Proceedings of IMECE2008

2008 ASME International Mechanical Engineering Congress andIMECE2008

Exposition
October 31-November 6, 2008, Boston, Massachusetts, USA
2008 International Mechanical Engineering Conference & Exposition
October 31 - November 6, 2008, Boston, Massachusetts, USA

IMECE2008-68840
IMECE2008-68840

Design of Experiments as Effective Design Tools

Mohamed Nasser and Badih Jawad

Lawrence Technological University
Mechanical Engineering Department
21000 West Ten Mile Road
Southfield, MI 48075

ABSTRACT to better understand a process and to determine how

input variations affect the outputs. Design of
Design of Experiment (DOE) provides a highly structured Experiments is a standard statistical technique used in
way to study the effects of multiple variables on product quality engineering, pharmaceuticals, manufacturing,
performance as well as efficient and effective methods and other industries to identify key factors and levels that
for determining the most significant factors and influence system performance and variability. This
interactions in a given design problem. Design of technique is especially useful when there is the need to
Experiments (DOE) is an off line quality improvement understand the interactions and effects of several
methodology that dramatically improves industrial system variables and an absence of concrete
products and processes. Input factors are varied in a information. In industry, designed experiments can be
planned manner to optimize output responses with used to systematically investigate the process or product
minimal variability. Design of Experiments is a standard variables that influence product quality.
statistical technique used in quality engineering,
pharmaceuticals, manufacturing, and other industries to Design of Experiments (DOE) is extensively used in
identify key factors and levels that influence system research and development, where a large percentage of
performance and variability. This technique is especially engineering resources are spent on solving optimization
useful when there is the need to understand the problems. Conducting as few experiments as possible
interactions and effects of several system variables and minimizes optimization costs. DOE techniques help
an absence of concrete information. In industry, reducing the cost, because DOE necessitates only a
designed experiments can be used to systematically small set of experiments. DOE is a logical, accurate
investigate the process or product variables that approach to engineering problem-solving that uses
influence product quality. statistical analysis to ensure the development of valid,
defensible, and sound engineering conclusions. In
addition, experimentations are performed under the
INTRODUCTION restriction of a minimal expenditure of money, time, and
engineering analysis runs. It is a purposeful change in
the inputs to a process or system, in order to
Design of Experiments (DOE) provides a highly observe and identify reasons for changes in the output.
structured way to study the effects of multiple variables Experiment results are analyzed using simple statistical
on product performance, as well as, efficient and techniques. Moreover, Design of Experiments (DOE)
effective methods for determining the most significant technique allows a designer to select data points
factors and interactions in a given design problem. intelligently in the design space for fitting a model and in
Design of Experiments (DOE) is a statistical technique, such a way as to produce an accurate and statistically
which allows the scientist to increase knowledge in order meaningful approximation.

