12 Computational Intelligence Techniques For New Product Design

Kit Yan Chan, C.K. Kwong, and Tharam S.
Dillon
Computational Intelligence Techniques for New Product Design
Studies in Computational Intelligence, Volume 403
Editor-in-Chief
Prof. Janusz Kacprzyk
Systems Research Institute
Polish Academy of Sciences
ul. Newelska 6
01-447 Warsaw
Poland
E-mail: kacprzyk@ibspan.waw.pl
Further volumes of this series can be found on our Vol. 393. Hung T. Nguyen, Vladik Kreinovich, Berlin Wu,
homepage: springer.com and Gang Xiang
Computing Statistics under Interval and Fuzzy
Vol. 383. Takayuki Ito, Minjie Zhang, Valentin Robu, Uncertainty, 2012
Shaheen Fatima, and Tokuro Matsuo (Eds.) ISBN 978-3-642-24904-4
New Trends in Agent-Based Complex Automated Vol. 394. David A. Elizondo, Agusti Solanas,
Negotiations, 2012 and Antoni Martı́nez-Ballesté (Eds.)
ISBN 978-3-642-24695-1 Computational Intelligence for Privacy
Vol. 384. Daphna Weinshall, Jörn Anemüller, and Security, 2012
and Luc van Gool (Eds.) ISBN 978-3-642-25236-5
Detection and Identification of Rare Audiovisual Cues, 2012
Vol. 395. Srikanta Patnaik and Yeon-Mo Yang (Eds.)
ISBN 978-3-642-24033-1
Soft Computing Techniques in Vision Science, 2012
Vol. 385. Alex Graves ISBN 978-3-642-25506-9
Supervised Sequence Labelling with Recurrent Neural
Networks, 2012 Vol. 396. Marielba Zacarias and
ISBN 978-3-642-24796-5 José Valente de Oliveira (Eds.)
Human-Computer Interaction: The Agency Perspective, 2012
Vol. 386. Marek R. Ogiela and Lakhmi C. Jain (Eds.) ISBN 978-3-642-25690-5
Computational Intelligence Paradigms in Advanced Pattern
Classification, 2012 Vol. 397. Elena Nikolaevskaya, Alexandr Khimich, and
ISBN 978-3-642-24048-5 Tamara Chistyakova
Programming with Multiple Precision, 2012
Vol. 387. David Alejandro Pelta, Natalio Krasnogor,
ISBN 978-3-642-25672-1
Dan Dumitrescu, Camelia Chira, and Rodica Lung (Eds.)
Nature Inspired Cooperative Strategies for Optimization Vol. 398. Fabrice Guillet, Gilbert Ritschard,
(NICSO 2011), 2011 and Djamel Abdelkader Zighed (Eds.)
ISBN 978-3-642-24093-5 Advances in Knowledge Discovery and Management, 2012
Vol. 388. Tiansi Dong ISBN 978-3-642-25837-4
Recognizing Variable Environments, 2012 Vol. 399. Kurosh Madani, António Dourado Correia,
ISBN 978-3-642-24057-7 Agostinho Rosa, and Joaquim Filipe (Eds.)
Vol. 389. Patricia Melin Computational Intelligence, 2012
Modular Neural Networks and Type-2 Fuzzy Systems for ISBN 978-3-642-27533-3
Pattern Recognition, 2012
Vol. 400. Akira Hirose
ISBN 978-3-642-24138-3
Complex-Valued Neural Networks, 2012
Vol. 390. Robert Bembenik, Lukasz Skonieczny, ISBN 978-3-642-27631-6
Henryk Rybiński, and Marek Niezgódka (Eds.)
Intelligent Tools for Building a Scientific Information Vol. 401. Piotr Lipiński and Konrad Świrski (Eds.)
Platform, 2012 Towards Modern Collaborative Knowledge, 2012
ISBN 978-3-642-24808-5 ISBN 978-3-642-27445-9
Vol. 391. Herwig Unger, Kyandoghere Kyamaky, Vol. 402. Theodor Borangiu, Andre Thomas, and Damien
and Janusz Kacprzyk (Eds.) Trentesaux (Eds.)
Autonomous Systems: Developments and Trends, 2012 Service Orientation in Holonic and Multi-Agent
ISBN 978-3-642-24805-4 Manufacturing Control, 2012
ISBN 978-3-642-27448-0
Vol. 392. Narendra Chauhan, Machavaram Kartikeyan,
and Ankush Mittal Vol. 403. Kit Yan Chan, C.K. Kwong, and Tharam S. Dillon
Soft Computing Methods for Microwave and Millimeter-Wave Computational Intelligence Techniques for New Product
Design Problems, 2012 Design, 2012
ISBN 978-3-642-25562-5 ISBN 978-3-642-27475-6
Kit Yan Chan, C.K. Kwong, and Tharam S. Dillon
Computational Intelligence
Techniques for New Product
Design
123
Authors
Kit Yan Chan Tharam S. Dillon
Curtin University of Technology Curtin University of Technology
Digital Ecosystems and Business Digital Ecosystems and Business
Intelligence Institute Intelligence Institute
Perth Perth
Australia Australia
C.K. Kwong
The Hong Kong Polytechnic
University
Department of Industrial and
Systems Engineering
Kowloon
Hong Kong SAR
ISSN 1860-949X e-ISSN 1860-9503

ISBN 978-3-642-27475-6 e-ISBN 978-3-642-27476-3
DOI 10.1007/978-3-642-27476-3
Springer Heidelberg New York Dordrecht London
Library of Congress Control Number: 2012930480
c Springer-Verlag Berlin Heidelberg 2012

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed. Exempted from this legal reservation are brief ex-
cerpts in connection with reviews or scholarly analysis or material supplied specifically for the purpose
of being entered and executed on a computer system, for exclusive use by the purchaser of the work.
Duplication of this publication or parts thereof is permitted only under the provisions of the Copyright
Law of the Publisher’s location, in its current version, and permission for use must always be obtained
from Springer. Permissions for use may be obtained through RightsLink at the Copyright Clearance
Center. Violations are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publi-
cation does not imply, even in the absence of a specific statement, that such names are exempt from the
relevant protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.
Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)
Preface
Over the recent years, applying computational intelligence techniques for product
design is a fast-growing and promising field. In this book, a wide range of compu-
tational intelligence techniques including fuzzy systems, evolutionary computa-
tions and neural network are discussed. How to implement these computational
intelligence techniques for product design is the core topic addressed in this book.
Fundamental concepts and essential analysis on computational intelligence tech-
niques are presented to offer systematic and effective tools for product design.
This book discusses common issues on product design including identification of
customer requirements in product design, determination of importance of cus-
tomer requirements for product design, determination of optimal design attributes
in new products, relating design attributes of new products and customer satisfac-
tion, integration of marketing aspects into product design, affective product de-
sign, as well as quality control of new products. Approaches for enhancement of
computational intelligence techniques which include hybridization of various
computational intelligence techniques and integration of statistical methods into
computational intelligence techniques are discussed. Case studies of product de-
sign in terms of development of real-world new products are included to illustrate
the design procedures, as well as the effectiveness of the computational intelli-
gence techniques. This book is organized as follows:
• Chapter 1 discusses fundamental concerns of product design, current is-
sues and needs for product design, as well as roles of product designers.
Existing problems in manufacturing new products and marketing the new
products are introduced. How and why computational intelligence methods
can help address these issues in terms of product design is discussed.
• Chapter 2 introduces commonly used computational intelligence methods
including evolutionary computation, swarm optimization, neural networks
and fuzzy systems etc, in order to address product design issues.
• Chapter 3 discusses a fuzzy weighting method to determine importance
weights of customer requirements of a new product. It intends to overcome
the limitation of the existing weight methods that ignore fuzziness for new
product design. A case study of a bicycle splash guard design is used to
evaluate the effectiveness of the fuzzy weighing method.
• Chapter 4 discusses a new fuzzy weighing method which is an enhanced
version of the method discussed in Chapter 3. It intends to improve the im-
precise ranking of customer satisfaction inherited from the previous works
based on the existing weight methods. A case study of hair dryer design is
presented in order to evaluate the effectiveness of the approach.
VI Preface
• Chapter 5 presents a genetic programming method in order to develop a

functional model which relates design attributes to customer requirements
for new products. It is intended to overcome the two main limitations of
previous modeling approaches: i) ignore nonlinearity of design attributes of
new products, which are unavoidable; and ii) are able to generate only
black-box models which are not preferred by product development terms.
A case study of digital camera design is used to evaluate the effectiveness
of the genetic programming approach.
• Chapter 6 introduces a hybrid fuzzy and genetic programming approach to
developing customer satisfaction models that relates customer requirements
to the design attributes of a new product. It is intended to address the fuzzy
nature of customer survey data, which are unavoidable. A case study of the
affective design of mobile phones is used to evaluate the effectiveness of
the hybrid fuzzy and genetic programming approach.
• Chapter 7 discusses a hybrid algorithm which integrates the mechanism of
fuzzy regression and generalized least square regression. The hybrid algo-
rithm intends to address the uncertainties of the development of house of
quality which contains both human fuzziness and randomness inherent in
the survey. A case study of packing machine design is used to evaluate the
hybrid algorithm.
• Chapter 8 presents an enhanced neuro-fuzzy approach for developing
customer satisfaction models to link customer requirements with design at-
tributes of a new product. The approach overcomes the limitation of the ex-
isting neural-fuzzy approaches that are implicit in nature. A case study of a
notebook computer design is used to demonstrate the operations of the en-
hanced neuro-fuzzy approach.
• Chapter 9 introduces a hybrid simulated annealing process which inte-
grates the mechanisms of experimental design methods to maximize cus-
tomer satisfaction of a new product by optimizing the design attributes of
the new product. A case study involving the optimization of a packing-
machine design is presented to demonstrate the effectiveness of the
approach.
• Chapter 10 discusses a hybrid evolutionary algorithm which is integrated
with a statistical method, namely orthogonal design, to optimize customer
satisfaction with a new product. A case study of the optimization of a car
door design is used to illustrate the effectiveness of the hybrid evolutionary
algorithm.
• Chapter 11 presents a hybrid algorithm which integrates the mechanisms
of fuzzy regression and genetic programming to develop models for manu-
facturing processes in order to improve the quality of the manufactured
product. It overcomes the existing fuzzy regression approaches which can-
not address the nonlinearity of manufacturing processes. A case study of a
solder paste dispenser is used to demonstrate the effectiveness of the hybrid
algorithm.
Preface VII
• Chapter 12 discusses a rule-extraction-based genetic algorithm approach

which aims to extract information in rule format from non-informative data
(Kwong et al. 2009b). It allows engineers to understand the behaviors of
the manufacturing systems, in order to enhance the quality of the manufac-
tured products. A case study of an epoxy dispensing process is used to
demonstrate the operations and procedures of the rule extraction approach.
• Chapter 13 summarizes the research on computational intelligence meth-
ods and product design issues. It also provides several suggestions for
future research in these areas.
Acknowledgements
We express our sincere thanks to Ms Bruna Pomella for proofreading the book,
and Professor Elizabeth Chang for providing wonderful research environment in
the Digital Ecosystems and Business Intelligence Institute including solid discus-
sions and the equipment used in conducting some of the experiments presented in
the book.
Contents
1 Integrated Product Design .............................................................................1

1.1 Introduction .............................................................................................1
1.2 Determination of Importance of Customer Requirements.......................4
1.3 Identification of New Product Opportunities...........................................9
1.4 Functional Modeling of the Relationships between Customer
Requirement and Design Attributes.......................................................11
1.4.1 Linear Modeling Methods.........................................................14
1.4.2 Nonlinear Modeling Methods ...................................................15
1.5 Maximization of Overall Customer Satisfaction and
Determination of Design Attribute Setting of a New Product ...............15
1.6 Development of Manufacturing Process Models for Quality
Prediction of Manufactured Products ....................................................19
1.7 Conclusion.............................................................................................21
References .......................................................................................................22
2 Computational Intelligence Technologies for Product Design..................25

2.1 Introduction..........................................................................................25
2.2 Modeling Approaches ..........................................................................26
2.2.1 Fuzzy Regression ....................................................................28
2.2.1.1 Tanaka’s Fuzzy Regression .....................................30
2.2.1.2 Peters’ Fuzzy Regression.........................................30
2.2.2 Neural Networks......................................................................33
2.2.2.1 Different Configurations of Neural Networks .........34
2.2.2.2 Learning Algorithms for Neural Network
Weights ....................................................................40
2.3 Stochastic Optimization Approaches ...................................................43
2.3.1 Simulated Annealing ...............................................................43
2.3.2 Evolutionary Algorithm...........................................................46
2.3.3 Particle Swarm Optimization...................................................48
2.4 Summary of This Chapter ....................................................................52
2.5 Application of Computational Intelligence Techniques to
Product Design within This Book ........................................................53
References .....................................................................................................55
X Contents
3 Determination of Importance of Customer Requirements Using

the Fuzzy AHP Method................................................................................59
3.1 Introduction..........................................................................................59
3.2 Hierarchical Structure for the Development of Customer
Requirements .......................................................................................60
3.3 Fuzzy Representation of Pairwise Comparison ...................................61
3.4 Fuzzy AHP...........................................................................................63
3.5 Case Study: Removable Mountain Bicycle Splashguard .....................65
3.5.1 Developing a Hierarchical Structure of Customer
Requirements for Bicycle Splash-Guard Design .....................65
3.5.2 Constructing Fuzzy Comparison Matrices ..............................66
3.5.3 Computing Importance Weights of Customer
Requirements...........................................................................68
3.7 Conclusion ...........................................................................................75
References .....................................................................................................76
4 An Enhanced Fuzzy AHP Method with Extent Analysis for

Determining Importance of Customer Requirements...............................79
4.1 Introduction..........................................................................................79
4.2 Overall Customer Satisfaction on Hair Dryer Design..........................79
4.2.1 Development of the Fuzzy Matrix...........................................80
4.2.2 Pairwise Comparison of Customer Requirements ...................81
4.2.3 Calculation of the Consistency Index and Consistency
Ratio ........................................................................................85
4.2.4 Determination of Weight Vectors for Customer
Satisfactions ............................................................................86
4.2.5 Comparison of Fuzzy Numbers...............................................87
4.3 Conclusion ...........................................................................................92
References .....................................................................................................92
5 Development of Product Design Models Using Classical

Evolutionary Programming .........................................................................95
5.1 Introduction ...........................................................................................95
5.2 Classical Genetic Programming ............................................................96
5.2.1 Model Representation ...............................................................98
5.2.2 Fitness Function ........................................................................99
5.2.3 Crossover and Mutation ..........................................................100
5.2.4 Selection and Convergence .....................................................101
5.3 A Case Study of Digital Camera Design .............................................102
5.4 Conclusion...........................................................................................107
References .....................................................................................................107
Contents XI
6 Development of Product Design Models Using Fuzzy

Regression Based Genetic Programming..................................................111
6.1 Introduction .........................................................................................111
6.2 Fuzzy Regression Based Genetic Programming..................................112
6.2.1 Specification of the Form of the Fuzzy Regression
Model ......................................................................................112
6.2.2 Determination of Fuzzy Coefficients ......................................113
6.2.3 Pseudocode of Algorithm........................................................113
6.2.3.1 Functional Model Representation............................115
6.2.3.2 Fitness Function ......................................................116
6.2.3.3 Evolutionary Operations..........................................117
6.3 An Illustrative Example.......................................................................117
6.3.1 Mobile Phone Design..............................................................117
6.3.2 Functional Model Development..............................................120
6.3.3 Optimization of Affective Design ...........................................124
6.4 Conclusion...........................................................................................125
References .....................................................................................................126
7 Generalized Fuzzy Least Square Regression for Generating

Customer Satisfaction Models ...................................................................129
7.1 Introduction........................................................................................129
7.2 Theoretical Background of Generalized Fuzzy Least Squares
Regression..........................................................................................130
7.3 Modeling Functional Relationships Using Generalized Fuzzy
Least-Squares Regression (GFLSR) ..................................................133
7.4 An Illustrative Case: Packing Machine Design..................................138
7.4.1 Establishing a HOQ for Packing Machine Design ................138
7.4.2 Normalizing Engineering Performance Values of
Engineering Characteristics...................................................138
7.4.3 Development of Functional Models Regarding QFD............140
7.5 Conclusion .........................................................................................142
References ...................................................................................................142
8 An Enhanced Neuro-fuzzy Approach for Generating Customer

Satisfaction Models.....................................................................................145
8.1 Introduction ........................................................................................145
8.2 An Enhanced Neural Fuzzy Network Approach................................145
8.2.1 Development of Neural Fuzzy Network Models...................146
8.2.2 Extraction of Significant Fuzzy Rules and the
Corresponding Internal Models Using a Proposed Rule
Extraction Method .................................................................148
8.3 Case Study: Notebook Computer.......................................................150
8.4 Conclusion .........................................................................................160
References ...................................................................................................161
XII Contents
9 Optimization of Customer Satisfaction Using an Improved

Simulation Annealing .................................................................................163
9.1 Introduction........................................................................................163
9.2 Development of Neighbourhood Function Based on Orthogonal
Experimental Design for Product Design Purposes ...........................164
9.2.1 Orthogonal Array Based Neighbourhood Function
(ONF) ....................................................................................164
9.2.2 An Improved Orthogonal Array Based Neighbourhood
Function.................................................................................166
9.3 A Case Study: Emulsified Dynamite Packing Machine.....................168
9.4 Conclusion .........................................................................................173
References ...................................................................................................174
10 An Enhanced Genetic Algorithm Integrated with Orthogonal

Design.........................................................................................................177
10.1 Introduction......................................................................................177
10.2 Orthogonal Array Based Crossovers................................................178
10.2.1 Orthogonal Crossover (OC)..............................................179
10.2.2 Main Effect Crossover (MC) ............................................182
10.3 Interaction Crossover (IC) ...............................................................184
10.4 A Case Study: Car Door Design ......................................................186
10.5 Conclusion .......................................................................................194
References ...................................................................................................195
11 A Nonlinear Fuzzy Regression for Developing Manufacturing

Process Models ..........................................................................................199
11.1 Introduction......................................................................................199
11.2 Nonlinear Fuzzy Regression ............................................................200
11.2.1 Model Representation ........................................................202
11.2.2 Fitness Function .................................................................203
11.2.3 Crossover and Mutation .....................................................204
11.2.4 Selection and Convergence ................................................204
11.3 Validation of Genetic Programming Based Fuzzy Regression
Approach to Modeling Manufacturing Processes ............................205
11.4 Conclusion .......................................................................................210
References ...................................................................................................211
12 Rule Extraction from Experimental Data for Manufacturing

Process Design ..........................................................................................213
12.1 Introduction ....................................................................................213
12.2 Fluid Dispensing for Microchip Encapsulation ..............................214
12.3 GA-Based Rule Discovery System.................................................215
12.3.1 Generation of Random Strings .........................................216
12.3.2 Fitness Evaluation.............................................................216
12.3.3 Selection and Convergence...............................................218
12.3.4 Crossover and Mutation....................................................219
Contents XIII
12.3.5 Rule Induction ..................................................................220

12.4 Results Verification........................................................................221
12.5 Conclusion......................................................................................226
References..................................................................................................226
13 Conclusion and Future Work ..................................................................229

13.1 Conclusions ....................................................................................229
13.1.1 Determination of Importance Weights for Customer
Requirements ..................................................................230
13.1.2 Development of Customer Satisfaction Models..............231
13.1.3 Optimization of Overall Customer Satisfaction ..............233
13.1.4 Development of Manufacturing Process Models for
Quality Prediction of Products ........................................233
13.2 Future Works..................................................................................234
13.2.1 Collection of Customer Survey Data Using Web
Mining.............................................................................234
13.2.2 Investigation of Innovative Computational
Intelligence Approaches..................................................235
References..................................................................................................235
Index ...................................................................................................................237
Chapter 1
Integrated Product Design
1.1 Introduction
The design of the late 1940’s automobile, the “Tucker 48”, is one of the most
proclaimed cases of failure in the annals of American industry, after the World
War II. With its Cyclops headlight which turned in tandem with the steering
wheel, its rear-mounted engine, as well as its aerodynamic sheet metal, the Tucker
48 model, demonstrated to the public in 1947, produced much excitement. How-
ever, the Tucker 48 was never given to a factory for manufacturing. Only 51 cars
were manufactured by hand, and they were all produced at enormous expense and
manpower. Only the engineering characteristics such as car speed and efficiency
of the engine were optimized or addressed by engineering personnel. Several im-
portant customer requirements such as low cost had not been considered. Existing
equipment, commonly used car components and available engineering skills were
inadequate for the large scale manufacture of the Tucker 48 with its relatively so-
phisticated design. It was hugely expensive to produce just one of them. These
cars were so costly that they were beyond the means of the general public. There-
fore, the development of Tucker 48 provides a valuable lesson that customer
needs, marketing issue and engineering constraints need to be considered in prod-
uct design stage.
More recently, product design has become part of the concurrent engineering
movement that blossomed in the 1980s. A good product design requires the
efficient and effective coordination of engineering, design, manufacturing and
marketing personnel. It is critical that product design incorporate the views of all
personnel during the earliest stages before manufacturing takes place. Figure 1.1
shows a product development model for supporting concurrent engineering that
involves marketing personnel, product designers, engineering personnel and man-
ufacturing personnel.
Marketing personnel are more concerned with market opportunities and posi-
tioning, choosing the right price and understanding ‘customer needs’ using market
surveys. Product designers are concerned more on the defining product characte-
ristics in terms of functionality, appearance, and design features. Engineering per-
sonnel emphasize the delivery of an optimal, feasible design, given a set of desired
K.Y. Chan et al.: Comput. Intell. Techniques for New Product Design, SCI 403, pp. 1–24.
springerlink.com © Springer-Verlag Berlin Heidelberg 2012
2 1 Integrated Product Design
performance targets, features and costs. Engineering personnel might be concerned

only with whether the engineering requirements are both feasible and satisfactory.
Manufacturing personnel mainly concern design of manufacturing processes and
their process parameter settings, as well as quality of manufactured products and
production cost. In short, these communities have different notions of the drivers
of success, optimization variables and the nature of constraints.
Fig. 1.1 Product design model
In general, marketing personnel aim to optimize market share, profits, and cus-
tomer satisfaction. They are also concerned with competitors’ current products in
the market. Engineering personnel focus mainly on the technical design of the
product. Product designers define product characteristics in terms of functionality,
appearance, and design elements.
As shown in Figure 1.2, the product designers develop product features and
specifications as well as create an outlook of a new product to fulfill customer re-
quirements. Based on the product specifications and the customer requirements
provided by marketing personnel, design (or engineering) specifications are de-
fined by the engineering personnel. A design specification normally consists of a
design attribute and its settings. These design specifications largely determine the
downstream product development activities and manufacturing processes. There-
fore, the establishment of design specifications is an important process in the de-
velopment of any new product.
1.1 Introduction 3
Marketing Engineering
Customer
personnel personnel
requirements
Engineering
requirements
Customer
Product designer
requirements
Design specification
Fig. 1.2 Design specification of the new product
In this chapter, a methodology is described, which aims to integrate marketing

and engineering for defining design specifications in order to maximize customer
satisfaction. This is an importance activity to include a multi-discipline review for
checking correctness, completion, and fulfillment of customer requirements of
new products.
The four main tasks comprising the methodology discussed in this chapter are
shown in Figure 1.3. First, we identify the new product opportunities in Section
1.2. Second, the importance of customer requirements of a new product will be
discussed in Section 1.3. Third, the generation of functional relationships between
customer requirements and design attributes of a new product, will be introduced
in Section 1.4. Fourth, in Section 1.5 the optimization of design attribute settings
of a new product will be introduced, in order to maximize overall customer satis-
faction with the product.
It is also important to consider manufacturing concerns in product design stage,
in order to achieve the manufactured products with better quality and lower pro-
duction cost. One common way to address the consideration is to predict the quali-
ty of manufactured products by developing manufacturing process models.
Section 1.6 presents the methodologies for developing manufacturing process
models, from which proper process parameter settings can be obtained and quality
of manufactured products can be predicted.
Identification of new product

opportunities
Determination of importance of
customer requirements
Modeling the functional relationships

between design attributes and
customer requirements
Maximization of customer
satisfaction
Design specifications
Fig. 1.3 A methodology for defining design specifications of new products
1.2 Determination of Importance of Customer Requirements

Customer requirements significantly affect the target values setting of design
attributes for a new product. Therefore, it is important to determine and prioritize
customer requirements which help to identify the optimal value setting of design
attributes for maximizing customer satisfaction.
For example, three customer requirements, low price, good hardness, and light
weight (as illustrated in Figure 1.4), are obtained for the development of a casing for a
notebook computer. These three customer requirements must be taken into considera-
tion when a product development team decides on the most appropriate material to be
used for the casing of the new notebook computer. If the product development team
uses stronger but heavier material for the casing, this is likely to increase the hardness
of the casing and prevent damage to the notebook computer if dropped. However, a
stronger but heavier material will mean an increase in both the cost of the material and
the weight of the case. Stretching the frontier and improving satisfaction relative to
primary needs is a higher priority such as increasing the hardness of the casing by us-
ing stronger material. However, customers may prefer a lighter notebook computer
with a less sturdy casing, which they may handle more carefully.
1.2 Determination of Importance of Customer Requirements 5
Good hardness
Lower Cost
Light weight
Customer
requirements
Determination of design attribute set-

tings for the casing of notebook
t
Appropriate materials for the casing

of the new notebook computer
Fig. 1.4 Design of notebook computer case
Therefore, before designing a new product, it is critical to acknowledge and

rank the importance of customer requirements, in order to maximize the overall
customer satisfaction with a new product. It is a reasonable hypothesis that cus-
tomers are able to state and prioritize most of their important requirements. A
product development team could save time and money by using the frequency of
customer statements as an indication of their level of importance. As shown in
Figure 1.5, four steps are usually involved in the determination of the importance
of customer requirements.
Step 1: Identification of Customer Needs. Step 1 determines the custom-

ers’ needs and concerns regarding a product. This must be done, before the
customer requirements of the product are ranked in terms of their importance.
Customer needs are usually expressed verbally, and their statements are
collected by means of focus groups or individual interviews. Individual face-
to-face interviews are usually more cost effective than a focus group of cus-
tomers. Griffin and Hauser (1993) mentioned that 20± 30 customers should
be interviewed in order to obtain 90± 95% of all possible customer needs.
However, conducting customer surveys based on mail or telephone responses
is usually not suitable for collecting qualitative data regarding customer
needs or wishes regarding a new product, because it is not easy to control the
scope of customers’ responses.
Step 1 Identification of
customer needs
CR1, CR2, CR3, CR4
Step 2 Relative importance Step 3 Competitive analysis

ratings based on against competitors’
customer survey products
CR1>CR3> CR4> CR2 CR1>CR2> CR3> CR4
Step 4 Final importance

ratings
CR1>CR3> CR2> CR4
Fig. 1.5 Determination of importance of customer requirements
Words collected during customer interviews are either too brief or too
complex to be directly interpreted as customer wishes or needs. The words
collected are usually organized in a tree-representation of hierarchical struc-
ture, which illustrates different levels of customer wishes and needs, in order
to assist in identifying real customer requirements. Those at a specific level
can be chosen as the final customer needs based on the situation. In an Af-
finity Diagram, natural and logical groups can be produced by arranging a set
of random data (Bossert 1991, Cohen 1995). For example, customer needs
can be produced or structured based on the approach of cluster analysis (Grif-
fin and Hauser 1993).
1.2 Determination of Importance of Customer Requirements 7
Step 2: Relative Importance Ratings Based on Customer Data. Following

the previous step, the customer input then consists of their rating the impor-
tance of their expressed needs or wishes. The most important customer needs
or wishes should be focused by the company, which should also disregard the
less important ones, in order to efficiently use its resources. To do this, one of
three different conventional methods can be used to measure or estimate the
level of importance of customer requirements:
• Relative importance ratings are provided by the customers. For example,
five-, seven- or nine-points can be used as scales. Higher resolution scales,
like the one-to-ten point scale and anchored scale, can also be used (Griffin
and Hauser 1993).
• First, all primary needs of the product are ranked by the customers with
twenty points. Then, the secondary needs, which are below each individual
primary need, are assigned with twenty points. Finally, each set of tertiary
needs, which are below each individual secondary need, are also assigned
with twenty points.
• The highest priority primary needs are assigned by the customer with
twenty points. After that, the other five primary needs are assigned by the
customers with twenty points. Then, the second priority needs under each
primary need are assigned with twenty points by the customers. Finally, the
tertiary needs under each secondary need are assigned with twenty points.
Customer
Customers survey Customer
data
Customers’ Web mining

Web-blog opinion (web 2.0)
Fig. 1.6 Collection of customer data
A commonly used approach for collecting this information is via customer

surveys using mail or telephone. Survey data given by focus groups or indi-
vidual interviews is usually not satisfactory for collecting a huge amount of
quantitative information, in order to determine the importance of customer
needs or wishes. Therefore, a completed survey, which involves a sufficient
number of interviews, needs to be conducted, in order to ensure the statistical
significance of results.
However, a significant amount of time is required to collect the data, and
also the size of the collected data is usually not enough. Therefore, as shown
in Figure 1.5, the Web data collection method, using Web 2.0 (Dillon et al.
2009), can also be used to collect customer opinion data from the Web and
then this customer opinion data is transformed into customer survey data.
Four mining approaches based on Web 2.0 are introduced:
 The sentiment-based approach classifies opinions into three
groups – positive, negative, and neutral. For example, a Weblog
entry “all Sony notebook computer cases are made with hard ma-
terial which is too heavy to carry” indicates a positive attitude to
lighter but less hard material.
 The item-based approach identifies distinct ‘items’ within the
same opinion context such as the page view, the blog entry, the
post message, etc, which often combines a number of different
products. Sometimes, a negative customer opinion does not neces-
sarily mean that this customer dislikes every aspect of the product,
and vice versa. For example, IBM notebook computer cases made
with hard material are heavy but may be more durable, since they
are not easily damaged.
 The feature-based approach models a product as a number of
sub-components, each of which is associated with a set of design
attributes that can be evaluated through expressions of opinion.
 The comparison-based approach invites opinions that often com-
pare the features of two similar products. For example, a customer
may write in a post: “The case of this notebook computer is much
harder than the other one”. From the market intelligence perspec-
tive, such a comparative opinion can be useful as it may provide
suggestions on product features and information about competi-
tors’ products. The customer survey data reflects customers’ opi-
nions distributed across the Web regarding competitive products
in the markets. Relative importance ratings can be extracted.
The exact choice of one of the methods for a product survey could vary for
different types of products. In addition, it may necessary to use a combina-
tion of the above methods.
Step 3. Evaluative Analysis of Competitors’ Products. In order to evaluate
a rival product in terms of the specified customer needs, potential customers
are asked to evaluate the performance of the company’s product and those of
the competitors’ products. Identifying strengths and weaknesses of the devel-
oped product, is also necessary, in order to obtain competitive advantages. To
do this, the potential customers are asked to rate the relative performance of
the products, which are developed by the company and the competitors. This
step is similar to Step 2. This information can be obtained based on customer
surveys using mail or telephone. As for Step 2, a particular customer group or
individual interviews are inappropriate for obtaining this information. The
reason is that these customers are able to evaluate the performance only of
the products, which are usually used by them or are familiar by them. There-
fore, a significant number of customers must be used for collecting this
information. Also, an analysis of the relative position of the company is es-
sential to update this information (Sullivan 1986, Bossert 1991, Cohen 1995).
1.3 Identification of New Product Opportunities 9
Also, Web 2.0 can be used to collect this information which will save time
in collecting data and the amount of data can be larger. It can be used to col-
lect customer opinion data from the Web and then this customer opinion data
is transformed into numerical data in order to rate the product’s performance
and the company’s main competitors.
Step 4. Final Importance Ratings. This step involves merging the relative
importance of customer requirements as described in Step 2, as well as the
competitive level of the company as described in Step 3. Then, the final im-
portance ratings of customer requirements are produced with respect to the
product, which is to be designed by the company. An indication of final rat-
ings of customer requirements can assist product designers and engineering
personnel to develop a product which is better to satisfy customer needs.
1.3 Identification of New Product Opportunities

The identification duct opportunities is important. It allows marketing personnel to
understand the marketing positions of various competitive products, identify poss-
ible new product opportunities, as well as determine the position of a new product
and clusters of competitive products.
It involves the following three steps. First, it identifies customer requirements
from examining competitive products. As the number of customer requirements is
usually more than two, the collected customer requirements cannot be depicted
with a two-dimensional graph. The second step is to transform multi-customer re-
quirements into two major components. Then, the third step is to construct a two-
dimensional perceptual map in order to graphically represent competing alterna-
tives in Euclidean space. This allows marketing personnel to identify regions of
new product opportunities.
For example, identification of opportunities of a new packing machine is consi-
dered. The matrix of the house of quality of this new packing machine which
shows the relationship between customer requirements and design attributes of the
new packing machine is used (Chen et al, 2004). In the packing machine design,
four major customer requirements are identified as “quality of packing” (CR1),
“efficiency of packing” (CR2), “packing noise” (CR3) and “rigidity of the ma-
chine” (CR4). Seven design attributes of the packing machine are also identified,
which are “precision of the moulding of the clip” (DA1), “precision of the pack-
ing” (DA2), “control force of the packing” (DA3), “efficiency of the packing”
(DA4), “hardiness of the pressing hammer” (DA5), “noise of the cam power
transmission” (DA6) and “height of the machine bed” (DA7).
Step 1) Competitive Product Analysis: Based on the matrix of the house of qual-
ity of this new packing machine, customer requirements of the competitive prod-
ucts as well as the importance of the customer requirements can be identified. Five
major competitive products are identified which are denoted as Comp1, Comp2,
Comp3, Comp4 and Comp5 respectively. They are identified by the matrix of the
house of quality of this new packing machine (Chen et al, 2004). Table 1.1 shows
the importance of the four specified customer requirements, CR1, CR2, CR3, and
CR4. Also, the benchmarks of the five major competitors regarding these four cus-
tomer requirements are shown.
Table 1.1 Benchmarks of the five major competitors
Competition benchmarking
Customer Importance Comp1 Comp2 Comp3 Comp4 Comp5 Min Max
require- of each cus-
ments tomer re-
quirements
CR1: y1 0.46 3.4 4 1.9 3.7 3.6 1 5
CR2: y2 0.28 3.1 3 1.8 2.9 3.9 1 5
CR3: y3 0.16 2.2 3.7 4.3 1.8 3.5 1 5
CR4: y4 0.10 1.6 3.7 3.3 3.7 4 1 5
Step 2) Transforming Customer Requirements Into Two Major Components:

Factor analysis (Johnson and Wichern, 1992) is then used to transform the cus-
tomer requirements into two major components. It is a statistical technique which
is commonly used to uncover the latent structure (dimensions) of a set of va-
riables. It reduces attribute space from a larger number of variables to a smaller
number of factors and as such is a "non-dependent" procedure. To apply factor
analysis for generating perceptual maps, firstly, the correlations among all cus-
tomer requirements (CRs) are calculated using benchmarking data of the matrix of
house of quality. After that, underlying factors are obtained based on the correla-
tion matrix from the house of quality. Factor loadings are then calculated which
represent the correlation coefficients between the variables and factors. Then, the
underlying factors are rotated. The purpose of the rotation is to make the output
more understandable and enable the factors to be interpreted more easily.
Using factor analysis, the two factors with the largest values of factor loading
were determined regarding the four customer requirements, CR1, CR2, CR3, and
CR4. They are denoted as f1 and f 2. The results of the factor analysis are shown in
Tables 1.2 and 1.3. Table 1 shows the rotated factor loadings and communalities
of each dimension of customer satisfaction, while Table 2 shows the factor score
coefficients of f 1 and f 2.
Table 1.2 Rotated factor loadings and communalities
Variable Factor 1 (f1) Factor 2 (f2) Communality

y1 0.949 0.041 0.903
y2 0.906 -0.055 0.824
y3 -0.535 -0.722 0.808
y4 0.219 -0.905 0.867
1.4 Functional Modeling of the Relationships 11
Table 1.3 Factor score coefficients
Variable Factor 1 (f1) Factor 2 (f2)

y1 0.464 -0.030
y2 0.449 -0.100
y3 -0.217 -0.508
y4 0.166 -0.695
From Table 1.2, it can be observed that f1 is dominated by CR1 and CR2 while f2
is dominated by CR3 and CR4. From Table 1.3, the relationships between the two
underlying factors and the four CRs can be written as follows:
f1 = 0.464y1 + 0.449y2 − 0.217y3 + 0.166y4,
f2 = −0.030y1 − 0.100y2 − 0.508y3 − 0.695y4,
where yi (i = 1 to 4) is the degree of satisfaction of CRi .
Step 3) Construction of Perceptual Map: Based on the above two models, a

perceptual map is constructed as shown in Figure 1.7. The positions of the five
major competitive products are shown on this map. There are three regions of
possible new product opportunities, which were identified as the high potential
product opportunities in which a successful new product would be developed. The
input regions are commonly in the forms of circles, ellipses and rectangles. Here it
is assumed that the form of all the regions is elliptical.
1.4 Functional Modeling of the Relationships between

Customer Requirement and Design Attributes
1.4 Functio nal Mo deling of t he Relationships
One of the main tasks of product design is to identify target value settings of de-
sign attributes of products for achieving a high degree, or even the maximum cus-
tomer satisfaction. The process is very complex, as it usually involves quite a
number of design attributes and customer requirements of products, and also the
interactions among the design attributes of products need to be considered. Prior
to obtaining maximum customer satisfaction, models for relating DAs and dimen-
sions of CR as well as those among Das, need to be developed.
The house of quality can also be used to develop the models for relating CRs
and DAs as well as a DA and other DAs based on the matrix of the house of quali-
ty. It can also be used to develop a relationship between a DA and other DAs. It
indicates the interrelationship of CRs to DAs of the new product. Based on the
house of quality, a structured methodology can be developed to specify the cus-
tomer requirements and evaluate the performances of competitive products in or-
der to satisfy all these CRs (Clausing and Hauser 1988). The use of the matrix of
2.5
2
1
1.5
Region 2 Region 1
1
Second Factor
0.5
4
0
2
-0.5
3
-1
Region 3 5
-1.5
-2
-2.5
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
First Factor
Fig. 1.7 Perceptual map of packing machines
the house of quality has gained extensive international support for product plan-
ning and development decisions. The fundamental tenet of the matrix of the house
of quality is to increase customer satisfaction by satisfying all CRs. When organi-
zations intend to meet all customer requirements, the internal conflict can be mi-
nimized. Also, the cycle time for development can be decreased, and the market
penetration can be increased. Hence, it results in higher revenues (Cohen 1995).
The matrix of the house of quality of a packing machine is considered as an
example as shown in Figure 1.8. In the QFD, it is assumed that the subjective
symbols “_, O” and their corresponding values can truly indicate significant rela-
tionships between the CAs and their associated DAs. “_” indicates that the rela-
tionship between DAi and CRj is weak. “O” indicates that the relationship between
DAi and CRj is strong. For example, the relationship between DAi and CRj is
strong and the relationship between DAi and CRj is weak. In the machine design,
there are four major customer requirements, CR1, CR2, CR3, and CR4, and six de-
sign attributes, DA1, DA2, DA3, DA4, DA5, and DA6.
The house of quality relating the customer requirements to the design attributes
is usually used to redevelop an old product or develop a new product. Subjective
conclusions of customer surveys can be used to determine the values of customer
preferences. The house of quality is normally analyzed in a relatively trivial ap-
proach by engineering personnel. Relationships between CRs and DAs are usually
determined by engineering personnel using their previous experience as well cus-
tomer inputs.
1.4 Functional Modeling of the Relationships 13
DA1 DA2 DA3 DA4 DA5 DA6 DA7

x1 x2 x3 x4 x5 x6 x7
ECs
CRs W Relation
CR1 y1 0.46 O O O - - - -
CR2 y2 0.28 - - O O O - -
CR3 y3 0.16 - - - - - O -
CR4 y4 0.10 - - - - - - O
-2 -2 -1
Unit m m N ns HRC dB M
Fig. 1.8 House of Quality for packing machine design
However, applications for product design are mostly not cooperative. Empirical
models that represent relationships between CRs and DAs are not only useful for
product development teams to understand between these two factors, but also op-
timization of products can be conducted based on the empirical models. These
empirical models are generated based on customer survey data or QFD as shown
in Fig. 1.9. Modeling methods including linear and nonlinear modeling methods
are used. They are detailed in Section 1.4.1 and Section 1.4.2 respectively.
Linear modeling Nonlinear modeling
Customer Modeling
survey data methodologies
DA1: x1 Customer
Design DA2: x2 requirement
Explicit model
attributes :
:
CR: y
DAn: xn
Fig. 1.9 Modeling methodologies for generating empirical models

