. The noises of
oped by using existing design theories. The first one is the Taguchi
design variables and design parameters are represented as zb ∈ R n
method, and it is the foundation of robust design (see Refs. 8–10).
and z p ∈ R 0 , respectively, where 0 is the number of parameters. The
Taguchi proposed methods of determining variables to make per-
noise factors zb and z p , are also called the uncontrollable factors,
formance insensitive to noises in the manufacturing process. The
in the sense that they cannot be controlled by the designer.1,8,9 The
concept of robustness was established by the Taguchi method. In a
noise factor may be an external, an internal, or a unit-to-unit noise
product design, robustness is obtained by appropriate modification
of a product. They are summarized in Table 1.28,29
of the Taguchi method. The second one is robust optimization, where
The control factors are selected as independent variables that
the robustness concept is added to conventional optimization.11−19
can be controlled by the designer and have significant impact on the
The objective and the constraint functions are redefined with ro-
problem characteristics. The control factors correspond to the design
bustness indices. The last one is axiomatic design, which can be
variable vector b, whereas the characteristic function is equivalent
exploited well in conceptual design. In axiomatic design, the inde-
to the objective function f of the optimization problem. The char-
pendence axiom is utilized to find excellent designs. When multiple
acteristic function is a function or response that is to be measured
designs are found by using the independence axiom, the best one
and improved on.
is selected by the information axiom, and the robustness concept is
When Eqs. (6) and (7) are substituted into Eqs. (1–5) to consider
employed in this process.2,20−22
the robustness in the design formulation, Eqs. (1–5) are rewritten as
In this research, various robust design methodologies are
follows: Find
overviewed based on the preceding three methods. The trend of
the research is surveyed, and the application is discussed. However, b ∈ Rn (8)
theoretical aspects are emphasized more rather than application as-
pects. The advantages and disadvantages of the methods are also to minimize
discussed. In the robust design area, some terminologies have been
used in a different way. Thus, those terminologies are precisely de- F(b, p, z) (9)
fined, and a unified viewpoint of robust design is proposed.
There are some methods that are similar to robust design. They are subject to
design methods considering reliability and uncertainty.23−25 Uncer-
tainties or noises are included in these methods as well. Optimization G j (b, p, z) ≤ 0, j = 1, . . . , r (10)
terminologies can be utilized to distinguish these methods from ro-
b L ≤ b ≤ bU (11)
bust design. In robust design, insensitiveness of the objective func-
tion is emphasized. In reliability design, reliability of constraints where F and G j are the functions considering the noises in f and
is important. The reliability of constraints is similar to the robust- g j , respectively. Generally, F and G j are defined as the functions
ness of the constraints in robust design. It is calculated exactly by a of the mean and the variance of f and g j , respectively.
probability theory in reliability design, whereas it is directly calcu- Robust design should not be developed by the deterministic con-
lated in robust design. Design with uncertainties is similar to robust cept. It should be developed by the stochastic concept because the
optimization. However, robustness of the objective function is not deterministic approach cannot include the noises. The performance
considered. The distinctions are made based on the methods of the
early stages. These days, the methods are fused, therefore, it may
not be possible to distinguish them in some applications. Methods Table 1 Variation sources and examples
considering reliability or uncertainties are not reviewed in this paper.
Factors Examples
II. Mean and Variance External Temperature and relative humidity
Load
When the optimization framework is used, a design problem can
Human error
be formulated as follows26,27 : Find Dust in the environment
Unit to unit Weight of each part
b ∈ Rn (1) Dimension tolerance
Differences by batch-to-batch
to minimize Thickness variations in coated products
Internal Wear
f (b) (2) (aging) Mileage of a car
Weathering of paint on a house
subject to Total current passed through a car battery
Plastic creep
h i (b) = 0, i = 1, . . . , q (3)
 PARK ET AL. 183

Table 2 Formulas of means and variances

Name Symbol Formulaa

1

k
Sample mean f¯ or (m f ) f¯ = fi
k
i =1
  1

k
Sample variance V or s 2f V = ( f i − m f )2
k −1
√ i =1
Sample standard sf sf = V
deviation
1

l
Population mean µ µ= fi
l
i =1
1

l
Population sample σ2 σ2 = ( f i − µ)2 Fig. 1 Examples of distribution of the response functions.
variance l
√ i =1
Population sample σ σ = σ2
standard deviation
a
Here k is the number of samples and l is the number of population.

measures are defined as the data to be measured experimentally or

mathematically, which represent the performances of a product or a
process. The performance measures are usually modified to form a
characteristic function. The performance measures could vary un-
der the same circumstance due to the distribution of noise factors.
Generally, the distribution is measured by the mean and the vari-
ance. The formulas of mean and variance are summarized in Table 2,
where f i is the ith characteristic function in experiment or the ith
response function in the design.30
When Eqs. (6) and (7) are considered, the mean µ f and the vari-
ance σ 2f of the characteristic function f are represented as10,31 Fig. 2 Probability density functions of two designs with design range.
  
   
µ f = E[ f (b)] = ··· f (b + zb , p + z p )u 1 z 1b · · · u n z nb
That is, it was regarded as a method for design of experiments rather
 p  p p p
than a design methodology.8−10
× v1 z 1 · · · v0 z 0 dz 1b · · · dz nb dz 1 ··· dz 0 (12) In Eqs. (8–11) of the design formulation, the original Taguchi
   method does not deal with the constraints in Eq. (10). It leads to
 2
σ 2f = E[ f (b) − µ f ]2 = ··· f (b + zb , p + z p ) − µ f unconstrained optimization problems. Although the design param-
eters p and the noise factors z are not explicitly defined in the design,
     p  p p p the Taguchi method can be used to determine the settings of control
× u 1 z 1b · · · u n z nb v1 z 1 · · · v0 z 0 dz 1b · · · dz nb dz 1 · · · dz 0 factors for robust design. To explain this in detail, let us take an ex-
(13) ample with three distributions of the objective function f as shown
p
in Fig. 1.
where u i (z ib ) and vi (z i ) are the probability density functions of In Fig. 1, µ A , µ B , and µC are the means of distributions A,
p
noise factors z ib and z i , respectively. In Eqs. (12) and (13), the noise B, and C, respectively. Then, robustness can be measured by the
factors are assumed to be statistically independent. If it follows the magnitude of variation. When the design is focused only on the
Gaussian distribution, the probability density function becomes as28 robustness of the performance, the candidate designs B and C are
    √    2  2 better than design A. However, the target value of the performance
u i z ib = 1 σbi 2π exp − bi − µbi 2σbi (14) generally exists in the design. The target value can be zero, infinite,
or a specific nominal value. Thus, the Taguchi method suggests a
where µbi and σbi are the mean and the standard deviation of the ith design with minimum variation around the target value. From the
design variable bi . viewpoint of the Taguchi method, design C is the best out of the
Robust design techniques have been developed by replacing the three candidate designs.
functions of the mean and variance into the functions of Eqs. (8–10). The procedure to obtain a robust design can be explained as fol-
In the next sections, robust design methodologies are examined. lows: The design candidates D and E with two arbitrary distributions
are represented in Fig. 2. The design range is equivalent to the al-
III. Taguchi Method lowable range, and the means of the two distributions coincide with
A. Taguchi Method the target value m of the design. A small part of candidate design
By the end of the 1940s, the Taguchi method had been developed D is outside the design range, whereas none of candidate design
by for quality improvement. His technique for quality engineering E is outside the design range. However, if any unexpected noise
is referred to as the Taguchi method or robust design.1,30 Whereas factors become active, the distribution of E has a larger probabil-
the Taguchi method was successfully applied at the Electrical Com- ity to be outside the design range than that of D. For instance, the
munications Laboratories of the Nippon Telephone and Telegraph unexpected noises are the tolerances of product sizes and variances
Co., AT&T Bell Laboratories was interested in it. In 1980, Phadke10 of the environment such as temperature, humidity, external forces,
invited Taguchi to AT&T Bell Laboratories. After being impressed manufacturing process, etc. Thus, it is clear that candidate design D
by the Taguchi method, Phadke published a textbook on the Taguchi is more robust than candidate design E.
method in 1989.10 Taguchi introduced a quadratic loss function to represent robust-
The Taguchi method has greatly contributed to quality improve- ness as (see Ref. 10)
ment of various designs. In early case studies, the Taguchi method
was applied to the process design rather than the product design. L( f ) = k( f − m)2 (15)
184 PARK ET AL.

