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

net/publication/268574606

Alternatives to Taylor Series Approximation for the Variance Estimation in

Robust Design

Conference Paper · September 2010

DOI: 10.2514/6.2010-9083

CITATIONS READS
0 689

4 authors:

Markus Rippel Seung-Kyum Choi

Georgia Institute of Technology Georgia Institute of Technology
6 PUBLICATIONS 34 CITATIONS 74 PUBLICATIONS 1,382 CITATIONS

SEE PROFILE SEE PROFILE

Farrokh Mistree Janet Allen

University of Oklahoma University of Oklahoma
593 PUBLICATIONS 10,963 CITATIONS 425 PUBLICATIONS 8,604 CITATIONS

SEE PROFILE SEE PROFILE

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

Sustainable Rural Develoment View project

Realization of Multistage Manufacturing Processes View project

All content following this page was uploaded by Farrokh Mistree on 05 September 2016.

The user has requested enhancement of the downloaded file.

 13th AIAA/ISSMO Multidisciplinary Analysis Optimization Conference AIAA 2010-9083
13 - 15 September 2010, Fort Worth, Texas

Alternatives to Taylor Series Approximation for the Variance

Estimation in Robust Design

Markus Rippel1 and Seung-Kyum Choi2

Systems Realization Laboratory
G. W. Woodruff School of Mechanical Engineering
Georgia Institute of Technology
Savannah, GA 31407, USA

Farrokh Mistree3
School of Aerospace and Mechanical Engineering
Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

University of Oklahoma
Norman, Oklahoma 73019, USA

and

Janet K. Allen4
School of Industrial Engineering
University of Oklahoma
Norman, Oklahoma 73019, USA

When applying robust design methods, designers prefer to find solutions with a performance that
is insensitive towards uncertainty and noise. It is common to use the variance of the performance as
a measure for robustness. In many robust design methods the Taylor series expansion of the model
is used as an inexpensive way to estimate the variance. In this paper, it is demonstrated that this can
lead to fatal errors, especially for nonlinear and multimodal design spaces. Following the demonstra-
tion of this problem, three alternative methods for variance estimation are proposed. Their accuracy
is assessed within the Robust Concept Exploration Method (RCEM) and recommendations for their
application are given.

Nomenclature
d i− , d i+ = deviation variable in a compromise DSP
f(x) = performance function
L = length of pressure vessel middle section
P = internal pressure
R = radius of pressure vessel
sc = maximum stress in cylindrical section
ss = maximum stress is spherical section
sy = material yield strength
SF = safety factor
T = thickness of the pressure vessel wall
V = volume of pressure vessel
W = weight of pressure vessel
Wi = weight of ith objective
x = random design variable

1
MSME Graduate, [email protected]
2
Assistant Professor, AIAA Member
3
Professor, L.A. Comp Chair and Director, AIAA Associate Fellow
4
Professor, John and Mary Moore Chair, [email protected], AIAA Senior Member (corresponding author)
1
American Institute of Aeronautics and Astronautics

Copyright © 2010 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved.
 y = performance function value
z = noise factor
Z = objective function
μ, E[] = mean Value
ρ = material density
σ = standard deviation
σ2, Var[] = variance

I. Frame of Reference

T HE goal when applying robust engineering design methods is to improve a system’s quality by reducing its sen-
sitivity to uncertainty that has influence on the performance of the product. This design philosophy was estab-
lished by Genichi Taguchi1, a Japanese industrial consultant. Some robust design methods utilize global response
surfaces like the Response Surface Methodology (RSM)2 for modeling the performance of engineering systems.
These surrogate models are then used for uncertainty and sensitivity analysis. Often the first order Taylor series ex-
Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

pansion is used to estimate the performance variance and thus the robustness of the system3. An example for a
method that utilizes both of these techniques is the Robust Concept Exploration Method (RCEM)4. This design
method, has been employed successfully for a simple structural problem and design of a solar powered irrigation
system5, a High Speed Civil Transport (HSCT)6, and a General Aviation Aircraft7. In addition, RCEM has been
extended to facilitate the design of complex hierarchical systems8 and product platforms9. However, the application
of the first order Taylor series expansion for variance estimation can lead to grave errors if performance is a nonlin-
ear and multimodal within the design space.
Alternatives to the Taylor series expansion for the estimation of system robustness have been developed3; 10; 11
but have not been assessed in a framework like RCEM. These techniques build the basis for the alternative methods
for variance estimation introduced and assessed here.
In this paper, first the RCEM approach and the limitations of the first order Taylor series are explained and dem-
onstrated. In Section III three alternative techniques for variance estimation are introduced and assessed. The most
promising approach is incorporated into a modified RCEM and evaluated in Section IV for the design of a robust
pressure vessel. The paper is concluded with remarks that address the significance of the research presented and
ideas for future investigation.

II. The RCEM Approach and the Limitations of the Taylor Series for Variance Estimation
A foundational component for the research presented in this paper is the Robust Concept Exploration Method
(RCEM) introduced by Chen, et al. 4-6. The limitations of using the Taylor series for robust design are demonstrated
and the alternatives are assessed within this method. Therefore, the RCEM is introduced briefly in the following.
This section is concluded with the explanation of the Taylor series for variance estimation and the demonstration of
its limitations.

A. The Robust Concept Exploration Method (RCEM)

RCEM is a domain-independent, systematic approach that is formulated by integrating statistical
experimentation and approximate models, robust design techniques and multi-objective decision support. The
method can be divided into three steps. The three steps of RCEM are4:

Step 1: Build response surface models to relate each response to all important control- and noise- factors using the Re-
sponse Surface Methodology (RSM).
Step 2: Derive functions of mean and variance of the responses based on the type of robust design applications.
Step 3: Use the compromise DSP to find the robust design solution.

After constructing the surrogate model in the first step, the first-order Taylor series is used to derive the function
for the estimation of the performance variance 3 in Step 2. In Step 3, the compromise Decision Support Problem
(cDSP)12 is used as a multi-objective decision support tool for determining the values of design variables that satisfy
a set of constraints and balance a set of (conflicting) goals. These goals contain both performance and robustness
objectives. Families of robust designs can be generated simply by changing the priority levels or weights of goals or
by changing the target values for design requirements - none of which require reformulation of the cDSP.

2
American Institute of Aeronautics and Astronautics
 Although the RCEM has been successfully applied to several engineering design problems, it has potential
deficiencies due to the nature of the RSM and the Taylor series approximation. The research presented in this paper
is focused on the latter problem. For further details on the limitations of RSM in RCEM refer to Ref. 13.

