World Academy of Science, Engineering and Technology

Vol:3 2009-03-25

Statistical Process Optimization

Through Multi-Response Surface Methodology
S. Raissi, and R- Eslami Farsani ∗

one and is optimized by adhering to the other constraints set

Abstract—In recent years, response surface methodology (RSM) has by the criteria. Many heuristic methodologies have been
brought many attentions of many quality engineers in different developed to resolve the multi-response problem. Cornell and
industries. Most of the published literature on robust design Khuri [4] surveyed the multi-response problem using a
methodology is basically concerned with optimization of a single
response or quality characteristic which is often most critical to
response surface method. Tai et al. [5] assigned a weight for
consumers. For most products, however, quality is multidimensional, each response to resolve the problem. Pignatiello [6] utilized a
so it is common to observe multiple responses in an experimental squared deviation-from-target and a variance to form an
situation. Through this paper interested person will be familiarize expected loss function for optimizing a multiple response
with this methodology via surveying of the most cited technical problem. Layne [7] presented a procedure capable of
papers. simultaneously considering three functions: weighted loss
International Science Index Vol:3, No:3, 2009 waset.org/Publication/14692

It is believed that the proposed procedure in this study can resolve

function, desirability function, and distance function. While
a complex parameter design problem with more than two responses.
It can be applied to those areas where there are large data sets and a providing a multi-response example in which Taguchi
number of responses are to be optimized simultaneously. In addition, methods are used, Byrne and Taguchi [8] discussed an
the proposed procedure is relatively simple and can be implemented example involving a connector and a tube.
easily by using ready-made standard statistical packages. Logothetis and Haigh [9] also discussed a manufacturing
process differentiated by five responses. In doing so, they
Keywords—Multi-Response Surface Methodology (MRSM), selected one of the five response variables as primary and
Design of Experiments (DOE), Process modeling, Quality optimized the objective function sequentially while ignoring
improvement; Robust Design.
possible correlations among the responses. Optimizing the
process with respect to any single response leads to non-
I. INTRODUCTION
optimum values for the remaining characteristics.
R ESPONSE Surface Methodology (RSM) is a well known
up to date approach for constructing approximation
models based on either physical experiments, computer
II. RESPONSE SURFACE METHODOLOGY
experiments (simulations) (Box et al., [1] ; Montgomery, [2]) Often engineering experimenters wish to find the conditions
and experimented observations. RSM, invented by Box and under which a certain process attains the optimal results. That
Wilson, is a collection of mathematical and statistical is, they want to determine the levels of the design parameters
techniques for empirical model building. By careful design of at which the response reaches its optimum. The optimum
experiments, the objective is to optimize a response (output could be either a maximum or a minimum of a function of the
variable) which is influenced by several independent variables design parameters. One of methodologies for obtaining the
(input variables). An experiment is a series of tests, called optimum is response surface technique.
runs, in which changes are prepared in the input variables in Response surface methodology is a collection of statistical
order to recognize the reasons for changes in the output and mathematical methods that are useful for the modeling
response (Montgomery & Runger [3]). RSM involves two and analyzing engineering problems. In this technique, the
basic concepts: main objective is to optimize the response surface that is
(1) The choice of the approximate model, and influenced by various process parameters. Response surface
(2) The plan of experiments where the response has to be methodology also quantifies the relationship between the
evaluated. controllable input parameters and the obtained response
The performance of a manufactured product often surfaces.
characterize by a group of responses. These responses in The design procedure of response surface methodology is as
general are correlated and measured via a different follows:
measurement scale. Consequently, a decision-maker must (i) Designing of a series of experiments for adequate and
resolve the parameter selection problem to optimize each reliable measurement of the response of interest.
response. This problem is regarded as a multi-response (ii) Developing a mathematical model of the second order
optimization problem, subject to different response response surface with the best fittings.
requirements. Most of the common methods are incomplete in (iii) Finding the optimal set of experimental parameters
such a way that a response variable is selected as the primary that produce a maximum or minimum value of
response.
R. Eslami Farsani is with Islamic Azad University, Tehran South Branch.