1.4.1 Linear Modeling Methods

A product’s overall desirability which depends upon the customer value functions
is suggested by Yoder and Mason (1995). Each CR, implicitly models the combi-
nation of the customer’s preferences and objectively captures those desires. In or-
der to develop an explicit model with respect to an individual CR, a first-order
linear model in term of each CR, is in the following form can be used:
CR = β 0 + β1 x1 + β 2 x2 + ... + β n xn (1.1)
The equation (1.1) can be derived by the method of least squares for each CR.
Sample data on customer assessments for various product configurations are used
for fitting the parameters in equation (1.1). Analysis of residuals and the amount
of curvature reveals the validity of a value function.
Various approaches have been attempted in the modeling including statistical
regression (Dawson and Askin 1999) and fuzzy linear regression (Fung et al 2006,
Chen and Chen 2006). The approach of multiple linear regression (Chuang and
Ma 2001, Kuang and Jiang 2008) has been used to model relationships between
design attributes and customer requirements. This approach is simple to apply but
it assumes that the design attributes in the regression are linear, and that the effect
of an independent design attribute is constant throughout the entire range of the
customer requirements. Under this assumption, more and more terms of design
attributes are included in the model in order to fit a wide range of customer re-
quirements. It increases the number of terms of design attributes in the model of
the customer requirements, and the terms of the design attributes involved cannot
be guaranteed to be significant for the design attributes. Thus, the resulting model
is more complex and more difficult to interpret when many design attributes are
involved (Han and Hong 2003).
As modeling the relationships based on a fuzzy linear regression can yield ex-
plicit models using a small number of data sets, it can be adopted to model the
relationships. In addition, fuzziness of the relationships between customer re-
quirements and design attributes as well as a design attribute and other design
attributes can be addressed by fuzzy regression properly. Details of the modeling
using fuzzy regression can be found in the authors’ publications (Chen et al.
2004). To address the fuzziness of the modeling, quite a few previous studies have
adopted the fuzzy set theory on modeling the relationship. Kim and Park (1998)
suggested a fuzzy regression approach to estimate the functional relationships in
QFD. Chen et al. (2004) proposed another fuzzy regression approach, based on
asymmetric triangular fuzzy coefficients, to develop models for the functional re-
lationships in QFD. The use of nonlinear programming to develop fuzzy regres-
sion models for modeling the functional relationships in QFD was proposed by
Chen and Chen et al (2006).
1.5 Maximization of Overall Customer Satisfaction 15
1.4.2 Nonlinear Modeling Methods

Linear regression models (with no higher orders and interaction terms) are not al-
ways appropriate to model customer requirements, since it cannot be assumed that
all design attributes are linearly independent of each other. Therefore, developing
a nonlinear model which is represented by a higher order polynomial form may be
necessary, in order to address some levels of nonlinearity. The model with higher
orders and interaction terms is represented as:
CR = β 0 + β1 x1 + ... + β n xn + β11 x12 + β 22 x2 2 ... + β nn xn 2 (1.2)
+ β 12 x1 ⋅ x2 + β 12 x1 ⋅ x3 ... + β n−1,n xn−1 ⋅ xn
Quite a number of studies have attempted to investigate the modeling of the rela-
tionship between customer requirement and design attributes, in order to address
nonlinearity. Chen et al. (2006) developed a prototype system for affective design
in which Kohonen’s self-organizing map neural network was employed to conso-
lidate the relationship between design attributes and customer requirement. Park
and Han (2004) adopted the fuzzy rule-based approach to build models relating
customer requirements to design attributes. Hsiao and Tsai (2005) proposed a me-
thod that enables an automatic product form search or product image evaluation
by means of a neural-network-based fuzzy reasoning genetic algorithm. The neur-
al-network-based fuzzy reasoning algorithm was applied to establish relationships
between the input form parameters and a series of adjectival image words. Liu et
al. (2007) proposed a fuzzy model to examine a customer satisfaction index in e-
commerce. They proposed a method to calculate the index based on a five-level
quantity table using fuzzy techniques. However, the model they developed was
implicit. Lin et al. (2007) proposed a fuzzy logic model to determine the consum-
er-oriented mobile phone form design. From their experimental results, they re-
ported that the fuzzy model outperformed two neural-network-based models in
terms of the root of mean square errors. Grigoroudis and Siskos (2002) developed
the multi-criteria satisfaction analysis (MUSA) method for measuring and analyz-
ing customer satisfaction. MUSA is a preference disaggregation model based on
the working principles of ordinal regression analysis. Using the survey data,
MUSA aggregated individual judgments into a collective value function so as to
quantify customer requirements. The model assumed that global or overall cus-
tomer requirements were measured solely with respect to a number of customer
attributes.
1.5 Maximization of Overall Customer Satisfaction

and Determination of Design Attribute Setting
of a New Product
1.5 Maximization of Overall C usto mer Satisfaction
The final task in product design is to maximize overall customer satisfaction

(OCS) which indicates the degree of satisfaction with which a new product can
fulfill all customer requirements. It can be maximized by optimizing the design
attributes of a new product. As illustrated in Figure 1.10, it depends on two fac-

tors – customer requirements of the products, CR1, CR2, CR3, …, CRm, as well as
their importance w1, w2, …w3, where there are m customer requirements of the
product.
Overall customer satisfaction

OCS
+
w1
w2 wm
CR1: y1 CR2: y2 ... … CRm: ym
The m customer requirements
Fig. 1.10 Relationship between customer requirements and overall customer satisfaction
A function for deriving overall customer satisfaction can be constructed by ag-

gregating various dimensions of customer requirements yi, i=1,…,m, and their im-
portance weights wi, i=1,…,m. This can be regarded as multiple attribute synthesis
at the system level. In most previous studies, the linear weighted sum method was
used to derive OCS, which can be expressed as equation (1.3).
m
OCS= ∑ wi yi (1.3)
i =1
However, in general it cannot be assumed that all customer requirements would

contribute to overall customer satisfaction at the same level. According to the
Kano model as shown in Figure 1.11, usually there are three major types of overall
customer satisfaction. The first, depicted by the diagonal line, represents explicit
requirements, which are directly related to OCS. The second type represents inno-
vations, as shown by the curved line in the upper left corner of the figure. If the
associated requirements are fulfilled, OCS will increase dramatically. The third
one represents “must-have” requirements, as shown by the lower right curve of the
figure. When these customer requirements are not met, the customers will be very
dissatisfied; but if they are met, they are taken for granted, and OCS remains
relatively stable.
1.5 Maximization of Overall Customer Satisfaction 17
High overall customer satisfaction
Satisfied customer
Unsatisfied customer requirements
requirements Satisfied
Low overall customer satisfaction
Fig. 1.11 Kano Model
To consider different kinds of customer requirements as depicted in the Kano

model in real-world problems, a new objective function of deriving OCS based on
a generalized weighted mean method (Kano, 1984) has been constructed and is
presented as follows:
∑ϖ ∑ϖ ∑ϖ
k l m 1
OSC = ( i y is ) s + ( i yi ) + ( i y it ) t (1.4)
i =1 i = k +1 i = l +1
yi with i=1,…, k denotes the customer requirement of “Attractive” type customer

requirements. yi with i=k+1,…,l denotes the customer requirement of “explicit”
type customer requirements. yi with i=l+1,…,m denotes the customer requirement
of “must-have” type customer requirements. The value of s and t are dependent
on the nature of the new product design.
Based on (1.4), an optimization model can be developed to determine a set of
levels of attainment of design attributes for a new product. The mechanisms of the
optimization model are summarized in Figure 1.12. It aims to maximize the OCS
by optimizing the design attributes for a new product which is restricted by the
product positioning constraints and engineering constraints.
An optimization model is formulated to determine a set of levels of attain-
ment of design attributes, x1, x2, …,xn, for a new product in order to maximize the
OCS subject to the product positioning constraints and other constraints. If
the λ th region of new product opportunities is selected from the perceptual
Optimal design attribute

setting
x1 x2 ……………………………. xn
…………………………….
Engineering Maximize the OCS based on Marketing

constraints the optimization model concerns
y1 , y2 ……………………………. ym
CR1 CR2 ……………………. CRm
Customer requirements
Fig. 1.12 An optimization model for determining an optimal design attribute setting of a
new product
product map, the optimization model of integrating marketing with engineering

for determining value setting of design attributes of new products can be formu-
lated using the following Kano model:
k l m
OCS = (∑ϖ i yis ) s + ( ∑ ϖ i yi ) + ( ∑ ϖ i yit )t
1
(1.5)
i =1 i = k +1 i = l +1
Subject to
yi = f i ( x ) , i=1,2,…,m (1.6)
φλ = pλ ( f ) ≤ 0 , λ = 1," ,η (1.7)
yimin ≤ yi ≤ yimax , i=1,2,…,m (1.8)
0 ≤ xj ≤ 1, j=1,2,…,n (1.9)
where (1.6) refers to the models of the functional relationships between customer
requirements and design attributes. (1.7) refers to the constraint of the position of
the new product; (1.8) and (1.9) refer to the ranges of customer satisfaction values
and levels of attainment of the j-th design attribute respectively.
The landscape of the optimization problem (1.5-1.9) is a constrained nonli-
near function. As it involves multi-optima characteristics, gradient methods, which
require gradient information to trace the local or global optimum, cannot be ap-
plied to solve the optimization problem (1.5-1.9). They may find only the sub-
optima of the problem. To solve this problem, computational intelligence methods
can be used such as simulated annealing, evolutionary computation and particle
swarm optimization, all of which have a rich literature for solving for the global
optimum of this kind of problem.
1.6 Development of Manufacturing Process Models 19
1.6 Developme nt of Manufacturing Process Models
1.6 Development of Manufacturing Process Models for Quality

Prediction of Manufactured Products
It is common that settings of some design attributes are used as the quality re-
quirements of manufactured products. To ensure that the specified quality re-
quirements of the manufactured product can be fulfilled, manufacturers need to
control variability at each of the many processing steps in the manufacturing
process. Also, all of the variables that impact on the desired quality of a new
product need to be understood and optimized to maintain tight control.
To do this, a manufacturing process model that relates the quality requirements
of the new product to the manufacturing process variables needs to be developed.
The model can help engineering personnel to predict the quality of manufactured
products in the product design stage and determine proper setting of process pa-
rameters. For example, injection moulding is an important process in the manufac-
ture of plastic products. The quality requirements of these products are controlled
by the settings of the process parameters. Thus, determination of an appropriate
setting of process parameters for injection moulding is important to the quality of
moulded plastic products. There are 12 process parameters to be considered: the
hydraulic diameter of runners, z1, the hydraulic diameter of gates, z2, the melt tem-
perature, z3, the mould temperature, z4, the injection pressure, z5, the holding pres-
sure, z6, the back pressure, z7, the clamping force, z8, the screw surface speed, z9,
the fill time, z10, the holding time z11, and the cooling time z12,. For this injection
moulding process, the quality of injection moulded plastic parts is denoted by the
model with the consideration of molding parameter setting which is illustrated in
Figure 1.13. Analytical models or empirical models can be used to represent the
manufacturing process, which relate the process parameters with the quality re-
quirements of products.
Analytical models are based on a physical understanding of the manufacturing
process, and rely on physical laws, typically a set of governing partial differential
equations. They are attractive because they provide a fundamental understanding
of the relationships between the input and output parameters. For example, various
analytical models have been developed for manufacturing processes, such as fluid
dispensing (Chen 2002, Li et al. 2001), injection moulding (Chiang et al. 1991),
and transfer moulding (Han et al. 2000), However, many manufacturing processes
are too complex to model accurately and analytically. Also, many idealizing
assumptions are required which is not practical in real-world applications.
Empirical modeling is a popular approach to the development of process
models, based on using experimental data. It can be divided into five approaches:
Statistical regression, artificial neural networks, fuzzy logic modeling technique,
fuzzy regression and evolutionary programming.
Quality requirements of the

plastic product
Process model for injection

moulding
z1 , z2 ……………………………. z12
Process parameter setting of
injection moulding
Fig. 1.13 Process model for injection moulding
Statistical Regression Method is a common empirical approach to the develop-

ment of such process models (Seber 2003). It is accurate only in the range in
which they are developed. In conventional regression analysis, deviations between
the observed values and the estimates are assumed to be due to random errors.
Thus, statistical techniques can be applied in order to make estimates and infe-
rences in regression analysis. Statistical regression models can be applied only if
the given data is distributed according to a statistical model, and the relationship
between dependent and independent variables is crisp. However, in some manu-
facturing processes, it is difficult to find probability distributions for dependent
variables. The deviations in such cases are due to the indefinite structure of the
system or to imprecise observations. The uncertainty in this type of process mod-
eling becomes fuzzy.
Artificial Neural Networks (Simpson 1989) have been used to develop process
models for various manufacturing processes, such as resistance spot welding
(Li et al., 2007) and transfer moulding (Tong et al., 2004). These networks have
the capability to transform a non-linear mathematical model into a simplified
black-box structure, and have the advantage of learning and generalization abili-
ties, as well as nonlinearity. Previous research has already confirmed that neural
networks are powerful tools for modeling nonlinear, complex, and noisy
processes.
A Fuzzy Logic Modeling Technique has been successfully applied to the develop-
ment of models for various manufacturing processes, such as the Flip-Chip bonding
process (Kang et al., 1993), vapor phase soldering (Xie et al. 1994), and the waterjet
depainting process (Babets and Geskin, 2000). The basic elements of a fuzzy logic
model are internal functions, membership functions, and outputs. The use of several
internal functions accounts for the fuzziness of the model. However, fuzzy logic mod-
eling approaches have the same limitation as have neural networks due to their simple
black-box nature. Because of their lack of transparency, sensitivity studies of process
parameters cannot be easily undertaken.
1.7 Conclusion 21
Fuzzy Linear Regression has the distinct advantage that a manufacturing process,
which has a high degree of fuzziness, can be modeled by using only a few or even in-
complete experimental data sets (Tanaka et al. 1982, Takagi and Sugeno 1985, Tanaka
and Watada 1988). An attempt was made by Schaiable and Lee (1997) to model the
vertical CVD process using the fuzzy linear regression method. Lai and Chang (1994)
applied fuzzy linear regression to model the die casting process. Ip et al. (2003a) intro-
duced the fuzzy linear regression to develop a process model for epoxy dispensing.
Modeling of transfer moulding using fuzzy linear regression was reported by Ip et al.
(2003b). Kwong and Bai (2005) have performed process modeling and optimization
using both fuzzy linear regression and fuzzy linear programming approaches. Three
different approaches of fuzzy linear regression were summarized in Chang and Ayyub
(2001). However, the existing fuzzy regression approaches cannot be used to develop
models which contain interaction terms or even higher order terms. In fact, the interac-
tion among process parameters and the nonlinear behavior of manufacturing processes
commonly exist. If interaction terms or higher order terms were to be integrated into
the approach of fuzzy regression, more accurate models could be developed.
Evolutionary Programming can be used to generate models with interaction terms
or higher order terms (Koza 1994). Lakshminarayanan et al. (2000) and Madar et al.
(2005) have demonstrated how evolutionary programming can be used to generate
models with interaction terms or higher order terms, and the least square algorithm is
then used to perform the associated parameter estimation of the models. However,
quite a number of manufacturing processes involve uncertainty, due to fuzziness.
Therefore, the evolutionary programming which is integrated with the least square re-
gression algorithm may not provide the most appropriate model, as it cannot address
uncertainty, which comes from fuzziness.
The characteristics of the above modeling approaches are summarized in Table 1.4.
Table 1.4 Characteristics of the modeling approaches
Statis- Artificial Fuzzy Fuzzy Evolutionary

tical neural logic linear programming
regres- networks modelling regres-
sion methods sion
Transpa- Explicit Black- Black-box Explicit Explicit
rency box
Handling Poor Good Good Poor Good
nonlinear
process
Handling Poor Good Good Good Poor
fuzzy data
1.7 Conclusion
In new product development, it is quite common for marketing personnel, product
designers, engineering personnel and manufacturing personnel to have different
goals and concerns which are usually addressed in isolation from one another.
This isolation typically would not result in optimal design decisions as the two
sets of goals and concerns should always be interrelated. Therefore, it is important
to integrate the concerns of marketing personnel with those of engineers when de-
fining design specifications. In addition, manufacturing concerns should also be
considered in product design stage such that the products with better quality and
lower production cost are produced.
This chapter introduces methodologies for product design, which aim to inte-
grate marketing with engineering for defining design specifications of new prod-
ucts, in order to maximize overall customer satisfaction. This is a critical product
design activity, which includes the multi-discipline review for checking correct-
ness, completion and fulfillment of customer requirements. To do this, the follow-
ing tasks were discussed. First, Section 1.2 introduced a methodology for identify-
ing new product opportunities; second, Section 1.3 discussed ways to identify the
importance of customer requirements of a new product; third, the generation of a
functional models for modeling the relationships between customer requirements
and design attributes of a new product was discussed in Section 1.4; fourth, Sec-
tion 1.5 presented methods for determining optimal design attribute settings of a
new product, in order to maximize overall customer satisfaction with the new
product. To address the consideration of manufacturing concerns in product de-
sign stage, Section 1.6 introduced methods for developing manufacturing models
from which proper setting of process parameters can be obtained and quality of
manufactured products can be predicted.
Since the above involve the modeling of the functional relationships and manu-
facturing processes, they may have highly nonlinear and fuzzy characteristics. Al-
so, these tasks involve procedures requiring the solving of difficult optimization
problems, which could have multi-optimua. Computational intelligence methods
including fuzzy systems, neural networks and evolutionary computations have rich
literature in handling these modeling and optimization tasks. Also, they have con-
tributed to optimizing and modeling a wide range of manufacturing processes,
which increasingly are required to be more robust, more flexible, more responsive,
more complex and more efficient. These computational intelligence methods will
be introduced in Chapter 2.
References
Askin, R.G., Dawson, D.W.: Maximizing customer satisfaction by optimal specification of
engineering characteristics. IIE Transactions 32(1), 9–20 (2000)
Babets, K., Geskin, E.S.: Application of fuzzy logic for modeling of waterjet depainting.
Machining Science and Technology 4(1), 81–100 (2000)
Bossert, J.L.: Quality Function Deployment, A Practitioner’s Approach. ASQC Quality
Press, Milwaukee (1991)
Chang, Y.H.O., Ayyub, B.M.: Fuzzy regression methods – a comparative assessment.
Fuzzy Sets and Systems 119, 187–203 (2001)
Chen, Y.: Fuzzy regression-based mathematical programming model for quality function
deployment. International Journal of Production Research 42(5), 1009–1027 (2004)
References 23
Chen, D.X.: Modeling and off-line control of fluid dispensing for electronics packaging.
PhD thesis, University of Saskatchewan (2002)
Chen, C.H., Khoo, L.P., Yan, W.: An investigation into affective design using sorting tech-
nique and Kohonen self-organizing map. Advances in Engineering Software 37(5), 334–
349 (2006)
Chiang, H.H., Hieber, C.A., Wang, K.K.: A unified simulation of the filling and postfilling
stages in injection molding, Part 1: formulation. Polymer Engineering and Science 31,
116–124 (1991)
Chuang, M.C., Ma, Y.C.: Expressing the expected product images in product design of mi-
cro-electronic products. International Journal of Industrial Ergonomics 27(4), 233–245
(2001)
Dawson, D., Askin, R.G.: Optimal new product design using quality function deployment
with empirical value functions. Quality and Reliability Engineering International 15(1),
17–32 (1999)
Dillon, T.S., Chen, W., Chang, E.: An abstract layered model for Web-inclusive distributed
computing leading to enhancing GRIDSpace with Web 2.0. Concurrency and Computa-
tion: Practice and Experience 21(5), 605–634 (2009)
Chen, Y., Chen, L.: A non-linear possibilistic regression approach to model functional rela-
tionships in product planning. International Journal of Advanced Manufacturing Tech-
nology 28(11-12), 1175–1181 (2006)
Clausing, D., Hauser, J.: The house of quality. Harvard Business Review 66(3), 63–73
(1988)
Cohen, L.: Quality Function Deployment. Addison-Wesley, Reading (1995)
Crawford, M., Benedetto, A.: New products Management. McGraw-Hill (2003)
Fung, R.Y.K., Chen, Y.Z., Tang, J.F.: Estimating the functional relationships for quality
function deployment under uncertainties. Fuzzy Sets and Systems 157(1), 98–120
(2006)
Grigoroudis, E., Siskos, Y.: Preference disaggregation for measuring and analyzing cus-
tomer satisfaction: the MUSA method. European Journal of Operational Re-
search 143(1), 148–170 (2002)
Griffin, A., Hauser, J.R.: The voice of the customer. Marketing Science 12(1), 1–27 (1993)
Han, R., Shi, L., Gupta, M.: Three-dimensional simulation of microchip encapsulation
process. Polymer Engineering and Science 40(3), 776–785 (2000)
Han, S.H., Hong, S.W.: A systematic approach for computing user satisfaction with product
design. Ergonomics 46(13), 1441–1461 (2003)
Hsiao, S.W., Tsai, H.C.: Applying a hybrid approach based on fuzzy neural network and
genetic algorithm to product form design. International Journal of Industrial Ergonom-
ics 35(5), 411–428 (2005)
Ip, C.K.W., Kwong, C.K., Bai, H., Tsim, Y.C.: The process modeling of epoxy dispensing
for microchip encapsulation using fuzzy linear regression with fuzzy intervals. Interna-
tional Journal of Advanced Manufacturing Technology 22, 417–423 (2003)
Johnson, R.A., Wichern, D.W.: Applied multivariate statistical analysis. Prentice Hall, En-
glewood Cliffs (1992)
Kang, S.Y., Xie, H., Lee, Y.C.: Physical and fuzzy logic modeling of a flip-chip thermo-
compression bonding process. Journal of Electronic Packaging 115, 63–70 (1993)
Kano, N.: Attractive quality and must-be quality. The Journal of the Japanese Society for
Quality Control 14(2), 39–48 (1984)
Kim, K., Park, T.: Determination of an optimal set of design requirements using house of
quality. Journal of Operations Management 16(5), 569–581 (1998)
Kwong, C.K., Bai, H.: Fuzzy Regression Approach to Process Modelingand Optimization
of Epoxy Dispensing. International Journal of Production Research 43(12), 2359–2375
(2005)
Koza, J.: Genetic Programming II: automatic discovery of reusable programs. MIT Press
(1994)
Kuang, J., Jiang, P.: Product platform design for a product family based on Kansei engi-
neering. Journal of Engineering Design 20(6), 589–607 (2008)
Lai, Y.J., Chang, S.I.: A fuzzy approach for Multiresponse optimization: an off-line quality
engineering problem. Fuzzy Sets and Systems 63, 117–129 (1994)
Lakshminarayanan, S., Fujii, H., Grosman, B., Dassau, E., Lewin, D.R.: New product de-
sign via analysis of historical databases. Computers and Chemical Engineering 24, 671–
676 (2000)
Li, H.X., Tso, S.K., Deng, H.: A concept approach to integrate design and control for the
epoxy dispensing process. International Journal of Advanced Manufacturing Technolo-
gy 17, 677–682 (2001)
Li, H.L., Chou, T., Chou, C.P.: Optimization of resistance spot welding process using Ta-
guchi method and a neural network. Experimental Techniques 31(5), 30–36 (2007)
Lin, Y.C., Lai, H.H., Yeh, C.H.: Consumer-oriented product form design based on fuzzy
logic: a case study of mobile phones. International Journal of Industrial Ergonom-
ics 37(6), 531–543 (2007)
Liu, X.: A fuzzy model for customer satisfaction index in e-commerce. Mathematics and
Computers in Simulation 77(5-6), 512–521 (2007)
Madar, J., Abonyi, J., Szeifert, F.: Genetic programming for the identification of nonlinear
input – output models. Industrial and Engineering Chemistry Research 44, 3178–3186
(2005)
Park, J., Han, S.H.: A fuzzy rule-based approach to modeling affective user satisfaction to-
wards office chair design. International Journal of Industrial Ergonomics 34(1), 31–47
(2004)
Schaiable, B., Lee, Y.C.: Fuzzy logic based regression models for electronics manufactur-
ing applications. Advances in Electronic Packaging 1, 147–155 (1997)
Seber, G.A.F.: Linear regression analysis. Wiley (2003)
Simpson, P.K.: Artificial neural systems. Pergamon Press, New York (1989)
Sullivan, L.P.: Quality function deployment. Quality Progress 19(6), 39–50 (1986)
Takagi, T., Sugeno, M.: Fuzzy identification of systems and its application to modeling and
control. IEEE Transactions on Systems, Man and Cybernetics 15(1), 116–132 (1985)
Tanaka, H., Uejima, S., Asai, K.: Linear regression analysis with fuzzy model. IEEE Trans-
actions on Systems, Man, and Cybernetics 12, 903–907 (1982)
Tanaka, H., Watada, J.: Possibilistic linear systems and their application to the linear re-
gression model. Fuzzy Sets and Systems 272, 275–289 (1988)
Tong, K.W., Kwong, C.K., Yu, K.M.: Intelligent process design system for the transfer
moulding of electronic packages. International Journal of Production Research 42(10),
1911–1931 (2004)
Xie, H., Lee, Y.C.: Process optimization using a fuzzy logic response surface method. IEEE
Transactions on Components Packaging and Manufacturing Technology – Part A 17(2),
202–210 (1994)
Yoder, B., Mason, D.: Evaluating QFD relationships relationships through the use of re-
gression analysis. In: Seventh Symposium on Quality Function Deployment, ASI &
GOAL/QPC, pp. 35–59 (1995)
Chapter 2
Computational Intelligence Technologies for
Product Design
2.1 Introduction
Chapter 1 defines product design as the transformation of a market opportunity in-
to a product available for sale made possible by product development technology.
This transformation is a complex process, as it draws upon and contributes to dif-
ferent domains. Moreover, it is not well formalized. Computational intelligence
algorithms fuse historical design information distributed in space and time into
coherent and understandable design knowledge (Kusiak and Salustri 2007). This
chapter introduces and discusses the recent computational intelligence methods
used for product design, which offer modeling methods and optimization algo-
rithms that are developed to design formalization and automation in terms of new
product development.
In order to understand customers’ needs accurately, huge volumes of customer
data are being collected by surveys or webs, and these are growing at unprece-
dented and incredibly fast rate from a number of other sources as well. These huge
amounts of data are usually difficult to analyze, but they are likely to be a useful
source of valuable knowledge and information which assists in the development of
new products which satisfy all customer requirements. Computational intelligence
provides modeling algorithms for translating the legacy data into an explicit or
implicit model, which can let designers understand the relationship between cus-
tomer requirements and the design attributes of products. Previously, quite a num-
ber of studies have attempted to build models to explain the relationship between
the design attributes of products and customer requirements using statistical mul-
tivariate analysis techniques. These approaches, however, have limitations due to
their inability to capture the fuzziness of consumer requirements, which appears in
customers’ survey data. Also, it is questionable whether the nonlinearity between
design attributes can be addressed by the linear statistical multivariate analysis
techniques.
Computational intelligence approaches including fuzzy systems and neural
networks which are excellent for modeling fuzzy and nonlinear characteristics will
be introduced in Section 2. Figure 2.1 shows that the model is first developed by
the computational intelligence algorithm based on the collected customers’ survey
data, which relates customer requirements to design attributes of products. After
26 2 Computational Intelligence Technologies for Product Design
the model has been developed, the optimal setting of design attributes of the prod-
uct have to be determined. Then, an effective optimization algorithm is used to
determine the maximum customer satisfaction by optimizing the design attribute
setting of products. Due to the non-linear nature of the relationship and the varied
customer requirements, this product design problem is frequently a multi-objective
and multi-optimization problem with multi-optima. Computational intelligence al-
gorithms including evolutionary algorithm, evolutionary programming, simulated
annealing and particle swarm optimization which are well-suited to addressing
multi-optimization problems, will be introduced in Section 2.3.
Customer sur- Modelling Customer requirements

vey data algorithm of the new product
Design attributes
Model for the relationship Optimization
between customer requirements algorithm
and design attributes Customer re-
quirements
Optimal design attribute

setting of the new product
Fig. 2.1 Mechanism for identifying design attribute setting of a new product based on cus-
tomer requirements
2.2 Modeling Approaches

As customer requirements are specified by customers who have human feelings,
fuzziness and nonlinearity often exist in customer survey data.
Before undertaking product design, surveys of potential customers are com-
monly carried out using the interview method. In a survey about product planning
for a new car design, the interviewer might ask potential customers whether the
noise level produced by car A is satisfactory. The answers of the potential custom-
ers mostly depend on human fuzzy feelings, meaning that an absolute answer can-
not be provided. There may have been a lot of unclear answers such as the noise
level of car A is ‘very quiet’. Some may answer that it is ‘quite noisy’ While oth-
ers consider it to be ‘not that quiet’ or ‘very noisy’. Therefore, fuzziness always
exists in survey data.
Also, the customer requirement regarding the noise levels is not linearly related
to the noise levels produced by the car. As shown in Figure 2.2, the customer
2.2 Modeling Approaches 27
might be satisfied with the noise level of the car if it is under a certain level. On
the other hand, customers might be totally dissatisfied if the noise is above a cer-
tain level. Therefore, the relationship between the customer requirement and the
design attribute is often not linear.
Customer
requirement
regarding
noise level
Noise levels of cars

Fig. 2.2 Relationship between noise levels of cars and customer requirement regarding
noise level
Due to the characteristics of fuzziness and nonlinearity, classical linear statis-

tical multi-regression may not generate an appropriate model to represent the
customer requirements and design attributes of a new product. Therefore, compu-
tational intelligence methods which are well-suited to handling fuzzy and nonli-
near characteristics need to be used to generate the models. The neural network
and fuzzy linear regression approaches are commonly used for product design
purposes.
The neural networks are able to transform a non-linear mathematical model into
a simplified black-box structure, which consists of input nodes and output nodes.
They also have the advantage of learning and generalization abilities for customer
requirements, which can be nonlinear or discontinuous. They are powerful for
modeling nonlinear and complex customer survey data for product design purpos-
es. On the other hand, fuzzy linear regression may not perform well on modeling
highly nonlinear characteristics for product design purposes. However, fuzzy li-
near regression has the distinct advantage in product design involving a customer
survey which is fuzzy by nature, and has only a few or even incomplete data sets.
Also, models developed by fuzzy linear regression are explicit in that the relation-
ships between product parameters, fuzziness of product parameters as well as im-
portance of product parameters, can be inferred from the models. Therefore, prod-
uct designers generally prefer to use the models in fuzzy linear regression form
rather than the models in neural network form, even if the latter can obtain more
accurate analysis results than those in fuzzy linear regression form.
Two commonly used methods, fuzzy regression and neural networks, are dis-
cussed in Section 2.2.1 and Section 2.2.2, respectively.
2.2.1 Fuzzy Regression

Tanaka et al. (1982) first introduced fuzzy linear regression analysis, where the
two factors, the degree of fitness and the fuzziness of data sets, are addressed. A
fuzzy linear regression model is commonly presented as follows:
y = f LR (x ) = A0 + A1 x1 + A2 x 2 + ... + A j x j + ... + AN x N = A x
~ ~ ~ ~ ~ ~ ~ ~
(2.1)
where x = [1, x1 , x 2 ,..., x N ] is a crisp vector of independent variables, and ~

T
y is
~ ~ ~ ~ ~
[ ]
the estimated fuzzy output. A = A0 , A1 , A2 ,... AN is a vector of fuzzy parameters
~
of the fuzzy linear regression model. A j is presented in the form of symmetric
triangular fuzzy numbers denoted by A j = (α j , c j ) , j = 0,1,2,…,N, where its
~
membership function is shown as below:
⎧ aj −α j
μ A~ j (a j ) = ⎨1 − c
⎪ ,α j − c j ≤ a j ≤ α j + c j , (2.2)
⎪
j otherwise,
⎩ 0
where α j is the central value of the fuzzy number and c j is the spread. There-
fore the fuzzy linear regression model can be rewritten as shown below:
y = (α , c ) + (α , c )x + (α , c )x + ... + (α , c )x
~
0 0 1 1 1 2 2 2 N N N
(2.3)
The estimated output ~

y can be obtained by using the extension principle
(Zimmermann 1985). The derived membership function of the fuzzy number ~
y
is:
⎧ y −α T x
⎪1 − , x ≠ 0,
⎪ cT x
⎪
μ ~y ( y ) = ⎨1, x = 0, y = 0, (2.4)
⎪0, x = 0, y ≠ 0,
⎪
⎪⎩
where x = ( x1 , x2 ,..., xN )
T
, the central value of y is α T x , and the spread
(range) of y is cT x .
Figure 2.3a shows the relationship between the samples and the models devel-
oped by fuzzy regression. It shows a fuzzy regression model which contains all
samples within its range, and is intended to illustrate all possibilities included in
the system. This indicates that it expresses all possibilities which are embodied by
the data samples and which exist in the system under consideration. Figure 2.3b
shows the relationship between the samples and the models developed by linear
regression. It shows that the model can fit only the samples of the data, but is
unable to indicate the robustness of the model to include a large proportion of the
samples of the data. The two mostly commonly used methods for generating fuzzy
regression models, the Tanaka et al. (1982) approach and the Peters approach
(1994), are discussed in Section 2,2.1.1 and Section 2.2.1.2, respectively.
α T x + cT x
α T x − cT x
αTx
Fig. 2.3a The fuzzy linear regression model
αTx
Fig. 2.3b The linear regression model

2.2.1.1 Tanaka’s Fuzzy Regression

Tanaka et al. (1982) formulated the fuzzy regression problem as the following
linear programming problem:
⎛ ⎞
∑ ∑x
N M
min J = ⎜⎜ c j ij
⎟⎟
j =0 ⎝ i =1 ⎠
∑ ∑c
N N
s.t. : α j xij + (1 − h ) j xij ≥ y i

j =0 j =0
∑α x ∑c
N N
j ij − (1 − h ) j X ij ≤ yi
j =0 j =0
c j ≥ 0, α j ∈ R, X i 0 = 1,
0 ≤ h ≤ 1, i = 1,2,..., M , j = 0,1,2,..., N ,
where J is the total fuzziness of the regression model. The h value, which is be-
tween 0 and 1, is referred to as the fitting degree of the fuzzy linear model to the
given data, and is subjectively chosen by decision makers. Constraints 2 and 3
mean that each observation y i has at lest an h degree of belonging to y i as:
μ yi ( y i ) > h (i = 1,2,..., M ) . Therefore, the goal of the linear programming prob-
lem is to determine the fuzzy parameters y i such that the total vagueness J is
minimized subject to μ yi ( y i ) > h (i = 1,2,..., M ) . It can be noted that fuzzy re-
gression intervals derived from Tanaka’s approach are determined by all the col-
lected data and the value h. Figure 2.3a shows a fuzzy regression model which
contains all samples within its range, and intends to illustrate all possibilities in-
cluded in the system. As a result, the model has larger possibilities than the system
should have because of the incorrect or scattered data. In other words, the bounds
of the interval are determined by the ‘worst’ data in the data set. Thus, the interval
is usually too wide to enable an accurate prediction.
2.2.1.2 Peters’ Fuzzy Regression

To further enhance the Tanaka et al. (1982) approach, Peters (1994) proposed a
new model with symmetrical triangular fuzzy number coefficients by extending
Tanaka’s approach. The dependent data y of Peters model is no longer inside or
outside the interval, but belongs to a certain range. Outliers are compensated for
by data which lies within the interval, and the estimated interval is determined by
using all the data. A new variable is introduced to represent the membership de-
gree to which the solution belongs within the set of ‘good’ solutions’, and a new
fuzzy linear programming problem is formulated as follows.
1 M
max λ = ∑ λi
M i =1
(2.5)
M N
s.t. (1 − λ ) p − ∑ ∑ a 0
S
j
xij ≥ − d 0 (2.6)
i =1 j=0
N N
(1 − λ ) p − ∑ a
i i
C
j
xij − ∑a S
j
xij ≤ yi , i = 1, 2, " , M (2.7)
j =0 j =0
N N
(1 − λ ) p + ∑ a
i i
C
j
xij + ∑a S
j
xij ≥ yi , i = 1, 2, " , M (2.8)
j=0 j =0
λi ≤ 1 , i = 1, 2, " , M (2.9)
a j ≥ 0 , j = 0,1, " , N
S
(2.10)
where pi is the width of the ‘tolerance’ interval of datum yi and λ is the arith-
metic mean of all λi . The parameter d 0 represents the desired value of the ob-
jective function and in most cases, d 0 will be given the value zero, where
∑∑ ( c ) x
M N
j ij
=0
i =1 j =0
is the desired value of the total vagueness; thus, a model as crisp as possible will
be obtained. The descriptions for these two parameters, d 0 and pi , are shown in
Figure 2.4. The fuzzy interval for yi is illustrated in Figure 2.4(b). Figure 2.4(a)
describes the fuzzy interval of the fuzzy regression model.
Fig. 2.4 Fuzzy intervals for the objective function and y i

Symmetric triangular fuzzy numbers are commonly used in the previous studies
associated with Tanaka’s and Peters’ fuzzy linear regression approaches to process
modeling. As shown in Figure 2.5(a), when fuzzy regression with symmetric coef-
ficients is applied, the regression line obtained may not be the best fit because of
αT x
Fig. 2.5(a) Fuzzy regression with symmetric triangular fuzzy coefficients
y
αT x
x
Fig. 2.5(b) Fuzzy regression with asymmetric triangular fuzzy coefficients
the existence of a large number of outliers. Many residuals and the highly asym-
metrical distribution of data points on both sides of the regression line would oc-
cur (Tanaka and Watada 1998). To make up for this deficiency, we extended the
Peters’ approach by introducing asymmetric triangular fuzzy coefficients. Figure
2.5(b) shows that the regression model can best fit all the sampling points, and the
outliers can be illustrated more clearly.
If the regression coefficients A j are asymmetric triangular fuzzy numbers, they
( L R
)
can be denoted as a j , a j , a j , where a j is the fuzzy center, a j is the left fuzzy
L
R
spread, and a j is the right fuzzy spread. A Peters’ fuzzy linear regression model
with asymmetric triangular fuzzy coefficients can be formulated as:
M
1
max λ =
M
∑λ i
(2.11)
i =1
(1 − λ ) p − ∑ ∑ ( a )x
M N
+ aj ≥ −d 0
L R
s.t. 0 j ij
(2.12)
i =1 j=0
N N
(1 − λ ) p − ∑ a j xij − ∑ a j xij ≤ yi , i = 1, 2, " , M

L C L
i i
(2.13)
j =0 j =0
N N
(1 − λ ) p + ∑ a j xij + ∑ a j xij ≥ yi , i = 1, 2, " , M

R C R
i i
(2.14)
j=0 j =0
λi ≤ 1 , i = 1, 2, " , M (2.15)
a j , a j ≥ 0 , j = 0,1, " , N
L R
(2.16)
L R
where the regression parameters p0 , pi and pi are also determined in a con-
text-dependent way according to the decision maker’s experience and knowledge.
x = ( x , x ,..., x ) , the central value of ~
y is α T x , and the spread (range) of
T
1 2 N
~
y is α T x .
2.2.2 Neural Networks