1 Copyright © 2008 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/30/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use
 the class of Latin hypercube measures. Owen [28]
There are two types of experimental designs, "classical" supported using orthogonal arrays as appropriate
experimental designs and "space filling" designs. The full designs for computer experiments.
factorial, the Central Composite Design (CCD) and Box-
Behnken (BB) are used extensively for experimental Design of Experiments principles offer a systematic and
designs. Some space filling designs such as Latin efficient ways of evaluating a design space, provide
Hypercubes, Nearly Orthogonal Latin Hypercube information for design variable screening, analyze
Sample (NOLHS), Orthogonal Array (OA), and design variable impact, and detect significant design
Orthogonal Array-based Latin hypercube, among others, variable interactions [20, 25, 26]. Some practitioners use
are the competing design types. Classic DOE textbooks experimental design methods as alternate methods for
[13, 14, 26] and comprehensive textbook [21, 30] cover design optimization. Regardless of the distinction, DOE
a brand range of topics and provide a detailed list of can be used as an essential part of the design
references. Booker [3] summarized the difference optimization procedures
between classical experimental designs and new space
filling designs very well.
BACKGROUND
Many researchers [8, 33] argue that classical
experimental designs, such as the CCD and BB designs, Design of Experiments has been used by many
are not well-suited for sampling deterministic computer engineers and scientists for a long time and numerous
experiments, which have no random error. For this experimental design techniques exist. A comprehensive
reason, the classical notions of experimental blocking, review of these techniques and their use in engineering
randomization, and replication are irrelevant [32, 41], design is addressed in Simpson et al [36]. Below is a
hence, designs for deterministic computer experiments brief description of different types of the most used
should fill the space, as opposed to possess properties experimental designs.
for estimating the variability in the data [4]. Numerous
space filling experimental designs have been developed 1) Box-Wilson includes the center point and axial
in an effort to provide more efficient and effective means point and corner point distributed around an n-
for sampling deterministic computer experiments. dimensional sphere circumscribing the cuboid
Numerous space filling experimental designs exist defined by the input minima and maxima .
including minimax and maximin designs, Integrated
Mean Squared Error (IMSE) designs, maximum entropy 2) Full-factorial design produces a uniform grid with
designs, orthogonal arrays, scrambled nets, latin user specified density covering the input
hypercubes, and randomized grids to name few [18]. parameter space. This is the most basic
Latin hypercube designs were introduced for use with experimental design. The size of full –factorial
computer codes and compared to random sampling and design experiment increases exponentially with
stratified sampling [23]. Maximin and minimax designs the number of design variables, which may lead
were invented specifically for use with computer to an unmanageable number of experiments.
experiments [7, 15, 35]. Shewry and Wynn [31] used the Figure 1 illustrates an example of a full factorial
maximum entropy principle to develop designs for design in three variables, 23 = 8 experiments.
computer experiments. Similarly, Sacks et al [34] Figure 1 shows graphical representation of all
discussed entropy designs, in addition to IMSE designs runs. This practice provides extensive
and maximum mean squared error designs for use with information for accurate estimation and
deterministic computer experiments. Finally, a review of interaction effects. However, it is often deemed
several Bayesian experimental designs for linear and cost prohibitive due to the number of analyses
nonlinear regression models is given in Chaloner and required.
Verdinelli [6]. Koch et al [17] researched the application
of modified version of the central composite design
(CCD) by combining half-fractions of a faced centered
CCD with an inscribed CCD with a face-centered CCD.
Tang [39, 40] described the difference between general
Latin hypercubes and the orthogonal array-based Latin
hypercubes and determined that the latter is more
suitable for computer experiments. Park [29] introduced
optimal Latin hypercube designs for use with computer
experiments, which either maximize entropy or minimize
IMSE by distributing the points out over the region of
interest.. Beattie and Lin [2] offered a method to build
Latin hypercubes design using rotated factorial designs.
Fang and his co-authors [10, 11] used number-theoretic
techniques to create uniform designs for use with
computer-based experiments. Morris and Mitchell [24]
described maximin distance designs embedded within
Figure 1 - full factorial experimental design (8 points)
2 Copyright © 2008 by ASME
Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/30/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use
 additional points for each factor (called “star
3) Fraction-factorial design can be used when points”). Therefore, five levels are defined for
experiments are costly and the number of each factor. Using Central Composite Design
n
design points for a full factorial design is large. requires 2 + 2n+1 design point evaluations to
Fraction factorial design consists of a fraction of study n factors. Figure 3 illustrates the Central
a full factorial design. Figures 6 illustrates an Composite Design points for three factors.
example of a fraction factorial design in three Different CCDs are formed by varying the
3-1
variables, 2 = 4 experiments. Figure 2 shows distance from the center of the design space to
graphical representation of all runs. One specific the star points. CCD consists of the following:
family of fraction factorial design commonly used
for screening is the two level Plackett-Burman a) Ordinary central composite design
(P-B) designs. (CCD) - star points are placed a
distance of ± α(α> 1) from the center
with the cube points placed at ±1 from
the center.
b) Face Central Design (FCD) design - star
points are positioned on the faces of the
cube.
c) Inscribed Central Design (ICD) - star
points are positioned at ±1/α from the
center with the cube points placed at ±1.

In addition, combinations of the CCD, FCD, and ICD

are investigated based on the suggestions and
observations discussed in [17].