Table 3 Examples of S/N ratios

Characteristic type S/N ratios

Nominal the best With scale factor
η = 10 log(µ2 /σ 2 )
With adjustment factor
η = −10 log σ 2


n 
1
Smaller the better η = −10 log yi2
n


i =1
n
1 1
Larger the better η = −10 log
n yi2
i =1

Fig. 3 Block diagram of a product/process: P diagram.

where k is the constant to define the loss and m is the target value.
The expected value of the loss function is defined as10

Q = E[L( f )] = k[σ 2 + (µ − m)2 ] (16)

where µ is the mean of f and σ is the standard deviation of f .

The symbol f represents the response or objective function in a
general design, whereas the symbol y is used for f in the Taguchi
method. Robust design is the design minimizing the average loss, a) First step
which can select design candidate D in Fig. 2.
Figure 3 is called the P diagram and used to explain the proce-
dure of the Taguchi method from another viewpoint. The P diagram
represents the schematic relationship between the factors and the
product or process. The noise factor is the one that the designer
cannot control even though it causes variability. In a general design,
tolerances are considered as noise factors. In a parameter design,
control factors are determined to reduce the effect of the noise fac-
tors. Therefore, noise factors are not directly considered. On the
other hand, noise factors are directly controlled in a tolerance de-
sign. The parameter design is the stage to improve the quality of
a product or process without cost increase, whereas the tolerance
b) Second step
design does with cost increase. In this paper, the parameter design
is explained. Fig. 4 Two steps of Taguchi method.
When the target value of a response is given as shown in Fig. 3, the
There are two goals in performing robust design. One is to min-
Taguchi method determines the optimum setting of control factors
imize the variability produced by the noises factors. The other is
so that the variation of a response is minimized, although uncontrol-
to make the mean value close to the target value. The design to
lable factors (noise factor) exist. Equation (16) may be regarded as
attain one goal is not usually consistent with the one to attain the
an index to find a robust design. Suppose that we have a scale factor
other goal. To meet the two goals, Taguchi developed a two-step
s to adjust the current mean to the target value. The scale factor s is
optimization strategy. The first step is to reduce the variation, and
described as
the second step is to adjust the mean on the target. The procedure
s = m/µ (17) is shown in Fig. 4. The first step, shown in Fig. 4a, concentrates
on minimization of the variation, while the mean is ignored. In the
When the current mean is adjusted to the target value, the average second step (Fig. 4b), the mean is moved to the target value while
loss function of Eq. (16) is changed to10 sacrificing the improved variation somewhat.1,8−10
The first step is to find the optimum setting of control factors (de-
Q a = k[{µ(m/µ) − m}2 + {(m/µ)σ }2 ] = km 2 (σ 2 /µ2 ) (18) sign variables) to maximize the S/N ratio or to reduce the variation.
The orthogonal arrays are employed to assign the control factors to
To enhance additivity of the effect of the control factors, Eq. (18) is an experimental matrix. The best condition can be selected from the
transformed to10 full combinations of control factors. However, it is efficient to select
the smallest orthogonal array because full combinations are costly
η = 10 log10 (µ2 /σ 2 ) (19) in most cases. An orthogonal array is a simple experimental ma-
trix to determine the factor effects by ignoring the high interactions
Equation (19) is the ratio of the power of signal factors µ and the between factors. Since Rao developed orthogonal arrays in 1946,
power of noise factors σ . Thus, it is called the signal-to-noise (S/N) they have been widely utilized in the field of statistics and design of
ratio. Maximizing Eq. (19) is equivalent to minimizing Eq. (18). experiments.33
That is, a robust design is obtained by maximizing Eq. (19). The smallest size of an orthogonal array can be selected by using
As in the example of Fig. 2, the response with target value m the number of control factors and the number of levels.10 In a matrix
is referred as the nominal-the-best type characteristic. On the con- experiment, experiments for a row of the orthogonal array called
trary, the responses with the target values of zero and infinity are an inner array are repeatedly conducted to evaluate the effects of
referred as the smaller-the-better type and the larger-the-better type, noise factors.10,11,30,33 From the experiments for each row, the S/N
respectively. The S/N ratios of the smaller-the-better type and the ratios are calculated. Then, the optimum setting of control factors is
larger-the-better type can be derived by the similar procedure as for determined by performing analysis of means (ANOM) of the S/N
the nominal-the-best type. As the index of robustness, the S/N ratio ratios. A one-way table is utilized in the ANOM process. If the
of Eq. (19) is preferred to the average loss function of Eq. (16). The average loss of Eq. (16) is defined as the characteristic function to
examples of S/N ratios are summarized in Table 3.10,32 find a robust design, the effect on the variation is confounded by the
 PARK ET AL. 185