International Scholarly and Scientific Research & Innovation 3(3) 2009 247
 World Academy of Science, Engineering and Technology
Vol:3 2009-03-25

(iv) Representing the direct and interactive effects of which factors or variables are likely to be important in
process parameters through two and three dimensional response surface study. It is usually called a screening
plots. experiment. The objective of factor screening is to reduce the
If all variables are assumed to be measurable, the response list of candidate variables to a relatively few so that
surface can be expressed as follows: subsequent experiments will be more efficient and require
y = f ( x1 , x 2 ,....x k ) (1) fewer runs or tests. The purpose of this phase is the
identification of the important independent variables.
The goal is to optimize the response variable y . It is
Phase 1: The experimenter’s objective is to determine if the
assumed that the independent variables are continuous and current settings of the independent variables result in a value
controllable by experiments with negligible errors. It is of the response that is near the optimum. If the current settings
required to find a suitable approximation for the true or levels of the independent variables are not consistent with
functional relationship between independent variables and the optimum performance, then the experimenter must determine
response surface. Usually a second-order model is utilized in a set of adjustments to the process variables that will move the
response surface methodology. process toward the optimum. This phase of RSM makes
k k k considerable use of the first-order model and an optimization
y = β0 + ∑ β i xi + ∑ β ii xi2 + ∑ β ij xi x j +ε (2) technique called the method of steepest ascent (descent).
i =1 i =1 i =1 Phase 2: Phase 2 begins when the process is near the
where ε is a random error. The β coefficients, which optimum. At this point the experimenter usually wants a
model that will accurately approximate the true response
International Science Index Vol:3, No:3, 2009 waset.org/Publication/14692

should be determined in the second-order model, are obtained

by the least square method. In general (2) can be written in function within a relatively small region around the optimum.
matrix form. Because the true response surface usually exhibits curvature
near the optimum, a second-order model (or perhaps some
Y = bX + E (3)
higher-order polynomial) should be used. Once an appropriate
where Y is defined to be a matrix of measured values, X to
approximating model has been obtained, this model may be
be a matrix of independent variables. The matrixes b and E
analyzed to determine the optimum conditions for the process.
consist of coefficients and errors, respectively. The solution of
This sequential experimental process is usually performed
(3) can be obtained by the matrix approach.
within some region of the independent variable space called
(
b = XT X )
−1
XTY (4) the operability region or experimentation region or region of
interest.
where X is the transpose of the matrix X and (XTX)-1 is the
T

inverse of the matrix XTX.

IV. MULTI-RESPONSE PROBLEM OVERVIEWS
The mathematical models were evaluated for each response
by means of multiple linear regression analysis. As said Optimization of the multi-response problem is a challenge
previous, modeling was started with a quadratic model to optimize output responses all together. Among the
including linear, squared and interaction terms. The significant simultaneous optimization methods, most of the authors used
terms in the model were found by analysis of variance the approaches that combine all the different response
(ANOVA) for each response. Significance was judged by requirements into one composite requirement. Hence, the
determining the probability level that the F-statistic calculated compromise solution is obtained in a much simpler way. A
from the data is less than 5%. The model adequacies were simple weighting method was found in Ilhan et al. [10], as
checked by R2, adjusted-R2, predicted-R2 and prediction error applied in an electrochemical grinding (ECG) process. Zadeh
sum of squares (PRESS). A good model will have a large [11] normalized each response and then gave a simple weight
predicted R2, and a low PRESS. After model fitting was for each response. The discussion regarding the assignments
performed, residual analysis was conducted to validate the of weights can be found in [12].
assumptions used in the ANOVA. This analysis included Myers and Carter [13] proposed an algorithm for obtaining
calculating case statistics to identify outliers and examining the optimal solutions of the dual-response surface system
diagnostic plots such as normal probability plots and residual (DRSM). Their method assumed that the DRSM includes a
plots. primary response and a constraint response which both of
Maximization and minimization of the polynomials thus them can be fitted as a quadratic model.
fitted was usually performed by desirability function method, Lee-Ing Tong et al. [14] used the signal to noise (SN) ratio
and mapping of the fitted responses was achieved using and system sensitivity are used to assess the performance of
computer software such as Design Expert. each response. They performed principal component analysis
(PCA) on SN values and system sensitivity values to obtain a
III. THE SEQUENTIAL NATURE OF THE RESPONSE SURFACE set of uncorrelated principle components, which are linear
METHODOLOGY combinations of the original responses. Additionally, they
Most applications of RSM are sequential in nature and can used of variation mode chart to interpret the variation mode
be carried out based on the following phases. (or principal component variation) resulting from PCA. They
Phase 0: At first some ideas are generated concerning suggested that based on engineering requirements, engineers
can determine the optimization direction for each principal