Figure 2 - 23 Fraction Factorial Experimental Design (4

points)

4) Plackett-Burman (PB) design: R.L. Plackett and

J.P. Burman published their paper "The Design
of Optimal Multifactorial Experiments” in 1946.
Plackett-Burman (PB) designs are used for
screening a large number of factors to a smaller
set of important factors for further
experimentation. PB design includes main
effects and, in general, does not evaluate
interactions. However, the effects for all the
factors are placed in a Parato chart and
important factors are chosen for further
evaluation. PB design is a very economical
Figure 3 Central Composite Design for Three Factors
design and the number of runs is based on
[10]
Hadmard matrix, which is a multiple of 4 (rather
than a power of 2), i.e., n= 4, 8, 12, 16. One of
6) Box-Behnken (B-B) design is an independent
the main advantages of PB designs is the
quadratic design that it does not contain an
reduced number of runs to evaluate large
embedded factorial or fractional factorial design.
numbers of factors. The PB design in n runs
In this experimental design, the treatment
may be used for an experiment that includes up
combinations are at the midpoints of edges of
to n-1 factors (all degree of freedom are utilized
the design space and at the center. These
to estimate the main effects). For example, the
designs are rotatable (or nearly rotatable) and
PB design in 16 runs may be used for an
are an important alternative to Central
experiment includes up to 15 factors.
Composite Design, because they require three
levels of each factor [5]. Figure 4 provides a
5) Central Composite Design (CCD): CCD design
graphical sketch of the experimental layout of a
was developed specifically for fitting second-
Box-Behnken design with three factors.
order response surface methodology models
[27]. CCD is a statistically based technique in
which a 2-level full-factorial experiment is
augmented with a center point and two
3 Copyright © 2008 by ASME
Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/30/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use
 probability region (where r is the number of runs
and n is the number of input variables). An
advantage of using Latin Hypercubes over the
Orthogonal Arrays is that more points and more
combination can be studied for each factor.
Figure 6 shows the Latin Hypercube concept for
two factors (X1, X2) and 9 design points. This
matrix has nine design points for the two factors,
but there are nine levels for each factor as well,
allowing higher order polynomial models to fit
the data of greater assessment of nonliterary.

Figure 4 Box-Behnken Design for Three Factors [19]

7) Orthogonal Arrays (OA) has been used since as

early as the 1940’s by Plackett and Burman and
then was popularized by Taguchi and Konishi
[38], who developed a family of 2- and 3- level
orthogonal arrays to study interaction effects.
Orthogonal Array avoid a costly full-factorial
experiment in which all combination of all inputs
at different levels are analyzed and, instead,
perform a fraction factorial experiment, which is
carefully selected to maintain orthogonality
(independence) among the various factors and Figure 6 Three- level Latin Hypercube [10]
certain interactions. OA provides balanced (full .
factorial) designs [1]. In Figure 5 a standard 9) Optimal Latin Hypercubes (introduced by Park,
three Orthogonal Array is shown. This matrix 1994) [29] is a modified Latin Hypercube, in
has nine design points and only three levels for which combination of each factor is optimized,
each factor. rather than randomly combined. The design
matrix in which the point is spread as evenly as
possible within the design space is defined by
the lower and upper level of each factor. Figure
7 shows the Optimal Latin Hypercube concept
for two factors (X1, X2) and 9 design points.
With this matrix, the nine design points cover
nine levels of each factor and are spread evenly
within the design space. Optimal Latin
Hypercube offers the best opportunity to model
the true behavior of the response across the
range of factors.

Figure 5 Three- level Orthogonal Array [10]

8) Latin Hypercubes: The first type of design

proposed for computer experiments was Latin
Hypercubes[22]. A Latin Hypercube is a matrix
of m row and n columns where m is the number
of levels being analyzed and n is the number of
input variables. Each of the n columns includes
the level 1, 2, ,,,m, arbitrary permuted and the
m columns are randomly matched for the Latin Figure 7 Optimal Latin Hypercube Configuration [10]
Hypercube. Latin Hypercube provides an
orthogonal arrays that randomly samples the 10) Nearly Orthogonal Latin Hypercube Sampling
n
entire design space broken down into r equal – (NOLHS): Similar to Latin Hypercube, NOLHS