effect on the mean. Therefore, the S/N ratio adjusted to the target of experiments in the inner array and Nout is the number of exper-
value of Eq. (19) or the S/N ratio defined as Eq. (20) has been utilized iments in the outer array) times. This may result in an inefficient
as the characteristic in the nominal-the-best type problem, and costly process. Welch et al.,42 Vining and Myers,43 Shoemaker
and Tsui,44 and Kwon and Moon45 indicated shortcomings and sug-
η = −10 log10 σ 2 (20) gested the combined array technique. A few researchers40−42 devel-
oped the approximate models of the mean and the variance based on
When significant interactions exist, they should be considered in the response surface models. However, the reliability of the robust
selecting the orthogonal array and assigning the factors. However, it design determined from the approximate models strongly depends
is not easy to identify the significant interactions among the control on the reliability of the approximate models.
factors in design. This can be a drawback of using orthogonal arrays. Third, research that deals with more than one characteristic func-
Thus, the optimum setting determined by performing ANOM should tion has been performed because the traditional Taguchi method
be followed by a confirmation experiment. provides robust design for a single characteristic function. Design
A computer simulation model provides the same output with the problems with more than one characteristic function are often met.
same input because it is a deterministic method. Although it is a To obtain robust design in the problem with multiple characteris-
disadvantage, the computer simulation model has an advantage in tic functions, Phadke10 determined the robust optimum by the de-
that it can define the design parameters p and the noise factors z signer’s decision on tradeoffs. Shiau46 and Tai et al.47 introduced
exactly. In this case, the Taylor series expansion, the Monte Carlo the weighting factors to unify each characteristic function. How-
simulation, or an outer array can be utilized to obtain the variations ever, these methods strongly depend on the designer’s intuition.
(see Ref. 10). An outer array is an orthogonal array, which is used Tong and Su48 determined the weighting factors by using the fuzzy
for a row of the orthogonal array (inner array). The outer array has theory to avoid the described ambiguity.
multiple cases of experiments for a row; therefore, it accommodates Pignatiello49 derived the loss function of multiple characteristic
the repeated experiments of an experiment case. functions, and Tsui50 suggested robust design by deriving the ex-
The second step is to set the mean to the target value by selecting pectation value of Pignatiello’s loss function. In addition, Wu51 sug-
a control factor called the scale factor. The scale factor should only gested the following procedure: 1) The optimum levels correspond-
have an effect on the mean, while the S/N ratio is maintained. In ing to each characteristic function are calculated. 2) The percentage
general experiments, time can be one of the scale factors. However, it reduction of quality loss is determined by recording the change in
is difficult to find a scale factor in a general design. Furthermore, no the S/N ratio for each optimum. 3) The optimum is determined as the
scale factor may exist, and this prevents the designer from applying levels with maximum percentage reduction and minimum approxi-
the Taguchi method directly to the design. mated loss function. Recently, Liao and Chen52 suggested the data
envelopment analysis ranking to consider multiple characteristic
B. Research Topics of the Taguchi Method functions. Similar to the optimization problem with a multiobjec-
The Taguchi method has been widely applied to various fields tive function,53 the method has both advantages and disadvantages.
due to its simple and explicit procedures.1,8−10,34−37 Even beginners Thus, the designer should deliberately select an appropriate method
can easily use the Taguchi method. Especially, the Taguchi method for one’s problem.
has been regarded as an excellent tool in a design process using Finally, the Taguchi method cannot include the constraints of
experimental approaches. Based on the concepts of the traditional Eq. (10). To include the constraints and to obtain a robust design
Taguchi method, researchers of design methodologies have tried to for the problem without the scale factor, Lee and Park12 and Lee31
use Taguchi’s outstanding ideas in their fields. However, there are suggested the following procedures. 1) The deterministic optimum
some conflicts between the traditional Taguchi method and the de- is calculated. It is assumed that the target value is obtained in this
sign methodologies. Thus, research has been performed to overcome process. 2) Near the optimum, the candidate design variables are
the difficulties. assigned to an orthogonal array, which is called the inner array.
First, many researchers35,37−40 have pointed out the possibility of 3) Each row of the inner array generates the outer array considering
using the S/N ratio in an incorrect manner. Leon et al.38 found the the noises, and the experiments are performed. 4) From the results
condition of the transfer function in which the S/N ratio of Eq. (19) of the outer array, the characteristic function of the inner array is
is effective. It is that the S/N ratio could be a performance measure determined as
independent of adjustment (PerMIA) only when a scale factor ex-
(b) = σ y + P(b) (21)
ists and the standard deviation is proportional to the mean. Box35
and Box et al.39 suggested a more general transformation approach
where P(b) is the penalty function that is the maximum violation
and proved that the variance of a transformed variable becomes a
of the experiments in the outer array. 5) The ANOM of the charac-
PerMIA. Their research is developed based on the assumption that
teristic of Eq. (21) is performed to estimate the characteristic at an
a scale factor exists.
arbitrary design. 6) For each combination of the candidate design
The performance criterion such as the S/N ratio is only effective
variables, the characteristic is estimated. 7) After the estimators are
when factors influencing the mean are separated from the factors
arranged by ascending order, the optimum is selected as the set with
influencing the variance. However, the effects on the mean con-
the lowest estimator and P(b) = 0. However, this method selects a
found with the effects on the variance when the factors influencing
robust design with limited candidate designs. As already mentioned,
the mean cannot be separated. The same difficulties exist with the
recent research has been performed to supplement the deficiencies
S/N ratios of the smaller-the-better and the larger-the-better type
of the traditional Taguchi method.
characteristics.40 Thus, one should identify whether the S/N ratio
can give a robust design before solving the problem.
Phadke10 and Fowlkes and Creveling30 tried to obtain the validity IV. Robust Optimization
of the S/N ratio by examining the qualification of the characteristic A. Robust Optimization
function. They suggested some guidelines. However, the guidelines Aforementioned robust design problems have been solved based
are quite strict, and it is difficult to determine whether they are on well-developed mathematical optimization techniques during the
satisfied. Montgomery40 recommended the use of Eq. (20) instead 15 years.42,43,54−56 Because the Taguchi method uses an orthogonal
of Eq. (19) as the S/N ratio, and Lee et al.11 used the weighted array and design variables are defined in a discrete space, it is dif-
multiobjective function for robust design. Wu41 also evaluated the ficult to treat a wide design range. Design in a discrete space may
performance criteria of the S/N ratio, mean standard deviation, and have an advantage; however, design in a continuous space is often re-
variance. quired as well. Generally, a design may have many design constraints
Second, research has been performed to reduce the number of ex- that the Taguchi method may not resolve. Because optimization
periments in the traditional Taguchi method. The traditional method techniques can easily handle the constraints, robust optimization
requires as many experiments as Nin × Nout (where Nin is the number based on optimization techniques has been studied.57 In a general
186 PARK ET AL.

Note that the objective function in Eq. (23) is defined in terms

of the mean and the standard deviation, that is, the variance. Equa-
tion (24) means that all of the constraints should be satisfied despite
the noises in design variables and parameters. Therefore, robust
optimization methods have been efficiently developed to solve the
problem defined as Eqs. (22–25).11,15,58,63−65 Optimization with un-
certainties is similarly defined as Eqs. (22–25).

B. Robustness of Objective Function

Robustness of the objective function has been emphasized more
than that of constraints because an insensitive design of the objec-
tive function is pursued. The mean µ f and the variance σ f of the
objective function are approximated as follows:

µf ∼
= f (µb , µ p ) (26)
Fig. 5 Robustness of the objective function.
n 

2 0 

2
∼ ∂f ∂f
σ 2f = σb2i + σ p2i (27)
i =1
∂bi i =1
∂ pi

If b and p are assumed to have normal distribution, the mean µ f of

Eq. (26) can use the value of Eq. (2) during the optimization process.
Note that Eqs. (26) and (27) are meaningful when the interactions
of the design variables and the design parameters are negligible and
deviations of them are small.10,66
The variance is made of the sensitivity, that is, first derivative,
as shown in Eq. (27). If the objective function includes the sen-
sitivity, second-order derivatives of the objective are required in
the optimization process. However, calculation of the second-order
derivatives of the objective function is so expensive that it is not
recommended. Because the constraints of Eq. (24) are represented
in terms of the variance, second-order derivatives of the objective
are also needed. Thus, robust optimization without second-order
derivatives is recommended.
In the simplest method, the objective function is defined by the
sensitivity and the second derivatives are directly evaluated. This
method is quite expensive but reliable. It has some limits in that
Fig. 6 Change of feasible region of constraints.
the solution can include local information. Belegundu14 obtained
robustness by minimizing the sensitivity information. Jung and
design problem, we can hardly separate the mean and variance of Lee67,68 have shown that robust optimization can be achieved for
design variables. Optimization techniques can obtain design vari- mass minimization because the second derivative can be easily cal-
ables by considering the mean and the variance simultaneously. The culated for the mass of a structure. However, robustness of mass may
goal of robust optimization is to solve these problems. not be important in engineering design. Han and Kwak69 have di-
Robust optimization is formulated as Eqs. (8–11). The functions rectly used the second-order sensitivity for an objective function that
F and G are defined to enhance robustness, and the robustness, is the natural frequency of a structure. As computer power increases,
of functions is defined as follows: 1) Robustness of the objective calculation of the second-order sensitivity is viable for mathemati-
function12,58 can be achieved by reducing the change of the objec- cal or small-scale engineering problems, but this is difficult to apply
tive with respect to the changes of tolerances for the design variables. to large-scale problems.
Figure 5 presents an objective function that has a minimum at the Many methods have been proposed to avoid second-order sen-
peak point A, where the objective function can drastically change sitivity calculation. Sundaresan et al.15,65 proposed the sensitivity
by small variations of the design variables. Thus, the local mini- index (SI) to incorporate the evaluation of constraints with the
mum point B or C is better for an insensitive design. 2) Robustness worst combination of design variables in an orthogonal array. Their
of constraints12,16−19 means that all of the constraints are satisfied method can reduce the computing cost. Su and Renaud64 evaluated
within the range of tolerances for the design variables, that is, the the first-order sensitivity of SI automatically. Also, the objective
feasible region is reduced as shown in Fig. 6. Probabilistic consider- function can be approximated as an explicit regression model. Chen
ation of the objective function and constraints is known as reliability et al.56 have shown that the second-order sensitivity of the objective
optimization59 or stochastic optimization.60−62 They will be briefly function can easily be evaluated if the objective function is approxi-
mentioned later because we mainly focus on the review of robust mated by the response surface method (RSM).70 Also approximation
design. of the objective function using the RSM has been often used in ro-
Now let us consider the formulation of robust optimization as bust optimization. However, the RSM has the risk to converge to a
follows: Find sensitive design because the sensitive part of the objective function
can be ignored in the approximation process. The Taguchi method
b ∈ Rn (22)
can be applied to robust optimization as explained in Sec. III. The
to minimize mean and the variance are evaluated by using an outer orthogonal
array and minimized by the one-way table. This method does not
[µ f (b, p) σ f (b, p)] (23) need sensitivity information.32,71,72
subject to Robust optimization has also been developed with uncertain-
ties. Du and Chen proposed the propagating model uncertainty
g j (b + zb , p + z p ) ≤ 0 j = 1, . . . , r (24) by applying the extreme condition approach and the statistical
approach.73 To integrate robust design with multidisciplinary design
b L ≤ b ≤ bU (25) optimization, efficient uncertainty analysis has been developed from
 PARK ET AL. 187