International Scholarly and Scientific Research & Innovation 3(3) 2009 248
 World Academy of Science, Engineering and Technology
Vol:3 2009-03-25

component using the variation mode chart. Finally, technique applying neuro-fuzzy model. ANOVA is carried out on the
for order preference by similarity to ideal solution (TOPSIS) MRS values to identify the key factors/interactions having
applied to derive the overall performance index (OPI) for significant effect on the overall process. Finally, optimal
multiple responses. The optimal factor/level combination can setting of the control factors is decided by selecting the level
be determined with the maximum OPI value and therefore, having highest value of MRS.
simultaneously reduces the quality variation and brings the
mean to the target value. V. DESIRABILITY FUNCTION
Onur Koksoy and Tankut Yalcinoz [15] presented a The desirability function was originally developed by
methodology for analyzing several quality characteristics Harrington [22] to simultaneously optimize the multiple
simultaneously using the mean square error (MSE) criterion responses and was later modified by Derringer and Suich [23]
when data are collected from a combined array. They to improve its practicality. The desirability function approach
proposed a genetic algorithm based on arithmetic crossover is one of the most frequently used multi-response optimization
for the multi-response problem in conjunction with a techniques in practice. The desirability lies between 0 and 1
composite objective function based on the individual MSE and it represents the closeness of a response to its ideal value.
functions of each response. If a response falls within the unacceptable intervals, the
Lee-Ing Tong et al. [16] proposed procedure used the desirability is 0, and if a response falls within the ideal
desirability function and dual-response-surface method to intervals or the response reaches its ideal value, the
optimize the multi-response problems in a dynamic system. desirability is 1. Meanwhile, when a response falls within the
They established a regression model to obtain the sensitivity tolerance intervals but not the ideal interval, or when it fails to
International Science Index Vol:3, No:3, 2009 waset.org/Publication/14692

and quality variation for each experimental run and the reach its ideal value, the desirability lies between 0 and 1. The
desirability function is used to obtain a total measurement for more closely the response approaches the ideal intervals or
the multiple responses. Next, the dual-response-surface ideal values, the closer the desirability is to 1. According to
method was used to obtain a set of possible optimal factor– the objective properties of a desirability function, the
level combinations. The optimal factor–level setting proposed desirability function can be categorized into the nominal-the-
to maximize total desirability. best (NB) response, the larger-the-better (LB) response and
Liao and Chen [17] proposed data envelopment analysis the smaller-the-better (SB) response. Interested persons can
ranking (DEAR) approach to optimize multi-response follow the expressed relevant desirability functions in [101].
problem. The author states that Taguchi method can only be The proposed desirability function transforms each response to
used to optimize single response problems and PCA, although a corresponding desirability value between 0 and 1. All the
considered to solve multi-response problem, itself has desirability can be combined to form a composite desirability
shortcomings. The new approach is capable of decreasing function which converts a multi-response problem into a
uncertainty caused by engineering judgment in the Taguchi single-response one. The desirability function is a scale-
method and overcoming the shortcomings of PCA. invariant index which enables quality characteristics to be
In order to overcome the single response optimization compared to various units. In such method the plant manager
problem of Taguchi method, Liao [18] proposed an effective can easily determine the optimal parameters among a group of
procedure called PCR-TOPSIS that is based on process solutions.
capability ratio (PCR) theory and on the theory of order Kun-Lin Hsieh et al. [24] believed that when desirability
preference by similarity to the ideal solution (TOPSIS) to values lies more close to 0 or 1 may lead to a bad model’s
optimize multi-response problems. additive. To solve this problem, they referred to Taguchi
Hsu [19] presents an integrated optimization approach suggestion in using the Omega (Ω) transformation which is
based on neural networks, exponential desirability functions. employed to transfer the data into an additive mode. Ω
Fung and Kang [20] used Taguchi method and PCA to transformation’s philosophy is to simultaneously maximize
optimize the given process. Initially Taguchi method was used the average of the system and minimize the variation via S/N.
followed by PCA to correspond to multi-response cases, for This transformation transfer the desirability data lying in [0,1]
transforming the correlated friction properties to a set of
uncorrelated components and evaluating the principal to the range of . This transformation can resolve
components. The appropriate number of the principle the problem by summing up the control factor’s effect when
components, and the influence of the number on the optimum the data lie outside the interval [0,1].
process condition, was subsequently studied by extracting
more than one principal component and integrating it into a VI. SPECIAL CASE: DUAL-RESPONSE SURFACE METHOD
comprehensive index. In practical cases, there are many situations where the
Jiju Antony et al. [21] used artificial inteligent tool (neuro- researchers encounter to multi-responses. In such cases
fuzzy model) and Taguchi method of experimental design to surveying two or more response variables are critical.
tackle problems involving multiple responses optimization. Over the last few years in many manufacturing
They proposed a single crisp performance index called Multi- organizations, multiple response optimization problems were
Response Statistics (MRS) as a combined response indicator resolved using the past experience and engineering judgment,
of several responses. MRS is computed for every run by which leads to increase in uncertainty during the decision-