4 Copyright © 2008 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/30/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use
 is a matrix of m row and n columns where m is because experimental units (plots) located in
the number of levels being analyzed and n is the each block are more or less uniform. When all
number of input variables. NOLHS is multi-level treatments are found within each block in a
screening design in which the design design, the design is called complete block
parameters are equally spaced throughout the design. Sometimes it is difficult, inconvenient, or
design region and every experiment is unique. impossible to apply all treatments in each block
The number of runs is equal to the number of resulting incomplete blocks. For instance, there
design variables multiplied by the number of can be a block in which the available number of
design levels. NOLHS designs have good plots may not be enough to apply all treatments.
space-filling and orthogonality properties for For example, in an agricultural field experiment,
small or moderate number of design variables. the size of a block of land may be too small to
demarcate the required number of plots within
11) Hammersley Sequence Sampling (HSS). HSS is the block. In engineering experiments, if five
a low discrepancy sequence and it requires less different brands of tires are to be tested with a
runs than Latin Hypercube to maintain the same number of cars and one car can accommodate
level of accuracy in estimating prediction. A low only 4 tires, then the result is an incomplete
discrepancy means a uniform distribution of block. Therefore, when all treatments cannot be
points in the design space. Sampling efficiency applied within each block, incomplete block
is degraded in Latin hypercube because designs are used.
correlation exists between the columns. In HSS,
sampling is very efficient because correlations DOE DESIGN SELECTION
between columns are almost zero [16].

12) D-optimal design is the most popular computer- Design engineers may be puzzled over which DOE
generated experiments design. Computer- techniques to use. Table 1 shows the suggested DOE
techniques based on the purpose of the DOE study.
generated designs are experimental designs
that are based on a particular optimality
criterion. One popular criterion is D-optimality, Table 1 - DOE Selection Techniques [9]
which seeks to maximize X'X, the determinant
of the information matrix XX of the design. This Purpose of DOE Study Suggested Technique
criterion results in minimizing the generalized
variance of the parameter estimates according Efficiency– a lot of Orthogonal Arrays, Latin
to a pre specified model. Therefore, D-optimal
design involves a search for an experimental information in few runs Hypercubes
design such that its overall prediction error on Optimal Latin
model parameters is smallest. Computer-
generated experimental designs, such as the D- Hypercubes, Nearly
optimal design, have some advantages over
traditional response surface designs such as the Orthogonal Latin
central composite design and Box-Behnken Hpercubes Sampling
design. One major advantage is much greater
flexibility in selecting response surface model (NOLHS)
types and the number of experimental runs.
Accuracy- a broad spread Latin Hypercubes,
13) Uniform Designs Uniform designs have been
of data points for Optimal Latin Hypercubes
used since 1980. If uniformity property is added
to the fraction factorial design, then the design constructing higher fidelity Central Composite
will be become uniform designs, which provide
uniformly scattered design points in the approximation Design, Full- Factorial
experimental field. If the experimental field is
Box-Behnken
finite, uniform design is similar to Latin
Hypercubes. When the experimental field is
continuous, uniform design is different from Latin Flexibility Full-Factorial, Nominal
Hypercubes. The main difference is that in a
uniform design, point is selected from the center design
of the cells, whereas, in Latin Hypercubes points
Center point design
are randomly selected from cells. For more
information on the uniform designs and their
application, see reference [11].

14) Block design: In basic experimental designs, all

treatments are applied within each block,

5 Copyright © 2008 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/30/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use
 WHY THE USE OF DESIGN OF EXPERIMENTS Experiments (DOE) through the design cycle will yield
(DOE) true and last competitive advantages.

It is well documented and proven within the defense and

the commercial industries that the cost of fixing design REFERENCES
defects due to faulty or incomplete design requirements
increases exponentially as the design advances through
1. Barton, R. R., 1994, “Metamodeling: A State of
the design development process. Experts estimate that
the Art Review,” Proceedings of the1994 Winter
fixing design defects once design has entered the
Simulation Conference, Lake Beuna Vista, FL,
"operational" phase is 10 to 100 times more costly than
IEEE.
doing so at the earliest phases of the design process.
Moreover, correcting errors after product is sold can
cause economic catastrophe for manufacturers due to 2. Beattie, S. D. and Lin, D. K. J., 1997, "Designing
recall and warranty costs. Computer Experiments: Rotated Factorial
Designs," Technical Report No. 97-06,
Currently, quality methods can be characterized as after- Department of Statistics, The Pennsylvania
the-fact practices, since they use lagging information to State University, University Park, PA.
developmental activities such as bench tests and field
data. Companies that follow these practices usually
suffer from high development costs, longer time to
3. Booker, A. J. (1996). Case studies in design and
analysis of computer experiments. In
market, lower quality levels, and a marginal competitive
Proceedings of the Section on Physical and
edge. Building prototypes of complex systems is
Engineering Sciences, pages 244-248.
expensive and it is difficult and time consuming to
American Statistical Association.
develop them. Consequently, it is extremely beneficial to
know as much as possible about the mechanism and to
optimize its dynamic characteristics before the first 4. Booker, A.J., “Design and Analysis of Computer
physical prototype is built. Department of Defense Experiments,” AIAA paper 98–4757.
(DOD) spend Billion of Dollars annually beyond what is
necessary to support its quality assurance approach.
Historically, numerous acquisition programs have had
5. Box, G. E. P. and D. W. Behnken (1960). “Some
New Three- Level Designs for the Study of
quality problems in production, because robust design
Quantitative Variables.” Technometrics, 2, pp.
techniques were absent from the design phase. Major
455–475.
military programs encountered major manufacturing
problems because they went forward with unstable
designs and relied on inspections to find and rework 6. Chaloner, K. and Verdinelli, I., 1995, "Bayesian
defects once in production. To effectively address these Experimental Design: A Review," Statistical
issues, the Design of Experiment should be employed as Science, Vol. 10, No. 3, pp. 273-304
a crucial analysis tool in the early stage of the design
development process.
7. Cressie, N. A. C., 1993, Statistics for Spatial
Data, Revised, John Wiley & Sons, New York.