The mechanisms of artificial neural networks are inspired by the biological infor-
mation processing mechanism (Baba and Jain 2001). They are developed to
represent a mapping relationship between a set of inputs and a set of outputs.
Neural networks are intended to simulate a biological system, which uses a set of
interconnections or interrelations of simple net nodes that are analogous to biolog-
ical neurons. They are adaptive information processing systems that can automati-
cally learn operational or generalization capabilities, in order to adapt to the
current learning environments. Therefore, neural networks can learn from past
experience and can generalize from observed or learnt cases, where the structure
of the required mapping algorithm may be an unknown, and also the tolerance to
faulty input information is required.
The processing elements within the neural networks are interrelated in a speci-
fied way. The performance and the effectiveness of a trained neural network de-
pend on the neural network weights to which they are also related as the correla-
tions between the processing elements. Figure 2.6 shows the configuration of the
simple feed-forward neural network. It mainly consists of three basic layers, out-
put layer, input layer and hidden layer. Information flows in a forward direction,
where the inputted data is first presented to the input layer. Then it passes on to
the hidden layer and finally it passes on to the output layer. The data flow of the
neural network is distributed and is processed in parallel ways. There are two im-
portant factors which determine the behaviour of a neural network. They are the
optimal configuration of the neural networks and the optimal weights within the
neural networks. Different configurations of neural networks are described in Sec-
tion 2.2.2.1. Approaches for determining the right weights between the input and
output layers are given in Section 2.2.2.2.
Fig. 2.6 Configuration of the simple feed-forward neural network
2.2.2.1 Different Configurations of Neural Networks

Figure 2.7 shows the configuration of a feed-forward three-layer fully-connected
neural network. The output which relates with the inputs of the feed-forward
three-layer fully-connected neural network is formulated by:
⎛ nh ⎛ nin ⎞ ⎞
yh = tf h2 ⎜ ∑ vgh tf g1 ⎜ ∑ wig zi − b1g ⎟ − bh2 ⎟ , h = 1, 2, …, nout (2.17)
⎜ g =1 ⎟
⎝ ⎝ i =1 ⎠ ⎠
where zi is the input variable with i = 1, 2, …, nin ; the number of input nodes is
denoted by nin ; the number of hidden nodes is nh in which the bias node of the
feed-forward three-layer fully-connected neural network is excluded; the weight of
the interrelation between the g-th hidden nodes and the i-th input nodes is denoted
by wig with g = 1, 2, …, nh ,; the weight between the h-th output node and the
g-th hidden node is denoted by vgh ; the biases for the hidden nodes and output
nodes are denoted by b and b , respectively; tf g1 ( ⋅) and tf h2 ( ⋅) denote the
1
g
2
g
transfer functions in the hidden nodes and output nodes respectively. The com-
monly used transfer functions are the logarithmic sigmoid function (logsig), hy-
perbolic tangent sigmoid function (tansig), and linear function (pureline), which
are shown in Figure 2.7:
1
logsig( η ) = ∈ [ 0 1] , η ∈ ℜ , (2.18)
1 + e−η
− 1 ∈ [ −1 1] , η ∈ ℜ ,
2
tansig( η ) = (2.19)
1 + e −2η
pureline( η ) = η , η ∈ ℜ , (2.20)
logsig function
1 tansig function
1
0.9
0.8
0.8
0.6
0.7
0.4
0.6
0.2
0.5
0
0.4
-0.2
0.3 -0.4
0.2 -0.6
0.1 -0.8
0 -1
-5 -4 -3 -2 -1 0 1 2 3 4 5 -5 -4 -3 -2 -1 0 1 2 3 4 5
n n
Fig. 2.7a ‘logsig’ function Fig. 2.7b ‘tagsig’ function
pureline function
1
0.8
0.6
0.4
0.2
-0.2
-0.4
-0.6
-0.8
-1
-5 -4 -3 -2 -1 0 1 2 3 4 5
n
Fig. 2.7c ‘pureline’ function

The total number of parameter ( n para ) of the feed-forward three-layer fully-

connected neural network, which is required to be optimized, is given by,
n para = ( nin + 1) × nh + ( nh + 1) × nout (2.21)
So far, different kinds of neural network configurations have been developed. The
adaptive linear element is a kind of single-layer feed-forward neural network
(Widrow and Lehr 1962). The input-output relationship of the adaptive informa-
tion processing systems is linear. Functional-link networks (Pao 1989) are other
single-layer neural networks that are able to handle linearly non-separable tasks
using the appropriate enhanced input representation. However, a single-layer
neural network finds it difficult to handle the nonlinear separation problem. The
multi-layer feed-forward neural network (Widrow and Lehr 1990) has become the
most popular network topology. It is intended to simulate nonlinear separation
functions from a compact domain to an arbitrary accuracy, which are smooth and
continuous (Widrow and Lehr 1990, Ham and Kostanic 2001). Several different
types of multi-layer feed-forward neural networks have been developed.
z1 tf12 (⋅) y1
w11 v11
w12 tf11 (⋅) vnh n1
z2 tf 22 (⋅) y2
wnin 1 tf n1h (⋅) v1nout
tf n2out (⋅)
z nin vnh nout y nout
wnin nh
2
b
1
b11 -1 bn2out
bn1h
-1
Fig. 2.8 Architecture of a feed-forward three-layer fully-connected neural network (or a

three-layer-feed-forward neural network)
Self-organizing map is one of the most effective networks for the calculations
of numerical aero-dynamic flow. However, extensive training is required by this
network. The purpose of self-organizing mappings is that patterns of high dimen-
sion are transformed into single dimensional or multi-dimensional patterns. The
self-organizing mapping can determine clusters with respect to data and structure.
It performs an ordered or topology preserving mapping, in order to reveal existing
similarities in the data. The topology preserved with the self-organizing mapping
is not required to be specified or pre-defined as a physical arrangement. It can cor-

respond to a set of statistical features in terms of the data.
In the self-organizing mapping, nodes are configured in a two-dimensional grid
(Cove 1993). It is, however, not impossible to use one or more than one dimen-
sion. This grid of nodes can be treated as a Kohonen layer, where all these nodes
within the input layer are all interconnected to be the nodes within the feature
map. All feedbacks are constrained to lateral interconnections with immediate
neighbouring units in the layer, as illustrated in Figure 2.9, where all links have an
associated weight between inputs and layer nodes. The net input into each neuron
is the same as the weighted sum of the inputs. Training of the net is performed by
changing these weights from an initial distribution, which is pre-defined with re-
spect to the presentation of each vector in terms of input patterns. The final weight
matrix is the final overcome of the training process. The network structure of this
figure is discussed in (Dillon and Niebur 1996).
Fig. 2.9 Neural network architecture of self-organizing feature map
Hopfield network (Hopfield 1982) is a single-layer feedback network, which

can be implemented on a large scale. Normally, its inputs are binary. The inputs
and outputs of the Hopfield network take on only the values of +1 and -1. The
weights in the Hopfield network are fixed and reflect the information that is stored
in the network. Feedback is an essential operation for the Hopfield network. It es-
sentially indicates that the input of neurons in the network is fed from the
weighted output from each neuron in the network. This results in the current input
to the neuron being a sum of the weighted external input and the weighted output
from every neuron. The concept of feedback is illustrated in Figure 2.10 for the
Hopfield network consisting of two neurons.
The updating technique of the network is another of its important operations, Its
purpose is to determine an appropriate time for the network to change its outputs
by modifying the inputs. The neurons sample their inputs in the discrete stochastic
network, based on time samples which are generated randomly. The network is
assumed to be asynchronous, if the sampling time of each neuron is independent
with respect to all other neurons.
This modified input now leads to a new set of outputs. These are again fed
back. This process is repeated until the system settles down to one of the stable
states that correspond to the current input. An important feature of the weight ma-
trix is that the weights are symmetric, i.e. the weight for the feedback from the
output of neuron (1) to the input of neuron (2) is the same as the weight for the
feedback from the output of neuron (2) to input of neuron (1). Furthermore,
the weight for the feedback from a neuron to its own input is zero, i.e. w11=w22=0.
Fig. 2.10 Hopfield network
Radial basis function network (Moody and Darken 1989) is the feed-forward
network shown in Figure 2.11.
The hidden units in the radial basis function net, which correspond to a set of
functions, are linked with the input nodes. The Euclidean distance between the in-
put vector x and a central point of the input space, c j , is used as the input of the
(
function g j x − c j ) , where is denoted as the Euclidean distance. Hence, the
maximum value of c j in terms of the function g j can be determined. As each
output node yi is fully connected with the hidden nodes by the corresponding
weights wij, the combination of the radial basis functions produces the output yi,
which is denoted as:
( )
m
yi = ∑ wij g j x − c j
j =1
Fig. 2.11 Illustration of radial basis function net
An activation function, namely Gaussian function, is usually used as the radial

basis function, where the mean value of c j is the centre. Its size is proportional to
the variances which are predefined for all nodes:
(
gj x −cj ) = exp ( x − c j
2
4σ 2 )
with
σ = d / 2N
where N is the number of RBF units. The maximum distance between the corres-
ponding centres is denoted by d. This radial basis function network is a hybrid
network. It integrates the mechanisms of both the unsupervised and supervised
learning schemes. This network provides a fast learning speed but requires a large
memory.
Pruning neural network (Castellano 1997) is one type of neural network that
has a large size initially and then decreases in size with the removal of unneces-
sary weights and/or units, as shown in Figure 2.12. Certain links/nodes will be re-
moved after the training and the other link weights will be tuned to compensate for
the loss of those links/nodes. However, once a link/node has been deleted, it can-
not be reborn in the rest of the training process. The decision regarding which
link/node is to be deleted is based on a sensitivity formula. Unlike the mechanism
for pruning a neural network, a growing neural network (Martinelli et al. 1990)
starts with a small network and gradually increases in size until it can well fit the
training sample data. Its size is increased by adding nodes and adding networks
between nodes.
This section introduces readers to a brief mechanism and basic operations of
various types of neural networks. For more detailed operations, the readers can re-
fer to (Dillon and Niebur 1996).
……..
Input set
……..…...
Hidden set
Output set
Pruning
……..
Input set
Hidden set ……..…...
Output set
Fig. 2.12 Pruning neural network
2.2.2.2 Learning Algorithms for Neural Network Weights
The learning algorithms for neural network weights can be divided into two main
types: gradient based learning methods and stochastic learning methods.
Gradient Based Learning Methods:

As shown in Figure 2.13, this method basically initiates the weight of the neural
networks and determines the errors which indicate the differences between the ac-
tual observations and the output of the neural networks. Then it estimates an ap-
propriate change of weights based on the errors, and the weights are updated by
the changes of weights. The neural network weights stop changing when the error
is small enough. The error correction and gradient methods are used to find a set
of optimal network weights of the neural networks.
Initialize the neural

network weight
Compute the error

regarding the difference
between the observation
and the estimates
Is the error Yes

Stop training
small enough?
No
Adjust the neural net-
work weights
Fig. 2.13 Operations for gradient-based learning algorithms for neural networks
These gradient-based learning methods have been demonstrated to be very suc-

cessful in training or developing multi-layered neural networks The performance
of neural networks is not given by the reinforcement, which depends on how the
neural networks can be accomplished for a task. The errors between actual obser-
vations and estimates are fed as the inputs to the neural network, in order to adjust
the connections between the layers. Then, the performance of the neural networks
continues to improve. Figure 2.14 shows the common mechanism of the gradient-
based learning method for seeking a solution in a multi-optima landscape. The
gradient-based learning method first initializes a solution randomly and then,
based on the gradient of the landscape, the solution moves down from the initial
point step-by-step to an optimal point which has zero gradient. If its initial solu-
tion is at A1, then it reaches the final solution at A2, which is only a local opti-
mum. If its initial solution is at B1, then it reaches the final solution at B2, which is
the global optimum within the searching domain.
There are many approaches for gradient searches, which have different me-
chanisms regarding users’ requirements in terms of computational time and solu-
tion qualities. There are some examples of error correction methods (Widrow and
Lehr 1990, Ham and Kostanic 2001), including perception learning rule algorithm,
the α-LMS algorithm and May’s rule algorithm. These algorithms adjust the
Gradient based learning

from A1 to A2
y
A1 Gradient based learning
from B1 to B2
B1
Local optimum
A2
B2 Global optimum
x
Fig. 2.14 Gradient-based learning for a multi-optima landscape
network parameters, in order to tune the network output error to match the corres-
ponding input pattern. Some error correction methods can only be used to adjust
separable linear functions. The gradient methods (Andes et al. 1990, Widrow and
Lehr 1990, Ham and Kostanic 2001), such as the Madaline rule I, Madaline rule
II, Madaline rule III, and back-propagation techniques, adjust the neural network
weights using the gradient characteristics of the function landscape. They reduce
the learning error over all input patterns. Many back-propagation algorithms,
which are integrated with the mechanism of momentum (Haykin 1999), variable
learning rate (Haykin 1999), and conjugate gradient algorithm (Moller 1993), have
been developed in order to improve the convergence rate of the learning.
Stochastic Learning Methods:

As mentioned above, the main limitation of the gradient-based learning methods is
that the learning function must be differentiable and continuous. The final solu-
tion is likely to be trapped in a suboptimum, when they are used to solve problems
which are multimodal (for example the learning rules are network structure
dependent). Stochastic learning methods such as evolutionary algorithms, simu-
lated annealing and particle swarm optimizations, can be used to overcome this
limitation. These stochastic algorithms are more appropriate for performing op-
timizations, where the searching domain is complex, non-differentiable, large,
poorly-defined, or multimodal. As shown in Figure 2.15, the stochastic algorithm
first initializes a solution randomly. Then the solution jumps to another point
based on the two factors of gradient of the landscape and the random operation.
As random operations are involved, the stochastic algorithm has the chance to es-
cape from a local optimum (point A) until a global optimum is reached (point B).
The details of the stochastic optimization algorithms are given in Section 2.3.
2.3 Stochastic Optimization Approaches 43
Starting point of the

search
y
A Finishing point of the
search
Local optimum
Jump from the local optimum B Global optimum

x
Fig. 2.15 Stochastic methods for performing optimization in a multi-optima landscape
2.3 Stochastic Optimization Approaches

Solving product design problems requires analyzing or processing a huge amount
of customer survey data, which involves complex human behaviors which involve
fuzziness and nonlinearity. Classical optimization approaches which involve ana-
lyzing the gradient of the problems’ landscapes are often inadequate for finding a
reasonable solution. This necessitates extending the set of stochastic algorithms to
cope with the problems. Using the stochastic algorithms, the range of problems
that are solvable with a given limited amount of time and space into well-defined
classes is a very intricate task, but it can help incredibly to save time and money
spent on the algorithms’ design (Kokash 2005). In the following, we discuss the
commonly used stochastic algorithms, simulated annealing, evolutionary algo-
rithms and particle swarm optimization, which aim to solve difficult optimization
problems. Stochastic algorithms are intended to find the optima within the search-
ing domain, where the objective function can be a-larger-the-better or a-smaller-
the-better function. The objective function is used to evaluate the solution quality
that has been generated to address the problems of a product design.
2.3.1 Simulated Annealing

Simulated annealing is motivated by an analogy to annealing in solids (Metropo-
lis, 1953). The algorithm of simulated annealing simulates the behaviour for cool-
ing material from a heat bath, a process which is called “annealing”. When a solid
is heated to melting point, and is then cooled off, the structural properties of the
solid are related to its cooling rate. If the temperature of the liquid decreases at a
slow enough rate, large crystals can be produced. However, some crystals may be
imperfect, when the liquid cools down quickly. Metropolis’s algorithm simulated
a cooling process of a particle in the material. The algorithm simulates the cooling
process whereby the temperature decreases gradually. The process continues, until
the temperature saturates to a steady frozen state.
For simulated annealing, Kirkpatrick et al. (Kirkpatrick, 1983, Aydin and Fo-
garty 2004) use the notion of the Metropolis’s algorithm to solve difficult optimi-
zation problems. The approach uses simulated annealing to search for a feasible
solution, until the global or local optimum is searched. As shown in Figure 2.16,
simulated annealing is a decision-making process or system which is based on
probabilistic operations. The searching parameter, namely temperature, is used to
determine whether or not to accept an uphill movement for minimisation problems
(or downhill movement for maximization problems). Starting with solution, sn, at
the beginning of the n-th iteration, we consider sn’ as the solution moved to in the
n-th iteration, which is used to give the qualified solution sn+1 for the (n+1)-th it-
eration. sn′ is generated based on the movement of sn, which is within the region of
the neighbourhood function. Some random mechanisms are contained in the
neighbourhood function. The new qualified solution, sn+1, is determined by the fol-
lowing operation:
⎧ sn′ Δs is less than 0
⎪
⎪ ⎛ −Δs ⎞
sn +1 = ⎨ sn′ exp ⎜ ⎟ is less than rn
⎪ ⎝ tn ⎠
⎪ s not the above both cases.
⎩ n
where Δs=sn′-sn , rn is generated randomly. It is used to make a stochastic decision
for accepting the new solution as the current solution. The level of temperature at
nth iteration is denoted by tn , which is decreased through the optimization process
based on a pre-defined decreasing function, which is denoted by,Τ =f(tn).
Move accepted with probability
Hill climbing
Hill climbing
Fitness
values
Hill climbing
Final temperature
Number of iterations
Fig. 2.16 Solving the multi-optima problem using simulated annealing

As a result, sn′ , qualifies as the new solution of the next iteration, when sn′ is
either better or poorer than the old one (sn), and at the same time, the stochastic
rule is satisfied in order to promote sn′ to be the new solution, sn+1. Satisfying the
specified stochastic rule is one the mechanisms behind simulated annealing. It can
be considered to be the probabilistic decision. It is designed to avoid the optimiza-
tion process, which cannot move out from local minima (or local maxima) (Aydin
and Fogarty 2004). Taking this action under the constraints of a Δs at temperature
tn has the probability of e-Δs/tn. At each step, the temperature is decreased by T.
Based on the Boltzman distribution, the chance of accepting a new solution in
terms of a large decreased solution quality decreases exponentially, until the tem-
perature reaches zero degree. It is intended to move the final solution into a near
optimal location, when zero degree is approaching.
As shown in Figure 2.17, a simulated annealing optimization [Laarhoven and
Aarts 1987, D.C. Kozen 1992] starts with an initial solution to the problem, which
is also the Best solution found so far. The temperature is defined at the initial tem-
perature, ti. This solution becomes the Current solution and the Parent or active
solution. The number of n attempts is set to 0.
n is incremented by one and is tested to evaluate whether it has reached the
maximum number of attempts in terms of the current temperature. The current
temperature is evaluated again, if this is true. If it is equal to the final temperature,
tf, the process is terminated and both the final solution and the Best solution gener-
ated by the process are assumed to be obtained. If the current temperature is above
the final temperature, it is reduced by using a cooling schedule. Then the number
of Monte Carlo attempts, n, is reset to 1.
The Parent solution is modified to generate a New solution, if the number of
Monte Carlo attempts at this temperature is reached, or the temperature is de-
creased. The Monte Carlo step is constituted. If the energy of the New solution is
lower than the one of the Parent, it is evaluated again to validate whether it is the
Best solution, which has been discovered so far. If it is the Best solution, it is kept
individually. It becomes the new Parent solution for the next Monte Carlo step,
whether it is the Best or not. Then, the Parent solution is updated, and the Current
solution is updated.
If the energy of the New solution is higher than the Parent's one by an amount
-ds/kT
dE, the Boltzmann probability is calculated by e , where k is Boltzmann's con-
stant and T is the current temperature. If this probability is greater than a random
number, which is randomly generated between 0.0 and 1.0, this New solution is
accepted as the Parent solution for the next iteration, as well as the Current solu-
tion. Conversely, the New solution is rejected and the Current/Parent solution
stays on the same, if the Boltzmann probability is less than the random number.
Otherwise, they are the same ones for the next iteration.
Initial solution, sn
Current solution, sn
tn with n=0
Parent solution, sn’
n=n+1
Y Y
Is tn= tf n> nmax?
Store as
No No best
Modify parameters solution
Reduce temp
New Y
solution
Return the
Best
solution N
solution
?
Is energy
Y
lower?
Y
Is Ran<exp
(-ds/kT) ?
Fig. 2.17 Operations of the simulated annealing
2.3.2 Evolutionary Algorithm

The evolutionary algorithm is a population-based stochastic searching method
which has been increasingly and widely used for solving a number of real-world
practical problems (Back et al 1997). It employs the evolutionary processes as a
major mechanism or approach for a computational problem solving systems.
The search process of evolutionary algorithms is inspired by the biological evo-

lution process which is applied to the solution domain. Evolutionary algorithms
can be extensively used to tackle a broad variety of problems in an extremely di-
verse array of fields. They have been extensively applied to problems whose solu-
tion space is irregular, too large or highly complex. Such problems are difficult to
solve using conventional optimization procedures which cannot be guaranteed to
search for the global optimum. They are robust in delivering global optimal solu-
tions and helping to overcome the limitations of traditional methods.
As shown in Figure 2.18, an evolutionary algorithm first generates the first
population based on random population operations. Fitness evaluation computes
the fitness of all chromosomes, according to their goodness in the environment
with respect to the fitness function. Fitness evaluation can be as simple as compu-
ting a simple mathematical function or arithmetical operation. It can be as com-
plex as simulating a process of a system or an industrial operation. After this, se-
lection is conducted based on two steps: selection and survival of chromosomes.
Selection of chromosome strings determines who the parents are and the number
of children of each parent. Then the children are reproduced by crossover, which
exchanges genes between the selected parents. Also, mutation further changes the
genes of the children. The children’s fitness is then computed based on the fitness
function. The final step is the survival step, which decides the children who can
Create initial population

of chromosomes
Fitness evaluation for the

population
Select and reproduce

population
Replace the old popula-

tion with new reproduced
Solution is Yes
good? Return the solution
No
Fig. 2. 18 Operations of evolutionary algorithm

survive in next generation. Although evolutionary algorithms are a simplified me-

chanism based on evolutionary processes for biologists, they are sufficient for
generating satisfactory, robust and effective solutions for solving many difficult
optimization problems.
Due to the simplicity of the evolutionary algorithm, the evolutionary algorithm
has steadily attracted attention as a novel search technique for a number of prob-
lems, whether searching for better solutions in function optimization problems or
discovering better solutions to machine learning problems [Goldberg 1989]. It has
also demonstrated its capability in solving many difficult optimization problems,
especially if the problem cannot be differentiable, or the problem is difficult to be
formulated as a mathematical form.
The structure of the classical evolutionary algorithm is as follows:

i=0; %the first generation%
Ф(i)=initialize_pop (i); % The i -th population %
F(i)=fitness_eval (P(i)); % Fitness of the population at the t-th genera-
tion%
Repeat
i = i +1; % next generation %
Ф(i)=selections(Ф(i-1), F(i-1));
Ф(i)=crossover(Ф(i));
Ф(i)=mutation(Ф(i));
F(i)=fitness_eval (P(i));
Until (done)
The above procedures outline a classical evolutionary algorithm. A population

Ф(i) of chromosome strings is first initialized. Then Ф(i) is evolved from the
(0)-th generation to (i+1)-th generation based on the evolutionary operations in-
cluding crossover, mutation, selection and fitness evaluation. For many applica-
tions of evolutionary algorithms, the population size is a constant value over all
generations.
2.3.3 Particle Swarm Optimization

Typically, an intelligent multi-agent system or process consists of a set of simple
agents, which interact locally with each other and within the agents’ environmen-
tal condition. Local interrelationships and interactions between agents usually
cause the emergence of a global position, even if no well-defined movement struc-
ture is set on how each agent behaves individually. Many biological or even
human creatures such as fish schools and bird flocks explicitly display well orga-
nized order as an intelligent swarm. The behavior of the swarms changes shape,
direction, speed and location. However, their movements are always behavior as a
single interrelated swarm (Couzin et al. 2002). As shown in Figure 2.19, five main
properties of behaviors can be observed (Das and Konar 2008): i) Homogeneity:
Birds in a swarm have the same behavioral model. The swarm moves without a
leader, even though temporary leaders seem to appear; ii) Locality: Swarm-mates
influence the motion of each bird, where vision is considered to be the most im-
portant sense for swarm organization; iii) Collision Avoidance: Birds avoid col-
liding with nearby swarm mates; iv) Velocity Matching: Birds attempt to match
the velocity of nearby swarm mates; v) Flock Centering: Birds intend to stay
close to nearby swarm mates.
Five basic principles define the mechanism of the particle swarm intelligence
(Das and Konar 2008). First, particles in swarms are able to handle space and time
computations. Second, particles in swarms are able to follow the best particle with
respect to the movements of the swarms. Third, particles in swarms do not con-
duct their activities along their own excessive, but narrow channels. Fourth, par-
ticles in swarms do not adapt their mode of behavior with respect to a bit of time,
when the environmental conditions changes slightly. Fifth, particles in swarms are
able to change their own moving behavior, when they recognize that the time is
appropriate for change.
Particle swarm optimization namely PSO is a recent and popular stochastic op-
timization method which is inspired by animal movement behaviours such as bird
or fish movement (Kennedy and R. Eberhart 1995). Comparing the effectiveness
with the other commonly used stochastic optimization methods, such as genetic
algorithms and simulated annealing, PSO has better searching performance for
solving real-world industrial problems (Kennedy and Eberhart 2001). Better solu-
tion quality with faster and more robust convergence speeds can usually be ob-
tained by PSO.
Homogeneity
with respect to
particles
Collective
Swarm flock Locality with
global be-
Centering havior with- respect to the
in swarm swarm
Velocity Collision
matching be- avoidance be-
tween par- tween par-
ticles of the ticles of the
swarm swarm
Fig. 2.19 Behaviors of swarms (Das and Konar 2008)

The operations of PSO are shown in Figure 2.20 and as follows:

{
i=0 % iteration number
Initialize the first swarm X(i) % X(i) denotes swarm at iteration t
Evaluate the first swarm by f(X(i)) % f(⋅) denotes the fitness function
while (termination condition is not reached)
{
i = i +1
Update v(t) and x(t) using (2.29) and (2.30) respectively
if v(i)>vmax
{
v(i)= vmax
}
if v(i)<−vmax
{
v(i)= − vmax
}
Replace the new swarm X(i -1) by a new swarm X(i)
Evaluate the i -th swarm by f(X(i))
}
}
In the PSO, we denote the swarm at the i-th iteration as X(i). The number of parti-
cles in the swarm is denoted by γ . The dimension of a particle is denoted by κ .
We denote the p-th particle at the i-th iteration as x p ( i ) ∈ X ( i ) , where
x jp ( i ) ∈ x p ( i ) with p = 1, 2,... , γ and j = 1, 2,… , κ . The PSO first initializes
the particles of the swarm randomly. Then it evaluates each particle of the swarm
based on a pre-defined fitness function. From iteration i to i +1, the swarm evolves
by repeating the operations. It aims to minimize (or maximize) the fitness values
(cost values) of particles through the iterations.
In the i-th generation, the position x jp ( i ) and the velocity v jp ( i ) of the j-th
element at the p-th particle can be computed based on equation (2.22) and (2.23)
respectively [Zhao et al. 2005]:
v jp (t ) = k ⋅ (w ⋅ v jp (t − 1) + ϕ1 ⋅ rand ())⋅ ( pbest j − x jp (t − 1)) + ϕ 2 ⋅ rand () ⋅ (gbest j − x jp (t − 1))
(2.22)
x (t ) = x (t − 1) + v (t )
p
j
p
j
p
j
(2.23)
where
pbest = ⎣⎡ pbest1 pbest2 ,... pbestκ ⎦⎤
gbest = ⎡⎣ gbest1 gbest2 ,... gbestκ ⎤⎦
j = [1.. κ ]
pbest denotes the best position of a particle found so far; gbest denotes the posi-
tion of the best particle among all the particles found so far; the acceleration
t<=1
Create an initial swarm

of particles X(t)
Evaluate fitness of each

particle xp(t)
Update velocity v(t)

based on (2.29)
Update position p(t)

based on (2.30)
Solution is Yes
good? Return the solution
No
t=t+1
Fig. 2.20 Operations of the particle swarm optimization
constants are denoted by φ1 and φ2 ; the inertia weight factor is denoted by w; the
constriction factor is denoted by k . It is determined by the stability analysis us-
ing equation (2.31). It aims to ensure that the PSO can converge maturely to a
global solution (Eberhart and Shi 2000); and rand() generate a random number
within the range between 0 to 1. k is a mathematical function in terms of φ1 and
φ2 , which is defined based on the following equation:
2
k= (2.24)
2 − φ − φ 2 − 4φ
where φ = φ1 + φ2 and φ > 4 .

In the PSO, pbest and gbest are used to prevent the particles from moving in
the same direction, but they are used to make the particles move toward the posi-
tions of pbest and gbest, over the generations. The balance between the global and
local explorations for the searching domain can be specified by selecting an
appropriate inertia weight w. Also, w can be defined with the following dynamical
equation (Kennedy and Eberhart 2001):
wmax − wmin
w = wmax − ×i (2.25)
T
where i denotes the current iteration; T denotes the total number of iteration; as
well as wmax and wmin denote the highest and lowest values of the inertia
weights respectively. In (2.22), the velocity of the particle is restricted within a
maximum value vmax . vmax determines the resolution of the domain regions
which need to be searched within the current and the final positions. Particles are
not likely to move out from local optima, if vmax is too small. Otherwise, parti-
cles are likely to miss the global optimum, if vmax is too high. Based on experi-
ence in using the algorithm, vmax was often pre-defined within 10% to 20% of
the moving domain of the element with respect to the corresponding dimensions.
2.4 Summary of This Chapter

In this chapter, a wide range of computational techniques including evolutionary
computations, neural networks, fuzzy regression and so on are discussed to deter-
mine how they can be implemented for solving product design problems. Funda-
mental concepts and an essential discussion of computational intelligence
techniques are presented in order to offer systematic and effective tools for prod-
uct design. Common issues of product design including optimization and model-
ing are covered.
Literature regarding product design and computational intelligence from the
past and present shows that classical computational intelligence approaches
address a wide range of product design issues. In general, reasonable results for
particular product designs can be obtained by applying classical computational in-
telligence approaches. However, such approaches might not achieve the best result
in terms of a particular product design problem. To arrive at a better or more accu-
rate solution, the integration of appropriate techniques into a classical computa-
tional intelligence approach is usually required. Recent research shows that the
effectiveness of classical computational intelligence approaches can be enhanced
if these are integrated with statistical methods or other computational intelligence
approaches. The resulting systems are called hybrid computational intelligence
methods which, according to the literature, are usually able to obtain higher quali-
ty solutions than those obtained by the classical computational intelligence ap-
proaches regarding particular product design problems. Also, they are the main
focus of this book which aims to introduce different hybrid computational intelli-
gence methods for solving different product design problems.
2.5 Application of Computational Intelligence Techniques to Product Design 53
2.5 A pplicatio n of Co mputatio na l I ntelligence Techniques to Product Desig n
2.5 Application of Computational Intelligence Techniques to

Product Design within This Book
The remaining chapters present applications of hybrid computational intelligence
methods for addressing different product design issues. The topics covered range
from the identification of customer requirements in product design, ranking of im-
portance of customer requirements for product design, determination of design
attributes in new product, integration of marketing into product design, to affec-
tive product design. Approaches for the hybridization of various computational
intelligence techniques and integration of statistical methods into computational
intelligence techniques are discussed. Simulation and experimental results regard-
ing the development of real-world new products are included to illustrate the de-
sign procedure and the effectiveness of the hybrid computational intelligence
techniques and their enhancement. A brief overview of the applications of compu-
tational intelligence techniques to product design presented in Chapters 3 to 12 is
given below.
Chapter 3 discusses a new version of AHP integrated with fuzzy systems to
determine importance weights of customer requirements (Kwong and Bai 2002).
The hybrid AHP intends to overcome the limitation of the traditional AHP that
cannot address the vagueness and uncertainty due to human judgment that is
fuzzy. A case study of bicycle splash guard design is used to evaluate the effec-
tiveness of the hybrid AHP approaches.
Chapter 4 discusses a new hybrid AHP which is an enhanced version of hybrid
AHP discussed in Chapter 3 (Kwong and Bai 2003). It uses the extent analysis
method and the principles of comparison for fuzzy numbers to derive the impor-
tance weights of customer requirements. It intends to improve the imprecise rank-
ing of customer satisfaction inherited from the previous works based on the
conventional AHP and the fuzzy AHP discussed in Chapter 3. A case study of hair
dryer design is presented in order to evaluate the effectiveness of the approach.
Chapter 5 presents a genetic programming method in order to develop a func-
tional model which relates design attributes to customer requirements for new
products (Chan et al. 2009b). It is intended to overcome the two main limitations
of previous modeling approaches, statistical methods and computational intelli-
gence methods: i) statistical methods ignore nonlinearity between design attributes
of new products, which are unavoidable; and ii) computational intelligence me-
thods, such as neural networks or neural fuzzy networks, that can generate only
black-box models which are not preferred by product development terms. The rea-
son is that product development terms usually want to understand the relationships
between design attributes and customer requirements. A case study of digital cam-
era design is used to evaluate the effectiveness of the genetic programming
approach.
Chapter 6 introduces a hybrid fuzzy and genetic programming approach to de-
veloping customer satisfaction models that relates customer requirements to the
design attributes of a new product (Chan et al. 2010a). It is intended to address the
fuzzy nature of customer survey data. In fact, human feeling is usually fuzzy but
has not been considered in previous approaches. A case study of the affective
design of mobile phones is used to evaluate the effectiveness of the hybrid fuzzy
and genetic programming approach.
Chapter 7 discusses a hybrid algorithm which integrates the mechanism of
fuzzy regression and generalized least square regression (Kwong et al. 2010). The
hybrid algorithm intends to address the uncertainties of the development of house
of quality which contains both human fuzziness and randomness inherent in the
survey. Neither of these two uncertainties has been addressed by existing ap-
proaches. A case study of packing machine design is used to evaluate the hybrid
algorithm.
Chapter 8 presents an enhanced neuro-fuzzy approach for developing custom-
er satisfaction models to link customer requirements with the design attributes of a
new product (Kwong et al. 2009a). The approach intends to overcome the limita-
tion of the existing neural-fuzzy modeling approaches that are implicit in nature
and do not allow the extraction of information about the customer satisfaction
models. This approach allows explicit information about the customer satisfaction
models to be indicated. A case study of a notebook computer design is used to
demonstrate the operations and procedures of the enhanced neuro-fuzzy approach.
Chapter 9 introduces a hybrid simulated annealing process which integrates
the mechanisms of experimental design methods to maximize customer satisfac-
tion of a new product by optimizing the design attributes of the new product (Chan
et al. 2009a). A case study involving the optimization of a packing-machine de-
sign is presented to demonstrate the effectiveness of the approach.
Chapter 10 discusses a hybrid evolutionary algorithm which is integrated with
a statistical method, namely orthogonal design, to optimize customer satisfaction
with a new product (Chan et al. 2010c). A case study of the optimization of a car
door design is used to illustrate the effectiveness of the hybrid evolutionary
algorithm.
Chapter 11 presents a hybrid algorithm which integrates the mechanisms of
fuzzy regression and genetic programming to develop models for manufacturing
processes in order to improve the quality of the manufactured product (Chan et al.
2009b). The approach overcomes the existing fuzzy regression approaches which
cannot address the nonlinearity of manufacturing processes. A case study of a
solder paste dispensing process is used to evaluate the effectiveness of the hybrid
algorithm.
Chapter 12 discusses a rule-extraction-based genetic algorithm approach
which aims to extract information in rule format from non-informative data
(Kwong et al. 2009b). It allows engineers to understand the behaviors of the man-
ufacturing systems ,in order to enhance the quality of the manufactured products.
A case study of an epoxy dispensing process is used to demonstrate the operations
and procedures of the rule extraction approach.
Chapter 13 summarizes the research on computational intelligence methods
and product design issues. It also provides several suggestions for future research
in these areas.
References 55
References
Aydin, M.E., Fogarty, T.C.: A simulated annealing algorithm for multi-agent systems: a
job-shop scheduling application. Journal of Intelligent Manufacturing 15, 805–814
(2004)
Andes, D., Widrow, B., Lehr, M., Wan, E.: MRIII: A robust algorithm for training abalogy
neural network. In: Proc. Int. J. Conf. on Neural Networks 1990 (IJCNN1990), Wash-
ington, USA, vol. 1, pp. 533–536 (1990)
Baba, N., Jain, L.C.: Computational intelligence in games. Springer (2001)
Bäck, T., Schwefel, H.P.: An overview of evolutionary algorithms for parameter optimiza-
tion. Evol. Comput. 1, 2–23 (1993)
Bäck, T.: Evolutionary Algorithms in Theory and Practice. Oxford, NY (1996)
Back, T., Hammel, U., Schwefel, H.P.: Evolutionary computation: comments on the history
and current state. IEEE Transactions on Evolutionary Computation 1(1) (April 1997)
Beyer, H.G.: The Theory of Evolution Strategies. Springer, NY (2001)
Castellano, G., Fanelli, A.M.: An iterative pruning algorithm for feedforward neural net-
works. IEEE Trans. Neural Networks 8(3), 519–531 (1997)
Chan, K.Y., Kwong, C.K., Tsim, Y.C.: Improved orthogonal array based simulated anneal-
ing with interaction analysis between variables for design optimization. Expert Systems
with Applications 36, 7379–7389 (2009a)
Chan, K.Y., Kwong, C.K., Tsim, Y.C.: A fuzzy nonlinear regression based on genetic pro-
gramming to modeling manufacturing processes. International Journal of Production
Research 48(7), 1967–1982 (2009b)
Chan, K.Y., Kwong, C.K., Dillon, T.S., Fung, K.Y.: An intelligent fuzzy regression ap-
proach for affective product design that captures nonlinearity and fuzziness. Journal of
Engineering Design 22(3), 523–542 (2010a)
Chan, K.Y., Kwong, C.K., Wong, T.C.: Modelling customer satisfaction for product devel-
opment using genetic programming. Journal of Engineering Design 22(1), 601–613
(2010b)
Chan, K.Y., Kwong, C.K., Tsim, Y.C., Aydin, M.E., Fogarty, T.C.: A new orthogonal array
based crossover, with analysis of gene interactions, for evolutionary algorithms and its
application to car door design. Expert Systems with Applications 37(5), 3853–3862
(2010c)
Cove, P.: IEEE 2nd International Workshop on Emerging Technologies and Factory Auto-
mation: Design and Operations of Intelligent Factories (1993)
Couzin, I.D., Krause, J., James, R., Ruxton, G.D., Franks, N.R.: Collective memory and
spatial sorting in animal groups. Journal of Theoretical Biology 218, 1–11 (2002)
Das, S., Konar, A.: Swarm intelligence in production management and engineering, Hand-
book of Computational Intelligence in Manufacturing and Production Management, pp.
345–365 (2008)
Daubechies, I.: The wavelet transform, time-frequency localization and signal analysis.
IEEE Trans. Information Theory 36(5), 961–1005 (1990)
Dillon, T.S., Niebur, D.: Neural networks applications in power systems. International Se-
ries in Intelligent Systems and their Applications. CRL Publishing (1996)
Eberhart, R.C., Shi, Y.: Comparing inertia weights and constriction factors in particle
swarm optimization. In: Proc. Congress on Evolutionary Computing, vol. 1, pp. 84–88
(2000)
Goldberg, D.E.: Genetic Algorithms in Search, Optimization and Machine Learning. Addi-
son Wesley Longman, Inc., United States of America (1989)
Ham, F.M., Kostanic, I.: Principles of Neurocomputing for Science & Engineering.
McGraw-Hill (2001)
Haykin, S.: Neural Network: A Comprehensive Foundation. Prentice-Hall (1999)
Huang, Y.C., Huang, C.M.: Evolving wavelet networks for power transformer condition
monitoring. IEEE Trans. Power Delivery 17(2), 412–416 (2002)
Holland, J.H.: Adaptation in natural and artificial systems. University of Michigan Press,
Ann Arbor (1975)
Hopfield, J.J.: Neural networks and physical systems with emergent collective computa-
tional abilities. Proc. Nat. Acad. Sci. USA 79, 2554–2558 (1982)
Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proc IEEE Int. Conf. Neural
Networks, vol. 4, pp. 1942–1948 (1995)
Kennedy, J., Eberhart, R.: Swarm Intelligence. Morgan Kaufmann Publishers (2001)
Kokash, N.: An introduction to heuristic algorithms (2005),
http://dit.unitn.it/~kokash/documents/Heuristical
Krause, J., Ruxton, G.D.: Living ingroups. Oxford University Press (2002)
Kirkpatrick, S., Gelatt, C.D., Vecchi, M.P.: Optimization by Simulated Annealing.
Science 220, 671–680 (1983)
Kozen, D.C.: The design and analysis of algorithms. Springer (1992)
Kwong, C.K., Bai, H.: A fuzzy AHP approach to the determination of importance weights
of customer requirements in quality function deployment. Journal of Intelligent Manu-
facturing 13, 367–377 (2002)
Kwong, C.K., Bai, H.: Determining the importance weights for the customer requirements
in QFD using a fuzzy AHP with an extent analysis approach. IIE Transactions 35, 619–
626 (2003)
Kwong, C.K., Wong, T.C., Chan, K.Y.: A methodology of generating customer satisfaction
models for new product development using a neuro-fuzzy approach. Expert Systems
with Applications 36(8), 11262–11270 (2009a)
Kwong, C.K., Chan, K.Y., Tsim, Y.C.: A genetic algorithm based knowledge discovery
system for the design of fluid dispensing processes for electronic packaging. Expert Sys-
tems with Applications 36, 3829–3838 (2009b)
Kwong, C.K., Chen, Y., Chan, K.Y., Luo, X.: A generalized fuzzy least-squares regression
approach to modelling functional relationships in QFD. Journal of Engineering De-
sign 21(5), 601–613 (2010)
Kusiak, A., Salustri, F.A.: Computational intelligence in product design engineering: re-
view and trends. IEEE Transactions on Systems, Man and Cybernetics –Part C: Applica-
tion and Reviews 3(5), 766–788 (2007)
Pao, Y.H.: Adaptive Pattern Recognition and Neural Networks. Addison-Wesley, Reading
(1989)
Peters, G.: Fuzzy linear regression with fuzzy intervals. Fuzz Sets Sys. 63, 45–55 (1994)
Van Laarhoven, P.J.M., Aarts, E.H.L.: Simulated Annealing: Theory and Applications. D.
Reidel Publishing Co. (1987)
Leung, F.H.F., Lam, H.K., Ling, S.H., Tam, P.K.S.: Tuning of the structure and parameters
of neural network using an improved genetic algorithm. IEEE Trans. Neural Net-
works 14(1), 79–88 (2003)
Mallat, S.G.: A theory for multiresolution signal decomposition: the wavelet representation.
IEEE Trans. Pattern Analysis and Machine Intelligence 11(7), 674–693 (1989)
References 57
Martinelli, D., Prina-Ricotti, L., Ragazzini, S., Mascioli, F.M.: A pyramidal delayed per-
ceptron. IEEE Trans. Circuits Syst. 37, 1176–1181 (1990)
Metropolis, N., Rosenbluth, A.W., Rosenbluth, M., Teller, A.H., Teller, E.: Equation of
State Calculations by Fast Computing Machines. Journal of Chemistry and Physics 21,
1087–1092 (1953)
Moller, M.F.: A scaled conjugate gradient algorithm for fast supervised learning. Neural
Networks 6(4), 525–533 (1993)
Moody, J., Darken, C.: Fast learning in networks of locally-tuned processing units. Neural
Comput. 1, 281–294 (1989)
Tanaka, H., Uejima, S., Asai, K.: Linear regression analysis with fuzzy model. IEEE Trans.
Sys. Man Cybern. SMC-12, 903–907 (1982)
Wai, R.J., Chang, J.M.: Intelligent control of induction servo motor drive via wavelet neur-
al network. Electric Power System Research 61(1), 67–76 (2002)
Widrow, B., Lehr, M.A.: 30 years of adaptive neural networks: Perceptron, madaline, and
backpropagation. Proceedings of the IEEE 78(9), 1415–1442 (1990)
Whitley, D., Starkweather, T., Bogart, C.: Genetic algorithms and neural networks: Opti-
mizing connections and connectivity. Parallel Computing 14, 347–361 (1990)
Yao, X.: Evolving artificial networks. Proceedings of the IEEE 87(7), 1423–1447 (1999)
Zimmermann, H.J.: Fuzzy sets theory and its applications. Kluwer, Boston (1985)
Zhao, B., Guo, C.X., Cao, Y.J.: A multiagent-based particle swarm optimization approach
for optimal reactive power dispatch. IEEE Trans. Power Systems 20(2), 1070–1078
(2005)
Chapter 3
Determination of Importance of Customer
Requirements Using the Fuzzy AHP Method
3.1 Introduction
As discussed in Chapter 1, it is important to determine the importance of customer
requirements prior to product design since this could greatly affect the target value
setting of engineering characteristics, which are related to the overall customer re-
quirements of the product.
In the past, a conventional Analytic Hierarchy Process (AHP) was commonly
used to determine importance weights for customer requirements which are crucial
to product design. For example, the pairwise comparisons for each level with re-
spect to the goal of customer requirements can be conducted by using a nine-point
scale (Saaty 1980). Each pairwise comparison represents an estimate of the priori-
ties of the customer requirements being compared. The nine-point scale expresses
preferences between options as: equally, moderately, strongly, very strongly, or
extremely preferred. As shown in Table 3.1, these preferences are translated into
pairwise weights of 1, 3, 5, 7 and 9, respectively, with 2, 4, 6, and 8 as interme-
diate values. The pairwise comparison ratios are in crisp real numbers. However,
customer requirements always contain ambiguity and multiplicity of meaning. The
descriptions of customer requirements are usually vague and expressed linguisti-
cally. Furthermore, it is also recognized that human assessment of qualitative
attributes is always subjective and thus imprecise. Therefore, conventional AHP
cannot adequately capture customer requirements explicitly and accurately deter-
mine the importance weight of customer requirements.
This chapter discusses a new version of AHP to determine importance weights
of customer requirements. A fuzzy AHP which is integrated with fuzzy scales is
discussed. The fuzzy AHP is intended to overcome the limitation of the traditional
AHP that cannot address the vagueness and uncertainty inherent in human judg-
ment. Based on a pairwise comparison technique, the comparison matrices of the
fuzzy AHP are built with a set of triangular fuzzy numbers. This fuzzy AHP is
aimed to improve the imprecise ranking of customer requirements which is
produced by the conventional AHP.
K.Y. Chan et al.: Comput. Intell. Techniques for New Product Design, SCI 403, pp. 59 – 77.
60 3 Determination of Importance of Customer Requirements
Table 3.1 Nine pairwise weights (Saaty 1980)
Intensity of Definition Explanation