International Scholarly and Scientific Research & Innovation 3(3) 2009 249
 World Academy of Science, Engineering and Technology
Vol:3 2009-03-25

making process. simultaneously optimizing a multi-response problem.

Myers and Carter [25] proposed an algorithm for obtaining
the optimal solutions of the dual-response surface method VII. MODEL ADEQUACY CHECKING
(DRSM). Their method assumed that the DRSM includes a To verify the derived mathematical model of each response,
primary response, y p and a constraint response, y s . Both model adequacy is always necessary to:
1. Examine the fitted model to ensure that it provides an
y p and y s can be respectively fitted as a quadratic model as adequate approximation to the true system;
follows: 2. Verify that none of the least squares regression
k k k assumptions are violated. There are several techniques for
y p = β0 + ∑ β i xi +∑ β ii xi2 +∑ β ij xi x j +ε p checking model adequacy.
Residual Analysis: The residuals from the least squares fit,
i =1 i =1 i =1

k k k defined by ei = y i − yi , i = 1, 2,…, n, play an important role
ys = β 0 + ∑ β i xi +∑ β ii xi2 +∑ β ij xi x j +ε s in judging model adequacy. Many response surface analysts
prefer to work with scaled residuals, in contrast to the ordinary
i =1 i =1 i =1 (4)
where the β ’s and γ ’s represent the unknown coefficients, least squares residuals. These scaled residuals often convey
more information than do the ordinary residuals.
and εP and εs denote the random errors, respectively. The The standardizing process scales the residuals by dividing
random errors are assumed to possess a normal distribution with them by their average standard deviation. In some data sets,
International Science Index Vol:3, No:3, 2009 waset.org/Publication/14692

2 residuals may have standard deviations that differ greatly.

mean 0 and variance σ .
There is some other way of scaling that takes this into account.
The DRSM attempts to obtain a set of X, which can Let’s consider this.
  
optimize y p subjected to the constraint y s = c , where C is a The vector of fitted values yi corresponding to the
constant.
observed values yi is
The desirability function simultaneously optimize the