a computational viewpoint. The system uncertainty analysis and the where d(zb , z p ) is the joint probability density function of proba-
concurrent subsystem uncertainty analysis methods are developed bilistic variables zb and z p . Evaluation of Eq. (34) is very difficult;
to improve the computational efficiency in highly coupled analyses. therefore, an approximate equation can be used to reduce the com-
Then the effort for the system level is significantly reduced.74 Putko puting cost.
et al. considered the geometric uncertainty and flow parameter un- To approximate Eq. (34), the moment matching formulation can
certainty for robust design using the quasi-one-dimensional Euler be used.77,78 In the moment-matching method, the probability is
computational fluid dynamics code.75,76 simplified by assuming that the constraints have normal distribution.
It is one of the most popular methods. If the constraints have normal
C. Robustness of Constraints distribution, Eq. (33) becomes
As mentioned earlier, the constraints can be transformed into  −g j (b, p)/σg j   φj (pj ) 
Eq. (24) in the case when noises exist. Treatment of constraints 1 −θ 2 1 −t 2
varies according to the detailed definition of Eq. (24). The research √ exp dθ ≥ √ exp dt
−∞ 2π 2 −∞ 2π 2
trends are as follows. (35)
where φ j ( p j ) is the variable for the probability p j of the normal
1. Reduction of the Feasible Region distribution and is given as
The feasible region is reduced by noises as shown in Fig. 6. In  
other words, the constraints in Eq. (27) must be satisfied even though φ j ( p j ) = g j (b + zb , p + z p ) − g j (b, z) σg j (36)
we have zb and z p . Thus, tolerance design methods have been applied
to robust optimization. From the inequality of Eq. (35), Eq. (33) can be modified as
With the noises of constraints, Eq. (24) can be written as16,60
g j,new ≡ g j + φ j ( p j )σg j ≤ 0 (37)
g j,new ≡ g j + kσg j ≤ 0 (28)
Equation (37) provides the base for Eq. (28). Equation (37) can be
where k is a user-defined constant depending on the design purpose
employed in robust optimization; however, Eq. (28) is used more
and σg j denotes the standard deviation of the constraint g j that can
often.
be approximated from Eq. (4) as follows10,60 :
n 
∂g j 2
0 


∂g j 2 D. Optimization Methods
σg2j ∼
= σb2i + σ p2i (29) The optimization scenario is the key point of robust optimiza-
i =1
∂bi i =1
∂ pi tion. Robust optimization has employed nonlinear programming,
but second-order derivatives are needed because of the variance
If the noise is given by a range, the worst case of g j,new and Eq. (25) in the objective function. Although the second-order derivatives
becomes are quite expensive, we may still use nonlinear programming with
n 

 0 

 second-order derivatives. However, some methods have been devel-
 ∂g j  b   ∂g j  p  oped to alleviate the use of the second-order derivatives.31,60,71,72
  zi + k j   zi
p
g j,new = g j + k bj (30)
i =1
∂bi i =1
∂ pi Optimization methods have been employed to accommodate the
multiobjective function. The multiobjective function can deal with
b L ≤ b + zb ≤ bU (31) uncertainties.79,80 Here some multiobjective methods used for robust
optimization are introduced.
p
where |z ib | and |z i |
denote the maximum values of the tolerance
p
ranges and k bj and k j are user-defined constants depending on the 1. Weighted Sum Method
design purpose. Equations (29) and (30) have the first derivatives In Sec. IV.C, we considered the robustness of constraints during
of constraints. Thus, the second-order derivatives of constraints are robust optimization where a feasible region is reduced. As shown in
needed if mathematical programming is employed as an optimizer. Eq. (23), there are multiple objective functions in robust optimiza-
It is similar to the case of the objective function. tion. Because general optimization considers one objective fnction,
The design point b∗ of Fig. 6 is the optimum point when the design various approaches are introduced to define a single objective func-
variables have the tolerances whose ranges are box-type domains tion from multiobjective functions. The most common approach
as shown in Fig. 6. Here g j,new is defined to satisfy the constraints in multiobjective optimization is the weighted sum method. Equa-
within the ranges as follows: tion (23) can be rewritten as12,31,81,82
g j,new = maximum{g j (b), ∀b ∈ the hypersurface of Fig. 6} (32)
   
f new = α µ f (b, p) µ∗f + (1 − α) σ f (b, p) σ ∗f (38)
In addition to the methods mentioned, g j,new can be defined based
where µ∗f and σ ∗f are base values for the mean µ and standard
on the theory of tolerance design, but the concept is very similar.
deviation σ of the objective function, respectively. They are used
for normalization and usually have the starting values of the opti-
2. Probabilistic Analysis mization process. The weighted sum method is inefficient, but it is
Probabilistic analysis is adopted from the theories of reliability popular due to simple and easy application.
optimization. It is also being used in the design that considers un-
certainties. Satisfaction of constraints can be written as probabilistic 2. Compromise Decision Support Problem
forms for normal distribution of noise59,60,66,67 :
One type of multiobjective optimization method is goal program-
P[g j (b + zb , p + z p ) ≤ 0] ≥ P j,0 , j = 1, . . . , r (33) ming. Designers determine the design goal, transform the goal into
goal constraints, and then optimize them with real constraints by
where P[·] denotes probability such that P j,0 represents the proba- means of nonlinear programming. A goal constraint is another ex-
bility of satisfaction for constraints, whereas the noise zb or z p is of pression that an objective function must achieve the design goal.
normal distribution. If probabilistic distributions of noises are given, The compromise decision support problem is a modified version
Eq. (33) can be rewritten as follows: of goal programming for robust optimization.47,56,58,83−85 A com-
promise decision support problem is a hybrid multiobjective opti-
p[g j (b + zb , p + z p ) ≤ 0] mization technique that incorporates both traditional mathematical
 programming and goal programming. The system constraints must
= d(zb , z p ) dzb dz p , j = 1, . . . , r (34) be retained without compromise while objectives are properly com-
g j (b + z b , p + z p ) ≤ 0 promised.
188 PARK ET AL.