importance
1 Equally important Contributions of the two
activities are equal
3 Moderately important Judgments of the two
activities are slightly
different
5 Strongly important Judgments of the two
activities are strongly
different
7 Very strongly important One activity dominate
the other one
9 Extremely important One activity over anoth-
er is of the highest poss-
ible order of affirmation
2, 4, 6, 8 Intermediate values be- When compromise is
tween the two judgments needed
3.2 Hierarchical Structure for the Development of Customer

Requirements
The application of AHP includes the following three steps to evaluate complex
multi-attribute alternatives involving subjective criteria:
Step 1: A general decision problem is decomposed in a hierarchical fashion in-
to sub-problems which are easier to comprehend and evaluate;
Step 2: The priority of each element at each level of the decision hierarchy is
determined;
Step 3: The priorities to determine the overall priorities of the decision alterna-
tives are synthesized.
All customer requirements have to be structured into different hierarchical levels,
in order to apply AHP in prioritizing the importance of customer requirements.
Figure 3.1 shows an example of a three-level hierarchy of customer requirements
of a new product, which aims to determine the “overall customer satisfaction”
with respect to all customer requirements of the product. The lowest level of the
three-level hierarchy is called the attribute level, which consists of 16 customer
requirements, CS1 to CS16. The middle level is called the category level which
consists of 7 categories: performance, manufacture accuracy, appearance, time,
cost, standards and safety. Under the appropriate categories, all relevant customer
requirements are listed so as to form the lowest level of the hierarchy. The upper
level is the goal which is the overall customer satisfaction, which is determined by
the 7 categories.
3.3 Fuzzy Representation of Pairwise Comparison 61
Fig. 3.1 An example of a 3-level hierarchy for customer requirements (Kwong and Bai 2002)
3.3 Fuzzy Representation of Pairwise Comparison

The hierarchy of customer requirements has to be established before performing
the pairwise comparison of AHP. After constructing a hierarchy in order to deter-
mine the overall customer requirements, the next step is to ask decision makers to
compare the elements at a given level on a pairwise basis to estimate their relative
importance in relation to the element at the next immediate level. With the con-
ventional AHP, the pairwise comparison is made using a set of ratio scales, which
consists of a set of real numbers. The nine-point scales are frequently used to ob-
tain the participants’ judgments or preferences from options such as ‘equally’,
‘moderately’, ‘strongly’, ‘very strongly’, or ‘extremely preferred’. Even though
the discrete scale of 1 to 9 has the advantage of simplicity and ease of use to indi-
cate the participants’ judgments or preferences, it does not take into account the
uncertainty or fuzziness associated with the mapping of a person’s perception (or
judgment) to a single real number. In fact, uncertainty or fuzziness is unavoidable
in a person’s perception or judgment.
~ ~
Therefore, nine triangular fuzzy numbers, 1 to 9 , are used to represent subjec-
tive pairwise comparisons of customer requirements in order to capture the vague-
ness. They seek to indicate the participants’ judgments or preferences including
extremely negative, very strongly negative, strongly negative, moderately nega-
tive, equally, moderately positive, strongly positive, very strongly positive, or sex-
tremely positive respectively.
The fuzzy number is defined as a special fuzzy set F = {( x, μ F ( x)), x ∈ R} ,
where x takes its values on the real line, R : −∞ < x < +∞ and μ F ( x ) is a conti-
nuous mapping from R to the close interval [0,1]. A triangular fuzzy number,
~
which is denoted as M = (a, b, c) with a ≤ b ≤ c , has the following triangular-
type membership function:
⎧ 0 x<a
⎪x−a
⎪ a≤ x≤b
⎪
μ M ( x) = ⎨ b − a
⎪c − x b≤ x≤c
⎪c−b
⎪ 0 x>c
⎩
Alternatively, by defining the interval of confidence level α, the triangular fuzzy
number can be characterized as:
∀α ∈ [0,1]
~
M α = [aα , cα ] = [(b − a)α + a,−(c − b)α + c]
Four basic operations of positive fuzzy numbers described by the interval of con-
fidence (Kaufmann 1991) are defined as:
Addition:
~ ~
M ⊕ N = [mαL + nαL , mαR + nαR ]
Subtraction:
~ ~
MΘN = [mαL − nαL , mαR − nRα ]
Multiplication:
~ ~
M ⊗ N = [mLα nLα , mαR nαR ]
Division
~ ~
MΟ
/ N = [mαL / nαR , mαR / nαL ]
~ ~
where ∀m L , m R , n L , n R ∈ R + , M α = [mαL , mαR ], Nα = [nαL , nαR ], α ∈ [0,1] .
~ ~
The triangular fuzzy numbers, 1 to 9 , are utilized to improve the convention-
al nine-point scaling scheme used when human uncertainty or fuzziness is
involved. Based on the triangular fuzzy numbers, the imprecision of human qua-
litative assessments can be considered. Figure 3.2 shows the corresponding
~
membership functions of the five triangular fuzzy numbers, 1 , 3 , 5 , 7 and 9
which are the odd scaling numbers.
3.4 Fuzzy AHP 63
μ M (x) Very
Equally Moderately Strongly Extremely
Strongly
~ ~ ~ ~ ~
1 3 5 7 9
1
0.5
0 1 2 3 4 5 6 7 8 9 11
~ ~~~~
Fig. 3.2 The membership functions of the five triangular fuzzy numbers 1 , 3 ,5 ,7 ,9
3.4 Fuzzy AHP
In fuzzy AHP, triangular fuzzy numbers are utilized to improve the scaling
scheme in the judgment matrices, and interval arithmetic is used to solve the fuzzy
eigenvector ((Cheng and Mon, 1994). For these processes, the mechanisms of Saa-
ty’s AHP (Saaty, 1980) method are used and this is a commonly used method for
determining fuzzy eigenvectors of fuzzy matrices. Based on the Saaty’s AHP
(Saaty, 1980) method, the eigenvector corresponding to the largest eigenvalue of
the pairwise comparisons matrix provides the relative priorities of the factors, and
preserves ordinal preferences among the alternatives. Therefore, the eigenvector
component is larger than that of the other, if one alternative is preferred to another.
A vector of weights obtained from the pairwise comparisons matrix reflects the
relative importance of the various attributes of the new product. The four steps of
the computational procedure of this fuzzy AHP are as follows:
Step1: Comparing the Performance Score
~ ~~~~
Triangular fuzzy numbers ( 1 , 3 ,5 ,7 ,9 ) are used to indicate the relative strength
of each pair of elements in the same hierarchy.
Step2: Constructing the Fuzzy Comparison Matrix
By using triangular fuzzy numbers, via pairwise comparison, the fuzzy judgment
~
matrix A (aij) is constructed as shown below:
⎡ 1 a~12 a~13 " a~1( n −1) a~1n ⎤

⎢ a~ 1 ~
a 23 " a~2( n−1) a 2 n ⎥⎥
~
⎢ 21
~ ⎢ # # # # # # ⎥
A=⎢ ⎥
⎢ # # # " # # ⎥
⎢a~( n −1)1 a~( n −1) 2 a~( n −1)3 " 1 a~( n −1) n ⎥
⎢ ~ ⎥
⎣⎢ a n1 a~ n2 a~n 3 " a~n ( n −1) 1 ⎦⎥
where
⎧⎪1, i= j
a~ij = ⎨~ ~ ~ ~ ~ ~ −1 ~ −1 ~ −1 ~ −1 ~ −1
⎪⎩1 , 3, 5, 7 , 9 or 1 , 3 , 5 , 7 , 9 , i ≠ j
Step3: Solving Fuzzy Eigenvalues

~
A fuzzy eigenvalue, λ , is a fuzzy number solution to
~ ~
A~x =λ ~
x (3.1)
~
where A is a n × n fuzzy matrix containing fuzzy numbers a~ij and ~ x is a non-
~
zero n × 1 fuzzy vector containing fuzzy numbers x i .
To perform fuzzy multiplications and additions using interval arithmetic and α-
cut, equation (1) is equivalent to
[aiα1l x1αl , aiα1u x1αu ] ⊕ " ⊕ [ainl

α α α α
x nl , ainu x nu ] = [λxilα , λxiu
α
] (3.2)
~
where A = [ a~ij ] , ~ x t = (~
x1 , " , ~ xαi = [ xαil , xαiu ] and
x n ) , ~a ijα = [aαijl , aαiju ] , ~
λα = [λlα , λuα ] for 0 < α ≤ 1 and all i, j , where i=1,2...n, j=1,2,…,n.
~
The degree of requirement for the judgment matrix A is estimated by the index
of optimism μ. A larger value of the index μ indicates a higher degree of optim-
ism. The index of optimism is a linear convex combination (Lee, 1999) defined as:
aîjα = μ aiju
α
+ (1 − μ )aijlα , ∀μ ∈ [ 0,1] (3.3)
When α is fixed, the following matrix can be obtained after setting the index of
optimism, μ, in order to estimate the degree of requirement.
α
⎡ 1 aˆ12 " aˆ1αn ⎤
⎢ ⎥
~ aˆ α " aˆ α2 n ⎥
A = ⎢ 21
1
(3.4)
⎢ # # % # ⎥
⎢ α α
⎥
⎢⎣ aˆ n1 aˆ n 2 " 1 ⎥⎦
3.5 Case Study: Removable Mountain Bicycle Splashguard 65
The eigenvector is calculated by fixing the μ value and identifying the maximal
eigenvalue.
Step 4: Determining the Total Weights
By synthesizing the priorities over all levels, the overall importance weights of
customer requirements are obtained by varying the α value.
3.5 Case Study: Removable Mountain Bicycle Splashguard
In this section, a case study is used to illustrate the fuzzy AHP approach to determine
the importance weights of customer requirements. The design of a removable moun-
tain bicycle splashguard shown in Figure 3.3, which is detailed in (Ullman 1992), is
used as an example.
Fig. 3.3 Mountain bicycle splashguard
3.5.1 Developing a Hierarchical Structure of Customer

Requirements for Bicycle Splash-Guard Design
A four-level hierarchy of customer requirements for the splashguard design was
constructed as shown in Figure 3.4. The upper level of the hierarchy represents the
overall customer requirements. The second and third levels are the main categories
and the sub-categories respectively. The second level consists of three main cate-
gories: functional performance, C1, special constraints, C2, and appearance, C3,
respectively.
Fig. 3.4 A hierarchy of customer requirements for bicycle splashguard design
The third level consists of seven sub-categories which are under the category de-
scribed in the second level. All 19 customer requirements, CS1 to CS19, are described
at the lowest level, which is the classification of the sub-categories. For example, the
first seven customer requirements, CS1 to CS7, belong to the sub-category, S1. There-
fore, the customer requirements of ‘easy to attach’, ‘easy to detach’, ‘fast to attach’,
‘fast to detach’, ‘can attach when the bike is dirty’ and ‘can detach when the bike is
dirty’ belong to the sub-category ‘attach/detach’.
3.5.2 Constructing Fuzzy Comparison Matrices

~ ~
Triangular fuzzy numbers, 1 to 9 , are used to express the preference in the pairwise
comparisons. By using geometric means of the pairwise comparisons, the fuzzy
comparison matrices (FCM) for each level are obtained.
The four sub-categories in the third level, S1, S2, S3, and S5 are described as
follows. For example, the matrix of S1 represents the first six customer require-
ments, CS1, CS2, CS3, CS4, CS5, and CS6, of which the interactions between
customer requirements are illustrated below.
CS CS CS CS CS CS
1 2 3 4 5 6
⎡1 ~ ~ ~ ~ ~⎤
CS1 3 5 7 5 7
⎢~ −1 ~ ~ ~ ~⎥
CS 2 ⎢3 1 1 3 1 3⎥
⎢~ ~ ~ ~ ~⎥
CS 5 −1 1 1 3 1 3
S1: FCM 1 = 3 ⎢~ −1 ~ −1 ~ −1 ~ −1 ⎥
CS 4 ⎢7 3 3 1 3 1⎥
CS 5 ⎢~ ~ −1 ~ ~⎥
⎢~5 1
~ −1
1
~ −1
3 1
~ −1
3⎥
CS 6 ⎢⎣ 7 −1 3 3 1 3 1 ⎥⎦
CS7 CS8
⎡1 ~⎤
CS 1
S2: FCM 2 = 7 ⎢~ −1 ⎥
CS8 ⎢⎣ 1 1 ⎥⎦
CS9 CS10 CS11 CS12 CS13 CS14

~ ~ ~ −1 ~ ~ −1 ⎤
CS 9 ⎡ 1 1 5 3 1 3
⎢~ ~ ~ −1 ~ −1 ⎥
CS10 ⎢ 1 −1 1 3 5 1 3 ⎥
~ ~ −1 ~ ~ −1 ~ −1 ⎥
CS ⎢ 5 −1 3 1 9 −1 3 7
S3: FCM 3 = 11 ⎢ ~ ~ ~ ~ ~⎥
CS12 ⎢ 3 5 9 1 5 1⎥
CS13 ⎢ 1 −1
~ ~ ~ −1 ~ −1 ⎥
⎢~ 1 3 5 1 3 ⎥
~ ~ ~ −1 ~
CS14 ⎢⎣ 3 3 7 1 3 1 ⎥⎦
CS16 CS17
~⎤
CS ⎡ 1 1
S5: FCM 4 = 16 ⎢~ −1 ⎥
CS17 ⎣⎢ 1 1 ⎦⎥
The three categories at the second level, C1, C2, and C3 are described as
follows. For example, the matrix of C1 represents the first three sub-categories, S1,
S2, and S3 of which the interactions between sub-categories are illustrated.
S1 S2 S3
~ −1 ~ −1 ⎤
S1 ⎡ 1 3 1
⎢~ ~ ⎥
C1: FCM 5 = S2 ⎢ 3 1 1 ⎥
~ ~ −1
S3 ⎢ 1 1 1 ⎥
⎣ ⎦
S4 S5
~
S4 ⎡ 1 3 ⎤
C2: FCM 6 = ⎢~ −1 ⎥
S5 ⎢⎣ 3 1 ⎥⎦
S6 S
7
~
S6 ⎡ 1 1 ⎤
C3: FCM 7 = ⎢~ −1 ⎥
S7 ⎢⎣ 1 1 ⎥⎦
At the upper level, the overall customer requirement, G, consists of three catego-
ries, C1, C2 and C3 as follows. Interactions between C1, C2 and C3 are illustrated in
the following matrix.
C1 C2 C3
~ ~⎤
C1 ⎡ 1 5 7
⎢~ ~⎥
G: FCM 8 = C2 ⎢ 5 −1 1 1⎥
~ ~
C3 ⎢⎣ 7 −1 1 −1 1 ⎥⎦
3.5.3 Computing Importance Weights of Customer Requirements

The lower limit and upper limit of the fuzzy numbers with respect to the α can be
defined as follows by applying the equation (3.2):
~ ~ ~ −1 1 1
1α = [1, 3 − 2α ] , 3α = [1 + 2α , 5 − 2α ] , 3α = [ , ],
5 − 2α 1 + 2α
~ ~ 1 1 ~
5α = [3 + 2α , 7 − 2α ] , 5α −1 = [ , ] , 7α = [5 + 2α , 9 − 2α ] ,
7 − 2α 3 + 2α
~ −1 1 1 ~ ~ −1 1 1
7α = [ , ], 9α = [7 + 2α , 11 − 2α ] , 9α = [ , ]. (3.5)
9 − 2α 5 + 2α 11 − 2α 7 + 2α
For example, let α=0.5 and μ=0.5, and substitute the above expression with the
fuzzy comparison matrices, FCM1 to FCM8, all the α-cuts fuzzy comparison
matrices can be obtained as follows:
⎡ 1 [1,2] [4,6] [6,8] [4,6] [6,8]⎤

⎢[1 / 4,1 / 2] 1 [1,2] [2,4] [1,2] [2,4]⎥⎥
⎢
⎢ [1 / 6,1 / 4] [1, 2 ] 1 [2,4] [1,2] [2,4]⎥
S1: FCM1α = ⎢ ⎥
⎢[1 / 8,1 / 6] [1 / 4,1 / 2] [1 / 4,1 / 2] 1 [1 / 4,1 / 2] 1 ⎥
⎢ [4,6] [1 / 2,1] 1 [2,4] 1 [2,4]⎥
⎢ ⎥
⎣⎢[1 / 8,1 / 6] [1 / 4,1 / 2] [1 / 4,1 / 2] 1 [1 / 4,1 / 2] 1 ⎦⎥
⎡ 1 [1,2]⎤
S2: FCM 2α = ⎢ ⎥
⎣[1 / 2,1] 1 ⎦
⎡ 1 [1,2] [4,6] [1 / 4,1 / 2] [1 / 4,1 / 2]⎤

[1,2]
⎢ [1 / 2,1] 1 [2,4] [1 / 6,1 / 4] 1 [1 / 4,1 / 2]⎥⎥
⎢
⎢[1 / 6,1 / 4] [1 / 4,1 / 2] 1 [1 / 10,1 / 8] [1 / 4,1 / 2] [1 / 8,1 / 6]⎥
S3: FCM 3α = ⎢ ⎥
⎢ [2,4] [4,6] [8,10] 1 [4,6] [1,2] ⎥
⎢ [1 / 2,1] 1 [2,4] [1 / 6,1 / 4] 1 [1 / 4,1 / 2]⎥
⎢ ⎥
⎢⎣ [2,4] [2,4] [6,8] [1 / 2,1] [2,4] 1 ⎥⎦
⎡ 1 [1,2]⎤
S5: FCM 4α = ⎢ ⎥
⎣[1 / 2,1] 1 ⎦
⎡ 1 [1 / 4,1 / 2] [1 / 2,1]⎤
C1: FCM 5 α ⎢
= ⎢[2,4] 1 [1,2] ⎥⎥
⎢⎣ [1,2] [1 / 2,1] 1 ⎥⎦
⎡ 1 [2,4]⎤
C2: FCM 6α = ⎢ ⎥
⎣[1 / 4,1 / 2] 1 ⎦
⎡ 1 [1,2]⎤
C3: FCM 7α = ⎢ ⎥
⎣[1 / 2,1] 1 ⎦
⎡ 1 [4,6] [6,8]⎤
G: FCM 8 α ⎢
= ⎢[1 / 6,1 / 4] 1 [1,2] ⎥⎥
⎢⎣[1 / 8,1 / 6] [1 / 2,1] 1 ⎥⎦
The equation (3.3) and the MATLAB package (Harman 1997) are used to calcu-
late eigenvectors for all comparison matrices, from which the importance weights
of individual customer requirements can be obtained. For example,
FCM 50.5 can be obtained as shown below after applying the equation (3.3).
⎡1.000 3.000 1.000 ⎤

FCM 50.5 = ⎢⎢0.375 1.000 0.375⎥⎥
⎢⎣1.000 3.000 1.000 ⎥⎦
Let FCM 50.5 = A . Eigen values of the matrix A can be calculated as follows by
solving the characteristic equation of A, det(A-λI)=0.
λ1 = 3.081, λ2 = −0.081, λ3 = 0.000
As the value of λ1 is the largest, the corresponding eigenvectors of A can be calcu-

lated as follows by substituting λ1 into the equation, AX=λX.
X1=(0.6852,0.2469,0.6852)T
After normalization, the importance weights of the sub-categories of the customer

requirement, S1, S2, and S3, can be determined as shown below.
C1: ⎡⎣WS 1 ,WS 2,WS 3, ⎤⎦ = [0.4237, 0.1527, 0.4237]
Using a similar calculation, the importance weights of C1 to C3, S4 to S7 and CS1

to CS19 can be determined as shown below.
[WCS1 ,WCS 2 ,WCS 3 ,WCS 4 ,WCS 5 ,WCS 6 ]

S1:
= [0.4082, 0.1386, 0.1351, 0.0551, 0.1351, 0.0829]
S2: [WCS1 ,WCS 8 ] = [0.5858,0.4142]

[WCS 9 ,WCS10 ,WCS11 ,WCS12 ,WCS13 ,WCS14 ]
S3:
= [0.1564,0.1058,0.0376,0.3330,0.1058,0.2616]
S5: [WCS16 ,WCS17 ] = [0.5858, 0.4142]

C2: [WS 4 , WS 5 ] = [0.4142, 0.5858]
C3: [W S 6 , W S 7 ] = [0.4142, 0.5858]
G: [WC1, WC 2,WC 3, ] = [0.7387, 0.1533, 0.1080]
Based on the above results, the total importance weights of individual customer
requirements can be calculated by using the following equations as shown in
Table 3.1. The results are shown in Table 3.2.
Table 3.1 The equations used for calculating the total importance weights of individual
overall customer requirements
TWS1 = WC1 • WS1 , TWS 2 = WC1 • WS 2 ,

TWS 4 = WC 2 • WS 4 , TWS 5 = WC 2 • WS 5 ,
TWS 7 = WC 3 • WS 7 , TWCS1 = WC1 • WS1 • WCS 1 ,
TWCS 3 = WC1 • WS 1 • WCS 3 , TWCS 5 = WC1 • WS 1 • WCS 5 ,
TWCS 6 = WC1 • WS1 • WCS 6 , TWCS 7 = WC1 • WS 2 • WCS 7 ,
TWCS 9 = WC1 • WS 3 • WCS 9 , TWCS10 = WC1 • WS 3 • WCS10 ,
TWCS12 = WC1 • WS 3 • WCS12 , TWCS13 = WC1 • WS 3 • WCS13 ,
TWCS 15 = WC 2 • WS 4 , TWCS16 = WC1 • WS 5 • WCS 16 ,
TWCS18 = WC 3 • WS 6 , TWCS19 = WC 3 • WS 7
TWS 3 = WC1 • WS 3 ,
TWS 6 = WC 3 • WS 6 ,
TWCS 2 = WC1 • WS 1 • WCS 2 ,
TWCS 4 = WC1 • WS 1 • WCS 4 ,
TWCS 8 = WC1 • WS 2 • WCS 8 ,
TWCS11 = WC1 • WS 3 • WCS 11 ,
TWCS14 = WC1 • WS 3 • WCS14 ,
TWCS 17 = WC 2 • WS 5 • WCS17
,
In the pairwise comparisons of AHP, triangular fuzzy numbers were introduced

to improve the scaling scheme of the Saaty's method. The central value of a fuzzy
number is the corresponding real crisp number. The spread of the numbers is the
estimation from the real crisp numbers. Equation (3.3) defines how the estimated
number, âij , reacts to the real crisp number by adjusting the index of optimism, μ.
The μ indicates the degree of optimism, which could be determined by the design
team. If the real crisp number is overestimated (μ>0.5), the value of âij is higher
than the central value. If it is underestimated (μ<0,5), the value of âij is lower
than the central value.
Table 3.2 Importance Weights of Customer requirements for Bicycle Splashguard Design
Category Subcategory Attribute

Functional performance Attach/Detach Easy to attach (0.1278)
(0.7387) (0.3130) Easy to detach (0.0575)
Fast to attach (0.0423)
Fast to detach (0.0172)
Can be attached when bike is dirty
(0.0423)
Can be attached when bike is dirty
(0.0260)
Interface with bike Not mar (0.0661)

(0.1128) Not catch water, etc. (0.0467)
Structural integrity Not rattle (0.0489)

(0.3130) Not wobbly (0.0331)
Not bend (0.0118)
Long life (0.1042)
Lightweight (0.0331)
Not release accidentally (0.0819)
Spatial constraints Fit (0.0635) Most bikes (0.0635)
(0.1533)
Not interfere (0.0898) With lights & generator (0.0526)
With brakes (0.0372)
Appearance Shape (0.0447) Streamlined (0.0447)
(0.1080) Colour (0.0633) Popular colour (0.0633)
By setting μ value as 0.05, 0.5, and 0.95 respectively (reflecting the pessimistic,
the moderate and the optimistic situations), three graphs as shown in Figure 3.5,
Figure 3.6 and Figure 3.7 respectively were generated by using the MATLAB
package with the α varying from 0 to 1. From the graphs, mutual comparisons can
be performed for the most uncertain situation (α=0) to the most certain situation
(α=1), from which the relative importance of the customer requirements
(CS1~CS19) can be ascertained. For example, from the three graphs, the impor-
tance weight of customer requirement CS7 is less than the one of the customer re-
quirement CS15 under the most uncertain comparison (α=0) and highly optimistic
situation (μ=0.95). For the pessimistic situation (μ=0.05), the importance weight
of customer requirement CS7 is larger than the one of CS15. For the moderate situ-
ation (μ=0.50), the importance weights of the customer requirements CS7 and
CS15 are very close.
Weights μ=0.05
0.12
0.11
1
0.1
12
0.09 7
0.08
14
0.07
16
19
15
0.06
2
8
0.05
9
3, 5
17
0.04
18
0.03 10, 13
6
0.02
4
11
0.01
0
0.1 0.2 0.3 0.4 0.5 α 0.6 0.7 0.8 0.9 1
Fig. 3.5 The pessimistic, the moderate and the optimistic situations with μ = 0.05
Weights μ=0.5
0.14
0.13
1
0.12
0.11
12
0.1
0.09
0.08 14
7
0.07
15
19
0.06
2
16
0.05 9 8
18
0.04 3, 5
17
0.03 10,13
6
0.02
4
0.01 11
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
α
3.7 Conclusion 75
Weights μ=0.95
0.14
1
0.13
0.12
0.11
12
0.1
0.09
14
0.08
0.07
15
19
0.06 18
7
2
0.05 8
9
3, 5
0.04
16
10,13
17
0.03 6
0.02
4
11
0.01
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
α
These results indicate that for designing the removable mountain bicycle
splashguard, the customer requirements in term of ‘Easy to attach’ and ‘Long life’
are the most important. Product designers need to focus on the design of the shape
and the material, in order to satisfy the customer requirement ‘Easy to attach’. Al-
so, suitable material needs to be considered in order to satisfy the customer re-
quirement of ‘Long life’. Sturdier material should be used. Some other customer
requirements such as ‘Fast to detach’, ‘Not bend’, ‘Light-weight’ and ‘Not wob-
bly’ are not so important. Product designers may not need to pay as much atten-
tion to addressing these when designing or developing the removable mountain
bicycle splashguard.
3.7 Conclusion
This chapter presents a fuzzy AHP approach to determining the importance
weights of customer requirements in product design. In the methodology, triangu-
lar fuzzy numbers were introduced into the conventional AHP, in order to improve
the imprecise ranking of customer requirements. This approach has the following
major advantages over the conventional AHP method:
 Fuzzy numbers are preferable to extend the range of a crisp comparison

matrix of the conventional AHP method, as human judgment in the
comparisons of customer requirements is fuzzy by nature.
 The adoption of fuzzy numbers can allow design team number of QFD to
have freedom of estimation regarding the overall customer requirements
goal and actual situations. Judgment can range from optimistic to pessi-
mistic.
 The fuzzy AHP intends to overcome the limitation of the conventional
AHP that cannot address the vagueness and uncertainty of human judg-
ment. This fuzzy AHP aims to improve the imprecise ranking of custom-
er requirements which is developed based on the conventional AHP, in
order to provide more accurate estimations of important weights.
A case study of the design of a bicycle splashguard is used as an example to illu-
strate the application of the Fuzzy AHP method in determining the importance
weights of customer requirement for QFD. The overall results show that the com-
bination of fuzzy decision-making with AHP could become a useful tool for prod-
uct design.
Chapter 4 will discuss an enhanced version of the fuzzy AHP which uses the
extent analysis method and the principles of comparison of fuzzy numbers to de-
rive the importance weights of customer requirements. It seeks to improve the im-
precise ranking of importance weights of customer requirements inherited from
the conventional AHP and the fuzzy AHP discussed in this chapter.
References
Akao, Y.: Quality Function Deployment: Integrating Customer Requirements into Product
Design. Productivity Press, Cambridge (1990)
Armacost, R.T., Componation, P.J., Mullens, M.A., Swart, W.W.: An AHP Framework for
Prioritizing Custom Requirements in QFD: An Industrialized Housing Application. IIE
Transactions 26(4), 72–79 (1994)
Aswad, A.: Quality Function Deployment: A Systems Approach. In: Proceedings of the
1989 IIE Integrated Systems Conference on Institute of Industrial Engineers, Atlanta,
GA, pp. 27–32 (1989)
Boender, C.G.E., de Grann, J.G., Lootsma, F.A.: Multicriteria decision analysis with fuzzy
pairwise comparison. Fuzzy Sets and Systems 29, 133–143 (1989)
Buckley, J.J.: Fuzzy hierarchical analysis. Fuzzy Sets and Systems 17, 233–247 (1985)
Chan, L.K., Kao, H.P., Ng, A., Wu, M.L.: Rating the Importance of Customer Needs in
Quality Function Deployment by Fuzzy and Entropy Methods. Int. J. Prod. Res. 37(11),
2499–2518 (1999)
Che, A., Lin, Z.H., Chen, K.N.: Capturing Weight of Voice of the Customer Using Artifi-
cial Neural Network in Quality Function Deployment. Journal of Xi’an Jiaotong Univer-
sity 33(5), 75–78 (1999)
Cheng, C.H., Mon, D.L.: Evaluating Weapon System by Analytical Hierarchy Process
Based on Fuzzy Scales. Fuzzy Sets and Systems 63, 1–10 (1994)
Griffin, K., Hauser, J.R.: The Voice of the Customer. Marketing Science 12(1), 1–27
(1993)
References 77
Gustafsson, A., Gustafsson, N.: Exceeding Customer Expectations. In: The Sixth Sympo-
sium on Quality Function Deployment, Novi, Michigan (1994)
Harman, T.L.: Advanced Engineering Mathematics Using MATALAB. PWS Pub. Co.,
Boston (1997)
Hauser, J.R., Clausing, D.: The House of Quality. Harvard Business Review 66, 63–73
(1988)
facturing 13, 367–377 (2002)
Lai, Y.J., Ho, E.S.S.A., Chang, S.I.: Identifying Customer Preferences in Quality Function
Deployment Using Group Decision-Making Techniques. In: Usher, J., Roy, U., Parsaei,
H. (eds.) Integrated Product and Process Development, pp. 1–28. Wiley, New York
(1998)
Lee, A.R.: Application of Modified Fuzzy AHP Method to Analyze Bolting Sequence of
Structural Joints. UMI Dissertation Services. A Bell & Howell Company (1999)
Lu, M.H., Madu, C.N., Kuei, C., Winokur, D.: Integrating QFD, AHP and Benchmarking
in Strategic Marketing. Journal of Business & Industrial Marketing 9(1), 41–50 (1994)
Saaty, T.L.: The Analytic Hierarchy Process. McGraw-Hill, New York (1980); Ullman,
D.G.: The Mechanical Design Process. McGraw-Hill Inc. (1992)
Van Laarhoven, P.J.M., Pedrycz, W.: A fuzzy extension of Saaty’s priority theory. Fuzzy
Sets and Systems 11, 229–241 (1983)
Zakarian, A., Kusiak, A.: Forming Teams: An Analytical Approach. IIE Transactions on
Design and Manufacturing 31(1), 85–97 (1999)
Chapter 4
An Enhanced Fuzzy AHP Method with Extent
Analysis for Determining Importance
of Customer Requirements
An Enhanced Fuzzy AHP Method w ith Extent A nalys is
4.1 Introduction
This chapter discusses the implementation of the enhanced fuzzy Analytic Hie-
rarchy Process (AHP), which is the improved version of the fuzzy AHP discussed
in Chapter 3 for the determination of importance of customer requirements. Simi-
lar to the latter, the enhanced fuzzy AHP converts the linguistic assessment of
customer requirements to triangular fuzzy numbers, which are used to build the
pairwise comparison matrix of AHP. Then the enhanced fuzzy AHP uses the ex-
tent analysis method and the principles of comparison of fuzzy numbers to derive
weight vectors. This improves the hitherto imprecise ranking of importance
weights of customer requirements inherited from the previous works which used
the conventional AHP and the fuzzy AHP discussed in Chapter 3. The enhanced
fuzzy AHP with extent analysis refers to the "extent" to which an object satisfies a
goal and where "satisfied extent" is defined by means of triangular fuzzy numbers.
The weight vectors of the fuzzy AHP can be calculated using extent analysis and
the principles of comparison of fuzzy numbers. Compared to eigenvectors which
are used to calculate weight vectors in the conventional AHP, the enhanced fuzzy
AHP is simple and easy to implement for the purpose of prioritizing customer sa-
tisfaction of quality function deployment (QFD). A case study of a hair dryer de-
sign is used to illustrate the effectiveness of the enhanced fuzzy AHP.
4.2 Overall Customer Satisfaction on Hair Dryer Design