B. First Order Taylor Series in Robust Design and its Limitations

In RCEM, the first-order Taylor series is used to estimate the output variance, which is used to formulate and
assess the robustness of the system to be designed. In this section the underlying mathematics and the thereon based
limitations are explained.
In robust design the variance of the performance under uncertainty is used as a measure of robustness. Variance
is defined as the expected squared difference between the mean and the random variable X:

[
Var( X ) = σ X2 = E ( X − μ ) 2 ] (1)
Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

The variance has the following property:

Var (cX 1 + dX 2 ) = c 2 Var ( X 1 ) + d 2 Var ( X 2 ) + 2cd Cov( X 1 , X 2 ) (2)

where c and d are constant factors, X1 and X2 are random variables and Cov(X1, X2) is their covariance. If the two
random variables are independent, the covariance is equal to zero and the third term of Eq. (2) can be ignored. The
first-order Taylor series expansion of the surrogate model or explicit performance function f about the point (a,b) is
defined as:

⎛ ∂f ( a , b ) ⎞ ⎛ ∂f ( a, b) ⎞
f ( x1 , x 2 ) ≈ f ( a, b) + ⎜ ⎟( x1 − a ) + ⎜ ⎟( x 2 − b ) (3)
⎝ ∂x1 ⎠ ⎝ ∂x 2 ⎠

The variance of the performance f(x1, x2) can be calculated using the property given in Eq. (2). Assuming that x1
and x2 are Gaussian with the variances σ x21 and σ x22 , the performance variance can be derived as follows:

2 2
⎛ ∂f (a, b) ⎞ ⎛ ∂f (a, b) ⎞
Var[ f ( x1 , x 2 )] ≈ Var[ f (a, b)] + ⎜ ⎟ Var ( x1 − a) + ⎜ ⎟ Var ( x2 − b) (4)
⎝ ∂ x1 ⎠ ⎝ ∂x2 ⎠

With Var[ f (a, b)] = 0 , Var ( x1 − a) = Var ( x1 ) and Var ( x2 − b) = Var ( x2 ) respectively, the estimation for the per-
formance variance using first order Taylor series expansion is:

2 2
⎛ ∂f ⎞ 2 ⎛ ∂f ⎞ 2
σ y2 ≈ ⎜ ⎟ σ x1 + ⎜ ⎟ σ x2 (5)
⎝ ∂x1 ⎠ ⎝ ∂x 2 ⎠

Equation (5) can be generalized for n design variables x and m noise factors z:

2 2
n
⎛ ∂f ⎞ 2 m
⎛ ∂f ⎞ 2
σ y2 ≈ ∑ ⎜ ⎟ σ xi + ∑ ⎜ ⎟ σ zi (6)
i =1 ⎝ ∂x i ⎠ i =1 ⎝ ∂z i ⎠

In this equation, simply the square of the derivative at the point of interest is multiplied with the variance of the
design variable or noise factor. This means that the steeper the function is at this point, the greater the expected
variance. This approach works as long as the function is nearly linear. When dealing with nonlinear, multimodal
functions, as is the case in many robust design problems, this approach may lead to grave errors. The biggest
problem occurs at local optima; here the derivative is equal to zero, which leads to a performance variance
estimation of zero, while the actual variance can be very high if it is a peak optimum. This is the reason why an
exploration method using the first order Taylor series expansion will converge to a local optimum (if there is one)

3
American Institute of Aeronautics and Astronautics
 instead of converging in a robust area. In Figure 1, an example function is presented with a local maximum, a local
minimum and a robust area marked. When applying a robust design method, the designer would expect the method
to return a design solution in area (c), whereas the method which utilizes the first-order Taylor series expansion for
variance estimation will converge in area (a) or (b) depending on whether a high or a low value of y is preferred. An
example of such an error is presented in the following section.

C. Example for the Presented Limitation

The example function used in this paper is taken from Ref. 10, where a similar problem is addressed. The func-
tion is given in Eq. (7) and a plot is shown in Figure 1.

9
f ( x) = ∑ ai ( x − 900)(i −1) with x = [910...976] (7)
i =1
Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

a1 = – 659.23 a4 = 0.82691 a7 = – 3.2446 x 10-6

a2 = 190.22 a5 = – 0.021885 a8 = 1.6606 x 10-8
a3 = – 17.802 a6 = 0.0003463 a9 = – 3.5757 x 10-11

This function is chosen so that it is representative for an engineering problem, which RCEM and other robust
design methods are designed for. The example is not complex; however, it has characteristics, many concept explo-
ration methods have difficulties with, it has (a) a global performance maximum, (b) a global minimum and (c) a
robust area, which are illustrated in Figure 1.
In this section, the original RCEM, with first-order Taylor series is applied to an illustrative design problem. A
design solution is considered better the lower value of the
performance function value and the lower the variance at this GIVEN
point are. By assigning a high weight to the robustness goal, Performance function f(x): Eq. (7)
the designer would then expect the robust solution to be in Derived estimations for variance σ f(x)2: Eq.
area (c) of Figure 1. However, with the example it will be (6)
shown that the exploration algorithm will find the optimum Variance of variable σx2 = 4 (Gaussian)
(b) as the “robust solution”. This is because, for this formu-
lation, this point is the best solution, since it has the best per- FIND
formance and the Taylor series leads to a variance estimation Design variable: x
of zero (maximum robustness). For this demonstration a Deviation variables: d i− , d i+ i = 1,2
variance of the design variable x of 4 is assumed. The target
value for the performance is set to 10 and the goal for the SATISFY
performance variance is 0.5. The weights assigned to the GOALS
goals are 0.1 and 0.9 respectively. This way it is assured, 1
that robustness is the dominating criteria for selecting a f ( x)
Performance goal: + d1− − d1+ = 1
satisficing solution. The cDSP for the example is given in 1
Figure 2. This cDSP is solved using a Java implementation 10
1
120
σ 2f ( x )
Robustness goal: + d 2− − d 2+ = 1
1
100
0 .5
(a)
BOUNDS
910 ≤ x ≤ 976
80

d i− , d i+ ≥ 0
(c)
60 i = 1,2
y

(b) d i− ⋅ d i+ = 0 i = 1,2
40

20
MINIMIZE
Z = W1 ⋅ d1− + W2 ⋅ d 2− with d1+ , d 2+ = 0,
0
910 920 930 940 950 960 970 980 W1 = 0.1, W2 = 0.9
x

Figure 1. Example Function with (a) Maximum, Figure 2. cDSP for Illustrative Expample
(b) Minimum and (c) Robust Area Problem
4
American Institute of Aeronautics and Astronautics
 Table 1. RCEM Results for Illustrative Example

Estimated Actual Deviation

Run Initial Deviation (LHS)
Solution
# Point
d1− d 2− d1− d 2−
1 x = 910 931.9484 0.2837 0 0.3211 0.5659
2 x = 912 911.9925 0.9107 0 0.9040 0.9619
3 x = 920 931.9526 0.2837 0 0.3211 0.5699
4 x = 940 931.9528 0.2837 0 0.3211 0.5699
5 x = 960 931.9528 0.2837 0 0.3211 0.5699
6 x = 976 931.9488 0.2837 0 0.3211 0.5657
Figure 3. Convergence Plot of Run #4

of DSIDES12 that allows solving cDSPs with explicit goal functions.

Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

The results of the robust design exploration for different starting points are given in Table 1. For evaluation of
the results, accurate reference values for the deviation function are generated using Latin hypercube sampling. For
experiment run #4 the convergence plot is shown in Figure 3, which allows for the conclusion that the algorithm
converged smoothly to the solution.
When comparing the results with Figure 1, it can be seen that the solutions do not lie in the robust area, but in the
local optima. For most of the starting points the algorithm converges in the global minimum at x = 931.9499. From
the results presented it can be also seen that the estimated deviation from the variance/robustness goal in the solution
points calculated with the Taylor series is significantly different from the actual deviation, which is calculated using
Latin hypercube sampling with 10,000 points. These results are exactly as expected; the use of the first-order Taylor
series expansion for variance estimation leads to the grave misconception that solutions in local optima are the most
robust ones.
The findings presented in this section raise the demand for alternative techniques for estimating the performance
variance and thus the robustness of an engineering system under uncertainty. Some of those alternatives are intro-
duced, explained and tested in the next section.

III. Alternative Variance Estimation Techniques

The work presented in this paper was guided by the question: How can the estimation of the performance vari-
ance be improved, especially around local optima? For answering this question, three possible formulations are
introduced in this section and their effectiveness is tested on a simple illustrative example. Upon the evaluation of
the results, the most promising formulation is chosen and will be used to modify the original RCEM. The three
methods are:
1. Variance estimation using second order Taylor Series Expansion (Second Derivative Method)
2. Variance estimation using first order Taylor Series Expansion at the nominal value and points around the
mean (Multiple Derivative Method)
3. Evaluation of the function values at and around the nominal value (Multiple Point Method)
Similar approaches can be found in literature, e.g., Ref. 10 and 11, however, a comprehensive comparison of
their effectiveness, especially for their application in a robust concept exploration method, is missing.

A. Alternative Methods
In the following, the three alternative techniques for variance estimation and their underlying mathematics are
explained. Their effectiveness is assessed thereafter.

1. Second Derivative Method

The Second Derivative Method is derived from the second order Taylor Series Expansion. In the previous sec-
tion, it was shown that the first order Taylor series fails as a good approximation, since it does not account for the
nonlinearity of the performance function. By considering the second derivative of the performance function, the
second order Taylor Series might have the potential to eliminate this problem. The cross product terms are ignored
in order to simplify the calculation in Eq. (8).

5
American Institute of Aeronautics and Astronautics
 2
n
⎛ ∂f ( xi0 ) ⎞ 1 n ⎛⎜ ∂ f ( xi0 ) ⎞⎟
y = f (x) ≈ f ( x10 ,..., x n0 ) + ∑ ⎜⎜ ⎟⎟( xi − xi0 ) + ∑ ( xi − xi0 ) 2 (8)
i =1 ⎝ ∂xi ⎠ 2 i =1 ⎜⎝ ∂xi2 ⎟⎠

With the assumption of independent variables as well as the properties of the variance in Eqs. (1) and (2), the
performance variance can be derived to be:

2 2 2 2 2
n
⎛ ∂f ⎞ 2 m
⎛ ∂f ⎞ 2 1 n ⎛ ∂ f ⎞ 4 1 m ⎛ ∂2 f ⎞ 4
σ y2 ≈ ∑ ⎜ ⎟ σ xi + ∑ ⎜ ⎟ σ zi + ∑ ⎜⎜ 2 ⎟⎟ σ xi + ∑ ⎜⎜ ⎟⎟ σ zi (9)
i =1 ⎝ ∂x i ⎠ i =1 ⎝ ∂z i ⎠ 4 i =1 ⎝ ∂x i ⎠ 4 i =1 ⎝ ∂z i2 ⎠

for the design variables, x, and noise factors, z.

2. Multiple Derivative Method

Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

The idea for developing the Multiple Derivative Method is to consider not only the nominal input values but also
the points around the mean. Therefore, the first order Taylor Series approximation is used for the variance predic-
tion (Eq. (6)) and is evaluated at three points for each design variable and noise factor. This formulation prevents
the estimation of performance of zero at local optima.

2 2
⎡ n
⎛ ∂f ⎞ 2 m
⎛ ∂f ⎞ 2 ⎤
σ y2 ≈ W1 ⎢∑ ⎜ ⎟ σ xi + ∑ ⎜ ⎟ σ zi ⎥
⎢⎣ i =1 ⎝ ∂x i ⎠ i =1 ⎝ ∂z i ⎠ ⎥⎦
⎡ n ⎛ ∂f (..., xi + Δx i ,...) ⎞ 2 2 m ∂f (..., z + Δz ,...) 2
⎛ ⎞ 2⎤
+ W 2 ⎢∑ ⎜ ⎟ σ xi + ∑ ⎜ i i
⎟ σ zi ⎥ (10)
⎣⎢ i =1 ⎝ ∂x i ⎠ i =1 ⎝ ∂z i ⎠ ⎥⎦
⎡ n ⎛ ∂f (..., x i − Δx i ,...) ⎞ 2 2 m ∂f (..., z − Δz ,...) 2
⎛ ⎞ 2⎤
+ W3 ⎢ ∑ ⎜ ⎟ σ xi + ∑ ⎜ i i
⎟ σ zi ⎥
⎣⎢ i =1 ⎝ ∂xi ⎠ i =1 ⎝ ∂z i ⎠ ⎦⎥

In Eq. (10), the formulation of the Multiple Derivative Method is shown, which is basically a weighted average
of the first order Taylor Series approximation at the input means (nominal x and z values) and edge points (xi±∆xi,
zi±∆zi). Since the formulation consists of three terms it is reasonable to divide the input distribution into three arrays
to determine Wi, ∆xi, ∆zi
If Gaussian distributions are assumed for the uncertain input, according to Figure 4, the following parameter val-
ues for Eq. (10) can be suggested:
x i = μ xi and z i = μ zi with W1 = 0.7
Δx i = 2σ xi with W2 = W3 = 0.15
The middle section in Figure 4 is represented by the mean and covers about 70% of the probability density func-
tion. The outer sections are represented by xi+∆xi and xi-∆xi respectively. These sections cover about 15% of the
PDF each.