multiple responses and was later modified by Derringer and y = Xb = X( XT X) −1 XT y = Hy (6)
Such [23] to improve its practicality. The desirability function
−1
The n x n matrix H = X( X X) X is usually called
T T
approach is one of the most frequently used multi-response
optimization techniques in practice. The desirability lies the hat matrix because it maps the vector of observed values
between 0 and 1 and it represents the closeness of a response into a vector of fitted values. The hat matrix and its properties
to its ideal value. If a response falls within the unacceptable play a central role in regression analysis.
intervals, the desirability is 0, and if a response falls within the 
Since ei = y i − yi , there are several other useful ways to
ideal intervals or the response reaches its ideal value, the
desirability is 1. express the vector of residuals
Meanwhile, when a response falls within the tolerance e = y − Xb = y − Hy = (I − H)y (7)
intervals but not the ideal interval, or when it fails to reach its The “prediction error sum of squares” (PRESS) proposed in
ideal value, the desirability lies between 0 and 1. The more [26, 27], provides a useful residual scaling
closely the response approaches the ideal intervals or ideal 2
⎛ ei
n
⎞
values, the closer the desirability is to 1. According to the
objective properties of a desirability function, the desirability
PRESS = ∑
⎜⎜
1 − hii
⎟⎟
function can be categorized into three forms, nominal-the-best i =1 ⎝ ⎠ (8)
(NB), larger-the-better (LB) and smaller-the-better (SB). From [27], it is easy to see that the PRESS residual is just
The total desirability is defined as a geometric mean of the the ordinary residual weighted according to the diagonal
individual desirability: elements of the hat matrix hii . Generally, a large difference
D = (d1 × d 2 × ...× d k )
1
k between the ordinary residual and the PRESS residual will
(5)
th indicate a point where the model fits the data well, but a model
where D is the total desirability and di is the i desirability, i
built without that point predicts poorly.
= 1, 2, . . . , k. If all of the quality characteristics reach their
ideal values, the desirability di is 1 for all i. Consequently, the
total desirability is also 1. If any one of the responses does not
VIII. CONCLUSION
reach its ideal value, the desirability di is below 1 for that
response and the total desirability is below 1. If any one of the The RSM is one of the design of experiments (DOE)
responses cannot meet the quality requirements, the methods used to approximate an unknown function for which
desirability di is 0 for that response. Total desirability will then only a few values are computed. The RSM stems from science
be 0. The desirability function is a scale-invariant index which disciplines in which physical experiments are performed to
enables quality characteristics to be compared to various units. study the unknown relation between a set of variables and the
Therefore, the desirability function is an effective means of system output, or response, for which only a few experiment

International Scholarly and Scientific Research & Innovation 3(3) 2009 250
 World Academy of Science, Engineering and Technology
Vol:3 2009-03-25

values are acquired. These relations are then modeled using a [19] Hsu, C.M., “An integrated approach to enhance the optical performance
of couplers based on neural networks, desirability functions and tabu
mathematical model, called response surface. search”, International Journal of Production Economics, vol. 92, 2004,
There are many situations where the quality engineers pp. 241-54.
encounter to several correlated responses simultaneously. In [20] Fung, C.P. and Kang, P.C., “Multi-response optimization in friction
properties of PBT composites using Taguchi method and principle
such cases decision making on optimum set of parameters is a component analysis”, Journal of Materials Processing Technology, Vol.
complicated mathematical problem. In this paper an analysis 170 No. 3, pp. 602-10.
of the most cited methods proposed and the. [21] Jiju Antony, Raj Bardhan Anand, Maneesh Kumar, M.K. Tiwari,
Through this paper, readers could be familiar to multi- Multiple response optimization using Taguchi methodology and neuro-
fuzzy based model, Journal of Manufacturing Technology Management,
response optimization problem via the most cited methods. vol. 17, vol. 7, 2006, pp. 908-925.
The residual analysis method and the prediction error sum of [22] E. Harrington, The desirability function, Industrial Quality Control, vol.
squares (PRESS) proposed for evaluating the capability of the 21, no. 10, 1965, pp. 494–498.
[23] G. Derringer, R. Suich, Simultaneous optimization of several response
designed models. Researcher could follow standard variables, Journal of Quality Technology, vol. 12, no. 4, 1980, pp. 214–
optimization techniques such as the differentiation, the 218.
operation research method to set their process in optimum [24] Kun-Lin Hsieh, Lee-Ing Tong, Hung-Pin Chiu and Hsin-Ya Yeh,
Optimization of a multi-response problem in Taguchi’s dynamic system,
conditions. Computers & Industrial Engineering, vol. 49, 2005, pp. 556–571.
[25] Myers, R.H.; Carter, W.H., Jr. Response Surface Techniques for Dual
Response Systems. Technometrics 1973, vol. 15, no. 2, pp. 301–317.
[26] C.-P. Xu, S.-W. Kim, H.-J. Hwang, J.-W. Yun, Application of
REFERENCES statistically based experimental designs for the optimization of
[1] Box, G.E.P., Hunter, W.G., and Hunter, J.S., 1978, Statistics for exopolysaccharide production by Cordyceps milltaris NG3, Biotechnol.
International Science Index Vol:3, No:3, 2009 waset.org/Publication/14692