The enhanced fuzzy AHP with the extent analysis method is used to obtain the overall
customer satisfaction with a hair dryer design. Three different types of customers are
considered: males aged 18 to 35 years, females aged 18 to 35 years, and females aged
36 to 45 years. Interviews with these three groups of customers reveal a total of 18
customer requirements for hair dryer design (Wright 1998). All 18 customer require-
ments are categorized using an affinity diagram. Figure 4.1 shows a four-level hie-
rarchy of customer requirements of a hair dryer design. In the figure, the goal is the
“overall customer satisfaction.”
K.Y. Chan et al.: Comput. Intell. Techniques for New Product Design, SCI 403, pp. 79 – 93.
80 4 An Enhanced Fuzzy AHP Method with Extent Analysis
Fig. 4.1 A hierarchy of customer requirements of a hair dryer design
4.2.1 Development of the Fuzzy Matrix

The hierarchy of customer requirements is the subject of the pairwise comparison
of AHP. After constructing a hierarchy, decision makers are asked to compare the
elements at a given level on a pairwise basis in order to estimate their relative im-
portance in relation to the element at the immediately preceding level. In the con-
ventional AHP, the pairwise comparison is made using a ratio scale. A 9-point
scale is commonly used to show the participants’ judgment or preference between
options as equally, moderately, strongly, very strongly, or extremely preferred.
4.2 Overall Customer Satisfaction on Hair Dryer Design 81
Even though the discrete scale of 1 to 9 has the advantages of simplicity and ease
of use, it does not take into account the uncertainty associated with the mapping of
one’s perception (or judgment) to a number. As mentioned in Chapter 3, human
assessment of the relative importance of individual customer requirements is al-
ways subjective and imprecise. The linguistic terms that people use to express
their feeling or judgment are vague. Using objectives to define, and precise num-
bers to represent linguistic assessments, is not very reasonable, although it has
been widely adopted.
Whilst first advocated in 1965, the application of the fuzzy set theory
(Zimmermann 1996) has become an important means of dealing with ambiguity in
a system. The widely used triangular fuzzy numbers (Chan 1999) are used in this
chapter to represent a pairwise comparison of customer requirements. A fuzzy
number is a special fuzzy set F = {( x, μ F ( x)), x ∈ R} , where x takes its values on
the real line R1 : −∞ < x < +∞ and μ F ( x) is a continuous mapping from R1 to
the close interval [0,1]. A triangular fuzzy number is denoted as M = (l , m, u ) . Its
membership function μ M ( x) : R → [0,1] is equal to
⎧ x l
⎪ m − l − m − l , x ∈ [l , m],
⎪
⎪ x u
μM ( x) = ⎨ − , x ∈ [m, u ],
⎪ m − u m −u
⎪0, otherwise,
⎪
⎩ (4.1)
where l ≤ m ≤ u , l and u stand for the lower and upper value of the support of M,
respectively, and m is the mid-value of M. When l=m=n, it is a non-fuzzy number
by convention.
4.2.2 Pairwise Comparison of Customer Requirements

In order to take the vagueness of assessment on pairwise comparison of customer
requirements into consideration, triangular numbers M1, M3, M5, M7, and M9 are
used to represent the assessment from “equal to extremely preferred,” and M2, M4,
M6 and M8 are middle values. Figure 4.2 shows the triangular fuzzy numbers
M t = (lt , mt , ut ) where t=1, 2 …9 and where lt and ut are the lower and upper val-
ues of the fuzzy number Mt respectively; mt is the middle value of the fuzzy num-
ber Mt. δ is used to represent a fuzzy degree of judgment where ut − lt
= lt − ut = δ . A greater value of δ implies a higher degree of fuzziness in the
judgment. When δ = 0 , the judgment is a non-fuzzy number. Zhu et al. (1999) re-
ported that δ should be larger than or equal to 1/2. In this case study, the value
δ was set as 1.
μ M (x) Very
Equally Moderately Strongly Extremely
Strongly
M1 M3 M5 M7 M9
M2 M4 M6 M8
1
0.5
0 1 2 3 4 5 6 7 8 9
Fig. 4.2 The membership functions of the triangular numbers
First, participants in the focus group use the triangular numbers (M1to M9) to
express their preferences, given a number of options. For example, someone may
consider that the element i is very important compared with the element j under
certain criteria; he/she may set aij = (4, 5, 6) . Here, the customer considers that
there is a strong difference between i and j and hence the customer chooses mem-
bership function M5 which spans 4, 5 and 6 giving (4 ,5, 6). If the element j is
thought to be less important than element i, the pairwise comparison between j and
i could be presented by using the fuzzy number, aij = (1 / 6,1 / 5,1 / 4 ) .
Suppose that there are three people in the focus group, representing three different
customers of hair dryers. Then for C1 to C2, we obtain three different scores namely:
(1, 2, 3)
(2, 3, 4),
and (1, 1, 2).
Table 4.1 Pairwise comparison for category level
C1 C2 C3
C1 (1,1,1) (1, 2, 3) (1, 1, 2)
(2, 3, 4) (1, 1, 2)
(1, 1, 2) (1, 2, 3)
C2 (1/3, 1/2, 1/1) (1, 1, 1) (1, 1, 2)

(1/4, 1/3, ½) (1, 2, 3)
(1/2, 1/1, 1/1) (1, 1, 2)
C3 (1/2, 1/1,1/1) (1/2, 1/1,1/1) (1, 1, 1)

(1/2, 1/1,1/1) (1/3, 1/2, 1/1)
(1/3, 1/2, 1/1) (1/2, 1/1, 1/1)
When we do this for all C1 through to C3, we obtain the comparison in Table
4.1, which shows the results of their pairwise comparisons for the individual cate-
gory level.
Formula (4.1) is applied and a fuzzy matrix for the category level of the hie-
rarchy is obtained as follows:
C1 C2 C3
C1 ⎡ (1,1,1) (1.33, 2, 3) (1.00,1.33,2.33)⎤
⎢
FCM 1 = C 2 ⎢(0.33,0.50,0.75) (1,1,1) (1.00,1.33,2.33)⎥⎥
C3 ⎢⎣ (0.43,0.75,1.00) (0.43,0.75,1.00) (1,1,1) ⎥⎦
G:
For example, the fuzzy number of the pairwise comparison between C1 and C2 at
the first row of FCM1 can be calculated by referencing Table 4.1 as follows:
(1+2+1)/3 = 1.33,
(2+3+1)/3 = 2,
(3+4+2)/3 = 3
Similarly, fuzzy judgment matrices FCMn, for each level of the hierarchy of cus-
tomer requirements of a hair dryer design can be worked out.
The three categories in the second level, C1, C2, and C3 are described as
follows. For example, the matrix of C1 represents the first three sub-categories,
S1, S2, S3 and S4 of which the interactions between sub-categories are illustrated
below,
C1:
S1 S2 S3 S4
S1 ⎡ (1,1,1) ( 2.33,3.33,4.33) (0.43,0.75,3.03) (0.67,1.67,2.67)⎤
S 2 ⎢⎢ (0.23,0.3,0.43) (1,1,1) (0.2,0.25,0.37) (0.33,0.5,1.0) ⎥⎥
FCM 2 =
S 3 ⎢(0.33,1.33,2.33) (3,4,5) (1,1,1) (1.33,2.33,3.33) ⎥
⎢ ⎥
S 4 ⎣ (0.37,0.6,1.49) (1,2,3) (0.3,0.43,0.75) (1,1,1) ⎦
S5 S6 S7
S5 ⎡ (1,1,1) (0.67,1.67, 2.67) (1.33,2.33,3.33)⎤
⎢
FCM 3 = S 6 ⎢(0.37,0.6,1.49) (1,1,1) (1,1.33,2.33) ⎥⎥
S7 ⎣⎢(0.3,0.43,0.75) (0.43,0.75,1.00) (1, 1, 1) ⎦⎥
C2:
S8 S9 S10
S8 ⎡ (1, 1, 1) (0.37,0.6,1.49) (0.3,0.43,0.75) ⎤
FCM 4 = S9 ⎢⎢(0.67,1.67,2.67) (1,1,1) (0.43,0.75,1.00)⎥⎥
S10 ⎢⎣ (1.33,2.33,3.33) (1.00,1.33,2.33) (1,1,1) ⎥⎦
C3:
The seven sub-categories in the third level, S1, S3, S5, S6, S8, S9, and S10, are
described as follows. For example, the matrix of S1 represents the first six cus-
tomer satisfactions, CR1 and CR2, of which the interactions between customer
satisfaction are illustrated.
CR1 CR2
CR ⎡ (1, 1, 1) (0.33,1.33,2.33)⎤
FCM 5 = 1 ⎢ ⎥
CR2 ⎣(0.43,0.75,3.03) (1, 1, 1) ⎦
S1:
CR4 CR5
CR4 ⎡ (1, 1, 1) (0.37,0.6,1.49)⎤
FCM 6 = ⎢ ⎥
CR5 ⎣(0.67,1.67,2.67) (1, 1, 1) ⎦
S3:
CR7 CR8
CR7 ⎡ (1, 1, 1) (0.33,1.33,2.33)⎤
FCM 7 =
CR8 ⎢⎣(0.43,0.75,3.03) (1, 1, 1)
⎥
⎦
S5:
CR8 CR10
CR9 ⎡ (1, 1, 1) (0.3,0.43,0.75)⎤
FCM 8 =
CR10 ⎢⎣(1.33,2.33,3.33) (1, 1, 1)
⎥
⎦
S6:
CR12 CR13 CR14

CR12 ⎡ (1,1,1) (0.67,1.67, 2.67) (1.33,2.33,3.33)⎤
⎢
FCM 9 = CR13 ⎢(0.37,0.6,1.49) (1, 1, 1) (1,1.33,2.33) ⎥⎥
CR14 ⎢⎣(0.3,0.43,0.75) (0.43,0.75,1.00) (1,1,1) ⎥⎦
S8:
CR15 CR16
CR15 ⎡ (1, 1, 1) (0.33,1.33,2.33)⎤
FCM10 = ⎢ ⎥
CR16 ⎣(0.43,0.75,3.03) (1, 1, 1) ⎦
S9:
CR17 CR18
CR17 ⎡ (1, 1, 1) (1.33,2.33,3.33)⎤
FCM11 =
CR18 ⎢⎣(0.33,0.45,0.75) (1, 1, 1)
⎥
⎦
S10:
4.2.3 Calculation of the Consistency Index and Consistency

Ratio
Saaty’s AHP methodology provides a consistency index to measure any inconsis-
tency within the judgments in each comparison matrix as well as for the entire
hierarchy. The index can be used to indicate whether or not the targets can be
arranged in an appropriate order of ranking and how consistent the pairwise com-
parison matrices are. A defuzzification method of triangular fuzzy numbers was
employed to convert the fuzzy comparison matrices into crisp matrices, which the-
reafter are used for the investigation of the consistency.
The consistency index (C.I.) and the consistency ratio (C.R.) for a comparison
matrix can be computed with the use of the following equations.
C.I . = (λmax − n) / ( n − 1)
C.R. = (C.I . )100%

RI (n)
where, λmax is the largest eigenvalue of the comparison matrix, n is the dimension
of the matrix, and RI(n) is a random index depending on n as shown in Table 4.2.
Table 4.2 Consistency index of random matrices (Golden et al., 1989)
N 3 4 5 6 7 8 9
C.I.(n) 0.58 0.9 1.12 1.24 1.32 1.41 1.45
If the calculated C.R. of a comparison matrix is less than 10%, the consistency
of the pairwise judgment can be thought acceptable. Otherwise, the judgments ex-
pressed by the experts are considered to be inconsistent, and the decision maker
has to repeat the pairwise comparison matrix. A triangular fuzzy number, denoted
as M = (l , m, u ) , can be defuzzified to a crisp number as follows:
M _ crisp = (4m + l + n) / 6
Taking the comparison matrix FCM2 as an example, the corresponding crisp ma-
trix can be obtained as shown below:
S S S S
1 2 3 4
S1 ⎡1.0000 3.3300 1.0767 1.6700⎤
S2 ⎢ 0.3100 1.0000 0.2617 0.5550 ⎥⎥
FCM 2 _ CRISP = ⎢
S3 ⎢ 1.3300 4.0000 1.0000 2.3300 ⎥
⎢ ⎥
S4 ⎣ 0.7167 2.0000 0.4617 1.0000 ⎦
The largest eigenvalue of matrix FCM2, λmax , was calculated as 4.2193. The di-
mension of the matrix, n, is 4 and the random index, RI(n), is 0.9 by referencing
Table 2. Therefore, the consistency index and the consistency ratio of the matrix
can be calculated as follows:
C.I . = (λmax − n) / (n − 1) = (4.2193 − 4) / (4 − 1) = 0.0731
C.R. = (C.I . )100% = (0.0731 ) = 0.0812 ≤ 0.1

RI (n) 0.9
After calculating the consistency ratios of all the other comparison matrices, it was
found that they are all less than 10%. Therefore, the consistency of judgment in all
the comparison matrices is acceptable.
4.2.4 Determination of Weight Vectors for Customer

Satisfactions
To determine weight vectors for individual levels of a hierarchy of customer satis-
factions, the extent analysis method and the principles of comparison of fuzzy
numbers are employed (Chang et al. 1996). The extent analysis method is used to
consider the extent to which an object satisfies a goal, that is, the satisfied extent.
In the method, the "extent" is quantified by using a fuzzy number. On the basis of
the fuzzy values of extent analysis for each object, the fuzzy synthetic degree val-
ue can be obtained as follows:
Let X = {x1 , x2 , ..., xn } be an object set, and U = {u1 , u2 ,..., um } be a goal set. Ac-
cording to the extent analysis method, each object could be subjected to extent
analysis for every goal respectively. Therefore, m extent analysis values for each
object could be obtained as follows:
M g1i , M g2i ,..., M gmi , i = 1, 2,..., n
where all the M gji ( j = 1, 2,..., m) are triangular fuzzy numbers. The value of fuzzy
synthetic degree with respect to the i-th objects is defined as
−1
m ⎡ n m ⎤
Di = ∑ M gji ⊗ ⎢∑∑ M gji ⎥
j ⎣ i j ⎦ (4.2)
Based on the above definition, the fuzzy synthetic degree values of all elements in
k-th level can be calculated by using formula (4.2) based on the fuzzy judgment
matrix of the k-th level.
−1
⎛ n n
n ⎞
D = ∑ a ⊗ ⎜ ∑∑ aijk ⎟ , i = 1, 2,..., n
i
k k
ij
j =1 ⎝ i =1 j =1 ⎠ (4.3)
where, D kj is the fuzzy synthetic degree values of element i in the k-th level and
Ak = (aijk )nn is the fuzzy judgment matrix of the k-th level.
For the hair dryer design, with the use of formula (4.3), the fuzzy synthetic
degree values of all elements for the category level of the hierarchy can be calcu-
lated as shown below:
3 3
∑∑ a
i =1 j =1
ij = (1,1,1) + (1.33, 2, 3) + " + (1,1,1) = (7.52, 9.66,13.41)
3 3 −1
(∑ ∑ aij ) =(0.075, 0.104, 0.133)

i =1 j =1
∑a
j =1
1j = (1,1,1) + (1.33, 2,3) + (1.00,1.33, 2.33) = (3.33, 4.33,6.33)
Hence, the fuzzy synthetic degree values of the element C1, Dc1, can be calcu-
lated as follows:
3 3 3
Dc1 = ∑ a1 j ⊗ (∑∑ aij )−1 = (0.25, 0.45, 0.84)
j =1 i =1 j =1
Following a similar calculation, the fuzzy synthetic degree values of all elements
for the category level of the hierarchy can be obtained as shown below:
G: Dc1=(0.25,0.45,0.84)
Dc2=(0.17,0.29,0.54)
Dc3=(0.14,0.26,0.40)
4.2.5 Comparison of Fuzzy Numbers

The principles of comparison of fuzzy numbers were introduced to derive the
weight vectors of all elements for each level of the hierarchy with the use of fuzzy
synthetic values. Some definitions and a theorem of the principles for comparison
of fuzzy numbers are shown below (Zimmermann 1996, Zhu 1999).
Definition 1: M1 and M2 are two triangular fuzzy numbers. The degree of possibil-
ity of M 1 ≥ M 2 is defined as
V (M 1 ≥ M 2 ) = sup[min( μ M1 ( x), μ M 2 ( y ))]

x≥ y
.
Theorem: If M1 and M2 are triangular fuzzy numbers and denoted by (l1 , m1 , u1 ) and
(l2 , m2 , u2 ) respectively.
The necessary and sufficient condition of V ( M 1 ≥ M 2 ) = 1 is m1 ≥ m2 .
If m1 ≤ m2 , let V ( M 1 ≥ M 2 ) = hgt ( M 1 ∩ M 2 ) . Then

⎧ l2 − u1
⎪ , l2 ≤ u1 ,
V ( M 1 ≥ M 2 ) = μ (d ) = ⎨ (m1 − u1 ) − (m2 − l2 )
⎪0,
⎩ others
where d is the crossover point's abscissa of M1 and M2.

Definition 2: The degree of possibility for a fuzzy number to be greater than k
fuzzy numbers M i ( (i = 1, 2,..., k ) can be defined by
V (M ≥ M 1 , M 2 ,..., M k ) = min V ( M ≥ M i ), i = 1, 2,..., k
Let d ( pik ) = min V ( Sik ≥ S kj ) , where, pik is the i-th element of the k-th level,
j = 1, 2, ..., n; j ≠ i. The number of elements in the k-th level is n. Then the weight
vector of the k-th level is obtained as follows:
Wk' = (d ( p1k ), d ( p2k ),...., d ( pnk ))Τ
After the normalization, the normalized weight vector, Wk, becomes

Wk = ( w( p1k ), w( p2k ),...., w( pnk ))Τ
Regarding the example of hair dryer design, the following comparison results are
derived based on the theorem of the principles of comparison of fuzzy numbers in
order to calculate the weight vectors of the category level of the hierarchy.
V ( Dc1 ≥ Dc2 ) =
1
V ( Dc1 ≥ Dc3 ) =
1
(0.25 − 0.54)
V ( Dc2 ≥ Dc1 ) = = 0.65
(0.29 − 0.54) − (0.45 − 0.25)
V ( Dc2 ≥ Dc3 ) =
1
(0.25 − 0.40)
V ( Dc3 ≥ Dc1 ) = = 0.44
(0.26 − 0.40) − (0.45 − 0.25)
(0.17 − 0.40)
V ( Dc3 ≥ Dc2 ) = = 0.88
(0.26 − 0.40) − (0.29 − 0.17)
Based on the definition of the principles of comparison of fuzzy numbers, the

weight vector WG' of the category level of hierarchy can be calculated by using
the following formula:
d (C1 ) = min V ( Dc1 ≥ Dc2 , Dc3 ) = min{1,1} = 1.00
d (C2 ) = min V ( Dc2 ≥ Dc1 , Dc3 ) = min{0.65,1} = 0.65
d (C3 ) = min V ( Dc3 ≥ Dc1 , Dc2 ) = min{0.44, 0.55} = 0.44
WG ' = ( d (C1 ), d (C2 ), d (C3 )) Τ

=(1.00, 0.65, 0.44)
After the normalization, the normalized weight vectors of the category level are as
follows:
G: (Wc1, Wc2, Wc3)=(0.48, 0.31, 0.21)
Following a similar calculation, the weight vectors, Wci, Wgj and WCRk, can be
calculated where i=1 to 3, j=1 to 10 and k=1 to 18. Hence, the total weights of the
customer requirements can be derived as shown in Table 4.3 by using the follow-
ing equations.
TWS1 = WC1 • WS1 TWS 2 = WC1 • WS 2

, ,
TWS 4 = WC1 • WS 4 TWS 5 = WC 2 • WS 5
, ,
TWS 7 = WC 2 • WS 7 TWS 8 = WC 3 • WS 8
, ,
TWS10 = WC 3 • WS10 TWCR1 = WC1 • WS1 • WCR1
, ,
TWCR 4 = WC1 • WS 3 • WCR 4 TWCR5 = WC1 • WS 3 • WCR5
, ,
TWCR8 = WC 2 • WS 5 • WCR8 TWCR9 = WC 2 • WS 6 • WCR9
, ,
TWCR12 = WC 3 • WS 8 • WCR12 TWCR13 = WC 3 • WS 8 • WCR13
, ,
TWS 3 = WC1 • WS 3
,
TW S 6 = WC 2 • W S 6
,
TWS 9 = WC 3 • WS 9
,
TWCR 2 = WC1 • WS1 • WCR 2
,
TWCR 7 = WC 2 • WS 5 • WCR 7
,
TWCR10 = WC 2 • WS 6 • WCR10
,
TWCR14 = WC 3 • WS 8 • WCR14
,
Table 4.3 Importance weights of customer requirements for hair dryer design
Category Subcategory Attribute

Performance Quality Positive switch positions
(0.48) (0.168) (0.089)
Appropriate aesthetics (0.079)
Efficiency
(0.015) Dries hair quickly (0.015)
Reliability
(0.192) Seldom breaks down (0.078)
Lower risk to user (0.114)
Robustness
(0.105) Withstands rough treatment
(0.105)
Market competitive- Appearance Pleasing colour (0.077)

ness (0.149) Pleasing shape (0.070)
(0.31)
Portability Internal cable storage (0.032)
(0.102) Inbuilt power pack (0.070)
Price Low price (0.060)

(0.059)
Ergonomics design Easy controls Easy one-handed use (0.021)

(0.21) (0.044) Choice of air temperature set-
ting (0.015)
Comfortable to hold
(0.074) Choice of air speed setting
(0.008)
Size and weight
(0.095) Comfortable handgrip (0.038)
Well balanced (0.035)
Lightweight (0.078)
Appropriate size (0.016)
In the design of hair dryers, the product designers and engineering personnel
should pay more attention to two customer requirements, ‘lower risk to user’ and
‘withstands rough treatment’ in the new product design. The most insignificant
customer requirements are ‘Dries hair quickly’, ‘Choice of air speed setting’, and
‘Appropriate size’. Product designers and engineering personnel do not need to
pay too much attention to these customer requirements when designing or
developing the new hair dryers. One of a generated hair dryer design as shown in
Figure 4.3 which embeds the design features as shown in Table 4.4. However,
none of the design features is able to meet the significant customer requirements,
‘lower risk to user’ and ‘withstands rough treatment’. Therefore, The hair dryer
design may not be able to satisfy the customer expectations and the customers are
mostly concerned with the safety aspects and the sturdiness of the hair dryers.
Fig. 4.3 A hair dryer
Table 4.4 Design features of hair dryer design
Design feature 1 Hair dryer with detachable styling nozzle

Design feature 2 Hair dryer with two heat/speed settings and
detachable styling nozzle
Design feature 3 Hair dryer has all the punch you need in a
simple, small, sleek design
Design feature 4 It is easy for a right- or left-handed person to
use
Design feature 5 Heat / speed settings
Design feature 6 Image shown is indicative only; color may
differ slightly from the image shown.
4.3 Conclusion
This chapter extends the AHP technique discussed in Chapter 3 by developing an
enhanced fuzzy AHP with extent analysis to determine the importance weights of
customer requirements for product design. The enhanced fuzzy AHP with extent
analysis is demonstrably an effective method for calculating the importance
weights of customer requirements, as it can capture the vagueness of human
judgment, in order to improve the hitherto imprecise ranking of customer require-
ments inherited from the previous works based on the conventional AHP and the
previous fuzzy AHP discussed in Chapter 3. In addition, weight vectors of the en-
hanced fuzzy AHP can be calculated, through extent analysis and principles of
comparison of fuzzy numbers. It has been demonstrated that the enhanced fuzzy
AHP with extent analysis is a simple way to determine the weight vectors and is
easy to implement, and the tedious calculation of eigenvectors in conventional
AHP is no longer required. An example of a hair dryer design is used in this chap-
ter to illustrate the effectiveness of the approach.
In this chapter, the conversion of triangular fuzzy numbers is mostly used for
the customer linguistic assessments when developing the enhanced fuzzy AHP
with extent analysis, and the other more appropriate membership functions may
need to be used after analysis of the domains. In fact, triangular fuzzy numbers
may not be appropriate for all industrial applications. Further research is essential
to investigate the effectiveness of using various fuzzy numbers for a particular
application.
References
Design. Productivity Press, Cambridge (1990)
Armacost, R.T., Componation, P.J., Mullens, M.A., Swart, W.W.: An AHP Framework for
Prioritizing Custom Requirements in QFD: An Industrialized Housing Application. IIE
Transactions 26(4), 72–79 (1994)
Aswad, A.: Quality Function Deployment: A Systems Approach. In: Proceedings of the
1989 IIE Integrated Systems Conference on Institute of Industrial Engineers, Atlanta,
GA, pp. 27–32 (1989)
Chan, L.K., Kao, H.P., Ng, A., Wu, M.L.: Rating the Importance of Customer Needs in
Quality Function Deployment by Fuzzy and Entropy Methods. Int. J. Prod. Res. 37(11),
2499–2518 (1999)
Chang, D.Y.: Application of the Extent Analysis Method on fuzzy AHP. European Journal
of Operational Research 95, 649–655 (1996)
Che, A., Lin, Z.H., Chen, K.N.: Capturing Weight of Voice of the Customer Using Artifi-
cial Neural Network in Quality Function Deployment. Journal of Xi’an Jiaotong Univer-
sity 33(5), 75–78 (1999)
Fung, R.Y.K., Popplewell, K., Xie, J.: An Intelligent Hybrid System for Customer Re-
quirement Analysis and Product Attribute Targets Determination. International Journal
of Production Research 36, 13–34 (1998)
References 93
Golden, B.L., Harker, P.T., Wasil, E.E.: The Analytic Hierarchy Process: Applications and
Studies. Springer, Berlin (1989)
Griffin, H., Hauser, J.R.: The Voice of the Customer. Marketing Science 12(1), 1–27
(1993)
Gustafsson, A., Gustafsson, N.: Exceeding Customer Expectations. In: The Sixth Sympo-
sium on Quality Function Deployment, Novi, Michigan (1994)
Kaufmann, A.: Introduction to Fuzzy Arithmetic Theory and Applications. Van Nostrand
Reinhold, New York (1991)
Khoo, L.P., Ho, N.C.: Framework of a Fuzzy Quality Function Deployment System. Inter-
national Journal of Production Research 34, 299–311 (1996)
626 (2003)
Ho, E.S.S.A., Lai, Y.J., Chang, S.I.: An Integrated Group Decision-Making Approach to
Quality Function Deployment. IIE Transactions 31, 553–567 (1999)
Lu, M.H., Madu, C.N., Kuei, C., Winokur, D.: Integrating QFD, AHP and Benchmarking
in Strategic Marketing. Journal of Business & Industrial Marketing 9(1), 41–50 (1994)
Saaty, T.L.: The Analytic Hierarchy Process. McGraw-Hill, New York (1980)
Vanegas, L.V., Labib, A.W.: A Fuzzy Quality Function Deployment (FQFD) model for
driving optimum targets. International Journal of Production Research 39(1), 99–120
(2001)
Wright, I.C.: Design Methods in Engineering and Product Design. McGraw-Hill Publishing
Company, UK (1998)
Zhu, K.J., Jing, Y., Chang, D.Y.: A Discussion on Extent Analysis Method and Applica-
tions of fuzzy AHP. European Journal of Operational Research 116, 450–456 (1999)
Zimmermann, H.-J.: Fuzzy Set Theory and Its Applications, 3rd edn. Kluwer Academic
Publishers, Boston (1996)
Chapter 5
Development of Product Design Models Using
Classical Evolutionary Programming
Development of Product Desig n Models
5.1 Introduction
Chapters 3 and 4 discuss two AHP methods for determining the importance
weights for customer requirements of products. When the functional relationship
between each individual customer requirement and design attributes of a product
is available, the product development teams can maximize the overall customer
satisfaction with the new products by optimizing their design attribute settings.
As stated in Chapter 1, previous works have adopted various modeling tech-
niques to develop functional models for relating customer requirements to design
attributes, but those modeling techniques have one of two limitations: (a) they are
not able to generate functional models with nonlinear or high order terms that can
relate design attributes to customer requirements; (b) they can only generate func-
tional models which are black-box by nature in which no explicit information can
be exploited.. This chapter discusses a classical computational intelligence method
namely classical genetic programming (CGP) to generate functional models for re-
lating design attributes to customer requirements. The CGP is intended to com-
pensate for the two limitations (a) and (b). Genetic programming first initiates a
set of individuals in the form of branches of a tree, which can represent structures
of a functional model. Based on the tree representation, higher order and nonlinear
terms can be included in the functional models. Then the coefficients of all terms
in the functional model are determined based on an orthogonal least squares algo-
rithm. Based on the evolutionary operations, including crossover, mutation and
genetic selection, the genetic programming is intended to produce individuals
which can be used to represent a functional model which can relate design
attributes to customer requirements. The functional models thus developed are ex-
plicit, and consist of interaction terms and higher order terms in relating design
attributes to customer requirements. A case study of a digital camera design is
used to illustrate and evaluate the effectiveness of the genetic programming in ge-
nerating the functional models.
96 5 Development of Product Design Models
5.2 Classical Genetic Programming

Functional models which relate design attributes to customer requirements are de-
scribed as follows:
yi = fi(x1,x2,…xn) (5.1)
where yi, i = 1,2, … m, is the degree of satisfaction of the ith dimension of custom-
er requirement; xj, j=1,2,…n, is the level of attainment of the jth design attribute;
and the fi is a function of the relationship, which represent the customer
requirement and design attributes.
The pseudocode of the classical genetic programming (CGP) used in generating
the functional model (5.1) is shown below as well as the flowchart of CGP is
shown in Figure 5.1:
t=0
Initialize Ω(t)=[θ1(t), θ2(t),… θPOP(t)]
// Ω(t) is the population of the t-th generation.
// θi(t) is the i-th individual of Ω(t).
Assign parameters in all θi(t) by orthogonal least square method
Evaluate all θi(t) according to a fitness function
while (Terminational condition not fulfilled) do {
Parent Selection Ω(t+1)
Crossover Ω(t+1)
Mutation Ω(t+1)
Assign parameters in all θi(t+1) by orthogonal least square
method
Evaluate all θi(t+1)
Ω(t)= Ω(t+1)
t=t+1
}
The CGP starts with the creation of a random initial population of individuals
Ω(t) with POP individuals θi(t), while t=0. Each individual θi(t) is in the form of a
tree structure that can be represented in a polynomial function (5.2) for modeling
the functional relationship:
n
∑p ∑∑ p ∑ ...∑ p ∏ x
n n n n n
y i = p 0i + i
j1 x j1 + i
j1 j 2 x j1 x j2 + ... i
j1 ... j n jk
(5.2)
j1 =1 j1 =1 j 2 =1 j1 =1 j n =1 k =1
where yi is the degree of satisfaction of the i-th dimension of customer

requirements, and x jk is the jk -th design attribute.
5.2 Classical Genetic Programming 97
Set t <- 0
Initialize the population

Ω(t)=[θ1(t), θ2(t),…, θPOP(t)]
where the i-th individual θi(t) is in the
polynomial form illustrated as subsec-
tion 5.2.1
Assign coefficient to each individual

based on orthogonal least square
method
Evaluate all θi(t) based on the fitness

function (5.4)
Yes
Termination condition met
No
Selection the individuals from Ω(t) to
Ω(t+1).
Performing crossover on some of the

individuals on Ω(t+1).
Return the indi-

vidual with the Performing mutation on some of the
best fitness individuals on Ω(t+1).
Performing mutation on some of the

Fig. 5.1 The flowchart of the classical genetic programming

The coefficients of each individual θi(t) are determined using the orthogonal
least squares method (Billings et al. 1998, Chen et al. 1989). All individuals are
evaluated according to a defined fitness function, which is used to evaluate the
goodness-of-fit of modeling the functional relationship between customer re-
quirements and design attributes. The parent selection process uses the goodness-
of-fit of each individual to determine the selection of potential individuals for
performing crossover or mutation. Finally, the new individuals with the deter-
mined coefficients are evaluated using the fitness function in order to create a new
population Ω(t+1). The process continues until the pre-defined termination
condition is fulfilled. The major aspects of applying the CGP on modeling the
functional relationships are described below, in the following subsections:
5.2.1 Model Representation

The population member of the CGP represents the function fi as shown in (5.2). In
the CGP, one of the most popular methods for representing structures is by means
of hierarchical trees, which are composed of functions F and terminals T [Koza
1992]. The set of functions F contains the arithmetic operations, +, - and *, of
function (5.2), thus F is represented as F = { +, -, *}. The set of terminals T = {x,
p} contains the design attribute set x={ x1, x2, … xn } of the function (5.2) and the
coefficient set p={p0, p1, p2, … p n } of the function, where nt is the number of
t
terms of the function. A potential solution is depicted as a tree with branches,

which consists of operations (internal nodes of the tree) F from the function set
and arguments (terminal nodes of the tree) from the terminal set T. For example,
Figure 5.2 shows an example of a hierarchical tree that expresses the following
polynomial, which consists of nonlinear terms and high order terms:
(x1*x1) - (x2*x2) + (x1*x2*x4)
which is equivalent to:
x12 – x22 + x1·x2 ·x4
The coefficient set p = (p0, p1, p2 and p3) is ascertained after determining the struc-
ture of the polynomial, where the number of coefficients of the polynomial is 4.
Therefore, the completed function can be represented as follows.
p0 + p1 ·x12 – p2 ·x22 + p3 ·x1·x2·x4.
The coefficients p0, p1, p2 and p3 are determined using an orthogonal least square
algorithm (Billings et al. 1988, Chen et al. 1989), which has been demonstrated to
be effective in determining coefficients in a linear-in-parameters model generated
by the CGP (Madar et al. 2005). Details of the orthogonal least square algorithm
can be found in [Billings et al. 1988, Chen et al. 1989].
- *
* * *
x1 x 1 x2 x2 x1 x2 x4
Fig. 5.2 An example of a hierarchical tree
5.2.2 Fitness Function
The fitness function is based on the mean absolute error (MAE), which reflects the
differences between the calculated degree of customer satisfaction based on the
model and the actual degree of customer satisfaction based on the data sets. The
MAE of the jth individual can be calculated based on (5.3).
ntrain
y(k ) − F j (x(k ))
∑
1
MAE j = 100% × (5.3)
n train k =1 y (k )
where Fj is the model represented by the jth individual, y(k) is the degree of cus-
tomer satisfaction of the kth data set. x(k) = [x1(k), x2(k),… xn(k)] are the levels of
attainment of design attributes of the kth data set, and ntrain is the number of data
sets used for developing the model. MAE in (5.3) is commonly known as an indi-
cator of training errors of a model that reflects how well the model fits the data
sets. However, a model of the functional relationship between customer require-
ment and design attributes may contain a lot of unnecessary and complex terms. A
complex or over-parameterized model with a large number of parametrical terms
reduces the transparency and interpretation of the model. To prevent the CGP
from generating models that are too complex, a fitness function is designed to bal-
ance the trade-off between the reduction of complexity and model accuracy. In
this research, penalty terms are introduced into the fitness function of the CGP
(McKay et al. 1997) and the fitness of the jth individual denoted as:
(1 − MSE )
fitness j =
j
(1 + exp(c (L − c )))
(5.4)
1 j 2
where fitnessj is the fitness value, Lj is the number of arithmetic operations of the
model represented by the jth individual, and c1 and c2 are both penalty terms.
5.2.3 Crossover and Mutation
Like other evolutionary algorithms, the two main evolutionary operators, crossov-
er and mutation, are used in CGP. The crossover operation produces a pair of
offspring that inherit characteristics from both parents by selecting a random node
in each of the hierarchical tree structures of the parents (as shown in Figure 5.3a)
and exchanging the associated sub-expressions of the hierarchical tree structures
(as shown in Figure 5.3b). Because of the dynamic representation used in CGP,
the parents are typically of a different size, shape and content. The process of
mapping the genotype onto the phenotype does not correspond to a one-to-one re-
lationship. Therefore, the resulting offspring can be expressed by more than one
different tree structure and can allow diversification of the population.
Mutation is performed by randomly selecting a node that can be an internal or
terminal node, and by replacing the associated sub-expression with a randomly
generated sub-expression. For example, Figure 5.4 shows that the arithmetic
operation of a minus is selected and is mutated to a sum.
Parent 1 Parent 2
* -
+ - * +
x1 x 2 x3 x4
* * +
x1 x 2 x3 x4 x5 x6
Fig. 5.3(a) Random selection of a sub-expression before crossover

Offspring 1 Offspring 2
* -
+ - * +
x3 x4 x1 x2
+ * *
x5 x6 x3 x4 x5 x6
Fig. 5.3(b) Offspring produced by the crossover
Before mutation After mutation
- -
* - * +
x1 x2 x3 x1 x2 x3
Fig. 5.4 Offspring produced by the mutation
5.2.4 Selection and Convergence

After the operations of crossover and mutation, individuals from the current popu-
lation with relatively better fitness defined in (5.4) are selected to serve as parents
for the next generation. The roulette-wheel approach, which is one of the most
common selection methods used for selecting individuals to perform reproduction
operations in evolutionary algorithms [Goldberg et al. 1989], is used for the selec-
tion of individuals. The fitness of the j th individual is assigned a value fitness j ,
and the selection probability value, prob j , is defined as:
fitness j
prob j = Popsize
(5.5)
∑j=1
fitness j
where Popsize is the population size of the CGP. Equation (5.5) shows that the
individual with a larger fitness value has a higher probability of being selected.
After the selection, the population evolves and improves iteratively until the
number of generations is equal to a pre-defined number of generations. Otherwise,
the CGP goes on to the next evolutionary iteration.
5.3 A Case Study of Digital Camera Design

An example of digital camera design (Kwong et al. 2007) is used to illustrate the
method of using the CGP in modeling the relationship between the customer
requirement (CS) and design attributes (DAs) as defined in (5.2). In this example,
two dimensions of CS are used for illustrative and validation purposes, which are
“Photo quality” (CS1), and “Take distant image” (CS2). The associated design at-
tributes of the two CS are “Max. Resolution Support” (DA1), “Optical Zoom”
(DA2), “Aperture Exposure Control” (DA3), “LCD size” (DA4), “Storage Media
Support” (DA5), and “Weight” (DA6). A lead user survey of 15 competitive digital
cameras was conducted. The brand names of the digital cameras are not disclosed
here. The two customer requirements in terms of the six design attributes are
recognized. Table 5.1 shows the survey results and values of the DAs of corre-
sponding 15 competitive digital cameras.
Table 5.1 The data sets of the digital camera design of CS and DAs
Customer Design attributes (DAs)

require-
ment (CS)
Brands of digital y1 y2 x1 x2 x3 x4 x5 x6
cameras
A 4 3 5 3 3 1.85 2 323
B 1 1 2 1 0 1.6 3 90
C 4 5 4 10 3 2.2 2 337
D 5 5 3.2 10 3 1.5 2 441
E 3 3 3.1 3 0 1.5 1 180
F 2 3 4.23 3 0 1.5 1 165
G 2 3 4 3 0 1.5 2 185
H 5 4 6.3 6 3 1.8 3 590
I 2 3 3.24 3 0 1.6 2 150
J 2 3 5.1 3 0 2.5 1 200
K 3 3 1.95 3 3 1.5 2 210
L 5 3 5 3 3 2.5 1 705
M 5 5 8 7 3 1.8 3 906
N 2 3 1.92 3 2 2 1 300
O 4 4 4 4 3 1.8 1 490
5.3 A Case Study of Digital Camera Design 103
The CGP was implemented using MATLAB. The CGP parameters used with
reference to [Madar et al. 2005] are as follows: population size = 50, maximum
number of evaluated individuals = 5000, generation gap = 0.9, crossover rate =
0.5, mutation rate = 0.5, probability of changing terminal via non-terminal = 0.25,
c1=0.5 and c2=30. Since the CGP is a stochastic method, different results could be
obtained from different runs. To evaluate its overall performance, 30 runs on the
GP were carried out. The mean of the 30 runs was calculated for each modeling.
To compare the effectiveness of the CGP in modeling CS with those of
statistical linear regression (LR) (Madar et al. 2005) and fuzzy regression (FR)
[Fung 2006, Chen et al. 2005], the same data sets were used to develop CS models
based on LR and FR. Table 5.2 shows the models developed for CS1 ‘photo qual-
ity’ and CS2 ‘take distant image’ based on the three methods. It can be seen that
the interaction terms exist in the CS1 model and second order terms exist in the
CS2 model developed based on the GP, but do not exist in the models developed
based on LR and FR.
Table 5.2 Models developed for CS1 ‘photo quality’ and CS2 ‘take distant image’
Methods CS1 models CS2 models