Figure 4. Confidence Intervals of a Normal Distribution (modi-

fied from Ref. 14)
6
American Institute of Aeronautics and Astronautics
 3. Multiple Point Method

k ⎛ ( si1 ) 2 + ( si 2 ) 2 + w( si 3 ) 2 ⎞
∑ ⎜⎜ 3
⎟⎟
i =1 ⎝ ⎠
k ⎛ ( y − y − ) 2 + ( y − y + ) 2 + w( y − − y + ) 2 ⎞
= ∑ ⎜⎜ i i i i i i
⎟⎟
i =1 ⎝ 3 ⎠
1 k
= ∑ ( f ( x1 , x 2 ,..., xi ,..., xn ) − f ( x1 , x2 ,..., ( xi − Δxi ),..., x n ) ) 2 (11)
3 i =1
1 k
+ ∑
3 i =1
( f ( x1 , x2 ,..., xi ,..., xn ) − f ( x1 , x 2 ,..., ( xi + Δxi ),..., x n ) ) 2

1 k
Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

+ ∑ w( f ( x1 , x2 ,..., ( xi − Δxi ),..., xn ) − f ( x1 , x2 ,..., ( xi + Δxi ),..., xn )) 2

3 i =1

The Multiple Point Method was introduced in Ref. 10 and is based on the Sensitivity Index (SI) 11. In this
method, the weighted average of the performance difference between the points xi and xi±∆xi is calculated as in Eq.
(11) in order to estimate the level of robustness. In this equation, w is the weight for (si3)2. In Ref. 10 it is set to
0.25 because it is expected that

si 3 = ( yi− − yi+ ) = 2( yi − yi− ) = 2( yi − yi+ ) = 2 si1 = 2 si 2 (12)

and it is proposed the three terms should contribute equally to the goal formulation. However, no suggestions are
made on how to choose ∆xi. In this paper, 2σxi is used as ∆xi. Assuming that the inputs are Gaussian, this means
that 95.6% of the input values lie between xi+∆xi and xi-∆xi (see Figure 4). Using 2σxi for ∆xi means that using Eq.
(11) will lead to an estimation of 4σy2 and therefore the equation is modified to:

k ⎛ ( si1 ) 2 + ( si 2 ) 2 + w( si 3 ) 2 ⎞
σ y2 ≈ ∑ ⎜⎜ ⎟⎟
i =1 ⎝ 12 ⎠
k ⎛
( y − yi− ) 2 + ( yi − yi+ ) 2 + w( yi− − yi+ ) 2 ⎞
= ∑ ⎜⎜ i ⎟⎟
i =1 ⎝ 12 ⎠
1 k
= ∑ ( f ( x1 , x2 ,..., xi ,..., xn ) − f ( x1 , x2 ,..., ( xi − Δxi ),..., xn ) ) 2 (13)
12 i =1
1 k
+ ∑
12 i =1
( f ( x1 , x2 ,..., xi ,..., xn ) − f ( x1 , x2 ,..., ( xi + Δxi ),..., xn ) ) 2

1 k
+ ∑ w( f ( x1, x2 ,..., ( xi − Δxi ),..., xn ) − f ( x1, x2 ,..., ( xi + Δxi ),..., xn ) )2
12 i =1

with ∆xi=2σxi. Using Eq. (13), the value of the performance variance can be estimated and therefore this formula-
tion can be used to substitute the old robustness formulation in RCEM.
In this section, three methods for estimating the robustness are introduced. Whether these methods have the
potential to improve RCEM will be tested in the next section, where their performances are compared to Latin
hypercube samplings (LHS).

B. Assessment of the Alternative Variance Estimation Techniques

The three methods, which are introduced in the previous section, are assessed in this section using the example
function in Eq. (7). The ability of the three formulations to estimate the performance variance, and by this the
robustness of a design solution, are compared to reference values generated with the simulation of a Latin hypercube
sample. The findings are confirmed by evaluating the effectiveness of the three robustness formulations in a modi-

7
American Institute of Aeronautics and Astronautics
 fied RCEM framework. These results are again compared to Table 2: Comparison of the Accuracy in
those of Latin hypercube samples. The goal is to identify the Variance Estimation of the New Formulations
most promising formulation, which will be introduced to the
modified RCEM. Test with:
Method
In Section II, it is demonstrated that the local minimum σ x2 = 4 σ x2 = 9
of the example function is the critical point of the perform- Latin Hypercube Sampling σy2 = 1.32 σy2 = 7.52
ance function. The original RCEM fails in solving the Second Derivative Method σy2 = 0.08 σy2 = 0.41
Multiple Derivative Method σy2 = 3.19 σy2 = 18.68
design problem because its estimation of the performance
Multiple Point Method σy2 = 1.67 σy2 = 9.44
variance at this point is wrong. This is why, in this section,
the variance estimations of the three new methods in this
point are compared to the estimation generated from a Latin hypercube sample with 100,000 sample points. For the
comparison, Gaussian input distributions are assumed with standard deviations of 2 (σx2=4) and 3 (σx2=9) respec-
tively. The test point is the local minimum at x = 931.95. In Table 2, the performance variance estimations, σy2, for
the three new formulations and the normal distributed Latin hypercube sample are presented.
Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

From the results, it is seen that the Multiple Point Method

GIVEN
returns the variance estimates that are closest to the Latin
Performance function: Eq. (7)
hypercube outcomes, although they are still not very
accurate. Considering that the Multiple Point Method only Estimation for mean μ f ( x ) = f ( x)
needs three function calls, whereas the Latin hypercube Formulation for variance estimation σ 2f ( x ) :
sampling is done with 100,000 function calls, the Multiple
Point Method seems to be a promising augmentation to the a) Latin hypercube sampling
RCEM framework. The estimations of the Second b) Second Derivative Method
Derivative Method are too low to be considered a good c) Multiple Derivative Method
estimate. This is due to the formulation’s similarity to the d) Multiple Point Method
first order Taylor series approximation, which is equal to 0 at Variance of design variables σ x2 = 4
this point. The influence of the second derivative is too little
to be able to correct this result. In the formulation of the FIND
Multiple Derivative Method, the influence of the derivatives Design variable: x
at xi±2σ lead to an estimate that is too high in both examples. Deviation variables: d i− , d i+ i = 1,2
The results generated here are confirmed in the following.
The three new formulations are applied in a modified RCEM
SATISFY
framework for the estimation of the performance variance.
GOALS
In the example, the illustrative robust design problem is
1
solved with these methods. The cDSP for the example is f ( x)
given in Figure 5. Performance goal: + d1− − d1+ = 1
1
The results generated using a) Latin hypercube sampling 10
are presented first, in order to provide a basis for discussing 1
the results of the three methods to be tested. The results are σ 2f ( x )
generated using 10,000 LHS points to determine the Robustness goal: + d 2− − d 2+ = 1
1
performance mean and variance. In Table 3 and Table 4, the 0.5
results of the two scenarios are presented together with the BOUNDS
estimated deviation variables and function values. 910 ≤ x ≤ 976
The results are summarized in Figure 6, in which the d i− , d i+ ≥ 0 i = 1,2
locations of the solutions are shown. These results are taken − +
di ⋅ di = 0 i = 1,2
as reference values for the remaining section. Using 10,000
samples per algorithm iteration on a single design variable,
allows the assumption that these values are very similar to MINIMIZE
the actual values and results. Z = W1 ⋅ d1− + W2 ⋅ d 2− d1+ , d 2+ = 0
From these results it can be seen that the location of the Secnario 1 Scenario 2
solution depends on how the designer formulates the W1 = 0.5 W1 = 0.1
preferences through the weighting parameters. However, it W2 = 0.5 W2 = 0.9
also depends on the setting of the goal values. If, for
example, one goal value is set relatively far away from the Figure 5. cDSP for Illustrative Example
actual feasible performance, its deviation will only change Problem
insignificantly compared to an objective, for which the goal
is set close to the feasible area. This leads to a higher weight
8
American Institute of Aeronautics and Astronautics
 120
Table 3. Results for RCEM using the Latin
Performance Function
100 Solutions Using LHS
Hypercube Sampling (Scenario 1)