The objective function in a compromised decision support prob- 2) The information axiom is to minimize the information content
lem is a deviation function x(di− , di+ ) in terms of deviation variables. of the design.
Therefore, a robust optimum can be found by minimizing the devi- The independence axiom states that the independence of the func-
ation variables while the constraints are satisfied. The compromise tional requirements must always be maintained by an appropriate
decision support problem with t multiobjective functions is formu- choice of design parameters (DPs). For example, the design equa-
lated as follows: Find tions may be classified into the following three kinds when there are
two functional requirements:
b ∈ Rn (39)  
 
FR1 X 0 DP1
to minimize =
FR2 0 X DP2
      
x = w1 h 1 d1− , d1+ , w2 h 2 d2− , d2+ , . . . , wt h t dt− , dt+ (40)  
 
FR1 X 0 DP1
=
subject to FR2 X X DP2

gi (b, p) < 0, i = 1, . . . , r (41)  

 
FR1 X X DP1
= (45)
FR2 X X DP2
d +j d −j
f
f j (b, p) + − = gj (42)
In the first case of Eq. (45), the design matrix is diagonal. Each of
the functional requirements can be independently satisfied by one
d −j · d +j = 0 (43) DP. This is called an uncoupled design. In the second, the matrix
is triangular. The independence of functional requirements can be
guaranteed if the DPs are determined in an appropriate sequence.
d −j ≥ 0, d +j ≥ 0, j = 1, . . . , t (44) This is called a decoupled design. The third has a full design matrix
and results in a coupled design. It should be avoided because any
f
where w j is the weight and gj
is the goal that the jth objective choice of DPs will not satisfy the independence axiom.
function achieves. Here d −j and d +j are the deviation variables that The information axiom provides a metric of selecting the best
represent the distance between the target level and the actual at- design when there are multiple ones that satisfy the independence
tainment of the goal and indicate the degrees of shortage and over- axiom. The best value for selected DPs can be obtained by minimiz-
achievement of the goals, respectively. As a result, we can achieve ing information content. Information content I is defined in terms
the goal as much as possible by means of minimizing these devia- of the probability p of satisfying a given functional requirement as
tions. In contrast to traditional goal programming, the compromise
decision support program uses real design constraints as shown in I = log2 (1/ p) (46)
Eqs. (39–44).
Various deviation functions are defined in a compromise deci- Information content is defined in terms of the probability of suc-
sion support program. The most general and simplest form is the cess, which is calculated by specifying the design range and the
weighted sum of deviations according to the level of requirement to system range. The design range and the system range of a design
achieve each goal.84 The distance function considering the weighted with one functional requirement are shown in Fig. 7. The informa-
deviation vector has been used as a deviation function.85 tion is defined as
When a compromise decision support program is used in robust
I = Acr /Asr (47)
optimization, both the robustness and feasibility of constraints are
satisfied and various design goals are achieved by compromising where Acr is the common range and Asr is the system range. To
goals. Moreover, deviation variables provide the degrees of shortage minimize the information content, 1) the variance of the system must
and overachievement of the goals. This method can obtain a Pareto be small and 2) the bias must be eliminated to make the system range
solution for a nonconvex problem (see Ref. 85). within the design range. This policy is similar to the Taguchi method.
That is, the minimization of the information content is similar to the
3. Physical Programming robust design concept.
Messac,86 Messac and Wilson,87 and Messac and Ismail-Yahaya88 Robust design with axiomatic design is conducted after the inde-
have proposed an effective multiobjective optimization method. It pendence axiom is satisfied. Information of each FR is minimized.
is called physical programming. It classifies objectives and includes Therefore, it is one FR-one DP problem as follows20 :
them in the constraint set according to preference. The preference
is expressed by a utility function called an aggregate preference FR = A · DP (48)
function. For instance, the weighted objective function in Eq. (38)
is an example. This approach has been used in robust design with a Suppose that there are two candidate DPs in Eq. (48) and that
multiobjective function.89 the relationships between the FR and the DP are shown in Fig. 8.
In addition to the described methods, the difference between the The gradient of a design is called the stiffness of the design. The
maximum and minimum of the objective function has been min- specified tolerance of FR can be more easily achieved by lowering
imized to achieve the robustness of an objective function.18 The
relationship between robustness and optimum sensitivity has been
studied.17,90,91 Chen et al. have studied the role of robustness in con-
current engineering to enhance productivity.91 In robust design, a
unified method has not yet been proposed, and large-scale problems
rarely have been solved. Many practical case studies are needed.

V. Robust Design in Axiomatic Design

Axiomatic design has been proposed by Suh,2,20 and it provides
a framework for robust design. The basic postulate of axiomatic
design is that there are two fundamental axioms. The axioms are
stated as follows:
1) The independence axiom is to maintain the independence of
the functional requirements (FRs). Fig. 7 Information content based on probability density function.
 PARK ET AL. 189

ied how to use the decision models to perform axiomatic design

and robust design using the compromise decision support problem.
Park has mentioned the differences and common aspects between
the two methods.97 As shown in Fig. 2, design E is better than design
D from the viewpoint of axiomatic design, but design D is better in
the viewpoint of the Taguchi concept because the Taguchi method
does not use a design range. From this study, Hwang and Park,21
Hwang,22 and Park97 developed the robustness index by using the
design range and the probability density function. If the mean val-
ues are inside the design range and the distributions are the normal
ones as shown in Fig. 9, the two methods select the same design
candidate because the information content and the loss function
have the same property in the robust design process. Consequently,
the Taguchi method and the information axiom of axiomatic de-
sign can come to the same conclusion if the independence axiom is
satisfied.
Fig. 8 Distribution of an FR from the two designs.
VI. Summary
Classification of robust design and its characteristics are surveyed.
As mentioned earlier, robust design has been developed in three
ways. Because there are numerous case studies for each method, the
case studies are not introduced in this paper. Basic ideas and develop-
ment procedures of each method are only introduced and explained.
Robust design is an excellent concept in that an insensitive or
a tolerant design can be obtained. Methods for robust design have
been actively developed with much expectation. The methods are
established although they are neither perfect nor universal. Robust
design has been addressed from the area of design of experiments
(DOE) where the Taguchi method can be easily applied. Thus, most
of the case studies are from DOE. Design engineers adopted the
concept in mechanical design. They developed theories, as intro-
duced in this paper. Many problems have been solved; however,
most of them are local or small-scale problems. Especially, exam-
ples from astronautics and aeronautics have those characteristics.
It is expected that examples from the practical community will be
Fig. 9 Distribution of the two designs with target value and design defined. Also, the benefits of robust design should be demonstrated
range. from the examples.
Three methods are introduced in this paper. The characteristics
the stiffness of the system (design A). Allowing
DP to be larger and future direction of each method are summarized as follows:
leads to a robust design. Therefore, the process of robust design20 1) The Taguchi method is quite excellent for DOE, and designers
is stated as 1) select the design that has a low stiffness to minimize are making efforts to adopt the robustness concept of the method
the variation of the system range and 2) minimize the bias between in general design. However, it is not easy to find a scale factor
the target and the mean from the selected design. that is exploited in the original Taguchi method. Also, direct use
As mentioned earlier, common aspects exist between the Taguchi of the S/N ratio is difficult. Although many modified methods have
method and axiomatic design in that a robust design is induced.1,8,9 been developed according to the characteristics of application, a
However, different characteristics are discovered as well. The rigorous mathematical method for design has not yet been found.
Taguchi method emphasizes how each parameter is determined af- Some hybrid methods with robust optimization are expected to solve
ter the conceptual design. That is, the value of each parameter is engineering design problems.
determined from the S/N ratio of each row in an orthogonal array. 2) Robust optimization is an effort to adopt optimization, which
Axiomatic design emphasizes which parameters are chosen in the is mathematically excellent, to achieve robustness. However, it was
conceptual stage because a choice of the DP is important as shown in found that a multiobjective function exists and second-order deriva-
Fig. 8. In other words, selection of a part out of multiple candidates tives are needed. Therefore, compromised solutions are pursued, or
is determined by axiomatic design. approximation methods are employed. In this case, questions may
Robust design in axiomatic design is obtained by minimizing the arise if we have to use mathematical optimization. Reliability op-
information content. Various methods have been developed for cal- timization and optimization with uncertainty have similar concepts
culation of the information content for robust design.2,20−22,92−94 In to that of robust optimization. However, what they pursue is some-
general, there is little interest in the information content of a cou- what different. They can also be joined together. For example, the
pled design because a coupled one basically means a bad design. uncertainty concept can be used in the objective function and the
On the other hand, the calculation of an uncoupled design is very reliability concept can be employed for the constraints of robust
simple. The total information content of an uncoupled design is the design. These days, application of the RSM is being actively devel-
sum of the information content of each functional requirement. To oped. The RSM has been used for approximation in robust design.
calculate the information content of a decoupled design, there are The RSM can be further investigated for robust design.
several approaches because the calculation is not simple. Suh20 used 3) In axiomatic design, application of the information axiom can
the conditional probability to calculate the information content of a achieve robust design. When the information content is minimized,
decoupled design. Oh (see Ref. 20) developed a graphical approach robust design is achieved and the loss function of the Taguchi method
when there are two functional requirements. Meanwhile, a general- is also minimized. Because axiomatic design is relatively new, there
ized method95−97 was developed by using the multi-integral in the are not many case studies. Also, the case studies mostly use the
FR or DP domain. This method may not be used frequently because independence axiom. Therefore, more research is needed for the
the calculation is very complicated. information axiom to achieve robust design. Robust optimization
Some researchers have been interested in the relationship between might be exploited in the decision making process in the application
axiomatic design and the Taguchi method. Bras and Mistree98 stud- of the information axiom.
190 PARK ET AL.