Statistical linear re- y1 = 2.07 + 0.120·x1 + y2 = 2.87 + 0.210·x1 +
gression (LR) 0.113·x2 + 0.315·x3 - 0.639·x4 0.313·x2 + 0.153·x3 - 0.569·x4
- 0.210·x5 + 0.00389·x6 - 0.454·x5 + 0.00042·x6
Fuzzy regression y1 = (2.2503, 1.8253) + (- y2 = (2.0940, 2.8422×10-14) +
(FR) 0.3736, 2.5993) ·x1 + (6.1867×10-1, 0) ·x1 +
(0.4667, 1.8900×10-11) ·x2 + (1.5433, 0) ·x2 + (1.0822×10-
(0.4992, 1.3983×10-11) ·x3 + 1
,0) ·x3 + (-3.6483×10-1,0) ·x4
(-2.0806×10-2, 3.9336×10-11) + (-5.6080×10-1, 2.8765) ·x5
·x4 - (7.3416×10-1, + (-6.6674×10-2,0) ·x6
2.9729×10-11) ·x5 + (1.4758,
1.8957×10-11) ·x6
Classical genetic y1 = 0.290505 + 0.734186·x1 y2 = 0.370303 + 1.135947·x2
programming (CGP) + 0.672701·x2 - 0.672701·x4 - 0.064511·x4 + 0.064511·x5
+ 0.152575·x1·x3 - - 0.064511·x22 - 0.064511·x52
0.152575·x1·x2
Fig. 5.5 Digital camera X

Fig. 5.6 Digital camera Y
The design attributes of Digital cameras X (Figure 5.5) and Y (Figure 5.6) are
shown in Table 5.3, where the terms are as follows: x1 -‘Max. Resolution Support’,
x2 -‘Optical Zoom’, x3 -‘Aperture Exposure Control’, and x4 -‘LCD size’ are shown
in Table 5.3. Based on the polynomial function developed by the CGP for CS1, the
degree of customer requirements can be calculated as illustrated in Table 5.3. It
indicates that the customer requirement in terms of CS1 for digital camera Y is
higher than that for digital camera X. This example demonstrates that the cus-
tomer requirement can be evaluated based on the polynomial function developed
by the CGP.
Table 5.3 The data sets of the digital camera design of CS and DAs
Brands of digital Customer requirement (CS) Design attributes (DAs)

cameras y1 x1 x2 x3 x4
X 1.6054 3 1 2 3
Y 4.1496 4 8 4 3
The N-fold cross validation was used to evaluate the CGP methods as com-
pared with statistical linear regression and fuzzy regression in modeling customer
requirements. The trials of the cross validation were repeated 8 times. For each
trial, 13 of the 15 data sets were used for model training while the remaining 2
data sets were used for validating the trained models. The two measures, training
error and validation error, were used to evaluate the effectiveness of the CGP
approach. The training error reflects how well the developed model can fit the
training data sets. The validation error reflects how well the developed model can
predict a response.
The 8 training errors based on the three methods, LR, FR and CGP, are shown
in Figures 5.7a and 5.7b. The x-axis of the figures indicates the data sets which
were used for validation. For example, the (7,11) means that the 7th and 11th data
sets were used for validation, and the rest of the data sets were used for model
training. It can be seen from the figures that CGP yielded the smallest training er-
rors in modeling both the CS1 ‘photo quality’ and CS2 ‘take distant image’.
5.3 A Case Study of Digital Camera Design 105
Figures 5.8a and 5.8b show the validation errors based on the methods for the CS1
‘photo quality’ and CS2 ‘take distant image’ respectively. It shows that the models
based on CGP yielded the smallest number of validation errors in both the CS.
Training error (CS1: photo quality)
10
9
Training errors (percentage)
8
7
6
5
4
3
2
1
0
(7,11) (4,6) (1,4) (1,11) (4,11) (3,15) (1,7) (1,13)
Data sets
LR FR GP
Fig. 5.7a Training errors of CS1 ‘photo quality’
Training error (CS2: take distant image)
70
Training errors (percentage)
60
50
40
30
20
10
0
(7,11) (4,6) (1,4) (1,11) (4,11) (3,15) (1,7) (1,13)
Data sets
LR FR GP
Fig. 5.7b Training errors of CS2 ‘take distant image’

Validation error (CS1: photo quality)
40
Validation error (percentage)
35
30
25
20
15
10
0
(7,11) (4,6) (1,4) (1,11) (4,11) (3,15) (1,7) (1,13)
Data sets
LR FR CGP
Fig. 5.8a Testing errors of CS1 ‘photo quality’
Validation error (CS2: take distant image)
30
Validation error (percentage)
25
20
15
10
0
(7,11) (4,6) (1,4) (1,11) (4,11) (3,15) (1,7) (1,13)
Data sets
LR FR CGP
Fig. 5.8b Testing errors of CS2 ‘take distant image’
Means of the eight validation errors of the three methods are shown in Table
5.4 and Table 5.5 respectively for CS1 and CS2 from which it can be found that the
CGP yielded the smallest number of validation errors. Also from the two tables, it
can be seen that the means of the fifteen training errors based on the CGP are the
smallest compared with those based on LR and FR.
Table 5.4 Means of training and validation errors for CS1 ‘photo quality’
Training error Validation error

Linear regression (LR) 10.7968 23.8521
Fuzzy regression (FR) 15.4051 27.1331
CGP 9.8506 17.4154
5.4 Conclusion 107
Table 5.5 Means of training and validation errors for CS1 ‘photo quality’
Training error Validation error

Linear regression (LR) 6.4523 19.7529
Fuzzy regression (FR) 7.9230 22.3664
GP 2.5284 15.8086
5.4 Conclusion
This chapter shows how classical genetic programming can be used to develop
models for relating customer requirement to design attributes, with the use of
small size data sets. Classical genetic programming is used to create structures of
models based on tree representation in which nonlinear terms and higher order
terms can be addressed when linking design attributes to customer requirements.
Then the orthogonal least squares algorithm is used to estimate the contribution of
each branch of the tree so as to identify the coefficients of the functional models.
Since nonlinear terms or higher order terms can be introduced on branches of trees
in classical genetic programming, models with nonlinear terms and higher order
terms can be produced. To evaluate the effectiveness of the proposed method, a
case study of a digital camera design for relating customer requirements to design
attributes was carried out. Functional models based on classical genetic program-
ming method were developed. The effectiveness of the functional models was
compared with that of the functional models, which were generated based on sta-
tistical linear regression and fuzzy regression. Results of the comparison show that
the functional models based on classical genetic programming produce fewer
training errors and fewer validation errors than those obtained by linear regression,
and fuzzy regression which ignore the nonlinearity of the relationships.
Since uncertainty due to fuzziness is unavoidable when modeling the func-
tional relationship between customer requirements and design attributes, Chapter 6
shows how genetic programming can be modified by integrating fuzzy regression
which can address fuzziness when modeling the relationships.
References
nique and Kohonen self-organizing map. Advances in Engineering Software 37, 334–
349 (2006)
wards office chair design. International Journal of Industrial Ergonomics 34, 31–47
(2004)
ics 35, 411–428 (2005)
Fung, R.Y.K., Popplewell, K., Xie, J.: An intelligent hybrid system for customer require-
ments analysis and product attribute targets determination. International Journal of Pro-
duction Research 36, 13–34 (1998)
Liu, X., Zeng, X., Xu, Y., Koehl, L.: A fuzzy model for customer satisfaction index in e-
commerce. Mathematics and Computers in Simulation 77, 512–521 (2007)
logic: A case study of mobile phones. International Journal of Industrial Ergonomics 37,
531–543 (2007)
tomer satisfaction: The MUSA method. European Journal of Operational Research 143,
148–170 (2002)
Grigoroudis, E., Litos, C., Moustakis, V.A., Politis, Y., Tsironis, L.: The assessment of us-
er-perceived web quality: Application of a satisfaction benchmarking approach. Euro-
pean Journal of Operational Research 187, 1346–1357 (2008)
You, H., Ryu, T., Oh, K., Yun, M.H., Kim, K.J.: Development of customer satisfaction
models for automotive interior materials. International Journal of Industrial Ergonom-
ics 36, 323–330 (2006)
Han, S.H., Yun, M.H., Kim, K.J., Kwahk, J.: Evaluation of product usability: development
and validation of usability dimensions and design elements based on empirical models.
International Journal of Industrial Ergonomics 26, 477–488 (2000)
quality. Journal of Operations Management 16, 569–581 (1998)
Chen, Y., Tang, J., Fung, R.Y.K., Ren, Z.: Fuzzy regression-based mathematical program-
ming model for quality function deployment. International Journal of Production Re-
search 42(5), 1009–1027 (2004)
nology 28, 1175–1181 (2006)
with empirical value functions. Quality and Reliability Engineering International 15,
17–32 (1999)
Koza, J.: Genetic Programming: On the Programming of Computers by Means of Natural
Evolution. MIT Press, Cambridge (1992)
(1994)
(2005)
Rodriguez-Vazquez, K., Fonseca, C.M., Fleming, P.J.: Identifying the structure of nonli-
near dynamic systems using multiobjective genetic programming. IEEE Transactions on
Systems, Man and Cybernetics – Part A: Systems and Humans 34(4), 531–545 (2004)
676 (2000)
Fung, R.Y.K., Chen, Y.Z., Tang, J.F.: Estimating the functional relationships for quality
function deployment under uncertainties. Fuzzy Sets and Systems 157, 98–120 (2006)
Kwong, C.K., Chen, Y., Bai, H., Chan, D.S.K.: A methodology of determining aggregated
importance of engineering characteristics in QFD. Computers and Industrial Engineer-
ing 53(4), 667–679 (2007)
Xu, D., Yan, H.S.: An intelligent estimation method for product design time. International
Journal of Advanced Manufacturing Technology 30, 601–613 (2006)
References 109
Billings, S., Korenberg, M., Chen, S.: Identification of nonlinear outputaffine systems using
an orthogonal least-squares algorithm. International Journal of Systems Science 19,
1559–1568 (1988)
Chen, S., Billings, S., Luo, W.: Orthogonal least squares methods and their application to
non-linear system identification. International Journal of Control 50, 1873–1896 (1989)
McKay, B., Willis, M.J., Barton, G.W.: Steady-state modeling of chemical processes using
genetic programming. Computers and Chemical Engineering 21(9), 981–996 (1997)
Chen, Y.Z., Fung, R.Y.K., Tang, J.F.: Fuzzy expected value modeling approach for deter-
mining target values of engineering characteristics in QFD. International Journal of Pro-
duction Research 43(17), 3583–3604 (2005)
opment using genetic programming 22(1), 55–68 (2011)
Chapter 6
Development of Product Design Models Using
Fuzzy Regression Based Genetic Programming
Development of Product Desig n Models Using F uzzy Regressio n
6.1 Introduction
To develop a functional model which relates customer requirements to the design

attributes of a new product, a set of customer survey data is usually used. As men-
tioned in Chapter 5, customer survey data is usually fuzzy in nature, as human
feeling is usually fuzzy, and also nonlinearities are unavoidable in the relation-
ships between customer requirements and design attributes. However, the devel-
opment of explicit functional models has not been addressed by previous studies
in modelling the relationships between customer requirements and design
attributes. Also, those previous modelling methods can address either nonlinearity
or fuzziness only. To overcome the deficiencies of the above approaches, this
chapter presents a fuzzy regression based genetic programming method, namely
FR-GP, to generate functional models which represent this nonlinear and fuzzy re-
lationships between customer requirements and design attributes.
In order to do this, FR-GP uses the existing work on fuzzy regression by first
utilizing a genetic programming method to construct branches of a tree
representing structures of a functional model where the nonlinearity of the func-
tional model can be addressed. The fuzzy regression algorithm is then used to de-
termine the fuzzy coefficients of the functional model. Since both linearity and
nonlinearity are represented in branches of the tree based on the FR-GP, fuzzy re-
gression models with both fuzzy linear and fuzzy nonlinear terms can be generat-
ed. FR-GP intends to eliminate the deficiencies of the above approaches which ig-
nore nonlinear terms of relations between customer requirements and design
attributes, as well as the fuzzy nature of data. Moreover, FR-GP produces explicit
functional models, which are usually recommended by engineering users. To eva-
luate the effectiveness of the FR-GP for modeling the relationship between cus-
tomer requirements and design attributes, we use affective design for mobile
phones. Results of the modeling were compared with those based on the existing
fuzzy regression methods (Tanaka et al. 1982, Peters 1994) and statistical regres-
sion (Seber 2003).
112 6 Development of Product Design Models Using Fuzzy Regression
6.2 Fuzzy Regression Based Genetic Programming

The FR-GP approach generates the functional model which relates design
attributes and customer requirements in a fuzzy regression form. The approach
needs to:
a) specify the form of the fuzzy regression model
b) determine the fuzzy coefficients which characterize the functional models
c) define the algorithm
6.2.1 Specification of the Form of the Fuzzy Regression Model

In the fuzzy regression model, interactions between design variables and high
order terms of design variables are represented in a form of a higher order high-
dimensional Kolmogorov-Gabor polynomial (Friedman 1991), which can be writ-
ten as
N N N N N d
y = fNR ( x ) = A0 + ∑ Ai1 xi1 + ∑∑ Ai1i2 xi1 xi2 + ...∑ ...∑ Ai1 ...id ∏ x j (6.1)
i1 =1 i1 =1 i2 =1 i1 =1 id =1 j =1
where y is the degree of customer satisfaction; xk is the k-th design variable with
k=1,2, …N; and A0 = (α 0 , c 0 ) , A1 = (α 1 , c1 ) , A2 = (α 2 , c 2 ) , …
~ ~ ~
AN = (α N , c N ) , A11 = (α 11 , c11 ) , A12 = (α 12 , c12 ) ,… ANN = (α NN , c NN ) ,…

~ ~ ~ ~
AN ...N = (α N ...N , c N ...N ) .

~
Kolmogorov-Gabor polynomials have been used widely to create general nonli-

near models by evolutionary methods (Nikolaev and Iba 2001), but no fuzzy
coefficient has been considered. The purpose of the proposed fuzzy regression
model is to overcome the traditional fuzzy regression (Tanaka et al. 1982, Takagi
and Sugeno 1985, Tanaka and Watada 1988) which considers only the linear
terms; therefore, only the first two terms of the fuzzy regression model (6.1) are
considered and the remaining terms are all ignored. It is also intended to supplant
the evolutionary methods whereby no fuzzy coefficient can be generated and thus
no fuzziness can be addressed.
The fuzzy regression model (6.1) can be rewritten as:
y = A '0 x '0 + A '1 x '1 + A '2 x '2 + ... A 'N NR x 'N NR (6.2)
( )
y = (c' 0 , α ' 0 )x' 0 +(c'1 , α '1 )x'1 +(c' 2 , α ' 2 )x' 2 +... c' N NR , α ' N NR x' N NR
or ~ (6.3)
where 1+NNR is the number of terms of design variables in (6.1), (6.2) and (6.3);
~ ~ ~ ~ ~ ~ ~ ~
A' 0 = A0 , A'1 = A1 , A' 2 = A2 ,… A ' N NR = AN ... N ; x' 0 = 1 , x'1 = x1 , x' 2 = x 2 , …
x' N NR = x1 ⋅ x 2 ⋅ ...x d ; and A' 0 = (c0 , α 0 ) , A'1 = (c'1 , α '1 ) , … A' N NR = (c' N NR , α ' N NR ) .
~ ~ ~
6.2 Fuzzy Regression Based Genetic Programming 113
~
A'i and x' i are called the fuzzy coefficient and the transformed design variable
respectively, where i=0,1,2,…NNR.
6.2.2 Determination of Fuzzy Coefficients

The vectors of the fuzzy coefficients are defined as:
( ) ( ( ))
A ' = A ' 0 , A '1 ,... A ' N NR = (c ' 0 , α ' 0 ), (c '1 , α '1 ),... c ' N NR , α ' N NR ,
~ ~ ~ ~
(6.4)
(
c' = c' 0 , c'1 ,...c' N NR ,) (6.5)
and
α ' = (α '0 , α '1 ,...α ' N NR ) . (6.6)
The vector of the variables of the transformed design attributes is defined as:
(
x' = x' 0 , x'1 , x' 2 ,...x' N NR . ) (6.7)
Using the vectors of the fuzzy coefficient and the vector of transformed design va-
riables, equation (6.2) can be rewritten as:
~ ~
y = A'⋅x 'T (6.8)
As some terms in (6.1) may be redundant, prudent selection of significant terms or

orders is advisable if a more economical and adequate model is desired. In the FR-
GP, genetic programming is used to generate the structure of a fuzzy regression
model.
6.2.3 Pseudocode of Algorithm

The flowchart of FR-GP is shown in Figure 6.1 and the pseudocode of the FR-GP
used is shown below:
Pseudocode for FR-GP
Step 1: t←0
Step 2: Initialize the population Ω(t)=[θ1(t), θ2(t),… θPOP(t)].
// Ω(t) is the population of the fuzzy regression at the t-th generation.
// θi(t) is the i-th individual of Ω(t) which represents the structure of the fuzzy regression model.
Step 3: Assign fuzzy coefficients to all terms in θi(t) using the fuzzy regression
algorithm of Tanaka et al. (1982).
Step 4: Evaluate all θi(t) based on the fitness function (6.14).
Step 5: Parent Selection Ω(t+1) ← Ω(t)
Step 6: Crossover Ω(t+1)
Step 7: Mutation Ω(t+1)
Step 8: t ← t+1
Step 9: Step 3
Step 10: Step 4
}
Step 1 Set t <- 0
Step 2 Initialize the population

Ω(t)=[θ1(t), θ2(t),…, θPOP(t)]
where the i-th individual θi(t) is in the
polynomial form. For example,
θi(t)=x1x2+x22
Step 3/ Step 9 Assign fuzzy coefficient to each

individual. For example,
θi(t)=A12x1x2+A22x22
Step 4/ Step 10 Evaluate all θi(t) based on the fitness

function (14)
Y Termination condition met
N
Step 5 Selection the individuals from Ω(t) to
Ω(t+1).
Step 6 Performing crossover on some of the

Return the
individual with the
best fitness Step 7 Performing mutation on some of the
Step 8 Performing mutation on some of the

Fig. 6.1 The flowchart of the intelligent fuzzy regression

6.2 Fuzzy Regression Based Genetic Programming 115
In Step 1, the generation number t is set to 0. In Step 2, FR-GP use the ap-
proach of genetic programming to create a random initial population Ω(t) with
POP individuals, where θi(t) is the i-th individual. Each individual θi(t) is in the
form of a tree representation, which can be used to represent the structure of the
fuzzy regression model as defined in equation (6.1). In Step 3, the fuzzy coeffi-
cients are assigned to each term of the individual θi(t) by applying Tanaka’s fuzzy
regression (Tanaka et al. 1982, Tanaka and Watada’s 1988). In Step 4, all individ-
uals are evaluated based on a defined fitness function which aims at evaluating the
goodness-of-fit of the fuzzy regression model. In Step 5, the parent selection
process uses the goodness-of-fit of each individual to determine the selection of
potential individuals for performing crossover in Step 6 and mutation in Step 7. In
Step 8, the new individuals with the determined fuzzy coefficients are evaluated
using the fitness function in order to create a new population Ω(t+1). In Step 8, the
generation number t is augmented by one.
The process continues until the pre-defined termination condition is fulfilled.
The major aspects of applying the FR-GP to generate the fuzzy regression model
are discussed in the following.
6.2.3.1 Functional Model Representation
In FR-GP, hierarchical trees, which are composed of functions in the set F and
terminals in the set T (Koza 1992), are used to represent the structure shown in
equation (1). F consists of two arithmetic operations, + and *, which exist in the
fuzzy regression model (1). T = {x, a } contains the design variable set x={ x1, x2,
… xN} of the fuzzy regression model and the fuzzy coefficient set a =
{a , a , a ,..., a } of the fuzzy regression model, where n is the number of de-
0 1 2 N NS
sign variables and NNS is the number of terms of the fuzzy regression model. The
structure of the fuzzy regression model is depicted as a labeled tree with ordered
branches, which consists of operations (internal nodes of the tree) F from the func-
tion set and arguments (terminal nodes of the tree) from the terminal set T. For
example, the i-th individual θi(t) represents the following structure of the fuzzy
regression model:
θi(t) = x12 – x22 + x1·x2 ·x4
The fuzzy regression model with fuzzy coefficients can be represented by:
a0 + a1 ·x12 – a2 ·x22 + a3 ·x1 x2 x4,
where a0 , a1 , a2 and a3 are the fuzzy coefficients. It can also be rewritten as:
a0 + a1 ·x’1 – a2 ·x’2 + a3 ·x’3,

2 2
where x’1= x1 , x’2= x2 and x’3= x1·x2·x4.
For a fuzzy regression model with NNR terms, the fuzzy coefficients, a0 , a1 ,
..., a NS , can be determined by solving the following linear programming problem:
⎛
(i ) ⎞⎟⎟
N NR
∑ ∑ x'
M
Minimize J= ⎜⎜ c' j j
(6.9)
j =0 ⎝ i =1 ⎠
where a j = ( a cj , a sj ) , M is the number of data sets, and x ' j (i ) is the j-th trans-
formed variable of the fuzzy polynomial model of the i-th data set, subject to:
N NR N NR
∑ a x ' ( i ) + (1 − h)∑ a
j =0
s
j j
j =0
c
j x ' j (i ) ≥ y (i ) (6.10)
N NR N NR
∑ a sj x ' j ( i ) + (1 − h)∑ a cj x ' j ( i ) ≤ y ( i )

j =0 j =0
(6.11)
a cj ≥ 0, a sj ∈ R, x0 ' ( i ) = 1 for all i, (6.12)
0 ≤ h ≤ 1, i = 1, 2, " M , j = 0, 1, 2, " N NR . (6.13)
J in equation (6.9) is the total fuzziness of the regression model. The value of h in
equation (6.10) and equation (6.11) is between 0 and 1. h refers to the degree to
which the fuzzy linear model fits the given data sets, and is subjectively chosen by
decision makers. Constraints (6.10) and (6.11) impose the restriction that the ob-
servation of the i-th data set y (i ) has at least h degrees of belonging to ~
y (i ) as
μ ~y (i ) ( y (i )) ≥ h (i = 1, 2, " , M ) . Therefore, the objective of solving the linear pro-
gramming problem (6.9-6.13) is to determine the fuzzy nonlinear parameters
a j = ( a cj , a sj ) such that the total vagueness J is minimized subject to
μ ~y (i ) ( y (i )) ≥ h (i = 1, 2, " , M ) .
6.2.3.2 Fitness Function
In FR-GP, the fitness function is defined as:

1 − RMAE j
fitness j =
(1 + exp(c (L − c 2 )))
(6.14)
1 j
where fitnessj is the fitness value of the j-th individual, Lj is the number of nodes
of the j-th individual, c1 and c2 are both penalty terms, and the mean absolute error
of the j-th individual RMAEj is defined as the following formulation:
y (k ) − F j ( x(k ))
∑
M
1
RMAE j = 100% × , (6.15)
M k =1 y (k )
where Fj is the fuzzy regression model represented by the j-th individual,

( y(k ), x(k )) = ( y(k ), (x1 (k ), x2 (k ),...x N (k ))) is the k-th training data set, N is the
6.3 An Illustrative Example 117
number of variables of the training data set, and M is the number of training data
sets used for developing the fuzzy regression model.
Equation (6.15) aims at fitting training data sets to the fuzzy regression model,
and it avoids generating fuzzy regression models with too many insignificant
terms. It is designed to find a balance between minimizing the number of terms
and maximizing model accuracy, since a fuzzy regression model which contains
many insignificant terms reduces its interpretation (Madar et al. 2005).
6.2.3.3 Evolutionary Operations
The evolutionary operations are mainly based on crossover, mutation, and selec-
tion of individuals. The crossover operation first selects a couple of parent
individuals from the population. Then it selects a random node in each of the hie-
rarchical tree structures of the selected parent individuals. The children individuals
are generated by exchanging the associated sub-expressions of the hierarchical
tree structures which are under the nodes of the selected parent individuals. The
crossover process maps the genotype onto the phenotype, which does not corres-
pond to the mapping of a one-to-one relationship. Because of this mapping, the
children individuals are usually different from the parents in size, shape and con-
tent. Therefore, the children individuals can be expressed by more than one differ-
ent tree structure, and diversifications in the population can be introduced. Muta-
tion is performed by randomly selecting a node and replacing the associated
sub-expression in the node by a randomly generated sub-expression.
After performing the crossover and mutation operations, children individuals
from the current population Ω(t) with relatively better fitness are selected to serve
as parent individuals for the next population Ω(t+1). The roulette-wheel approach
is one of the most common selection methods used for selecting children individu-
als to perform reproduction operations in FR-GP. The population evolves and im-
proves iteratively until a stopping condition is met. Otherwise, FR-GP goes on to
the next generation.
6.3 An Illustrative Example
6.3.1 Mobile Phone Design

The example of a mobile phone design is used to illustrate the FR-GP approach to
modeling the relationships between customer requirements and design attributes
defined in equation (6.2) in Section 6.2.1. In the highly competitive market of mo-
bile phones, the product designers provide the consumers with various styles for
different brands and different product series. To capture the trend of the recent
market in mobile phones, 32 recent mobile phones of various brands including
Nokia, Sony Ericsson and Motorola were selected as shown in Figure 6.2. Mor-
phological analysis was used to extract representative elements of mobile phones
as numerical data sets, for which both the shape profiles and the product compo-
nents of the mobile phones were used. Table 6.1 shows the 9 representative
elements as “top shape”, “bottom shape”, “side shape”, “function button shape”,
“number button style”, “length width ratio”, “thickness”, “layout” and “border and
frame”, which are denoted as x1, x2, x3, x4, x5, x6, x7, x8 and x9 respectively. They
were identified from the 32 mobile phone samples. Each representative element
has various types of form variations, which ranged from 3 to 6.
Table 6.1 Morphological analysis of the 32 representative mobile phone samples

The four most representative customer requirements regarding affective re-

－
sponse for mobile phone design, “simple complex” (S-C), “unique general” －
－－
(U-G), “high-tech classic” (H-C), and “handy bulky” (H-B) (Lai et al. 2004),
were collected from the 14 image word pairs for micro-electronic products
(Chuang and Ma 1999), and they were used to evaluate the images values of the
mobile phones. A survey was conducted using an online questionnaire to ask
about the appearance of mobile phones on the S-C, U-G, H-C, and H-B responses,
in which their degrees of image values are denoted as y1, y2, y3 and y4 respectively.
The ranking scale ranges from 1 to 5 in which 1 is the smallest and 5 is the largest.
Both front and side views of mobile phones were used to evaluate the product im-
ages since both front and side views of mobile phones influence the product image
of mobile phones. Figure 6.3 shows one of the illustrations of the mobile phones
in the survey in which 32 illustrations were given. Table 6.1 shows the morpho-
logical matrix of the 32 mobile phones samples based on the 9 representative ele-
ments. Also, it shows the means of the customer requirements regarding affective
responses S-C, U-G, H-C, and H-B with respect to 32 interviewers.
Fig. 6.2a The 32 representative mobile phone samples

Fig. 6.2b The 32 representative mobile phone samples (continued)
6.3.2 Functional Model Development

The objective of modeling is to relate the design attributes x1, x2, x3, x4, x5, x6, x7,
x8 and x9 to one of the affective responses y1, y2, y3 and y4. The modeling results
based on FR-GP are compared with those based on statistical regression (Seber
2003), Peters’ fuzzy regression (Peters 1994) and Takagi’s fuzzy regression (Ta-
kagi and Sugeno 1985). Evaluation of the effectiveness of the functional models is
carried out by investigating the mean of training errors as shown below:
1 M yk ( i ) − Pk ( x1 ( i ) , x2 ( i ) ,..., x9 ( i ) )
Re =
M
∑
i =1 yk ( i )
× 100% (6.16)
Fig. 6.3 The format of the questionnaire for each mobile phone sample
In equation (6.16) above, M is the number of data sets; yk ( i ) is the k-th cus-
tomer requirements regarding affective response of the i-th data set; x1 ( i ) ,
x2 ( i ) , … and x9 ( i ) are the i-th data set for the design attributes; and
Pk ( x1 ( i ) , x2 ( i ) , x3 ( i ) ,..., x9 ( i ) ) is the prediction of the k-th customer requirement
regarding affective response for the i-th data.
Table 6.2 Form design matrix of 32 mobile phone samples
Model
x1 x2 x3 x4 x5 x6 x7 x8 x9 y1 y2 y3 y4
no.
1 3 3 1 3 2 2 3 1 1 1.85 3.62 2.97 2.56

2 3 3 2 2 1 2 2 1 2 2.59 3.44 3.15 2.79
3 6 6 1 1 5 2 4 1 2 2.88 2.76 3.21 3.32
4 4 4 3 1 6 2 2 2 2 2.41 2.65 2.88 2.59
5 3 4 3 4 6 2 2 2 2 2.06 2.85 2.53 2.47
6 3 3 1 5 6 1 3 2 1 2.71 2.41 2.15 3.18
7 1 1 2 4 6 3 4 2 6 3.26 2.53 2.47 3.18
8 1 1 1 2 6 2 2 2 6 2.79 2.74 2.50 2.71
9 3 4 6 1 6 2 3 2 2 2.91 2.65 2.85 3.12
10 4 4 3 6 4 3 2 1 2 2.65 2.82 3.00 2.15
11 2 2 6 5 6 2 4 2 2 2.76 2.62 2.47 3.18
12 2 2 6 3 6 2 3 2 4 2.71 2.56 2.41 3.38
13 6 6 6 4 6 2 3 2 2 2.09 2.76 2.85 2.71
14 4 4 2 6 6 1 2 3 3 2.21 2.09 2.09 1.94
15 4 3 6 1 6 2 3 2 2 2.44 2.82 2.71 3.09
16 3 3 6 5 6 2 3 2 1 2.62 2.15 2.35 2.94
17 3 3 2 6 6 1 2 3 1 2.12 2.53 2.35 3.03
18 2 4 6 5 2 3 1 1 5 2.50 3.38 2.97 2.59
19 3 3 1 4 5 3 3 1 1 2.41 3.00 3.00 3.03
20 4 4 6 5 1 3 2 1 1 2.68 3.68 3.53 3.06
21 4 4 1 1 2 3 2 1 2 2.88 3.35 3.29 3.12
22 6 4 3 1 4 3 2 1 2 2.88 2.94 2.97 2.97
23 3 3 6 2 3 2 3 1 6 3.12 3.38 3.15 3.56
24 5 5 1 4 3 3 2 1 1 2.50 2.85 3.24 2.62
25 4 4 6 1 6 2 3 2 2 2.44 3.21 3.06 3.09
26 3 6 5 1 6 3 3 2 2 2.68 2.97 2.85 3.32
27 1 1 5 1 6 2 2 2 4 2.65 2.79 2.79 2.91
28 3 3 4 1 6 1 2 3 2 2.00 1.91 1.91 2.53
29 4 4 2 1 6 1 2 2 2 2.41 2.47 2.21 2.56
30 4 4 4 5 2 2 3 1 2 3.26 3.15 2.82 3.03
31 3 3 1 6 6 2 3 4 2 3.38 2.79 2.76 3.18
32 3 3 1 1 6 2 3 2 1 2.32 2.62 2.56 3.50
Table 6.3 Training errors (Re in percentage) of the four modeling methods
Me- Trained models Training

thod error (%)
S-C SR y = 1.2798 + 0.2506·x1 – 0.1986·x2 – 0.0272·x3 – 0.0010·x4 – 9.2081
0.4208·x5 + 0.4547·x6 + 0.6401·x7 + 0.3634·x8 + 0.2580·x9
TFR y = (1.21,0) + (0.50,0)·x1 + (-0.53,3.52)·x2 + (0.21,0)·x3 + 11.836
(0.22,0.99)·x4 + (-0.01,0)·x5 + (0.51,0)·x6 + (0.07,0)·x7 +
(0.18,0)·x8 + (0.29,0)·x9
PFR y = (0.87,0.07) + (0.11,0.03)·x1 - (0.05,0.03)·x2 + 9.2132
(0.02,0.12)·x3 + (0.04,0.04)·x4 - (0.01,0.00)·x5 +
(0.14,0.02)·x6 + (0.14,0.27)·x7 + (0.16,0.01)·x8 +
(0.29,0.01)·x9
IFR y = (0.04529,0.6639) + (2.0400,0.0453)·(x9 + x6·x7) 0.091471
U-G SR y = 3.3302 – 0.1971·x1 + 0.1514·x2 + 0.0480·x3 – 0.1195·x4 – 8.3444
0.9145·x5 + 0.2959·x6 + 0.1916·x7 + 0.0522·x8 - 0.0106·x9
TFR y = (3.43,0) + (-0.59,0)·x1 + (0.46,0)·x2 + (0.13,0.38)·x3 - 8.5260
(0.23,0)·x4 - (1.11,0)·x5 + (0.31,0)·x6 + (0.43,0)·x7 -
(0.01,2.98)·x8 + (0.00,0)·x9
PFR y = (3.71,0) - (0.19,0.01)·x1 + (0.14,0.00)·x2 + (0.02,0)·x3 - 7.6739
(0.07,0)·x4 - (0.20,0)·x5 + (0.09,0)·x6 + (0.13,0)·x7 -
(0.08,0.14)·x8 - (0.00,0)·x9
IFR y = (2.73,0.00) ·x5·x8 + (-0.18,1.20)·x8 + (1.01,0.00) 0.07278
H-C SR y = 2.2971 + 0.1752·x1 + 0.0767·x2 + 0.0321·x3 – 0.1197·x4 – 6.7134
0.5569·x5 + 0.6110·x6 + 0.1731·x7 + 0.0673·x8 - 0.0042·x9
TFR y = (2.40,0.01) + (0.00,0.00)·x1 + (0.05,0.00)·x2 + 6.3738
(0.02,0.01)·x3 + (-0.03,0.00)·x4 + (-0.09,0.00)·x5 +
(0.29,0.01)·x6 + (0.08,0.01)·x7 + (-0.05,0.08)·x8 + (-
0.02,0.01)·x9
PFR y = (2.46,0) + (-0.03,0.59)·x1 + (0.13,0)·x2 + (0.01,0)·x3 + (- 7.6739
0.11,0)·x4 + (-0.40,0)·x5 + (0.62,0)·x6 + (0.30,0.095)·x7 + (-
0.12,1.31)·x8 + (-0.10,0.48)·x9
IFR y = (1.12,0) ·x62·x5 + (-0.02,0.3143)·x6 + (1.04,0) ·x2 + 0.0582
(0.04,0)
H-B SR y = 2.2905 - 0.1365·x1 - 0.0710·x2 + 0.0776·x3 – 0.2116·x4 – 9.8682
0.3376·x5 + 0.2183·x6 + 0.9195·x7 + 0.2235·x8 - 0.0471·x9
TFR y = (1.77,1.50) + (0.02,0)·x1 + (-0.08,0)·x2 + (0.11,0)·x3 + (- 8.8092
0.23,0.94)·x4 + (0.04,0)·x5 + (0.29,0)·x6 + (0.82,0)·x7 +
(0.18,1.3757)·x8 + (0.07,0)·x9
PFR y = (2.33,0.041) + (-0.04,0)·x1 + (-0.01,0)·x2 + (0.029,0)·x3 + 8.2113
(-0.09,0.026)·x4 + (-0.06,0)·x5 + (0.03,0)·x6 + (0.41,0)·x7 + (-
0.09,0.09)·x8 + (0.00,0)·x9
IFR y = (3.6920,0) ·x4 + (-0.1531,0) ·x8 + (0.0165,0.2423) ·x82 + (- 0.0756
0.5189,0)·(x8 +x72 + x4·x7)
Table 6.4 Means and variances of the prediction errors of the affective responses
statistical Takagi’s Peters’ Intelligent

regression fuzzy re- fuzzy re- fuzzy re-
gression gression gression
S-C Mean (%) 12.825 17.794 12.473 6.7548
Var. (%) 106.31 136.68 76.801 34.693
U-G Mean (%) 9.279 11.319 10.953 5.8147
Var. (%) 31.849 50.093 45.178 9.1224
H-C Mean (%) 8.9346 9.1336 8.5013 4.9788
Var. (%) 32.941 36.819 33.346 11.505
H-B Mean (%) 11.216 13.923 12.473 8.3124
Var. (%) 48.739 73.713 76.801 52.673
The 32 interviewees’ numerical data and their results are shown in Table 6.2.
Using these, the proposed FR-GP was implemented using Matlab to generate
functional models to relate the customer requirements regarding (y1, y2, y3, y4) and
the design attributes regarding (x1, x2, x3, … x9). The parameters used in FR-GP
were shown as follows: population size = 50, number of generations = 100, gener-
ation gap = 0.9, probability of crossover = 0.5, probability of mutation = 0.5,
probability of changing terminal via no-terminal = 0.25, penalty factors in the fit-
ness function (c1=0.5 and c2=30), maximum depth of tree = 30. Since FR-GP is a
stochastic method, different results will be obtained from different runs. To eva-
luate its overall performance, 30 runs on FR-GP were carried out, and the mean of
the 30 runs was calculated. The functional models developed for the four customer
requirements regarding affective responses S-C, U-G, H-C, and H-B by using the
four methods, and the training errors (Re) of the developed functional models are
summarized in Table 6.3. It shows that the Re of the proposed FR-GP is smaller
than those of the statistical regression, Takagi’s fuzzy regression and Peter fuzzy
regression. This indicates that the proposed FR-GP can fit the data sets with the
smallest mean errors.
To perform cross validation the modeling performance of the FR-GP, two data
sets were randomly selected from the 32 data sets, as shown in Table 6.4, as test-
ing data sets and the remaining 30 data sets were used to develop the functional
models. Their prediction errors were calculated. The validations were repeated 32
times. Because FR-GP is a stochastic algorithm, we ran the FR-GP 30 times in
each validation test, and the mean of the 30 runs was calculated. Table 6.3 sum-
marizes the means and variances of the prediction errors of the S-C, U-G, H-C and
H-B for the four methods respectively. From the table, it can be seen that FR-GP
yields the smallest means of prediction errors and variances of prediction errors
for all S-C, U-G, H-C and H-B.
6.3.3 Optimization of Affective Design

For the maximization of customer requirements regarding affective design of the
mobile phones, it is necessary to determine a setting of design variables so as to
6.4 Conclusion 125
maximize the degree of satisfaction of the four customer requirements, S-C, U-G,
H-C and H-B. To determine the optimal design attribute setting, the multi-
objective function is formulated by maximizing the degree of satisfaction of the
four customer requirements regarding affective design, S-C, U-G, H-C and H-B.
Among the four methods, the FR-GP can achieve the best modeling results. Thus,
the functional models developed by the FR-GP are used and the multi-objective
function is formulated by:
⎧ y1 = (0.04529,0.6639) + (2.0400,0.0453) ⋅ ( x9 + x6 ⋅ x7 )
⎪
⎪ y2 = (2.7266,0.0025) ⋅ x5 ⋅ x8 + (-0.1769,1.1952) ⋅ x8 + (1.0102,0.0023)
⎪
max ⎨ y3 = (1.1201,0) ⋅ x6 ⋅ x5 +(-0.0206,0.3143) ⋅ x6 + (1.0444,0) ⋅ x2 + (0.0366,0) (6.17)
2
⎪ y = (3.6920,0) ⋅ x + (-0.1531,0) ⋅ x + (0.0165,0.2423) ⋅ x 2 +