80 Run Initial Deviation

Solution
# Point d1- d2- Z
60 1 x = 910 910 0.888254 0.984668 0.936461
y

2 x = 912 932.7186 0.320906 0.498915 0.409911

40
3 x = 920 932.7188 0.320907 0.498914 0.409911
4 x = 940 932.7178 0.320901 0.498921 0.409911
20
5 x = 960 932.7189 0.320908 0.498913 0.409911
0
6 x = 976 932.7183 0.320904 0.498917 0.409911
910 920 930 940 950 960 970
x

Figure 6. Location of the Solutions Using Table 4. Results for RCEM using the Latin
Formulation a) Latin Hypercube Sampling Hypercube Sampling (Scenario 2)
Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

Run Initial Deviation

on the later objective, although the actual weighting # Point
Solution
d1- d2- Z
parameters in the deviation function remain 1 x = 910 951.67 0.6486 0.2389 0.2804
unchanged. It is very important that the designer is 2 x = 912 951.67 0.6486 0.2389 0.2804
aware of this phenomenon when using the cDSP to 3 x = 920 951.67 0.6486 0.2389 0.2804
formulate and solve engineering design problems. In 4 x = 940 951.67 0.6486 0.2389 0.2804
5 x = 960 951.67 0.6486 0.2389 0.2804
this example, Scenario 1 represents a situation where
6 x = 976 951.67 0.6486 0.2389 0.2804
the designer assigns equal weights on performance and
robustness, which results in a solution close to the
minimum. This means that increasing the robustness would not outweigh the loss in performance when moving the
solution towards the robust area. This is different in Scenario 2, where the focus is on finding a robust solution with
only little weight on performance. In this case, the gain in robustness in the robust area is valued higher than the
loss of performance. The outlier at x = 910 can be explained with an unfortunate selection of the starting point; the
maximum at around x = 912 prevents the exploration algorithm from finding the points on its right side. This can be
seen in for the following examples as well.
In the following tests, the three approximation methods are assessed using the same example and identical sce-
narios. This will show which methods are capable of suggesting the best compromises for each scenario. In Sce-
nario 1, the methods should suggest a solution
close to the minimum at around x = 932 and in Table 5. Results for RCEM using the Second Derivative
Scenario 2, a solution around x = 952 would Method (Scenario 1)
confirm the capability of the method.
The results for using RCEM with variance Run Initial Estimated Deviation Actual Deviation
Solution
goal formulation b) Second Derivative Method # Point d1- d2- d1- d2-
are given in Table and Table 6. From the 1 x = 910 910 0.9023 0.9991 0.8883 0.9995
results it can be seen that the Second Derivative 2 x = 912 910 0.9023 0.9991 0.8883 0.9995
Method (second order Taylor series) does not 3 x = 920 932.14 0.2841 0.1049 0.3201 0.5762
lead the designer to the robust area (see Figure 4 x = 940 932.14 0.2841 0.1049 0.3201 0.5762
5 x = 960 932.14 0.2841 0.1049 0.3201 0.5762
6) in the second Scenario. Using a modified 6 x = 976 932.14 0.2841 0.1049 0.3201 0.5762
RCEM with the Second Derivative Method as
the robustness formulation leads to the same
Table 6. Results for RCEM using the Second Derivative
incorrect results as the original RCEM.
- Method (Scenario 2)
From the very small values of d2 it can be
seen that the Second Derivative Method gives a Run Initial Estimated Deviation Actual Deviation
very low estimate of the performance variance. Solution
# Point d1- d2- d1- d2-
This is because the mathematical formulation 1 x = 910 932.14 0.2841 0.1049 0.3201 0.5762
that is derived from the second order Taylor 2 x = 912 932.14 0.2841 0.1049 0.3201 0.5762
Series approximation is very similar to the 3 x = 920 932.14 0.2841 0.1049 0.3201 0.5762
4 x = 940 932.14 0.2841 0.1049 0.3201 0.5762
original formulation in RCEM (first order 5 x = 960 932.14 0.2841 0.1049 0.3201 0.5762
Taylor Series). Note that the second derivative 6 x = 976 932.14 0.2841 0.1049 0.3201 0.5762
in Eq. (9) has a weight of only 0.25 and the first
two terms are identical to the variance
estimation used in the original RCEM. Thus,

9
American Institute of Aeronautics and Astronautics
 the additional terms only have very little Table 7. Results for RCEM using the Multiple
influence on the performance variance Derivative Method (Scenario 1)
estimation and the formulation fails for the
same reason as the first order Taylor Series. Run Initial Estimated Deviation Actual Deviation
Solution
The results for using RCEM with variance # Point d1- d2- d1- d2-
goal formulation c) Multiple Derivative Method 1 x = 910 910 0.9023 0.9997 0.8883 0.9995
are given in Table 7 and Table 8. It can be seen 2 x = 912 910 0.9023 0.9997 0.8883 0.9995
3 x = 920 951.62 0.6484 0.2724 0.6484 0.2392
that, by using the Multiple Derivative Method,
4 x = 940 951.62 0.6484 0.2724 0.6484 0.2392
the designer would be led to solutions in the 5 x = 960 951.62 0.6484 0.2724 0.6484 0.2392
robust area in both scenarios. This means that 6 x = 976 951.62 0.6484 0.2724 0.6484 0.2392
the method leads to an estimation of the
variance in the robust area that is lower than the Table 8. Results for RCEM using the Multiple
actual variance. Furthermore, around the Derivative Method (Scenario 2)
minimum, the variance is estimated to be higher
Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