CONCLUSIONS 8. Currin, C., Mitchell, M., Morris, M., and

Ylvisaker, D., 1991, “Bayesian Prediction of
Deterministic Functions, with Applications to the
In summary, the use of Design of Experiments (DOE) to Design and Analysis of Computer Experiments,”
model input variation to predict output variation with the Journal of the American Statistical Association,
right tools provides engineers with ability to perform Volume 86, pp.953-963.
“what if” situation. which will enable design engineers to
evaluate alternative approaches and explores options
early in the design cycle to arrive to an optimum design 9. Dyn, N Levin, D. and Rippa, S, 1986, “Numerical
that meets product performance and customer Procedures for Surface Fitting of Scattered Data
requirements.. Through this process, design engineers by Radial Basis Functions” SIAM Journal of
will be able quickly to investigate many design Scientific and Statistical Computing, Vol. 7, No.
alternatives and evaluate numerous ideas that would not 2, pp. 639-659.
be practical to test in physical prototypes.
10. Engineers Software, Inc., iSIGHT Designer’s
Design of Experiments provides a disciplined approach Guide, Morrisville, NC, 1998.
to product development, yielding a great knowledge
about the product before commitment to any hardware
build. The successful application of the Design of 11. Fang, K-T., Lin, D. K. J., Winker, P. and Zhang,
Y., 2000, “Uniform Design: Theory and

6 Copyright © 2008 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/30/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use
 Application,” Technometrics, Vol. 42, pp.237-
248.
22. Madsen, J.I., Shyy, W., and Haftka, R.T.,
“Response Surface Techniques for Diffuser
Shape Optimization,” AIAA Journal, Vol. 38, No.
12. Fang, K. T. and Y. Wang (1994). Number- 9, pp. 1512–18, September 2000.
Theoretic Methods in Statistics. New York:
Chapman and Hall.

23. McKay, M. D, R. J. Beckman, and W. J.

Conover (1979). “A Comparison of Three
13. G.E.P Box, W.G. Hunter and J.S. Hunter Methods for Selecting Values if Impute Variables
“Statistics for experimenters”, John Wiley and
I the Analysis of Output from a Computer Code.”
Son, Inc,., New York (1978).
Technometrics, 21, pp. 239-245.

14. G.E.P Box, N.R. Draper “Empirical model- 24. Morris, M. D. and Mitchell, T. J. (1992).
building and Response surfaces”, John Wiley
Exploratory designs for computational
and Son, Inc,., New York (1987).
experiments. Journal of Statistical Planning and
Inference NO. 43, 1995, pp. 381-402
15. Johnson, M. E., Moore, L. M. and Ylvisaker, D.,
1990, "Minimax and Maximin Distance Designs,"
Journal of Statistical Planning and Inference,
25. Montgomery, D.C., Design and Analysis of
Experiments, Fourth Edition, John Wiley & Sons,
Vol. 26, No. 2, pp. 131-148.
New York, 1997.

16. Kalagnanam, J. R. and Diwekar, U. M., 1997, 26. Myers, R.H. and D.C. Montgomery (2002),
"An Efficient Sampling Technique for Off-Line
Response Surface Methodology: Process and
Quality Control," Technometrics, Vol. 39, No. 3,
Product Optimization Using Designed
pp. 308-319.
Experiments. New York: Wiley.