⎪ 4 4 8 8
⎪⎩ (-0.5189,0)·(x8 + x7 + x4 ⋅ x7 )
2
subject to: x1 , x2 , x3 , x4 , x5 , x9 ∈ {1, 2,..., 6} ;
x6 , x8 ∈ {1,2,3} ;
x7 ∈ {1,2,3, 4}
(6.17) is a Pareto-based multi-objective problem which can be solved by the

multi-objective genetic algorithm richly represented in the literature for solving
multi-objective problems (Knowles and Corne, 2000; Zitzler and Thiele, 1999).
Therefore, GA was used in this research. The GA optimization model aims at mi-
nimizing the objective function (6.17) from which an optimal design attributes (x1,
x2, x3, x4, x5, x6, x7, x8 and x9) setting can be obtained to achieve the maximum of
y1, y2, y3 and y4. A detailed description of the multi-objective genetic algorithm is
not presented in this chapter as it is not within the scope of the discussion. For de-
tails, the reader is referred to (Fonseca and Fleming 1993). The operations and the
parameters used in the genetic algorithm can be referred to (Chan et al. 2009). The
optimal design attribute setting is found as:
x1 = 3, x2 = 2, x3 = 6, x4 = 2, x5 = 1, x6 = 2, x7 = 4, x8 = 3 and x9 = 5
6.4 Conclusion
Existing modeling methods which relate customer requirements to the design
attributes of a new product have not addressed the development of explicit models
which represent the nonlinearity or fuzziness inherent in such functional models.
The chapter presents a modeling method, namely fuzzy regression based genetic
programming (FR-GP), to generate functional models which relate design
attributes to customer requirements in which both nonlinearity and fuzziness are
considered. In order to develop a methodology which is capable of addressing
both fuzziness and nonlinearity in fuzzy regression models, FR-GP uses the
genetic programming to first construct branches of a tree which represents the

nonlinear structures of the functional model. Then FR-GP uses the fuzzy regres-
sion algorithm to calculate the fuzzy coefficients of the functional model. The re-
sulting functional model is explicit, and consists of fuzzy nonlinear terms in relat-
ing customer requirements to design attributes.
A case study of the affective product design of mobile phones is used to illu-
strate the effectiveness of FR-GP. To validate this FR-GP approach in the model-
ing of the relationships between customer requirements and design attributes, it
has been applied to generate functional models of affective product design of mo-
bile phones. Four customer requirements and nine design attributes of mobile
phones were considered. This is compared with the other commonly used explicit
modeling methods: statistical regression and two fuzzy regression methods. The
results show that functional models with the smallest training errors are generated
by FR-GP rather than the ones generated by the other methods. This indicates that
FR-GP is more capable of modeling the survey data sets which are fuzzy and non-
linear. The validation results show that the smallest prediction errors and errors in
variance are also produced by FR-GP rather than by other methods. Better results
were achieved and can be explained by the fact that nonlinearity and fuzziness
were taken into consideration using FR-GP, which is not represented in the exist-
ing commonly-used methods.
In this chapter, uncertainty regarding fuzziness, which comes from customers’
preferences when selecting products, has been addressed by the FR-GP. However,
uncertainty regarding randomness, which comes from the elicitation of perfor-
mance data of competitive products, has not been addressed. In Chapter 7,
uncertainty regarding both fuzziness and randomness will be addressed by the ge-
neralized fuzzy least square regression for developing customer requirement
models.
References
Artacho, M.A., Ballester, A., Alcantara, E.: Analysis of the impact of slight changes in
product formal attributes on user’s emotions and configuration of an emotional space for
successful design. Journal of Engineering Design (2009)
Barnes, C., Lillford, S.P.: Decision support for the design of affective products. Journal of
Engineering Design 20(5), 477–492 (2009)
Chan, K.Y., Kwong, C.K., Tsim, Y.C.: Modelling and optimization of fluid dispensing for
electronic packaging using neural fuzzy networks and genetic algorithms. Engineering
Applications of Artificial Intelligence 23(1), 18–26 (2010)
349 (2006)
Chuang, M.C., Ma, Y.C.: Expressing the expected product images in product design of mi-
cro-electronic products. International Journal of Industrial Ergonomics 27(4), 233–245
(1999)
Cross, N.: Engineering design methods: Strategies for product design, 3rd edn. Wiley,
Chichester (2000)
References 127
Eggink, W.: A practical approach to teaching abstract product design issues. Journal of En-
gineering Design 20(5), 511–521 (2009)
Girard, S., Johnson, H.: Developing affective educational software products: Soremo, a new
method for capturing emotional states. Journal of Engineering Design 20(5), 493–510
(2009)
Desmet, P.M.A.: Designing emotions. Delft University of Technology, Delft (2002)
Diener, E., Lucas, R.E.: Subjective emotional well-being. In: Lewis, M., Haviland-Jones,
J.M. (eds.) Handbook of Emotions, 2nd edn., pp. 325–337. The Guilford Press, Ne-
wYork (2000)
Friedman, J.H.: Multivariate Adaptive Regression Splines. The Annals of Statistics 19(1),
1–141 (1991)
Fonseca, C.M., Fleming, P.J.: Genetic algorithms for multiobjective optimization: formula-
tion, discussion and generalization. In: Proceedings of the 5th International Conference
in Genetic Algorithms (1993)
Gu, Z., Tang, M.X., Frazer, J.H.: Capturing aesthetic intention during interactive evolution.
Computer-Aided Design 38, 224–237 (2006)
Han, S.H., Hong, S.W.: A systematic approach for computing user satisfaction with product
design. Ergonomics 46(13), 1441–1461 (2003)
Han, S.H., Yun, M.H., Kim, K., Kwahk, J.: Evolution of product usability: development
Hsiao, S.W., Liu, M.C.: A morphing method for shape generation and image prediction in
product design. Design Studies 23(5), 497–513 (2002)
ics 35, 411–428 (2005)
Jiao, J.: A Kansei mining system for affective design. Expert Systems with Applications 30,
658–673 (2006)
Jindo, T., Hirasago, K.: Application studies to car interior of Kansei engineering. Interna-
tional Journal of Industrial Ergonomics 19, 105–114 (1997)
Jordan, P.W.: Designing pleasurable products: an introduction to the new human factors.
Taylor & Francis, London (2000)
Kesteren, I., Bruijn, S., Stappers, P.J.: Evaluation of materials selection activites in user-
centred design projects. Journal of Engineering Design 19(5), 417–429 (2008)
Khalid, H.M.: Towards affective collaborative design. In: Smith, M.J., Salvendy, G., Har-
ris, D., Koubek, R.J. (eds.) Proceedings of HCI International 2001. Usability Evaluation
and Interface Design, vol. 1, Lawrence Erlbaum, Mahwah (2001)
Kim, K.J., Moskowitz, H., Koksalan, M.: Fuzzy versus statistical linear regression. Euro-
Kouprie, M., Visser, F.S.: A framework for empathy in design: stepping into and out of the
user’s life. Journal of Engineering Design 20(5), 437–448 (2009)
Knowles, J.D., Corne, D.W.: Approximating the non-dominated front using the pareto arc-
hived evolution strategy. Evolutionary Computation 8, 149–172 (2000)
(1994)
Kuang, J., Jiang, P.: Product platform design for a product family based on Kansei engi-
neering. Journal of Engineering Design (2008)
Lai, H.H., Lin, Y.C., Yeh, C.H.: Form design of product image using grey relational analy-
sis and neural network models. Computers & Operations Research 32(10), 2689–2711
(2004)
Lau, T.W., Hui, C.L., Ng, S.F., Chan, C.C.: A new fuzzy approach to improve fashion
product development. Computers in Industry 57, 82–92 (2006)
(2005)
Mugge, R., Schoormans, J.P.L., Schifferstein, N.J.: Emotional bonding with personalized
products. Journal of Engineering Design 20(5), 467–476 (2009)
Nagamachi, M.: Kansei Engineering: A new ergonomic consumer-oriented technology for
product development 15, 3–11 (1995)
Nurkka, P., Kujala, S., Kemppainen, K.: Capturing user’s perceptions of valuable expe-
rience and meaning. Journal of Engineering Design 20(5), 449–463 (2009)
Nikolaev, N.I., Iba, H.: Accelerated Genetic Programming of Polynomials. Genetic Pro-
gramming and Evolvable Machines 2, 231–257 (2001)
Norman, D.A.: Emotional design: why we love (or hate) everyday things. Basic Books,
New York (2004)
wards office chair design. International Journal of Industrial Ergonomics 34, 31–47
(2004)
Peters, G.: Fuzzy linear regression with fuzzy intervals. Fuzzy Sets and Systems 63, 45–55
(1994)
Petiot, J.F., Grognet, S.: Product design: a vectors field-based approach for preference
modeling. Journal of Engineering Design 17(3), 217–233 (2006)
Shimizu, Y., Jindo, Y.: A fuzzy logic analysis method for evaluation human sensitivities.
Tanoue, C., Ishizaka, K., Nagamachi, M.: Kansei engineering: a study on perception of ve-
hicle interior image. International Journal of Industrial Ergonomics 19, 115–128 (1997)
Yang, S., Nagamachi, M., Lee, S.: Rule based inference model for the Kansei engineering
system. International Journal of Industrial Ergonomics 24, 459–471 (1999)
You, H., Taebeum, R.Y.U., Kyunghee, O.H., Yun, M.H., Kim, K.J.: Development of cus-
tomer satisfaction models for automotive interior materials. International Journal of In-
dustrial Ergonomics 36, 323–330 (2006)
Zitzler, E., Thiele, L.: Multiobjective evolutionary algorithms: a comparative case study
and the strength pareto approach. IEEE Transactions on Evolutionary Computation 3(4),
257–271 (1999)
Chapter 7
Generalized Fuzzy Least Square Regression
for Generating Customer Satisfaction Models
Generalized F uzzy Least Square Regression
7.1 Introduction
Chapter 1 mentioned that quality function deployment (QFD) is a commonly used
method to support product planning. QFD utilizes four sets of matrices called
Houses of Quality (HOQ) to relate customer requirements to product planning,
parts deployment, process planning and manufacturing operations (Hauser and
Clausing, 1988). In essence, QFD is a systematic and graphical approach, intended
to help a design team understand a product’s essential requirements, internal capa-
bilities and constraints, and thereby helps it fulfill customer requirements. Cus-
tomer requirements acquired from markets are typically qualitative and usually
ambiguous in nature, especially for consumer products. Under QFD, customer
requirements are mapped into engineering characteristics. Engineering characteris-
tics might not be specific design details or solutions, but they should be measura-
ble. Target values of engineering characteristics, normally housed at the bottom of
a HOQ, provide definitive and quantitative technical specifications for new prod-
ucts. This involves a complex decision-making process with multiple variables
and in practice, it is normally accomplished in a subjective or heuristic manner.
In the development of HOQ, uncertainties relating to human fuzziness and ran-
domness always exist. To enhance the generalization capabilities of functional
models, it is essential to consider these two uncertainties when developing such
models.
Chapter 6 discussed a fuzzy regression approach to address uncertainty due to
fuzziness when developing models based on customer survey data. However, un-
certainty due to randomness has not been addressed. Also, previous research only
addressed fuzziness and randomness independently of one another, when develop-
ing these functional models. Even if the approach of fuzzy least-square regression
takes into account the two uncertainties of fuzziness and randomness, it is limited
to developing only functional models based on fuzzy observed data. Often, the
modeling of functional models may involve both crisp type and fuzzy type
observed data. This chapter discusses an approach using generalized fuzzy
130 7 Generalized Fuzzy Least Square Regression
least-squares regression (GFLSR) which is intended to address these two uncer-

tainties when modeling the functional models. Moreover, it can be used to develop
models of the functional relationships based on fuzzy observations and/or crisp
observations. A case study of an emulsified dynamite packing machine design is
used to illustrate the effectiveness of GFLSR.
7.2 Theoretical Background of Generalized Fuzzy Least

Squares Regression
To formulate a FLSR problem, it is necessary to define a distance measure
between two fuzzy numbers. Various types of fuzzy distance have been used in
previous studies (Diamond, 1998; D’Urso and Gastaldi, 2000; Chang, 2001; Xu
and Li, 2001). The definition introduced by Xu and Li (2001) is adopted and is
described as:
Suppose B and C are two fuzzy numbers, the distance between B and C ac-
cording to a function t (α ) is defined as:
( ) (
d B , C = ⎡ ∫ t (α ) d 2 Bα , Cα dα ⎤ )
1 2
(7.1)
⎢⎣ 0 ⎥⎦
where Bα and Cα denote the α − level sets of fuzzy numbers B and C respective-
( )
ly, d Bα , Cα is referred to a distance between the α − level sets of fuzzy num-
bers B and C , and t (α ) is an increasing function between [ 0,1] for which
t ( 0 ) = 0 and ∫ t (α ) dα = 0.5 .
1
In (7.1), Bα and Cα can be expressed respectively as:
Bα = ⎡⎣ BαL , BαU ⎤⎦ (7.2)
and
Cα = ⎡⎣CαL , CαU ⎤⎦ (7.3)
( )
Therefore, d Bα , Cα can be calculated as:
( )
d 2 Bα , Cα = ( BαL − CαL ) + ( BαU − CαU )
2 2
(7.4)
If B is a symmetric triangular fuzzy number, it can be denoted by ( bC , bS ) uni-

quely, where b C and bS are center value and spread value of B respectively. If
(
B and C are both symmetric triangular fuzzy numbers, d 2 B , C in (7.4) can be )
derived as:
7.2 Theoretical Background of Generalized Fuzzy Least Squares Regression 131
( )
d 2 B , C = ( bC − c C ) + ( bS − cS )
1
2 2
(7.5)
6
Given a set of fuzzy type observed data {(η1 , ω1 ) , (η2 , ω2 ) ,", (ηλ , ωλ ) ,", (ηπ , ωπ )} ,
λ = 1,..., π , find a fuzzy least-squares regression (FLSR) model with fuzzy para-
meters as shown below.
ηˆ = A0 + A1ω1 + ... + Aξ ωξ + Aφ ωφ (7.6)
where Aξ , ξ = 0,1,",ϕ , are fuzzy numbers. The model can be obtained by
minimizing the sum of squared error (SSE) between ηˆλ and ηλ , λ = 1,..., π , as
shown below.
π π
λ =1 λ =1
(
SSE = ∑ d 2 (ηˆλ , ηλ ) = ∑ d 2 A0 + A1ω1 + " Aξ ωξ " + Aφ ωφ , ηλ ) (7.7)
When Aξ , ξ = 0,1, " , ϕ , are all symmetric triangular fuzzy numbers. According
to (7.5), the form of SSE in (7.7) can be converted into (7.8).
2 2
π ⎡⎛ φ ⎞ ⎤ 1 π ⎡⎛ φ ⎞⎤
SSE = ∑ ⎢⎜ ∑ aξCωξ − ηλC ⎟ ⎥ + ∑ ⎢⎜ ∑ aξS ωξ − ηλS ⎟ ⎥ (7.8)
λ =1 ⎢
⎣ ⎝ ξ =0 ⎠ ⎥⎦ 6 λ =1 ⎢⎣⎝ ξ =0 ⎠ ⎥⎦
In (7.8), there are 2 (φ + 1) unknown parameters ( aξC , aξS ) , ξ = 0,1, " , φ . Accord-
ing to Xu and Li (2001) and Chang (2001), to derive a formula for the unknown
regression coefficients based on minimizing the total residual errors SSE, the
derivatives of (7.8) with respect to the 2 (φ + 1) unknown parameters need to be
obtained, set to zero, and solved for the 2 (φ + 1) unknowns. The procedures are
similar to those of statistical regression analysis.
However, the above procedures of generating a fuzzy least square regression
model have two limitations. First, the procedures can be applied only to modeling
problems which are based on fuzzy type observed data. When observed data is
crisp, the above procedures become a statistical least-squares regression analysis.
However, in modeling the functional relationships in QFD, both fuzzy type and
crisp type observed data are involved. Second, negative spread values of fuzzy
coefficients may be generated based on the above procedure. In fact, the spread
values of triangular fuzzy coefficients should not be less than zero, i.e., aξS ≥ 0 ,
ξ = 0,1, " , φ .
To overcome the aforementioned deficiencies, a generalized fuzzy least-
squares regression is proposed by enhancing the existing fuzzy least-squares
regression with two areas. First, the condition aξS ≥ 0 , ξ = 0,1,",φ , needs to be sa-
tisfied. Second, similar to the fuzzy linear regression proposed by Tanaka and
Watada (1988), the estimated coefficients are determined by minimizing SSE and
at the same time the h-level set of given input-output pairs should be included in
the h -level set of estimations which can be described as [ηλ ]h ∈ [ηˆλ ]h . Therefore,
the programming model for the existing fuzzy least square regression can be
converted into the following constrained programming model:
2 2
π ⎡⎛ φ ⎞ ⎤ 1 π ⎡⎛ φ ⎞⎤
Min SSE = ∑ ⎢⎜ ∑ aξCωξ − ηλC ⎟ ⎥ + ∑ ⎢⎜ ∑ aξS ωξ − ηλS ⎟ ⎥ (7.9a)
⎢⎝ ξ = 0
λ =1 ⎣ ⎠ ⎦⎥ 6 λ =1 ⎣⎢⎝ ξ =0 ⎠ ⎦⎥
Subject to
φ φ
∑
ξ
aξ ωξ − (1 − h )∑ aξ ωξ
=0
C
ξ =0
S
≤ ηλC − (1 − h )ηλS , λ = 1," , π (7.9b)
φ φ
∑
ξ
aξ ωξ + (1 − h )∑ aξ ωξ
=0
C
ξ =0
S
≥ ηλC + (1 − h )ηλS , λ = 1," , π (7.9c)
aξS ≥ 0 , ξ = 0,1,",φ (7.9d)
where 0≤h<1 denotes the degree of fitness of the estimated fuzzy least square re-
gression model and determines the range of the possibility distribution of fuzzy pa-
rameters. A physical interpretation of h is that ηλ is in the support interval of ηλ
with a degree of membership of at least h for all λ .
The above model (7.9a-7.9d) is a constrained nonlinear programming problem.
A constrained variable metric method or generalized reduced gradient method is
found suitable to solve this type of problem (Reklaitis GV et al., 1983). A simple
test has been conducted as shown below to investigate how the proposed GFLSR
could be used to solve the problem of generating negative spread values of fuzzy
coefficients. Table 1 shows a set of fuzzy type observed data for the test.
Table 7.1 Input-output data
No. 1 2 3 4 5 6 7 8
ωλ 21 15 15 9 12 18 6 12
ηλ (4, 0.2) (3, 0.5) (3.5, 0.35) (2, 0.4) (3, 0.45) (3.5, 0.7) (2.5, 0.38) (4, 1)
7.3 Modeling Functional Relationships Using GFLSR 133
Based on the existing fuzzy least square regression, the following model can be
generated.
ηˆ = (1.7500, 0.5725) + ( 0.1065, −0.0056 ) ω
From the above model, it can be found that one spread value of the estimated
coefficients is negative. This indicates that the above model is invalid.
Based on the generalized fuzzy least square regression, the following model
with h=0.5 can be obtained.
ηˆ = (1.7500,0.0000 ) + ( 0.1065,0.2454 ) ω
From the above model, it can be found that none of the spreads of estimated
coefficients is negative.
7.3 Modeling Functio nal Relations hips U sing GFLSR
7.3 Modeling Functional Relationships Using Generalized

Fuzzy Least-Squares Regression (GFLSR)
Assume that in a product design, m customer requirements are denoted by
CR i ( i = 1,2,", m ) , n engineering characteristics are denoted by EC j ( j = 1,2,", n )
and q competitive products are denoted by Compq ( q = 1,2,", p ) are considered.
The product planning process based on a HOQ is to determine target value setting
of engineering characteristics of a new product, xj (j = 1, 2, .., n), by maximizing
overall customer satisfaction subject to:
yi = fi ( x ) , i = 1, " , m (7.10)
x j = g j ( x j ) , j = 1,", n (7.11)
where
x = ( x1 , x2 , " xn )
T
x j = ( x1 , " , x j −1 , x j +1 , " , xn )
T
yi Customer perception of the degree of satisfaction of CR i , i = 1,2," ,m ,

x j Level of attainment of EC j , j = 1,2, " , n ,
fi Functional relationship between CR i and engineering characteristics,
i.e. yi = f i ( x1 , x2 , " xn ) ,
g j Functional relationship between EC j and other engineering characteristics,
i.e. x j = g j ( x1 ,", x j −1 , x j +1 ,", xn ) .
The above is a general functional model to determine the target values setting of
engineering characteristics. Additional constraints, such as cost constraint, can be
added. Various forms of objective function could be formulated for different sce-
narios. For details of developing objective functions, readers could refer to Kim et
al. (2000), Bai and Kwong (2003), and Chen et al. (2004).
To model the functional relationships f i ( i = 1,2,", m ) and g j ( j = 1,2,",n) us-
ing GFLSR, first, the following fuzzy regression model relating CRi to ECj needs
to be determined.
Yi = fi ( X , A i ) = Ai 0 + Ai1 x1 + " + Ain xn (7.12)
where Yi is the fuzzy output of the degree of customer satisfaction of the CRi,
X = ( x0 , x1 ,", xn ) is the real-valued input vector of the level of attainment of en-
T
gineering characteristics with x0 = 1 , and A i = ( A i 0 , A i1 , " , A in ) is a vector of

fuzzy numbers, and A ij has symmetric triangular membership functions defined
by Aij = ( aijC , aijS ) .
Given a number of p non-fuzzy input-output pairs ( xq , yiq ) , q = 1, " , p , a set
of fuzzy parameters Ai 0 , Ai1 ,", Ain will be determined such that (7.12) can best fit
the given data points. The xq is a vector of inputs and gives fulfillment of engi-
neering characteristics for Comp q , x q = ( x0 q , x1q ," , x jq ," , xnq ) where
x0 q = 1∀q , x jq is the degree of fulfillment of the ECj of Comp q , and yiq is the sa-
tisfaction level of performance of CRi of Comp q . According to the programming
model (7.9), the functional relationship fi ( i = 1,2,", m ) can be obtained by solv-
ing the following constrained programming model GFLSR-fi ( i = 1,2,", m ) .
2 2
⎡⎛ n p
⎞ ⎤ 1 p ⎡⎛ n ⎞⎤
Min SSE = ∑ ⎢⎜ ∑ aijC x j − yiC ⎟ ⎥ + ∑ ⎢⎜ ∑ aijS x j − yiS ⎟ ⎥ (7.13a)
⎢⎝ j = 0
q =1 ⎣ ⎠ ⎦⎥ 6 q =1 ⎣⎢⎝ j =0 ⎠ ⎦⎥
Subject to
n n
∑ aijC x jq − (1 − h )∑ aijS x jq ≤ yiqC − (1 − h ) yiqS , q = 0,1,", p

j =0 j =0
(7.13b)
n n
∑a
j =0
C
ij x jq + (1 − h )∑ aijS x jq ≤ yiqC + (1 − h ) yiqS , q = 0,1,", p
j =0
(7.13c)
aijS ≥ 0 , j = 0,1,", n (7.13d)

Hence, the model of functional relationship fi ( i = 1,2,", m ) derived based on the

model (7.13) can be expressed as
yi = fi ( x1 ,", xn ) = ∑ ( aijC , aijS )x j , i = 1, 2,", m

n
(7.14)
j =0
Similarly, one can determine the models for functional relationships

g j ( j = 1,2,",n) among engineering characteristics using GFLSR, which can be
expressed as
= g xj , A
X j j j( j0 )
= A + A x +"+ A x + A x +"+ A x
j1 1 j , j −1 j −1 j , j +1 j +1 jn n (7.15)
where X j is the fuzzy output of the degree of attainment of target level of the ECj,
x j = ( x0 , x1,", x j−1, xj+1,", xn ) with x0 = 1 is the real-valued input vector of the level
T
j j0 j1 (
= A , A ,", A , A ,", A is
of attainment of engineering characteristics, and A j , j −1 j , j +1 jn )
a set of symmetric triangular fuzzy coefficient to be determined by solving the
following constrained programming model GFLSR- gj ( j = 1,2,", n ) :
2 2
⎡⎛ n p
⎞ ⎤ 1 p ⎡⎛ n ⎞⎤
Min SSE = ∑ ⎢⎜ ∑ a Cjk xk − x Cj ⎟ ⎥ + ∑ ⎢⎜ ∑ aSjk x j − xSj ⎟ ⎥ (7.16a)
q =1 ⎣⎝ k = 0 ⎠ ⎦ 6 q =1 ⎣⎝ k =0 ⎠⎦
Subject to
n n
∑a j =0
C
jk x jq − (1 − h )∑ aSjk x jq ≤ x Cjq − (1 − h ) xSjq , q = 0,1,", p
j =0
(7.16b)
n n
∑a
j =0
C
jk x jq + (1 − h )∑ aSjk x jq ≤ x Cjq + (1 − h ) xSjq , q = 0,1,", p
j =0
(7.16c)
aSjk ≥ 0 , j = 0,1,", n (7.16d)
By solving the model (7.16), the models for the functional relationship
g j ( j = 1,2,", n ) can be obtained as shown below.
x j = g j ( x1 ,", xn ) = ∑ ( a Cjk , aSjk )x j , j = 0,1,", n

n
(7.17)
j =0
To measure the degree of interpretation, Wang and Tsaur (2000) have defined an
index of confidence, which is similar to the coefficient of determination R2 in
statistics. According to Wang and Tsaur (2000), the central line has the highest
ability to represent the given data in fuzzy regression with symmetric triangular
fuzzy number coefficients. In a fuzzy linear regression model, each xq has its
corresponding interval yˆ q with the bounds yˆ qL and yˆ qU . Let yq' be any one point
within the fuzzy interval ⎡⎣ yˆ qL , yˆ qU ⎤⎦ . When yq' is used to represent the fuzzy
interval, the square of the fuzzy interval can be obtained as shown below.
Sq = ( yq' − yˆ qL ) + ( yˆ qL − yq' )
2 2
(7.18)
The value of yq' can be obtained by minimizing Sq . Therefore, the derivation of

(7.18) with respect to yq' needs to be derived and set to zero. It can be expressed
as follows.
∂Sq
= 2 ( yq' − yˆ qL ) − 2 ( yˆ qU − yq' ) = 0 (7.19)
∂yq'
By solving (7.19), the value of yq' is given as:
yq' =
2
( yˆ q + yˆ qU ) = yˆ q = yˆqC
1 L
(7.20)
From the above, it can be observed that when yq' is equal to the mean value of the
fuzzy interval denoted as yˆ q , the square is minimized. For the symmetric cases,
yˆ q is equal to yˆ qC . It means that yˆ qC has the highest ability to interpret the given
data y q when the membership function of fuzzy coefficients is a symmetric trian-
gular type.
The total sum of squares of a GFLSR model given in (7.13) can be obtained as
shown below.
SST = ∑ ( yqC − yˆ qL ) + ∑ ( yˆ qU − yqC )

p p
2 2
q =1 q =1
(7.21)
= 2∑ y + ∑ ( yˆ ) + ∑ ( yˆ ) ( yˆ )
p p p p
C
q
L 2
q
U 2
q − 2∑ y C
q
L
q + yˆ U
q
q =1 q =1 q =1 q =1
According to (7.20) and (7.21), SST can also be derived as
SST = 2∑ ( yqC ) + ∑ ( yˆ qL ) + ∑ ( yˆ qU ) − 4∑ yqC yˆ qC

p p p p
2 2 2
q =1 q =1 q =1 q =1
= ∑ ( yˆ ) + ∑ ( yˆ ) − 2∑ ( yˆ ) + 2∑ ( yqC − yˆ qC )
p p p p
L 2 U 2 C 2 2
q q q
q =1 q =1 q =1 q =1
= ∑ ( yˆ ) + ∑ ( yˆ ) + 2∑ ( yˆ ) − 4∑ ( yˆ qC ) + 2∑ ( yqC − yˆ qC )
p p p p p
L 2 U 2 C 2 2 2
q q q (7.22)
q =1 q =1 q =1 q =1 q =1
= ∑ ( yˆ qL ) + ∑ ( yˆ qU ) + 2∑ ( yˆ qC ) − 2∑ yˆ qC ( yˆ qL + yˆ qU ) + 2∑ ( yqC − yˆ qC )
p p p p p
2 2 2 2
q =1 q =1 q =1 q =1 q =1
= ∑ ( yˆ ) − 2∑ yˆ yˆ + ∑ ( yˆ ) + ∑ ( yˆ ) − 2∑ yˆ yˆ + ∑ ( yˆ qC ) + 2∑ ( yqC − yˆ qC )
p p p p p p p
C 2 C L L 2 U 2 U C 2 2
q q q q q i q
q =1 q =1 q =1 q =1 q =1 q =1 q =1
= ∑ ( yˆ qC − yˆ ) + ∑ ( yˆ ) + 2∑ ( yqC − yˆ )
p p p
L 2 C 2 C 2
q
U
q − yˆ q q
q =1 q =1 q =1
If we denote the regression sum of squares as SSR and the error sum of squares as
SSE, the total sum of squares of a fuzzy regression interval SST can be divided in-
to two parts as shown below.
SST = SSR + SSE (7.23)
where SSR and SSE, respectively, are given as:
SSR = ∑ ( yˆ qC − yˆ qL ) + ∑ ( yˆ qU − yˆ qC )
p p
2 2
(7.24)
q =1 q =1
and
SSE = 2∑ ( yˆ qC − yqC )
p
2
(7.25)
q =1
Hence, the total sum of squares (SST) measures the total variation of yqC between
lower and upper bounds. The error sum of squares (SSE) estimates the difference
when we use yˆ qC to estimate yqC , whereas regression sum of squares (SSR)
represents the variation of yˆ qC with respect to lower and upper bounds.
Similar to the determination of the coefficients R2 in statistical regression, the
index of confidence (IC) for evaluating a GFLSR model can be defined as given
below based on the concept of SST, SSR and SSE.
SSR SSE
IC = = 1− (7.26)
SST SST
Because 0 ≤ SSE ≤ SST , it follows that 0 ≤ IC ≤ 1 . Therefore, the higher the val-
ue of IC the better yˆ qC represents yqC .
7.4 An Illustrative Case: Packing Machine Design

To illustrate the GFLSR approach to generate functional models, product planning
of packing machines is described below.
7.4.1 Establishing a HOQ for Packing Machine Design

According to the lead users’ survey, four major customer requirements of packing-
machines were found to “improve the quality of packing dynamite” (CR1), “in-
crease efficiency of packing” (CR2), “reduce the packing noise” (CR3) and
“increase the rigidity of the machine” (CR4). Customer perception of the degree of
satisfaction of each customer requirement is scaled from 1 (worst) to 5 (best).
Based on the design team’s experience and their knowledge of the product,
seven engineering characteristics were identified: “improving the precision of the
molding of the clip” (EC1), “improving the precision of the packing” (EC2),
“increasing the control force of the packing” (EC3), “improving the efficiency of
the packing” (EC4), “increasing the hardness of the pressing hammer” (EC5), “re-
ducing the noise of the cam power transmission” (EC6), and “reducing the height
of the machine bed” (EC7). These engineering characteristics are measured in
units of mm-2, mm-2, kgf, um-1, HRC, dB and m respectively. Five major competi-
tive products, Comp1, Comp2, Comp3, Comp4 and Comp5 were identified as well.
Engineering performance data of the competitive products have been collected.
Based on the design team’s experience and engineering judgment, the relation-
ships between the customer requirements and the engineering characteristics, as
well as those among the engineering characteristics, were identified as well. The
relationships are represented by using the symbol “●” in the HOQ of packing ma-
chines as shown in Figure 7.1. The negative and positive signs on the engineering
characteristics as shown in the figure indicate the improvement of their target
values downwards and upwards respectively.
7.4.2 Normalizing Engineering Performance Values of

Engineering Characteristics
The level of attainment x j ( j = 1,2,", n ) can be obtained by normalizing the engi-

neering performance values of ECs. It involves changing, the current engineering
performance value of the ECj, l j ( j = 1,", n ) , to the level of attain-
ment x j ( j = 1,", n ) such that 0 ≤ x j ≤ 1 ( j = 1, " , n ) . There are two types of en-
gineering performance values: upward improvement and downward improvement.
These two types of engineering performance values can be normalized according
to the equations (7.27) and (7.28) which are for upward and downward
improvement respectively:
7.4 An Illustrative Case: Packing Machine Design 139
l mj ax − l j
xj = (7.27)
l mj ax − l mj in
l j − l mj in
xj = (7.28)
l mj ax − l mj in
where l max
j and l min
j can be determined by the considerations of competition
requirement and technology feasibility (Zhou, 1998). l max
j is the maximum value,
and l min
j is the minimum value obtainable.
Fig. 7.1 House of Quality of packing machines
Thus, the engineering performance values of EC3, EC4 and EC5 of the five
competitive products are normalized by using (7.27), while the engineering
performance values of EC1, EC2, EC6 and EC7 are normalized by using (7.28). The
following shows the results of the normalization.
x1 x2 x3 x4 x5 x6 x7
Comp1 ⎡ 0.44 0.63 0.20 0.75 0.75 0.38 0.20 ⎤
Comp 2 ⎢⎢ 0.78 0.75 0.67 0.63 0.50 0.55 0.60 ⎥⎥
X = Comp3 ⎢ 0.33 0.38 0.33 0.25 0.50 0.88 0.40 ⎥
⎢ ⎥
Comp 4 ⎢0.67 0.50 0.47 0.62 0.25 0.25 0.60 ⎥
Comp5 ⎢⎣ 0.56 0.25 0.67 0.38 0.75 0.50 0.80 ⎥⎦
7.4.3 Development of Functional Models Regarding QFD

Once the normalization of the engineering performance values is done, modeling the
functional relationships between the customer requirements and the engineering
characteristics as well as those among the engineering characteristics using GFLSR,
can be performed. To compare the modeling results based on GFLSR with those based
on fuzzy linear regression, fuzzy linear regression models were also developed using
the same data sets. The model relating CRi and ECj is used for the comparisons. In or-
der to examine how h affects the results, several values for h were selected and their
corresponding results based on fuzzy linear regression and GFLSR are summarized in
Table 7.2 and Table 7.3 respectively.
Table 7.2 Solutions based on fuzzy linear regression models
h=0 h=0.3 h=0.5 h=0.7 h=0.9

C
a 10
0.88 0.88 0.88 0.88 0.88
S
a 10
0.37 0.53 0.75 1.25 3.74
a 1C1 5.98 5.98 5.98 5.98 5.98
a 1S1 0.00 0.00 0.00 0.00 0.00
a 1C2 -0.33 -0.33 -0.33 -0.33 -0.33
a 1S2 0.00 0.00 0.00 0.00 0.00
a 1C3 -1.37 -1.37 -1.37 -1.37 -1.37
a 1S3 0.00 0.00 0.00 0.00 0.00
IC 0.50 0.67 0.80 0.92 0.99
Table 7.3 Solutions based on GFLSR models
h=0 h=0.3 h=0.5 h=0.7 h=0.9

C
a 10
1.29 1.29 1.29 1.29 1.29
S
a 10
0.44 0.62 0.87 1.45 4.36
a 1C1 5.83 5.83 5.83 5.83 5.83
a 1S1 0.00 0.00 0.00 0.00 0.00
a 1C2 -0.89 -0.89 -0.89 -0.89 -0.89
a 1S2 0.00 0.00 0.00 0.00 0.00
a 1C3 -1.62 -1.62 -1.62 -1.62 -1.62
a 1S3 0.00 0.00 0.00 0.00 0.00
IC 0.61 0.77 0.86 0.95 0.99
From Table 7.2 and Table 7.3, it can be seen that, in both cases, the changes of the
h value do not affect the center value of each A ij but influence the values of spread.
Also, the larger the value of h, the larger the value of IC is. This indicates that when a
7.4 An Illustrative Case: Packing Machine Design 141
larger value of h is chosen, the prediction capability of the models based on the
GFLSR or fuzzy linear regression with symmetric triangular coefficients would in-
crease. Further, when the same value of h is used, the prediction capability of the
GFLSR model is better than that of the fuzzy linear regression model.
When GFLSR is used to model functional relationships, the value of h is sub-
jectively pre-selected by a design team based on its engineering knowledge.
Tanaka and Watada (1988) suggested that selection of the h value should be based
on the sufficiency of the collected data set. When the number of data sets is
sufficiently large, h =0 should be used and increased along with the decreasing vo-
lume of the collected data. On the other hand, Moskowitz and Kim (1993) also
suggested that when we are pessimistic about the collected data sets, a large value
of h could be chosen. In fact, in a real-world environment, it is very common that
a developed HOQ contains only a small number of data sets. Therefore, a larger
value of h should be selected for modeling the functional relationships in QFD. In
the case of the HOQ of packing machines, as the number of data sets is small, the
value of h was set as 0.5. By using the two constrained programming models,
GFLSR-fi and GFLSR-gj , the fuzzy coefficients of f1 , f 2 , f 3 , f 4 and g 2 , g 4 , g 6
were determined as shown in Table 7.4. Because x1 , x3 , x5 and x7 do not relate to
other engineering characteristics (see Figure 1), g1 , g 3 , g 5 and g 7 do not exist.
Table 7.4 f i and g j based on GFLSR (h=0.5)
Intercept y1 y2 y3 y4
(1.29,0.87) (-0.14,0.00) (0.99,0.00) (1.25, 0.68)

x1 (5.83,0.00)
x2 (-0.89,0.00)
x3 (-1.62,0.00) (2.35,0.00)
x4 (1.74,1.90)
x5 (1.93,1.93)
x6 (4.12,1.87)
x7 (3.87,0.81)
IC 0.86 0.91 0.89 0.91
x2 x4 x6
(0.11,0.00) (0.48,0.00) (0.98,0.15)
x1
x2 (0.63,0.23)
x3
x4 (0.75,0.75) (-0.90,0.35)
x5
x6 (-0.53,0.00)
x7
IC 0.93 0.9000 0.87
7.5 Conclusion
This chapter describes a generalized fuzzy least-square regression (GFLSR)
approach to generate a functional model that accounts for the modeling that
involves both fuzzy type and crisp type observed data. The GFLSR integrates the
property of central tendency in least squares with the possibilistic property. Using
the GFLSR approach, functional models that relate customer satisfaction to engi-
neering characteristics, as well as among engineering characteristics, are translated
into two constrained programming models. By means of these two models, fuzzy
coefficients of the GFLSR models can be obtained. The proposed approach has
been applied to modeling the functional models for packing machines. Modeling
results based on the GFLSR are compared with those based on fuzzy linear regres-
sion. The comparisons indicate that the generalization capability of the GFLSR
models is better than that of the fuzzy linear regression models.
GFLSR is designed for addressing uncertainty due to fuzziness and randomness
when conducting modeling based on QFD. However, the models developed by
GFLSR contain only linear terms; nonlinear terms cannot be generated in the
model. Therefore, nonlinearity cannot be addressed by GFLSR. In Chapter 8
which follows, we discuss the neural fuzzy approach which is good for modeling
nonlinear systems, in order to generate functional models in relating customer
requirements to design attributes for product design.
References
Design, translated by Glenn Mazur. Productivity Press, Cambridge (1990)
Bai, H., Kwong, C.K.: Inexact genetic algorithm approach to target values setting of engi-
neering requirements in QFD. International Journal of Production Research 41, 3861–
3881 (2003)
Chang, Y.H.O.: Hybrid fuzzy least-squares regression analysis and its reliability measures.
search 42, 1009–1027 (2004)
nology 28, 11–12, 1175–1181 (2005)
D’Urso, P., Gastaldi, T.: A least-squares approach to fuzzy linear regression analysis.
Computational Statistics and Data Analysis 34, 427–440 (2000)
17–32 (1999)
Diamond, P.: Fuzzy least squares. Information Science 46, 141–157 (1998)
Fung, R.Y.K., Chen, Y., Tang, J., Tu, Y.: Estimating functional relationships for product
planning under uncertainties. Fuzzy sets and Systems 157, 98–120 (2006)
References 143
Fung, R.Y.K., Tang, J.F., Tu, P.Y.L., Chen, Y.Z.: Modeling of quality function deployment
planning with resource allocation. Research in Engineering Design 14, 247–255 (2003)
Hauser, J.R., Clausing, D.: The house of quality, pp. 63–73. Harvard Business Review
(1998)
Kim, K.J., Moskowitz, H., Dhingra, A., Evans, G.: Fuzzy multicriteria models for quality
function deployment. European Journal of Operational Research 121, 504–518 (2000)
approach to modeling functional relationships in QFD. Journal of Engineering De-
sign 21(5), 601–613 (2010)
Moskowitz, H., Kim, K.J.: On assessing the H value in fuzzy linear. Fuzzy Sets and Sys-
tems 58, 303–327 (1993)
Moskowitz, H., Kim, K.J.: QFD optimizer: a novice friendly quality function deployment
decision support system for optimizing product design. Computers and Industrial Engi-
neering 33, 641–655 (1997)
Park, T., Kim, K.J.: Determination of an optimal set of design requirements using house of
Reklaitis, G.V., Ravindran, A., Ragsdell, K.M.: Engineering optimization. John Wiley, NY
(1983)
Tanaka, H., Watada, J.: Fuzzy linear systems and their application to the linear regression
model. Fuzzy Sets and Systems 27, 275–289 (1988)
Tang, J., Fung, R.Y.K., Xu, B., Wang, D.: A new approach to quality function deployment
planning with financial consideration. Computer and Operations Research 29, 1447–
1463 (2002)
Wang, H.F., Tsaur, R.C.: Insight of a fuzzy regression model. Fuzzy Sets and Systems 112,
355–369 (2000)
Wassermann, G.S.: On how to prioritize design requirements during the QFD planning
process. IIE Transactions 25, 59–65 (1993)
Xu, R., Li, C.: Multidimensional least-squares fitting with a fuzzy model. Fuzzy Sets and
Systems 119, 215–223 (2001)
Yen, K.K., Ghoshary, S., Roig, G.: A linear regression model using triangular fuzzy num-
ber coefficients. Fuzzy Sets and Systems 106, 167–177 (1999)
Zhou, M.: Fuzzy logic and optimization models for implementing QFD. Computers and In-
dustrial Engineering 35, 237–240 (1998)
Chapter 8
An Enhanced Neuro-fuzzy Approach
for Generating Customer Satisfaction Models
8.1 Introduction
In this chapter, a new methodology for generating customer requirement models
using the approach of neural fuzzy networks is discussed. Non-linear and explicit
customer requirement models can be developed using this approach. Unlike stan-
dard neural network models, which are black-box in nature, explicit information
can be extracted from neural fuzzy network models which are explicit models.
The neural fuzzy networks approach is intended to overcome the limitations of the
fuzzy regression approach (discussed in Chapter 6 and Chapter 7) which cannot
address strong nonlinearity of customer requirements. It can also overcome the li-
mitations of the genetic programming approach (discussed in Chapter 5) which
cannot address the fuzzy nature of customer requirements. It consists of a set of
fuzzy rules which relate design attributes to customer requirements of new prod-
ucts. Therefore, explicit information can be extracted from rules within the cus-
tomer satisfaction models, which are generated based on the neural fuzzy network
approach. We discuss a rule extraction method for obtaining significant rules to
indicate the appropriate ranges of design attributes, in order to achieve reasonable
customer requirements in terms of new products. Based on these significant rules,
an explicit customer satisfaction model can be constructed. Customer perception
of a new product can be understood more easily with the generated customer satis-
faction model. An example of a notebook computer design is used to illustrate the
methodology. To examine the effectiveness of the proposed methodology, statis-
tical regression was the method against which the results for the new fuzzy ap-
proach were benchmarked. Experimental results show that the approach of neural
fuzzy networks outperforms statistical regression methods in terms of mean abso-
lute errors and variance of errors. Also, explicit information are more likely to be
extracted from the neural fuzzy networks.
8.2 An Enhanced Neural Fuzzy Network Approach