than the actual performance variance, which Run Initial Estimated Deviation Actual Deviation
Solution
makes this method useful if the designer puts a # Point d1- d2- d1- d2-
high weight on the robustness when looking for 1 x = 910 951.69 0.6487 0.2722 0.6487 0.2392
a compromise. 2 x = 912 951.69 0.6487 0.2722 0.6487 0.2392
3 x = 920 951.69 0.6487 0.2722 0.6487 0.2392
The results for using RCEM with variance 4 x = 940 951.69 0.6487 0.2722 0.6487 0.2392
goal formulation d) Multiple Point Method are 5 x = 960 951.69 0.6487 0.2722 0.6487 0.2392
given in Table 9 and Table 10. The results 6 x = 976 951.69 0.6487 0.2722 0.6487 0.2392
present very promising. When comparing them
to the reference results in Table 3, Table 4 and Table 9. Results for RCEM using the Multiple Point
Figure 6, it can be seen that the solution for Method (Scenario 1)
both scenarios lie very close to the solutions
generated with LHS. The method leads to a Run Initial
Solution
Estimated Deviation Actual Deviation
good estimation of the variance in the robust # Point d1- d2- d1- d2-
area as well as in the area around an optimum. 1 x = 910 910 0.9023 0.9996 0.8883 0.9995
2 x = 912 910 0.9023 0.9996 0.8883 0.9995
This can be attributed to the mathematical
3 x = 920 932.96 0.2331 0.5633 0.3228 0.5074
formulation of this method, which accounts for 4 x = 940 932.96 0.2331 0.5633 0.3228 0.5074
nonlinearity by evaluating the performance 5 x = 960 932.96 0.2331 0.5633 0.3228 0.5074
differences between the mean and the tail ends 6 x = 976 932.96 0.2331 0.5633 0.3228 0.5074
of the input distribution as well as the
difference between both tail ends. Particularly, Table 10. Results for RCEM using the Multiple Point
this second aspect leads to good variance Method (Scenario 2)
estimations around local optima.
Comparing results of the three introduced Run Initial
Solution
Estimated Deviation Actual Deviation
variance estimation techniques, it can be # Point d1- d2- d1- d2-
concluded that the Multiple Point Methods is 1 x = 910 951.69 0.6487 0.2757 0.6487 0.2391
2 x = 912 951.69 0.6487 0.2757 0.6487 0.2391
most accurate and reliable when dealing with 3 x = 920 951.69 0.6487 0.2757 0.6487 0.2391
nonlinear, multimodal performance functions. 4 x = 940 951.69 0.6487 0.2757 0.6487 0.2391
It is proposed to use this method in a modified 5 x = 960 951.69 0.6487 0.2757 0.6487 0.2391
RCEM instead of the first order Taylor series 6 x = 976 951.69 0.6487 0.2757 0.6487 0.2391
expansion.
In the next section this modified RCEM is applied to a robust design problem for the validation of its effective-
ness and efficiency for design problems with
more than one design variables.

IV. Design of a Robust Pressure Vessel

The purpose of this example is the validation
of the modified RCEM. With the example, it is
intended to demonstrate the effectiveness of the
method for multi-objective robust engineering
design problems. The design of a pressure vessel
has been investigated previously by authors in
Figure 7. Schematic of the Cylindrical Pressure Vessel 15
10
American Institute of Aeronautics and Astronautics
 several areas of engineering design 15-18. It is very suitable for showing the effectiveness of the modified RCEM
because it fulfills the following requirements:

- all performance functions are known explicitly; no simulations are required

- the design has multiple design variables
- the design has multiple objectives

The design variables in this example are the radius of the vessel (R), the length of the middle section (L) and the
thickness of the vessel wall (T). A schematic of the pressure vessel is given in Figure 7.
The application requirements demand a vessel that is able to contain a gas under 1,000 psi of pressure with a
minimum volume of 20,000 in3 and a maximum vessel weight of 2,500 lb. The material is UNS G10100 HR, with a
yield strength of 26,000 psi and a density of 0.284 lb/in3. The bounds on the radius R are 6 inch and 36 inch respec-
tively. The middle section L of the vessel has to be between 6 inch and 48 inch and the thickness T between 1.5 and
6 inches. Due to space limitations, the overall radius (R+T) is required to be less than 40 inch and the overall length
Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

(L+2R+2T) less than 60 inch. Furthermore, the radius has to be at least 5 times greater than the thickness of the ves-
sel. Due to inaccuracy and uncertainty in the manufacturing process, deviations in the input values of the design
variables can be observed. For this example, these deviations are assumed to be Gaussian. The variance of the
radius and length of the middle section are 0.444 in2; the variance of the thickness is 0.111 in2. Industry regulations
specify that the design has to comply with a minimum safety factor, SF, of 2 for the maximum stress appearing in
both the cylindrical and spherical sections of the vessel (Eqs. (14) and (15)) (see Ref. 19).

⎡ ( R + T )2 + R 2 ⎤ ⎡ 2 R 2 + 2 RT + T 2 ⎤ s y
sc = P ⎢ 2 2 ⎥ = P ⎢ 2 ⎥ ≤ SF (14)
⎣(R + T ) − R ⎦ ⎣ 2 RT + T ⎦

R + 0.5T ⎞ s y
s s = P⎛⎜ ⎟≤ (15)
⎝ 2T ⎠ SF

The the performance objectives, in this example, are the volume (Eq. (16)) and the weight (Eq. (17)) of the pres-
sure vessel.

V ( R, L ) = π ( 43 R3 + R 2 L ) [in3 ] (16)

[ ]
W ( R, L, T ) = πρ 43 ( R + T )3 + ( R + T )2 L − ( 43 R3 + R 2 L ) [lb] (17)

With the Multiple Point Method the variance of the Volume and the Weight are calculated as shown in Eqs. (18)
and (19) respectively..

σ V2 ≈ 121 [V ( R, L ) − V (( R − 2σ R ), L )]2 + 121 [V ( R, L ) − V ( R, ( L − 2σ L ))]2

+ 121 [V ( R, L ) − V (( R + 2σ R ), L )]2 + 121 [V ( R, L ) − V ( R, ( L + 2σ L ))]2
(18)
+ 1
48
[V (( R − 2σ R ), L ) − V (( R + 2σ R ), L )]2
+ 1
48
[V ( R, ( L − 2σ L )) − V ( R, ( L + 2σ L ))]2

σ W2 ≈ 121 [W ( R, L, T ) − W (( R − 2σ R ), L, T )]2 + 121 [W ( R, L, T ) − W ( R, ( L − 2σ L ), T )]2

+ 121 [W ( R, L, T ) − W ( R, L, (T − 2σ T ))]2
+ 121 [W ( R, L, T ) − W (( R + 2σ R ), L, T )]2 + 121 [W ( R, L, T ) − W ( R, ( L + 2σ L ), T )]2 (19)

+ 121 [W ( R, L, T ) − W ( R, L, (T + 2σ T ))]2 + 1
48
[W (( R − 2σ R ), L, T ) − W (( R + 2σ R ), L, T )]2
+ 1
48
[W ( R, ( L − 2σ L ), T ) − W ( R, ( L + 2σ L ), T )]2 + 48
1
[W ( R, L, (T − 2σ T )) − W ( R, L, (T + 2σ T ))]2

11
American Institute of Aeronautics and Astronautics
 The design objectives (goals) in this example are:
- High vessel volume mean μV (target: 50,000 in3)
- Low vessel volume variance σV2 (target: 5,000,000 in6)
- Low vessel weight variance σW2 (target: 200,000 lb2)
Only the aforementioned conditions, bounds and constraints are considered in this design example and a con-
straint violation of 1% is allowed, which is the default setting in the Java implementation of DSIDES. All other
sources of variability like temperature and material imperfections are ignored.
The information given above is summarized in the formulation of a cDSP, which is presented in Figure 8. There
are two scenarios, for which the example is calculated. Scenario 1 represents the optimization for maximum vol-
ume, which is equivalent to assigning a weight of 1.0 to the volume
goal and weights of 0.0 to the robustness goals. This way, only the GIVEN
deviation to the volume goal is minimized. In Scenario 2, the System goal functions:
designer is asked to find a robust pressure vessel. This means that Volume V: Eq. (16)
2 2
the designer is looking for a compromise among the three goals. Variances of V and W: σV , σW
Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