Acknowledgments 30 Fowlkes, W. Y., and Creveling, C. M., Engineering Method for Robust

This research was supported by the Center of Innovative De- Product Design, Addison Wesley Longman, Reading, MA, 1995.
31 Lee, K. H., “Robust Optimization in Continuous and Discrete Design
sign Optimization Technology, which was funded by the Korea Spaces for Structures,” Ph.D. Dissertation, Dept. of Mechanical Design and
Science and Engineering Foundation. The authors are thankful to Production Engineering, Hanyang Univ., Seoul, Republic of Korea, Feb.
MiSun Park for her correction of the manuscript. 1996 (in Korean).
32 Park, S. H., Quality Engineering, Min-Young Sa, Seoul, Republic of
References Korea, 2000 (in Korean).
1 Taguchi, 33 Hedayat, A. S., Sloane, N. J. A., and Stufken, J., Orthogonal Arrays,
G., Chowdhury, S., and Taguchi, S., Robust Engineering,
McGraw–Hill, New York, 2000. Springer, New York, 1999.
2 Suh, N. P., Axiomatic Design: Advances and Applications, Oxford Univ. 34 Dehnad, K., Quality Control, Robust Design, and the Taguchi Method,

Press, New York, 2001. Wadsworth and Brooks/Cole, Pacific Grove, CA, 1989.
3 Box, G. E. P., and Fung, C., “Is Your Robust Design Procedure Robust?,” 35 Box, G. E. P., “Signal-to-Noise Ratios, Performance Criteria, and Trans-

Quality Engineering, Vol. 6, No. 3, 1993, pp. 503–514. formation (with discussion),” Technometrics, Vol. 30, No. 1, 1988, pp. 1–17.
4 Shigley, J. E., and Mischke, C. R., Mechanical Engineering Design, 5th 36 Ku, K. J., Rao, S. S., and Chen, L., “Taguchi-Aided Search Method for

ed., McGraw–Hill, New York, 1989. Design Optimization of Engineering Systems,” Engineering Optimization,
5 Dieter, G. E., Engineering Design, McGraw–Hill, New York, 2000. Vol. 30, No. 1, 1998, pp. 1–23.
6 Hatamura, Y., The Practice of Machine Design, Oxford Univ. Press, New 37 Nair, V. H., “Taguchi’s Parameter Design: A Panel Discussion,” Tech-

York, 1999. nometrics, Vol. 34, No. 2, 1992, pp. 127–161.

7 Hyman, B., Fundamentals of Engineering Design, Prentice–Hall, Upper 38 Leon, R. V., Shoemaker, A. C., and Kacker, R. N., “Performance Mea-
Saddle River, NJ, 1998. sures Independent of Adjustment,” Technometrics, Vol. 29, No. 3, 1987,
8 Taguchi, G., “Introduction to Quality Engineering: Designing Quality
pp. 253–285.
into Products and Processes,” Asian Productivity Organization, Tokyo, 1986. 39 Box, G. E. P., Bisgaard, S., and Fung, C. A., “An Explanation and
9 Taguchi, G., Systems of Experimental Design, Vols. 1 and 2, Kraus In-
Critique of Taguchi’s Contributions to Quality Engineering,” Quality and
ternational, New York, 1987. Reliability Engineering International, Vol. 4, No. 2, 1988, pp. 123–131.
10 Phadke, M. S., Quality Engineering Using Robust Design, Prentice– 40 Montgomery, D. C., Design and Analysis of Experiments, Wiley,
Hall, Englewood Cliffs, NJ, 1989. Singapore, 1991, pp. 417–433.
11 Lee, K. H., Eom, I. S., Park, G. J., and Lee, W. I., “Robust Design for 41 Wu, F.-C., and Chyu, C.-C., “A Comparative Study on Taguchi’s SN
Unconstrained Optimization Problems Using the Taguchi Method,” AIAA Ratio, Minimizing MSD and Variance for Nominal-the-Best Characteristic
Journal, Vol. 34, No. 5, 1996, pp. 1059–1063. Experiment,” International Journal of Advanced Manufacturing Technol-
12 Lee, K. H., and Park, G. J., “Robust Optimization in Discrete De-
ogy, Vol. 20, No. 9, 2002, pp. 655–659.
sign Space for Constrained Problems,” AIAA Journal, Vol. 40, No. 4, 2002, 42 Welch, W. J., Yu, T. K., Kang, S. M., and Sacks, J., “Computer Ex-
pp. 774–780. periments for Quality Control by Parameter Design,” Journal of Quality
13 Ramakrishnan, B., and Rao, S., “A Robust Optimization Approach Us-
Technology, Vol. 22, No. 1, 1990, pp. 15–22.
ing Taguchi’s Loss Function for Solving Nonlinear Optimization Problems,” 43 Vining, G. G., and Myers, R. H., “Combining Taguchi and Response
Advances in Design Automation, Vol. 32-1, 1991, pp. 241–248. Surface Philosophies: A Dual Response Approach,” Journal of Quality Tech-
14 Belegundu, A. D., and Zhang, S., “Robustness of Design Through Min-
nology, Vol. 22, No. 1, 1990, pp. 38–45.
imum Sensitivity,” Journal of Mechanical Design, Vol. 114, June 1992, 44 Shoemaker, A. C., and Tsui, K. L., “Response Model Analysis for Ro-
pp. 213–217. bust Design Experiments,” Communications in Statistics, Part B, Vol. 22,
15 Sundaresan, S., Ishii, K., and Houser, D. R., “A Robust Optimization
No. 4, 1993, pp. 1037–1064.
Procedure with Variations on Design Variables and Constraints,” Engineer- 45 Kwon, Y. M., and Moon, I. S., “Simultaneous Optimization of Mul-
ing Optimization, Vol. 24, No. 2, 1995, pp. 101–117. tiple Responses to the Combined Array,” Journal of the Korean Data and
16 Parkinson, A., Sorensen, C., and Pourhassan, N., “A General-Approach
Information Science Society, Vol. 13, No. 1, 2001, pp. 57–64 (in Korean).
for Robust Optimal-Design,” Journal of Mechanical Design, Vol. 115, No. 1, 46 Shiau, G. H., “A Study on the Sintering Properties of Iron Ores Using the
1993, pp. 74–80. Taguchi’s Parameter Design,” Journal of the Chinese Statistical Association,
17 Parkinson, A., “Robust Mechanical Design Using Engineering Models,”
Vol. 28, 1990, pp. 253–275.
Journal of Mechanical Design, Vol. 117, 1995, pp. 48–54. 47 Tai, C. Y., Chen, T. S., and Wu, M. C., “An Enhanced Taguchi Method for
18 Parkinson, D. B., “Robust Design by Variability Optimization,” Quality
Optimizing SMT Processes,” Journal of Electronics Manufacturing, Vol. 2,
and Reliability Engineering International, Vol. 13, No. 2, 1997, pp. 97–102.
19 Parkinson, D. B., “The Application of a Robust Design Method to Toler- No. 3, 1992, pp. 91–100.
48 Tong, L.-I., and Su, C.-T., “Optimizing Multi-Response Problems in the
ancing,” Journal of Mechanical Design, Vol. 122, No. 2, 2000, pp. 149–154.
20 Suh, N. P., The Principles of Design, Oxford Univ. Press, New York, Taguchi Method by Fuzzy Multiple Attribute Decision Making,” Quality and
1990. Reliability Engineering International, Vol. 13, No. 1, 1997, pp. 25–34.
49 Pignatiello, J. J., “Strategies for Robust Multiresponse Quality En-
21 Hwang, K. H., and Park, G. J., “Development of a Robust Design Pro-
cess Using a Robustness Index,” Transactions of Korean Society of Mechan- gineering Experiments,” Institute of Industrial Engineers Transactions,
ical Engineers(A), Vol. 26, No. 10, 2002, pp. 2073–2081 (in Korean). Vol. 25, No. 3, 1993, pp. 5–25.
22 Hwang, K. H., “A Robust Design Methodology Using a Robustness In- 50 Tsui, K.-L., “Robust Design Optimization for Multiple Characteristic