17. Koch, P., Mavris, D., and Mistree, F., “Multi- 27. Myers, R. H.; Montgomery, D.C: Response
Level, Partitioned Response Surfaces for
Surface Methodology: Process and Product
Modeling Complex Systems,” (St. Louis, MO),
th Optimization Using Design of experiments, John
Presented at the 7 AIAA/USAF/NASA/ISSMO
Wiley & Sons, New York, 1995.
Symposium on Multidisciplinary Analysis and
Optimization, AIAA, September 2-4, 1998, Vol.
3, pp1954-1968 , AIAA-98-4958 28. Owen, A. B., 1992, "Orthogonal Arrays for
Computer Experiments, Integration and
Visualization," Statistica Sinica, Vol. 2, pp. 439-
18. Koehler, J. R. and Owen, A. B., 1996, 452.
“Computer Experiments,” Handbook of Statistics
(Ghosh, S. and Rao, C. R., eds.), Elsevier
Science, New York, pp.261-308.
29. Park, J. S. (1994). “Optimal Latin Hypercube
Designs for Computer Experiments.” Journal of
19. Kai Yang and Bassem S. El-Haik, B. (2003). Statistical Planning and Interference, 39, pp. 95-
“Design for Six Sigma; A Roadmap for Product 111.
Development.” McGraw-Hill, New York.
30. Robert O. Kuehl (2000) Design of Experiments-
20. Kodiyalam, S.; Su Lin, J.; Wajek, B.A. 1998, Statistical Principles of Research Design and
nd
Design Of Experiments Based Response Analysis, 2 Edition.
th
Surface Models for Design Optimization. 39
AIAA/ASME/ASCE/AHS/ASC Structures,
Structural Dynamics and Materials Conference,
31. Saliby, E.., 1990, Descriptive Sampling: A Better
Approach to Monte Carlo Simulation”, Vol. 41,
Long Beach, CA, pp. 2718-2727.
No. 12, pp. 1133-1142

21. Law, A. M. and W.D. Kelton (2000) Simulation

rd
Modeling and Analysis, 3 Edition, McGraw-Hill. 32. Sacks, J.; Welch. W. J.; Mitchell, T.J.; Wynn,
New York.Lin, Y., 2000, Robust Design Goal H.P. 1989: Design and analysis of Computer
Formulations and Metamodeling Techniques, Experiments (with discussion). Statistical
MS Thesis, Georgia Institute of Technology, Science, 4(4), 409-435.
Atlanta, Georgia.

7 Copyright © 2008 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/30/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use
 33. Sacks, J. and Schiller, S., 1988, "Spatial
Designs," Statistical Decision Theory and
Related Topics (Gupta, S. S. and Berger, J. O.,
eds.), Springer-Verlag, New York, pp.385-399.

34. Sacks, J., Schiller, S. B. and Welch, W. J.,

1989b, "Designs for Computer Experiments,"
Technometrics, Vol. 31, No. 1, February, pp. 41-
47.

35. Shewry, M. C. and Wynn, H. P., 1987,

“Maximum Entropy Sampling,” Journal of
Applied Statistics, Vol.14, No.2, pp.165-170.

36. Simpson, T. W.; Peplinski, J.; Koch, P. N.; Allen,

J. K. 1997: On the Use of Statistics in Design
and the Implications for Deterministic Computer
Experiments. Design Theory and Methodology-
DTM'97. Sacramento, CA. ASME Paper No.
DETC97/DTM-3881.

37. Stephen R. Schmidt and Robert G. Launsby,

Understanding Industrial Designed Experiments,
4th Edition, Air Academy Press & Associates,
2005

38. Taguchi, G., and Konishi, S., Taguchi Methods

Orthogonal Arrays and Linear Graphs; Tools for
Quality Engineering, (1987).

39. Tang, B., 1993, "Orthogonal Array-Based Latin

Hypercube," Journal of the American Statistical
Association, Vol. 88, No. 424, pp. 1392-1397.

40. Tang, B., 1994, "A Theorem for Selecting OA-

Based Latin Hypercube Using a Distance
Criterion," Communications in Statistics, Theory
and Methods, Vol. 23, No. 7, pp. 2047-2058.

41. T.J. Scanter, 2003, the Design and Analysis of

Computer Experiments, Springer-Verlag.

8 Copyright © 2008 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/30/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use