This is realized by assigning the following weights to the three FIND

deviation variables: W1=0.45, W2=0.25, W3=0.30. The mindset Design variables: R, L, T
behind choosing these values is that the robustness should have a
Deviation variables: d i− , d i+ i = 1,2,3
larger weight (0.25+0.30=0.55) than the performance (0.45).
The cDSP is solved using a Java implementation of DSIDES. SATISFY
The DSIDES results of the modified RCEM for design scenario 1 CONSTRAINTS
are presented in Table 10. From the results it can be seen that the R + T ≤ 40 in
DSIDES algorithm is sensitive towards the selection of the initial L + 2 R + 2T ≤ 60 in
point. This is a known limitation of many gradient based search R ≥ 5T
algorithms and is not further investigated in this research. For this sy sy
design problem the best results could be achieved in run # 5. The sc ≤ , ss ≤
SF SF
quality of a particular solution can be read from the deviation value GOALS
Z. The smaller the deviation from the goal(s) the more satisfactory Volume:
is the solution. The fact that the algorithm converged in similar
μ V ( R, L ) − +
points supports the internal consistency of the results. 3
+ d1 − d1 = 1
In Table 11, the DSIDES results of the modified RCEM for 50,000in
Scenario 2 are given. Again, it can be seen that the convergence of Robustness:
the algorithm depends on the initial point. The internal consistency 5,000,000in
6
− +
of the results can be concluded from the observation that the results 2
+ d2 − d2 = 1
of the different runs are almost the same and all show a smooth σV
convergence. The design solutions with the lowest deviation from 200,000lb
2
− +
the goals are the results from runs # 1 and 2. However, here the 2
+ d3 − d3 = 1
volume constraint is violated to the maximum extent of 1%. The σ W
solutions of the remaining runs do not violate the volume con- BOUNDS
straint. 6.0 ≤ R ≤ 36.0 in
A comparison of the solutions of the runs # 5 for both scenarios 6.0 ≤ L ≤ 48.0 in
is given in Table 12. The results are as expected. The design solu- 1.5 ≤ T ≤ 6.0 in
tions for Scenario 2 have a lower variance of the volume and the d i− ⋅ d i+ = 0 ; d i− , d i+ ≥ 0 i = 1,2,3
weight than the designs in Scenario 1. Since the Scenario 2 designs
are a compromise between performance and robustness, the volume MINIMIZE
in those designs is lower, yet still within the given constraints. Z = W1d1− + W2 d 2− + W3 d 3−
The computational costs of these modified RCEM runs are
almost negligible. The runtime to test one initial point is less than
0.1 seconds. Scenario 1: Scenario 2:
From the presented results it can be seen that the method is very
efficient and that the Scenario 2 solutions are indeed more robust
than the Scenario 1 designs. For the validation of the effectiveness W1 = 1 W1 = 0.45
and the efficiency, however, some additional data is required. The W2 = 0 W2 = 0.25
accuracy of the results has to be verified and the efficiency has to W3 = 0 W3 = 0.30
be compared to alternative methods. In the following, the results of
the modified RCEM are verified by comparing them to Latin Figure 8. cDSP for the Design of the
hypercube samplings with a sample size of 100,000. For three Robust Pressure Vessel
12
American Institute of Aeronautics and Astronautics
 uncertain variables (R, L, T), this sample Table 10. Scenario 1 Results (modified RCEM)
size is considered large enough to get
results that are very close to the real Estimated Estimated
values. The DSIDES solutions for the JavaDSIDES Estimated Estimated
Run Initial point Variance of Variance of
Solution Volume Deviation
Latin hypercube method are presented # {R,L,T} in [in]
{R,L,T} in [in] [in3]
Volume Weight
Z
in Table 13. [in6] [lb2]
1 { 6, 6, 1.5} {18.72, 8.79, 1.5} 37,281 13,755,928 376,214 0.2543
Comparing Table 12 (modified
2 { 6, 25, 6} {18.72, 8.79, 1.5} 37,281 13,755,928 376,214 0.2543
RCEM) and Table 13 (LHS) it is 3 {20, 25, 1.5} {18.72, 9.27, 1.5} 37,809 14,030,351 383,336 0.2438
observed that the solutions are very 4 {20, 25, 4} {18.91, 8.42, 1.5} 37,913 14,054,754 383,268 0.2417
similar. The solutions are exactly the 5 {20, 48, 4} {18.91, 8.41, 1.5} 37,914 14,054,950 383,264 0.2417
same, only the estimations for the per- 6 {36, 6, 4} {18.73, 9.25, 1.5} 37,811 14,030,969 383,334 0.2438
formance variance are slightly different, 7 {36, 48, 6} {18.72, 9.27, 1.5} 37,809 14,030,324 383,335 0.2438
however with a difference of fless than
0.5 %, the new method can be consid- Table11. Scenario 2 Results (modified RCEM)
Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

ered very accurate.