dex Considering Design Space and Target Value,” Ph.D. Dissertation, Dept. Problems,” International Journal of Production Research, Vol. 37, No. 2,
of Mechanical Design and Production Engineering, Hanyang Univ., Seoul, 1999, pp. 433–445.
51 Wu, F.-C., “Optimization of Multiple Quality Characteristic Based on
Republic of Korea, Feb. 2003 (in Korean).
23 Ayyub, B. M., and Haldar, A., “Practical Structural Reliability Tech- Percentage Reduction,” International Journal of Advanced Manufacturing
niques,” Journal of Structural Engineering, Vol. 110, No. 8, 1984, Technology, Vol. 20, No. 10, 2002, pp. 749–753.
52 Liao, H.-C., and Chen, Y.-K., “Optimizing Multi-Response Problem in
pp. 1707–1724.
24 Youn, B. D., Choi, K. K., and Park, Y. H., “Hybrid Analysis Method the Taguchi Method by DEA Based Ranking Method,” International Journal
for Reliability-Based Design Optimization,” Journal of Mechanical Design, of Quality and Reliability Management, Vol. 19, No. 7, 2002, pp. 825–837.
53 Marler, R. T., and Arora, J. S., “Survey of Multi-Objective Optimization
Vol. 125, No. 2, 2003, pp. 221–232.
25 Gumbert, C. R., Newman, P. A., and Hou, G. J.-W., “High-Fidelity Methods for Engineering,” Structural and Multidisciplinary Optimization,
Computational Optimization for 3-D Flexible Wings and Colon; Part II— Vol. 26, No. 6, 2004, pp. 369–395.
54 Unal, R., Stanley, D. O., Engelund, W., and Lepsch, R., “Design for
Effect of Random Geometric Uncertainty on Design,” Optimization and
Engineering, Vol. 6, No. 1, 2005, pp. 139–156. Quality Using Response Surface Methods,” 6th AIAA/USAF/ISSMO Sym-
26 Arora, J. S., Introduction to Optimum Design, McGraw–Hill, New York, posium on Multidisciplinary Analysis and Optimization, Vol. 1, AIAA,
1989. Washington, DC, 1994, pp. 592–597.
27 Haftka, R. T., and Gurdal, Z., Elements of Structural Optimization, 55 Unal, R., Wu, K. C., and Stanley, D. O., “Structural Design Optimiza-

Kluwer Academic, 1992. tion for a Space Truss Platform Using Response Surface Methods,” Quality
28 Ronald, E. W., and Raymond, H. M., Probability and Statistics for Engineering, Vol. 9, No. 3, 1997, pp. 441–447.
56 Chen, W., Allen, J. K., Mistree, F., and Tsui, K. L., “Integration of Re-
Engineers and Scientists, 5th ed., Macmillan, New York, 1993.
29 Hademenos, G. J., Spiegel, M. R., and Liu, J., Mathematical Hand- sponse Surface Methods with the Compromise Decision Support Problem
book of Formulas and Tables: Based on Schaum’s Outline of Mathematical in Developing a General Robust Design Procedure,” ASME Design Automa-
Handbook of Formulas and Tables, edited by M. R. Spiegel and J. Liu, tion Conference, DE-Vol. 82-2, American Society of Mechanical Engineers,
McGraw–Hill, New York, 2001. New York, 1995, pp. 485–492.
 PARK ET AL. 191