Experimenters, John Wiley & Sons, NY. Appl. Biochem, vol. 36, 2002, pp. 127–131.
[2] Montgomery, D. C., Design and Analysis of Experiments, 2000, Wiley, [27] M.J. Anderson, H.P. Anderson, Applying DOE to microwave popcorn,
New York. 5th Edition. Process Ind. Quality, 1993, pp. 30–32.
[3] D.C. Montgomery and G.C. Runger, Applied Statistics and Probability
for Engineers, 1994, John Wiley and Sons, Inc, New York.
[4] J.A. Cornell, A.I. Khuri, Response surfaces: Designs and Analysis,
1987, Marcel Dekker, New York.
[5] C.Y. Tai, T.S. Chen, M.C. Wu, An enhanced Taguchi method for
optimizing SMT processes, Journal of Electronics Manufacturing, vol. 2,
1992, pp. 91–100.
[6] J.J. Pignatiello, Strategy for robust multi-response quality engineering,
IIE Transactions, vol. 25, no. 3, 1993, pp. 5–15.
[7] K.L. Layne, Method to determine optimum factor levels for multiple
responses in the designed experimentation, Quality Engineering, vol. 7,
no. 4, 1995, pp. 649–656.
[8] D.M. Byrne, S. Taguchi, The Taguchi approach to parameter design,
Quality Progress, vol. 20, 1987, pp. 19–26.
[9] N. Logothetis, A. Haigh, Characterizing and optimizing multi-response
processes by the Taguchi method, Quality and Reliability Engineering
International, vol. 4, no. 2, 1988, pp.159–169.
[10] R.E. Ilhan, G. Sathyanarayanan, R.H. Storer, T.W. Liao, Off-line multi
response optimization of electrochemical surface grinding by a multi-
objective programming method, International Journal Machine Tools
and Manufacture, vol. 32, no. 3, 1992, pp. 435–451.
[11] L. Zadeh, Optimality and non-scalar-value performance criteria, IEEE
Transactions on Automatic Control, vol. 8, no. 59, 1963, pp. 59–60.
[12] T. Murata, H. Ishibuchi, H. Tanaka, Multi-objective genetic algorithm
and its applications to flowshop scheduling, Computers and Industrial
Engineering, vol. 30, no. 4, 1996, pp. 957–968.
[13] Myers, R.H.; Carter, W.H., Jr. Response Surface Techniques for Dual
Response Systems. Technometrics, vol. 15, no. 2, 1973, pp. 301–317.
[14] Lee-Ing Tong, Chung-Ho Wang, Chih-Chien Chen, Chun-Tzu Chen,
Decision Aiding Dynamic multiple responses by ideal solution analysis,
European Journal of Operational Research, vol. 156, 2004, pp. 433–444.
[15] Onur Koksoy, Tankut Yalcinoz, Mean square error criteria to multi-
response process optimization by a new genetic algorithm, Applied
Mathematics and Computation, 175, 2006, pp. 1657–1674.
[16] Lee-Ing Tong, Chung-Ho Wang, Jer-Yiing Houng and Jiao-Yan Chen,
Optimizing Dynamic Multiresponse Problems Using the Dual-
Response-Surface Method, Quality Engineering, vol. 14, no. 1, 2001-02,
pp. 115–125.
[17] Hung-Chang Liao, Yan-Kwang Chen, A Data Envelopment Analysis
Method for Optimizing Multi-Response Problem In The Taguchi
Method, Computers and Industrial Engineering Special Issue on
Selected papers form the 29th.International Conference on Computers
and Industrial Engineering, vol. 46, no. 4, 2004, pp. 591-916
[18] Liao, H.C., “Using PCR-TOPSIS to optimize Taguchi’s multi-response
problem”, International Journal of Advanced Manufacturing
Technology, vol. 22, 2003, pp. 649-55.

International Scholarly and Scientific Research & Innovation 3(3) 2009 251