These results allow for the conclu- JavaDSIDES Estimated
Estimated Estimated
Estimated
sion that the Multiple Point Method is a Run Initial point Variance of Variance of
Solution Volume Deviation
# {R,L,T} in [in] 3 Volume Weight
superior variance estimation technique {R,L,T} in [in] [in ]
[in6] [lb2]
Z
to the Taylor series expansion and is 1 { 6, 6, 1.5} {15.38, 6.0, 1.5} 19,800 5,904,472 176,411 0.31009
thus is proposed to be used in RCEM. 2 { 6, 25, 6} {15.38, 6.0, 1.5} 19,800 5,904,472 176,411 0.31009
The Multiple Point Method is applicable 3 {20, 25, 1.5} {15.44, 6.0, 1.5} 20,000 5,983,542 178,477 0.31109
for nonlinear, multimodal design prob- 4 {20, 25, 4} {15.44, 6.0, 1.5} 20,000 5,983,542 178,477 0.31109
lems, where the Taylor series might lead 5 {20, 48, 4} {15.44, 6.0, 1.5} 20,000 5,983,542 178,477 0.31109
6 {36, 6, 4} {15.44, 6.0, 1.5} 20,000 5,983,542 178,477 0.31109
to grave errors. However, it has to be be 7 {36, 48, 6} {15.44, 6.0, 1.5} 20,000 5,983,542 178,477 0.31109
considered that this modified RCEM is
only capable of solving design problems
Table 12. Comparison of Scenario 1 and 2 Results (modified
where the performance function is
RCEM)
known explicitly or where a response
surface can be built.
Estimated Estimated
JavaDSIDES Estimated
Variance Variance
V. Closing Remarks Scenario Solution Volume
of Volume of Weight
In this paper, it is shown that the {R,L,T} in [in] [in3]
[in6] [lb2]
first order Taylor series approximation Scenario 1 {18.91, 8.41, 1.5} 37,914 14,054,950 383,264
for the estimation of a system’s robust- Scenario 2 {15.44, 6.0, 1.5} 20,000 5,983,542 178,477
ness is critical when used in an iterative
exploration algorithm like in RCEM.
Table 13. Comparison of Scenario 1 and 2 Results (LHS))
Although this has been realized by
researchers for other applications
before, it was not demonstrated for Estimated Estimated
JavaDSIDES Estimated
methods like RCEM. In fact, several Variance Variance
Scenario Solution Volume
students and researchers already worked 3 of Volume of Weight
{R,L,T} in [in] [in ] 6
on developing these methods without [in ] [lb2]
noticing this gap. Scenario 1 {18.91, 8.41, 1.5} 37,914 N/A N/A
With the development of the modi- Scenario 2 {15.44, 6.0, 1.5} 20,000 5,966,442 179,148
fied RCEM, several alternatives to the
criticized Taylor series are demonstrated and assessed. The Multiple Point Method, which is introduced to the
modified RCEM, is a modification of the Sensitivity Index. It is augmented for the specific estimation of the per-
formance variance given Gaussian distributed design variables and noise factors. These augmentations make it a
superior substitute to the Taylor series expansion for variance estimation.
It can be certainly argued that the methods for the variance estimation both in the original RCEM as well as in
the modified RCEM have little computational advantage over a brute force method like Monte Carlo or Latin hyper-
cube when an explicit function or an exact surrogate model is available. With today’s available computational
power, those brute force methods are very fast when applied to explicit functions. In future research, it might be
worth investigating in which cases estimation is advantageous and for which scenarios brute force methods are the
better choice. Here, the computational efficiency as well as the methods’ accuracy, especially in nonlinear, multi-

13
American Institute of Aeronautics and Astronautics
 modal design problems are the crucial factors that have to be evaluated. However, to date, the Taylor series expan-
sion is still used for variance estimation in robust design, and thus the findings presented in this paper are important.
By substituting the Taylor series expansion with the Multiple Point Method introduced in this paper these methods
can be improved significantly with the additional advantage that no derivative of the performance function is
needed.

Acknowledgments
Financial support from NSF grant DMI-060259 SACE: Statistics Aided Computer Experiments, Georgia Tech
Savannah and The Center for Computational Materials Design is gratefully acknowledged. During his studies at
Georgia Tech, Markus Rippel was also supported by the German Academic Exchange Service (DAAD) and the
Federation of German-American Clubs (VDAC). We are grateful to Matthew Marston for providing a Java version
of DSIDES and we acknowledge the help from Wei Chen.
Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

References
1
Taguchi, G., Introduction to Quality Engineering: Designing Quality into Products and Processes, Asian Productivity
Organization, Tokyo, 1986.
2
Myers, R. H. and Montgomery, D. C., Response Surface Methodology: Process and Product Optimization Using Designed
Experiments, Wiley, New York, 1995.
3
Phadke, M. S., Quality Engineering Using Robust Design, Prentice Hall, Englewood Cliffs, NJ, 1989.
4
Chen, W., Allen, J. K., Tsui, K.-L. and Mistree, F., "A Procedure for Robust Design: Minimizing Variations Caused by
Noise Factors and Control Factors," ASME Journal of Mechanical Design, Vol. 118, No. 4, 1996, pp. 478-85.
5
Chen, W. A Robust Concept Exploration Method for Configuring Complex Systems, Ph.D. Dissertation, G. W. Woodruff
School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA, USA, 1995.
6
Chen, W., Allen, J. K., Mavris, D. and Mistree, F., "A Concept Exploration Method for Determining Robust Top-Level
Specifications," Engineering Optimization, Vol. 26, No. 2, 1996, pp. 137-58.
7
Simpson, T. W., Chen, W., Allen, J. K. and Mistree, F., "Conceptual Design of a Family of Products through the Use of the
Robust Concept Exploration Method," 6th AIAA/USAF/NASA/ISSMO Symposium on Multidisciplinary Analysis and Optimiza-
tion, Bellevue, WA, 1996, AIAA Paper Number 96-4161.
8
Koch, P. N. Hierarchical Modeling and Robust Synthesis for the Preliminary Design of Large Scale Complex Systems, PhD
Dissertation, G. W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA, USA, 1997.
9
Simpson, T. W., Maier, J. R. A. and Mistree, F., "Product Platform Design: Method and Application," Research in
Engineering Design, Vol. 13, No. 1, 2001, pp. 2-22.
10
Lin, Y., Krishnapur, K., Allen, J. K. and Mistree, F., "Robust Design: Goal Formulations and a Comparison of Metamod-
eling Methods," ASME Design Engineering Technical Conferences, Las Vegas, Nevada, 1999, ASME, DETC99/DAC-8608.
11
Sundaresan, S., Ishii, K. and Houser, D. R., "A Robust Optimization Procedure with Variations in Design Variables and
Constraints," Engineering Optimization, Vol. 24, No., 1995, pp. 101-17.
12
Mistree, F., Hughes, O. F. and Bras, B. A., "The Compromise Decision Support Problem and the Adaptive Linear Program-
ming Algorithm," Structural Optimization: Status and Promise, edited by M. P. Kamat, AIAA, Washington, D.C., 1993, pp.
247-86.
13
Rippel, M. Improved Robustness Formulations and a Simulation-Based Robust Concept Exploration Method, Master's
Thesis, G. W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA, USA, 2009.
14
Kemp, J., "Standard Deviation," http: // en.wikipedia.org/w/index.php?title=Standard_deviation&oldid=324120010
(accessed 11/2009), 2005, Wikipedia, The Free Encyclopedia.

14
American Institute of Aeronautics and Astronautics
 15
Hernández, G. "Platform Design for Customizable Products as a Problem of Access in Geometric Space," Ph.D. Disserta-
tion, G.W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA, 2001.
16
Sandgren, E., "Nonlinear Integer and Discrete Programming in Mechanical Design Optimization," Journal of Mechanical
Design, Vol. 112, No. 2, 1990, pp. 223-29.
17
Lewis, K. and Mistree, F., "Collaborative, Sequential, and Isolated Decisions in Design," Journal of Mechanical Design,
Vol. 120, No. 4, 1998, pp. 643-52.
18
Simpson, T. W. "A Concept Exploration Method for Product Family Design," Ph.D. Dissertation, The G.W. Woodruff
School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA, 1998.
19
Roark, R. J. and Young, W. C., Formulas for Stress and Strain, McGraw-Hill, 1975.
Downloaded by UNIVERSITY OF OKLAHOMA on September 5, 2016 | http://arc.aiaa.org | DOI: 10.2514/6.2010-9083

15
American Institute of Aeronautics and Astronautics

View publication stats