57 d’Entremont, K. L., and Ragsdell, K. M., “Design for Latitude Using 78 Cornell, C. A., “A Probability Based Structural Code,” Journal of the
TOPT,” Advances in Design Automation, DE Vol. 14, American Society of American Concrete Institute, Vol. 66, No. 12, 1969, pp. 974–985.
Mechanical Engineers, New York, 1988, pp. 265–272. 79 Mohandas, S. U., and Sandgren, E., “Multiobjective Optimization Deal-
58 Chen, W., Allen, J. K., Tsui, K. L., and Mistree, F., “A Procedure ing with Uncertainty,” Proceedings of the 15th ASME Design Automation
for Robust Design: Minimizing Variations Caused by Noise Factors and Conference, DE-Vol. 19-2, American Society of Mechanical Engineers,
Control Factors,” Journal of Mechanical Design, Vol. 118, No. 4, 1996, 1989, pp. 241–248.
pp. 478–485. 80 Sandgren, E., 1989, “A Multi-Objective Design Tree Approach for Opti-
59 Rackwitz, R., and Christensen, P. T., Reliability and Optimization of mization under Uncertainty,” Proceedings of the 15th ASME Design Automa-
Structural Systems, Springer Verlag, 1992. tion Conference, DE-Vol. 19-2, American Society of Mechanical Engineers,
60 Rao, S. S., Engineering Optimization, Wiley, New York, 1996. 1989, pp. 249–255.
61 Siddall, J. H., “A New Approach to Probability in Engineering Design 81 Osyczka, A., Multicriterion Optimization in Engineering in FORTRAN
and Optimization,” Journal of Mechanisms, Transmissions, and Automation Programs, Ellis Horwood Ltd., West Sussex, England, U.K., 1984.
in Design, Vol. 106, No. 1, 1984, pp. 5–10. 82 Grandhi, R. V., and Bharatran, G., “Multiobjective Optimization of
62 Eggert, R. J., and Mayne, R. W., “Probabilistic Optimal Design Using Large-Scale Structures,” AIAA Journal, Vol. 31, No. 7, 1993, pp. 1329–1337.
Successive Surrogate Probability Function,” Journal of Mechanical Design, 83 Mistree, F., Hughes, O. F., and Bras, B. A., 1993, The Compromise De-
Vol. 115, No. 3, 1993, pp. 385–391. cision Support Problem and the Adaptive Linear Programming Algorithm,
63 Otto, K. N., and Antonsson, E. K., “Extensions to the Taguchi Method Structural Optimization: Status and Promise, edited by M. P. Kamat, AIAA,
of Product Design,” Journal of Mechanical Design, Vol. 115, No. 1, 1993, Washington, DC, 1993, pp. 247–286.
pp. 5–13. 84 Athan, T. W., and Papalambros, P. Y., “A Note on Weighted Criteria
64 Su, J., and Renaud, J. E., “Automatic Differentiation in Robust Opti- Methods for Compromise Solutions in Multi-Objective Optimization,” En-
mization,” AIAA Journal, Vol. 35, No. 6, 1997, pp. 1072–1079. gineering Optimization, Vol. 27, 1996, pp. 155–176.
65 Sundaresan, S., Ishii, K., and Houser, D. R., “A Procedure Using Man- 85 Chen, W., Wiecek, M. M., and Zhang, J., “Quality Utility-A Compro-
ufacturing Variance to Design Gears with Minimum Transmission Error,” mise Programming Approach to Robust Design,” Journal of Mechanical
Journal of Mechanical Design, Vol. 113, No. 3, 1991, pp. 318–324. Design, Vol. 121, No. 2, 1999, pp. 179–187.
66 Ang, A.H-S., and Tang, W. H., Probabilistic Concepts in Engineering 86 Messac, A., “Physical Programming Effective Optimization for Com-
and Design, Vol. I, Basic Principles, Wiley, New York, 1975. putational Design,” AIAA Journal, Vol. 34, No. 1, 1996, pp. 149–158.
67 Jung, D. H., and Lee, B. C., “Development of an Efficient Optimization 87 Messac, A., and Wilson, B., “Physical Programming for Computational
Technique for Robust Design by Approximating Probability Constraints,” Control,” AIAA Journal, Vol. 36, No. 2, 1998, pp. 219–226.
Transactions of Korean Society of Mechanical Engineers(A), Vol. 24, No. 12, 88 Messac, A., and Ismail-Yahaya, A., “Required Relationship Between
2000, pp. 3053–3060 (in Korean). Objective Function and Pareto Frontier Orders: Practical Implications,”
68 Jung, D. H., and Lee, B. C., “Development of an Optimization Tech- AIAA Journal, Vol. 39, No. 11, 2001, pp. 2168–2174.
nique for Robust Design of Mechanical Structures,” Transactions of Korean 89 Sobieski, J. J., Barthelemy, J., and Riley, K., “Sensitivity of Optimum
Society of Mechanical Engineers(A), Vol. 24, No. 1, 2000, pp. 215–224 (in Solutions of Problem Parameters,” AIAA Journal, Vol. 20, No. 9, 1982,
Korean). pp. 1291–1299.
69 Han, J. S., and Kwak, B. M., “Robust Optimal Design of a Vibratory Mi- 90 Vanderplaats, G. N., Numerical Optimization Techniques for Engi-
crogyroscope Considering Fabrication Errors,” Journal of Micromechanics neering Design, Vanderplaats Research and Development, Inc., Colorado,
and Microengineering, Vol. 11, No. 6, 2001, pp. 662–671. Springs, CO, 1984.
70 Myers, R. H., and Montgomery, D. C., Response Surface Methodology- 91 Chen, W., Allen, J. K., and Mistree, F., “A Robust Concept Exploration
Process and Product Optimization Using Designed Experiments, Wiley, New Method for Enhancing Productivity in Concurrent Systems Design,” Con-
York, 1995. current Engineering, Vol. 5, No. 3, 1997, pp. 203–217.
71 Hwang, K. H., Lee, K. H., Park, G. J., Cho, Y. C., Lee, B. Y., and 92 Shannon, C. E., and Weaver, W., The Mathematical Theory of Commu-
Lee, S. H., “Robust Design of the Vibratory Gyroscope with Unbalanced nication, Univ. of Illinois Press, Urbana, IL, 1949.
Inner Torsion Gimbal Using Axiomatic Design,” Proceedings of ICAD2002 93 Albano, L. D., and Suh, N. P., “The Information Axiom and its Implica-
Second International Conference on Axiomatic Design, Massachusetts Inst. tion,” Intelligent Concurrent Design: Fundamentals Methodology, Model-
of Technology, Boston, 2002. ing and Practice, Vol. 66-DE, American Society of Mechanical Engineers,
72 Hwang, K. H., Lee, K. H., Park, G. J., Lee, B. L., Cho, Y. C., and Lee, Nov. 1993, pp. 1–12.
S. H., “Robust Design of the Vibratory Gyroscope with Unbalanced Inner 94 El-Haik, B., and Yang, K., “The Components of Complexity in En-
Torsion Gimbal Using Axiomatic Design,” Journal of Micromechanics and gineering Design,” Institute of Industrial Engineers Transactions, Vol. 31,
Microengineering, Vol. 13, No. 1, 2003, pp. 8–17. No. 10, 1999, pp. 925–934.
73 Du, X., and Chen, W., “Methodology for Managing the Effect of Uncer- 95 Frey, D. D., Jahangir, E., and Engelhardt, F., “Computing the Infor-
tainty in Simulation-Based Systems Design,” AIAA Journal, Vol. 38, No. 8, mation Content of Decoupled Designs,” Research in Engineering Design-
2000, pp. 1471–1478. Theory Applications and Concurrent Engineering, Vol. 12, No. 2, 2000,
74 Du, X., and Chen, W., “Efficient Uncertainty Analysis Methods for pp. 90–102.
Multidisciplinary Robust Design,” AIAA Journal, Vol. 40, No. 3, 2002, 96 Park, G. J., Analytic Methods for Design Practice, Springer-Verlag (to
pp. 545–552. be published).
75 Putko, M. M., Taylor, A. C., Newman, P. A., and Green, L. L., “Ap- 97 Park, G. J., “Investigation of Taguchi’s Loss Function and the Probabil-
proach for Input Uncertainty Propagation and Robust Design in CFD Using ity of Success for Robust Design and Proposition of a New Index,” Hanyang
Sensitivity Derivatives,” Journal of Fluids Engineering, Vol. 124, No. 1, Univ., Internal Rept., Seoul, Republic of Korea, 2000.
2002, pp. 60–69. 98 Bras, B., and Mistree, F., “Concurrent Axiomatic and Robust Design
76 Putko, M. M., Newman, P. A., Taylor, A. C., and Green, L. L., “Ap- Using Compromise Decision Support Problems,” Concurrent Engineering,
proach for Input Uncertainty Propagation and Robust Design in CFD Using Vol. 8, 1994, pp. 17–31.
Sensitivity Derivatives,” AIAA Paper 2001-2528, June 2001.
77 Onwuiko, C. O., Introduction to Engineering Design Optimization, E. Livne
Prentice–Hall, Upper Saddle River, NJ, 2000. Associate Editor

View publication stats