The neural fuzzy network is a multi-layer feed-forward network in which a neural
network is regarded as the learning mechanism and fuzzy reasoning is used for the
146 8 An Enhanced Neuro-fuzzy Approach
mapping of inputs into an output. Usually, the neural fuzzy network models con-
sist of a large number of fuzzy rules, so it is difficult to extract explicit informa-
tion from the neural fuzzy network models which can be defined as black box type
models. In order to generate non-linear and explicit customer requirement models,
a new methodology based on fuzzy membership is proposed. The methodology
consists of the following four steps: (a) Collection of market survey data from cus-
tomers, which involves the market survey design for customers and collection of
customers’ survey data is not within the scope of this chapter. The reader can refer
to Malhotra (2004); (b) Generating neural fuzzy network models based on the cus-
tomers’ survey data is discussed in Section 8.2.1; (c) Extraction of significant
fuzzy rules from neural fuzzy network model using a rule extraction method is
discussed in Section 8.2.2; and (d) Development of a customer satisfaction model
by aggregating the internal models of the extracted significant fuzzy rules is dis-
cussed in Section 8.2.3.
8.2.1 Development of Neural Fuzzy Network Models

The collected customers’ survey data can be used as the training data for the neur-
al fuzzy network. Figure 8.1 shows a neural fuzzy network with four layers and
two inputs is shown. For example, while the two inputs, x1 and x2, have two lin-
guistic descriptions, low and high, two membership functions are used to represent
the two linguistic descriptions. μ i (x1 ) and λ j ( x 2 ) denote the membership func-
tion for the i-th linguistic description of x1 and the j-th linguistic description of x2
respectively, where i = 1…2 and j = 1…2. 4 membership functions for all inputs
as defined by 4 nodes exist in Layer 1 (L1). At L2, one rule is used to denote the
outcome for each combination of x1 and x2, each of which represented by a mem-
bership function, so the number of rules is 4. The fuzzy rule is denoted as follows:
Rij : IF x1 is μi AND x 2 is λ j , THEN f ij = pij x1 + qij x2 + rij
where μi and λ j are the membership functions of i-th and j-th linguistic de-
scriptions of x1 and x2 respectively, pij, qij, and rij are the parameters of the internal
models (fij) of the fuzzy rules (Rij). For each Rij, the firing strength is defined by
(8.1).
wij = μ i (x1 ) ⋅ λ j ( x2 ) (8.1)
wij
and wij = (8.2)
W
where i, j = 1,2 and W = ∑∑ w
i j
ij
.
In (8.1), wij indicates the degree to which Rij is satisfied. The connection
weight between L2 and L3 is wij as defined by (8.2) which is the normalized fir-
ing strength. If the value of wij is larger, Rij is more significant. At L3, the
8.2 An Enhanced Neural Fuzzy Network Approach 147
internal functional model of Rij is a first-order Sugeno fuzzy model (Sugeno,

1985), which is defined by (8.3).
f ij = pij x1 + qij x1 + rij (∀i = 1,...,2, j = 1,...,2) (8.3)
The final output at L4 is computed based on the summation of all incoming sig-
nals, which is formulated by equation (8.4). y is the linear combination of all nor-
malized firing strengths and the corresponding internal models of all the fuzzy
rules in the fuzzy neural network.
∑∑ ∑∑ w ( p x + q x + r )
2 2 2 2
y= wij f ij = ij ij 1 ij 1 ij
(8.4)
i =1 j =1 i =1 j =1
The neural fuzzy network parameters (pij, qij, and rij) are determined based on a
learning algorithm which aims to minimize the error between y and the collected
training data. Detailed description of the neural fuzzy network architecture is not
the scope of this chapter. Readers could refer to for a more comprehensive review
of neural fuzzy networks in (Wang and Elhag 2008).
Based on Jang’s learning algorithm (Jang, 1993), the parameters (pij, qij, and rij)
can be determined such that the error between the neural fuzzy network output and
the training data can be minimized. First the values of the fuzzy network parame-
ters pij, qij, and rij are generated randomly. Based on the random fuzzy network pa-
rameters, y can be determined and the consequent fuzzy network parameters can
be identified by the least squares method. Then all error signals propagate back-
ward and the premise fuzzy network parameters are updated by the gradient des-
cent method.
Fig. 8.1 Layout of a neural fuzzy network

8.2.2 Extraction of Significant Fuzzy Rules and the

Corresponding Internal Models Using a Proposed
Rule Extraction Method
Based on the developed neural fuzzy network, a set of fuzzy rules, which consist
of both significant and insignificant fuzzy rules, can be extracted. If both signifi-
cant and insignificant rules are included in the neural fuzzy network model, the
operations of the functional models are inefficient. The final output y is relatively
independent of insignificant rules compared with significant rules. If the insignifi-
cant fuzzy rules are ignored and only significant fuzzy rules are considered, the fi-
nal output can be an approximation to the original value, and the effectiveness of
the neural fuzzy network can be enhanced. Explicit information can be extracted
more easily from the neural fuzzy network, as fewer fuzzy rules are included.
A rule extraction method is discussed to distinguish significant fuzzy rules from
the whole set of fuzzy rules. It is intended to determine the active range of inputs
if the associated membership values are greater than or equal to a threshold value
( Φ ), i.e. μ (x1 ) ≥ Φ . The value of Φ is between 0 and 1 which is the range of the
membership values. If Φ is too small, all fuzzy rules are executed, which is un-
desirable. If Φ is too large, the number of extracted fuzzy rules is too small and it
may have no response for some data ranges. A validation check is required to
ensure the entire input range can be covered. If the validation check is not satisfac-
tory, a new Φ is chosen. It is relatively straightforward to adjust the threshold
value ( Φ ) and check whether the condition is valid or not. For most of the cases,
Φ = 0.01 must be good enough.
Based on the example in Section 8.2.1, the following algorithm shows the logic
in determining the active range of each linguistic description of x1, and the asso-
ciated membership values are denoted as:
μ1 (x1 ) ≥ Φ for R1,min ≤ x1 < R1,max
μ 2 (x1 ) ≥ Φ for R2 ,min ≤ x1 < R2 ,max
where Ri , min and Ri , max denote the minimum and the maximum values of the
fuzzy rule set Ri respectively for i = 1, 2.
For i = 1 to 2 //Total linguistic descriptions of x1

Count = 0;
For x1 = LB to UB //Range of x1 (Lower Bound to Upper Bound)
If µ i(x1) ≥ Ф //Threshold check
Count++; //Add one more entry to set R
Ri(Count) = x1; //A set R contains the active range of i-th linguis-
tic description of x1
End
End
End
8.2 An Enhanced Neural Fuzzy Network Approach 149
In this case, if R1,max ≥ R2, min , the whole range of x1 can be covered such that
R1,min = lower bound of x1 and R2 ,max = upper bound of x1 . Otherwise, a new Φ
is required. Next, the active membership function(s) for all possible input ranges
of x1 can be decided as shown in Fig. 8.2. From the figure, there are three possi-
ble input ranges of x1 : R1,min ≤ x1 ≤ R2 ,min , R2 ,min ≤ x1 ≤ R1,max , and
R1,max ≤ x1 ≤ R2, max . As overlap between the first input range and the active range of
μ1 (x1 ) exists, μ1 (x1 ) can be assumed to be active. Similarly, μ1 (x1 ) and μ 2 (x1 )
are assumed to be active, as overlap between the second input range and the active
ranges of both μ1 (x1 ) and μ 2 (x1 ) exists. μ 2 (x1 ) is active only in the last input
range. Suppose a pre-defined range of x1 is input, the associated fuzzy rule ( Rij )
is defined as significant, if the calculation of the firing strength ( wij ) involves the
active membership functions as defined by (8.1). The procedure can be applied to
all inputs. After the rule extraction, the customer satisfaction models consist of on-
ly the significant fuzzy rules.
R1,MIN R2,MIN R1,MAX R2,MAX
μ1 ( x1 )
μ2 ( x1 )
Input
ranges
First Second Third
Fig. 8.2 Membership functions of x1
8.2.3 Formulation of the Customer Satisfaction Models by

Aggregating Internal Models of Significant Fuzzy Rules
The formulation of wij needs to be derived, in order to generate non-linear and
explicit customer satisfaction models. Membership functions of all inputs are ge-
neralized based on a bell-shaped function. The example with two inputs discussed
in Section 8.2.1 is considered. Membership functions of different linguistic de-
scriptions of x1 and x2 are denoted as follows:
μ i (x1 ) =
1
2 bi
(8.5)
x −c
1+ 1 i
ai
λ j (x2 ) =
1
2t j
(8.6)
x − ui
1+ 2
sj
where i, j =1, 2. Based on (8.1), wij can be determined by:
ai2 bi ⋅ s j
2t j
wij =
− ci ) i + ai2bi (x2 − u j ) (x −uj )
(8.7)
ai2 bi ⋅ s j j + s j
2t 2t j
(x
1
2b 2t j
+ (x1 − ci )
2 bi
2
2t j
Based on the proposed rule extraction method, significant fuzzy rules and the cor-
responding internal models can be identified. Customer satisfaction models can be
obtained by integrating internal models of significant fuzzy rules as defined by
(8.4). A case study of a notebook computer is presented in the following section to
illustrate the proposed methodology for rule extraction in neural fuzzy networks
and generation of customer satisfaction models.
8.3 Case Study: Notebook Computer
The effectiveness of the proposed methodology is demonstrated by developing

customer satisfaction models based on customers’ survey data for notebook com-
puter design. To develop the customer satisfaction models, a market survey was
conducted among notebook computer users. Eighty completed questionnaires
were returned. The developed customer satisfaction models can be used to identify
the customer perception towards new consumer products. In the survey, the cus-
tomers were asked to rate six different brands of notebook computers with respect
to five customer requirements: quality, performance, user-friendliness, comfort-to-
carry, and attractiveness. For ease of presentation, the computers are denoted as
brands A – F, images of which are shown in Figure 8.3. The design attributes re-
garding the selected computers were: CPU, RAM, hard disk memory, LCD screen
size, battery life, weight, and price. A 5-point scale (linguistic description) was
used to assess the computers against five different customer attributes: 1-Very
Bad, 2-Bad, 3-Moderate, 4-Good, and 5-Very Good. A summary of the computers
is given in Table 8.1. While the customers’ survey data is available, the proposed
methodology can be implemented to generate the customer satisfaction models.
8.3 Case Study: Notebook Computer 151
Fig. 8.3(a) Computer A Fig. 8.3(b) Computer B
Fig. 8.3(c) Computer C Fig. 8.3(d) Computer D
Fig. 8.3(e) Computer E Fig. 8.3(f) Computer F

The development of customer satisfaction models regarding the customer re-

quirement“comfortable-to-carry” is described for illustrative purposes. The two
most highly relevant design attributes, which are related to “comfortable-to-
carry”, were identified as “LCD screen size (in inches)” and “weight (in kg)”. In
this connection, the two inputs of the model are x1 (LCD screen size) and x2
(weight). The output of the neural fuzzy network model is y, which represents the
customer requirement of “comfortable-to-carry”. In the customers’ survey, the
range of x1 is from 13.3 to 15.4 inches and the range of x2 is from 1.93 to 2.9
kg. First, based on the customers’ survey data, the neural fuzzy network was im-
plemented by the MATLAB software package. The membership is represented by
the bell-shaped function. x1 consists of 3 linguistic descriptions: Small (S), Me-
dium (M), and Large (L). x2 consists of 6 linguistic descriptions: Very Light
(VL), Light (L), Slightly Light (SL), Slightly Heavy (SH), Heavy (H), and Very
Heavy (VH). The membership functions of x1 and x2 are shown in Figure 4 and
Figure 5 respectively.
Table 8.1 Questionnaire for the selected notebook computers
Pro- Design attributes Comfor-

file CPU RAM Hard LCD Battery Weight Price able-to-
disk screen life (kg) carry
memory size (1 to 5)
A Intel 16GB 750 GB 13.3 9-Cell 1.90 1299 3

core inch Primary
i3/i5/i7 Battery
B Intel 2.80 160 GB 15.4 6-Cell 2.20 1599 4
Core 2 GHz inch Primary
Duo Battery
T9600
C Intel 2GHz 320 GB 15.6 6-Cell 0.82 849 5
Core i3 Primary
Battery
D Intel 4GB 320GB 13.3 9-Cell 17.3 1070 3
Core i5- Primary
560 Battery
E Intel 16GB 750GB 17.3 9-Cell 17.3 1499 5
core i7- Primary
2530 Battery
F Intel 2GB 250GB 13.3 6-Cell 2.4 1538 3
Core i3 Primary
350M Battery
After the implementation, 18 fuzzy rules ( Rij ) with i = 1, 2, 3 and j = 1, 2, …6,

were generated as:
R11 : IF x1 is μ1 AND x2 is λ1 , THEN f11 = 0.3259 x1 + 0.04729 x2 + 0.0245
R12 : IF x1 is μ1 AND x2 is λ2 , THEN f12 = 0.01859 x1 + 0.002721x2 + 0.001394
R14 : IF x1 is μ1 AND x2 is λ4 , THEN f14 = 0.02089 x1 + 0.003501x2 + 0.001473
R15 : IF x1 is μ1 AND x2 is λ5 , THEN f15 = 0.000705x1 + 7.7e-05 x2 + 5.26e-05
R16 : IF x1 is μ1 AND x2 is λ6 , THEN f16 = 0.001131x1 + 0.000249 x2 + 7.42e-05
R21 : IF x1 is μ2 AND x2 is λ1 , THEN f 21 = 0.01966 x1 + 0.00287 x2 + 0.001475
R23 : IF x1 is μ2 AND x2 is λ3 , THEN f 23 = 0.3159 x1 + 0.05348x2 + 0.0224
R25 : IF x1 is μ2 AND x2 is λ5 , THEN f 25 = 0.000558 x1 − 0.00046 x2 + 6.56e-05

R31 : IF x1 is μ3 AND x2 is λ1 , THEN f31 = 0.001501x1 + 0.000225 x2 + 0.0001106
R32 : IF x1 is μ3 AND x2 is λ2 , THEN f 32 = 0.000873x1 + 0.00015 x2 + 5.83e-05
R34 : IF x1 is μ3 AND x2 is λ4 , THEN f 34 = 0.5612 x1 + 0.09723x2 + 0.03646
R35 : IF x1 is μ3 AND x2 is λ5 , THEN f 35 = −0.0749 x1 − 0.02226 x2 − 0.004863
R36 : IF x1 is μ3 AND x2 is λ6 , THEN f36 = 0.2691x1 + 0.06077 x2 + 0.01748
The next step is to distinguish significant fuzzy rules by the rule extraction method
described in Section 8.2.2. x1 (LCD screen size) has 3 linguistic descriptions. For
each description, the active range of the membership function can be determined if
Φ = 0.01 for i = 1…3. Three expressions can be extracted as:
μ1 (x1 ) ≥ 0.01 for 13.3 ≤ x1 < 14.9

μ 2 (x1 ) ≥ 0.01 for 13.3 ≤ x1 < 15.4
μ 3 (x1 ) ≥ 0.01 for 13.8 ≤ x1 < 15.4
A validation check is then required to ensure the entire input range can be cov-
ered. As all of the above 3 membership functions can cover the whole input range
(13.3 – 15.4), the legitimate condition with μ1 (x1 ) ≥ 0.01 is met. Otherwise, a
new Φ is chosen. Next, the active membership function(s) for all possible input
ranges can be decided. As depicted in Section 8.2.2, we can obtain 3 different in-
put ranges of x1 : 13.3 ≤ x1 ≤ 13.8, 13.8 < x1 ≤ 14.9, and 14.9 < x1 ≤ 15.4.
μ1 (x1 ) and μ 2 ( x1 ) are active in the first input range. All membership functions
are active in the second input range. μ 2 ( x1 ) and μ 3 (x1 ) are active in the third in-
put range.
The same procedure can be carried out for x 2 (weight). Six expressions can be
extracted as:
λ1 (x2 ) ≥ 0.01 for 1.93 ≤ x2 < 2.23
λ2 ( x2 ) ≥ 0.01 for 1.93 ≤ x2 < 2.42
λ3 (x2 ) ≥ 0.01 for 2.01 ≤ x 2 < 2.64
λ4 ( x2 ) ≥ 0.01 for 2.1 ≤ x2 < 2.9
λ5 ( x2 ) ≥ 0.01 for 2.35 ≤ x2 < 2.9
λ6 (x2 ) ≥ 0.01 for 2.6 ≤ x2 < 2.9
The relationship between the input ranges and the active membership functions is
summarized in Table 8.2. The table shows that these membership functions are not
always active for the entire input ranges.
Table 8.2 The relationship between the input ranges and the active MFs
Input range μ1(x1) μ2(x1) μ3(x1)

13.3≤ x1≤13.8 V V
13.8< x1≤14.9 V V V
14.9< x1≤15.4 V V
Input range λ1(x2) λ2(x2) λ3(x2) λ4(x2) λ5(x2) λ6(x2)

1.93≤ x2≤2.01 V V
2.01< x2≤2.1 V V V
2.1< x2≤2.23 V V V V
2.23< x2≤2.35 V V V
2.35< x2≤2.42 V V V V
2.42< x2≤2.6 V V V
2.6< x2≤2.64 V V V V
2.64< x2≤2.9 V V V
*Note: V - Active.
Based on this information, significant fuzzy rules can be identified for different
input ranges. If the ranges of the inputs are 13.3 ≤ x1 ≤ 13.8 and 1.93 ≤ x2 ≤ 2.01,
only μ1 ( x1 ) and μ2 ( x1 ) are active for x1 and only λ1 ( x2 ) and λ2 ( x2 ) are ac-
tive for x2 (from Table 1). Therefore the customer satisfaction model of “comfort-
able-to-carry” can be formulated by including only the significant fuzzy rules as
follows:
∑∑ w ⋅ f
2 2
y= ij ij
= (w11 ⋅ f 11 + w12 ⋅ f12 + w21 ⋅ f 21 + w22 ⋅ f 22 ) = F (x1 , x2 )
i =1 j =1
∑∑ w
2 2
wij
where wij = ,W = ij
,
W i =1 j =1
y refers to the customer requirement of “comfortable-to-carry”, x1 is “LCD screen

size”, and x2 is “weight”. The customer satisfaction model contains 4 terms and it
also means that only 4 significant rules are considered. The formulation of wij (i =
1…3 and j = 1…6) can be described as Table 8.3.
Table 8.3 Formulation of wij
(0.0751)(8.8128e-05)
w11 =
(0.0751)(8.8128e-05) + 8.8128e-05( x1 − 13.3) 4 + 0.0751(x2 -1.93) 4 + ( x1 − 13.3) 4 ( x2 − 1.93) 4
(0.0751)(8.1451e-05)
w12 =
(0.0751)(8.1451e-05) + 8.1451e-05( x1 − 13.3) 4 + 0.0751(x2 -2.124) 4 + ( x1 − 13.3) 4 ( x2 − 2.124) 4
(0.0751)(1.0530e-04)
w13 =
(0.0751)(1.0530e-04) + 1.0530e-04( x1 − 13.3) 4 + 0.0751(x2 -2.323) 4 + ( x1 − 13.3) 4 ( x2 − 2.323) 4
(0.0751)(4.7462e-04)
w14 =
(0.0751)(4.7462e-04) + 4.7462e-04( x1 − 13.3) 4 + 0.0751(x2 -2.556) 4 + ( x1 − 13.3) 4 ( x2 − 2.556) 4
(0.0751)(2.7607e-04)
w15 =
(0.0751)(2.7607e-04) + 2.7607e-04( x1 − 13.3) 4 + 0.0751(x2 -2.747) 4 + ( x1 − 13.3) 4 ( x2 − 2.747) 4
(0.0751)(8.9630e-05)
w16 =
(0.0751)(8.9630e-05) + 8.9630e-05( x1 − 13.3) 4 + 0.0751(x2 -2.9) 4 + ( x1 − 13.3) 4 ( x2 − 2.9) 4
(0.0749)(8.8128e-05)
w21 =
(0.0749)(8.8128e-05) + 8.8128e-05( x1 − 14.35) 4 + 0.0749(x2 -1.93) 4 + ( x1 − 14.35) 4 ( x2 − 1.93) 4
(0.0749)(8.1451e-05)
w22 =
(0.0749)(8.1451e-05) + 8.1451e-05( x1 − 14.35) 4 + 0.0749(x2 -2.124) 4 + ( x1 − 14.35) 4 ( x2 − 2.124) 4
(0.0749)(1.0530e-04)
w23 =
(0.0749)(1.0530e-04) + 1.0530e-04( x1 − 14.35) 4 + 0.0749(x2 -2.323) 4 + ( x1 − 14.35)4 ( x2 − 2.323) 4
(0.0749)(4.7462e-04)
w24 =
(0.0749)(4.7462e-04) + 4.7462e-04( x1 − 14.35)4 + 0.0749(x2 -2.556) 4 + ( x1 − 14.35) 4 ( x2 − 2.556) 4
(0.0749)(2.7607e-04)
w25 =
(0.0749)(2.7607e-04) + 2.7607e-04( x1 − 14.35)4 + 0.0749(x2 -2.747)4 + ( x1 − 14.35)4 ( x2 − 2.747)4
(0.0749)(8.9630e-05)
w26 =
(0.0749)(8.9630e-05) + 8.9630e-05( x1 − 14.35) 4 + 0.0749(x2 -2.9) 4 + ( x1 − 14.35) 4 ( x2 − 2.9) 4
(0.0761)(8.8128e-05)
w31 =
(0.0761)(8.8128e-05) + 8.8128e-05( x1 − 15.4) 4 + 0.0761(x2 -1.93) 4 + ( x1 − 15.4) 4 ( x2 − 1.93)4
(0.0761)(8.1451e-05)
w32 =
(0.0761)(8.1451e-05) + 8.1451e-05( x1 − 15.4) 4 + 0.0761(x2 -2.124) 4 + ( x1 − 15.4) 4 ( x2 − 2.124)4
(0.0761)(1.0530e-04)
w33 =
(0.0761)(1.0530e-04) + 1.0530e-04( x1 − 15.4) 4 + 0.0761(x2 -2.323) 4 + ( x1 − 15.4) 4 ( x2 − 2.323) 4
(0.0761)(4.7462e-04)
w34 =
(0.0761)(4.7462e-04) + 4.7462e-04( x1 − 15.4) 4 + 0.0761(x2 -2.556) 4 + ( x1 − 15.4) 4 ( x2 − 2.556) 4
(0.0761)(2.7607e-04)
w35 =
(0.0761)(2.7607e-04) + 2.7607e-04( x1 − 15.4) 4 + 0.0761(x2 -2.747) 4 + ( x1 − 15.4) 4 ( x2 − 2.747) 4
(0.0761)(8.9630e-05)
w36 =
(0.0761)(8.9630e-05) + 8.9630e-05( x1 − 15.4) 4 + 0.0761(x2 -2.9) 4 + ( x1 − 15.4) 4 ( x2 − 2.9) 4
It is noted that the above calculation of W includes only the wij of significant
rules. Moreover, the customer satisfaction model is a function of x1 and x2 in
which the non-linear relationship between customer requirement, y, and the two
design attributes, x1 and x2, can be explicitly depicted. For 13.3 ≤ x1 ≤ 13.8 and
1.93 ≤ x2 ≤ 2.01, 4 significant fuzzy rules of the customer satisfaction model and
the associated network model are shown in Figures 8.6(a) and 8.6(b) respectively.
If 13.5 ≤ x1 ≤ 14.5 and 2.0 ≤ x2 ≤ 2.2 are input, the customer satisfaction model can
be formulated as follows:
y = y1 ∪ y 2 ∪ y3 ∪ y 4 ∪ y 5 ∪ y 6
where
∑∑ w ⋅ f
2 2
y1 = ij ij
for 13.3 ≤ x1 ≤ 13.8 and 1.93 ≤ x2 ≤ 2.01
i =1 j =1
∑∑ w ⋅ f
2 3
y2 = ij ij
for 13.3 ≤ x1 ≤ 13.8 and 2.01 < x2 ≤ 2.1
i =1 j =1
∑∑ w ⋅ f
2 4
y3 = ij ij
for 13.3 ≤ x1 ≤ 13.8 and 2.1 < x2 ≤ 2.23
i =1 j =1
∑∑ w ⋅ f
3 2
y4 = ij ij
for 13.8 < x1 ≤ 14.9 and 1.93 ≤ x2 ≤ 2.01
i =1 j =1
∑∑ w ⋅ f
3 3
y5 = ij ij
for 13.8 < x1 ≤ 14.9 and 2.01 < x2 ≤ 2.1
i =1 j =1
∑∑ w ⋅ f
3 4
y6 = ij ij
for 13.8 < x1 ≤ 14.9 and 2.1 < x2 ≤ 2.23
i =1 j =1
Therefore, we have:
∑∑ w ⋅ f
3 4
y= ij ij = (w11 ⋅ f11 + w12 ⋅ f12 + w21 ⋅ f 21 + w34 ⋅ f 34 ) = F (x1 , x 2 )

i =1 j =1
∑∑ w
3 4
wij
where wij = ,W = ij
W i =1 j =1
Fig. 8.6(a) Four significant fuzzy rules of the customer satisfaction model
Fig. 8.6(b) Associated network model
In this section, two validation tests were performed for the proposed rule ex-
traction method and the proposed methodology respectively. For the first test, 10
datasets were used to validate the proposed rule extraction method. Table 8.4 re-
ports the validation results. Column “Y” shows the model outputs based on neural
fuzzy networks without rule extraction while column “Y+RE” shows the model
outputs based on neural fuzzy networks with rule extraction. Column “% Δ 1” re-
ports the differences (in percentage) of model outputs between them. Column
“% Δ 2” reports the reductions (in percentage) of fuzzy rules executed by neural
fuzzy networks with rule extraction over neural fuzzy networks without rule ex-
traction. From the table, it can be seen that the differences of model outputs be-
tween neural fuzzy networks with and without rule extraction are very small. The
average deviation is 0.94%, but the number of fuzzy rules considered for generat-
ing the customer requirement models is reduced by an average of 60.56%. This
makes the process of model development more effective and the generation of
customer satisfaction models of customer requirement is less complex.
The second test was performed to examine the performance of the customer sa-
tisfaction models generated by the proposed methodology and compare this with
the customer satisfaction models generated by the statistical regression method.
Using the same market survey data, the customer satisfaction model based on the
statistical regression method was generated as follows:
y = 31.7592 − 1.8307 x1 − 11.474 x2 + 0.7289 x1 x2

Table 8.4 Justification results of the simplified neural fuzzy networks
Test data Y Y+RE %Δ1 % Δ2

1 3.0541 3.1029 1.59 77.78
2 3.2356 3.2644 0.89 50
3 2.5955 2.6145 0.73 50
4 2.5481 2.5878 1.56 77.78
5 2.1534 2.2005 2.19 66.67
6 3.2926 3.2999 0.22 33.33
7 3.1078 3.1123 0.14 50
8 1.8635 1.8806 0.92 66.67
9 1.0305 1.0373 0.66 66.67
10 1.1387 1.1445 0.51 66.67
Mean - - 0.94 60.56
Then, the actual values of two design attributes (LCD screen size and weight)
of the six different brands of computers were input to the customer satisfaction
models generated by the proposed methodology and statistical regression method
respectively. Table 8.5 shows the comparison results. Column “Average” reports
the average rating of satisfaction of the customer requirement towards “comfort-
to-carry” for different brands. Column “P_Method” reports the differences (in
percentage) between the outputs of the customer satisfaction model generated by
the proposed methodology and the average rating of satisfaction of customer re-
quirement. Column “SRM” reports the same between the outputs of the customer
satisfaction model generated by statistical regression method and the average rat-
ing of satisfaction of the customer requirement. The last column shows the im-
provement achieved by the proposed methodology over the statistical regression
method. If the value is positive, it means that the customer satisfaction models
based on the proposed methodology perform better. Otherwise, it is negative. To
statistically compare the two methods, Mean Errors (ME) and Variance of Errors
(VoE) are adopted as defined by (8.5) and (8.6) respectively.
y hA − y hM
∑
n
1
ME = ⋅ 100 (8.5)
n h =1 y hA
2
⎛ y hA − y hM ⎞
∑
n
1 ⎜
VoE = ⋅ 100 − ME ⎟ (8.6)
n −1 ⎜ A
yh ⎟
h =1
⎝ ⎠
A M
where n is the total number of datasets (in this case, n = 6), yh and yh are the
Actual values (from survey data) and the model outputs (the proposed approach or
statistical regression) for h-th dataset respectively. It is noted that the smaller the
values of ME and VoE, the more accurate the model is. For the proposed
methodology, all output errors are within 1% and ME = 0.45. For statistical re-
gression, ME = 1.73 and some errors are quite significant such as 3.32% and
2.52% for brand C and brand E computers respectively. But, some of its errors are
very small such as 0.02% for brand F computer. Although the statistical regression
method outperforms the customer satisfaction models that are based on the pro-
posed methodology for brands B and F computers, the average improvement of
customer satisfaction models generated using the proposed methodology over sta-
tistical regression remains positive, i.e. 1.28%. Also, the variation of errors for
statistical regression is quite large (VoE = 1.588). In contrast, the customer satis-
faction model based on the proposed methodology produces a very small variation
(VoE = 0.0476) and the errors are relatively insignificant (ME = 0.45).
Table 8.5 Customer satisfaction models generated based on the proposed methodology and
statistical regression
Test data Average P_Method SRM Δ (SRM-

(Brand) (%Δ) (%Δ) P_Method)
A 3.1372 0.21 1.99 1.78
B 2.9122 0.74 0.44 -0.3
C 2.7555 0.44 3.32 2.88
D 2.9995 0.21 2.11 1.9
E 2.9181 0.44 2.52 2.08
F 3.9753 0.65 0.02 -0.63
ME - 0.45 1.73 1.28
VoE - 0.0476 1.588 -
8.4 Conclusion
Customer satisfaction models play an important role in identifying the customer
perception towards consumer products. With the models, the relationships be-
tween design attributes and customer requirements could be formulated explicitly.
This chapter discussed an innovative neural fuzzy network approach to generate
Customer satisfaction models. The results of the second test indicates that the
models developed based on the proposed methodology outperforms the models
developed based on statistical regression in terms of mean absolute errors and va-
riance of errors. In addition, the models developed based on the proposed metho-
dology are explicit and are capable of modeling the non-linear relationships
between customer requirements and design attributes for new products.
This approach overcomes the limitations of the existing neural fuzzy network
approaches, which can generate only black-box models and no explicit informa-
tion can be extracted within the models generated by such approaches. Unlike
standard neural network models, with this innovative neural fuzzy network ap-
proach, explicit information can be extracted. This approach also overcomes the
limitations of the fuzzy regression approach (discussed in Chapter 6 and Chapter
7) which cannot address strong nonlinearity of the relationships between customer
requirements and design attributes. It can also overcome the limitations of the
References 161
genetic programming approach (discussed in Chapter 5) which cannot address the

fuzzy nature of the relationships. The next two chapters will discuss how to use
the developed customer satisfaction models to determine the optimal design
attribute settings of new products.
References
Bateni, S.M., Borghei, S.M., Jeng, D.S.: Neural network and neuro-fuzzy assessments for
scour depth around bridge piers. Engineering Applications of Artificial Intelligence 20,
401–414 (2007)
Chang, F.J., Chang, Y.T.: Adaptive neuro-fuzzy inference system for prediction of water
level in reservoir. Advances in Water Resources 29, 1–10 (2006)
349 (2006)
nology 28, 1175–1181 (2005)
search 42(5), 1009–1027 (2004)
Choi, K.S., Cho, W.H., Lee, S., Lee, H., Kim, C.: The relationships among quality, value,
requirement and behavioral intention in health care provider choice: A South Korean
study. Journal of Business Research 57(8), 913–921 (2004)
17–32 (1999)
Deng, W.J., Pei, W.: Fuzzy neural based importance-performance analysis for determining
critical service attributes. Expert Systems with Applications (in press),
doi:10.1016/j.eswa.2008.02.063
Fung, R.Y.K., Popplewell, K., Xie, J.: An intelligent hybrid system for customer require-
ments analysis and product attribute targets determination. International Journal of Pro-
duction Research 36, 13–34 (1998)
tomer requirement: The MUSA method. European Journal of Operational Research 143,
148–170 (2002)
Grigoroudis, E., Litos, C., Moustakis, V.A., Politis, Y., Tsironis, L.: The assessment of us-
er-perceived web quality: Application of a requirement benchmarking approach. Euro-
Han, S.H., Yun, M.H., Kim, K.J., Kwahk, J.: Evaluation of product usability: development
Hansemark, O.C., Albinsson, M.: Customer requirement and retention: The experiences of
individual employees. Managing Service Quality 14(1), 40–57 (2004)
ics 35, 411–428 (2005)
Huang, M.J., Tsou, Y.L., Lee, S.C.: Integrating fuzzy data mining and fuzzy artificial neur-
al networks for discovering implicit knowledge. Knowledge-Based Systems 19, 396–
403 (2006)
Jang, J.S.R.: ANFIS: Adapative-network-based fuzzy inference systems. IEEE Transactions
on Systems, Man and Cybernetics 23, 665–685 (1993)
Johnson, M.D., Nader, G., Fornell, C.: Expectations, perceived performance, and customer
requirement for a complex service: The case of bank loans. Journal of Economic Psy-
chology 17(2), 163–182 (1996)
Kelemen, A., Kozma, R., Liang, Y.: Neuro-fuzzy classification for the job assignment
problem. In: Proceedings of the IEEE International Joint Conference on Neural Net-
works, Honolulu, Hawaii, pp. 1831–1837 (2002)
Kim, K.J., Moskowitz, H., Dhingra, A., Evans, G.: Fuzzy multicriteria models for quality
Kwong, C.K., Chen, Y., Bai, H., Chan, D.S.K.: A methodology of determining aggregated
importance of engineering characteristics in QFD. Computers & Industrial Engineer-
ing 53, 667–679 (2007)
Kwong, C.K., Chan, K.Y., Wong, H.: Takagi-Sugeno neural fuzzy modeling approach to
fluid dispensing for electronic packaging. Expert Systems with Applications 34, 2111–
2119 (2008)
Kwong, C.K., Wong, T.C., Chan, K.Y.: A methodology of generating customer require-
ment models for new product development using a neuro-fuzzy approach. Expert Sys-
tems with Applications 36(8), 11262–11270 (2009)
logic: A case study of mobile phones. International Journal of Industrial Ergonomics 37,
531–543 (2007)
Liu, X., Zeng, X., Xu, Y., Koehl, L.: A fuzzy model for customer requirement index in e-
commerce. Mathematics and Computers in Simulation 77, 512–521 (2008)
Malhotra, N.K.: Marketing research: an applied orientation, 4th edn. Pearson, Upper Saddle
River (2004)
Myers, R.H.: Classical and modern regression with application, 2nd edn. PWS-KENT, Bos-
ton (1990)
Park, J., Han, S.H.: A fuzzy rule-based approach to modeling affective user requirement
towards office chair design. International Journal of Industrial Ergonomics 34, 31–47
(2004)
Sugeno, M.: Industrial applications of fuzzy control. Elsevier, Amsterdam (1985)
Wang, Y.M., Elhag, T.M.S.: An adaptive neuro-fuzzy inference system for bridge risk as-
sessment. Expert Systems with Applications 34, 3099–3106 (2008)
You, H., Ryu, T., Oh, K., Yun, M.H., Kim, K.J.: Development of customer requirement
models for automotive interior materials. International Journal of Industrial Ergonom-
ics 36, 323–330 (2006)
Zeithaml, V.A.: Service quality, profitability, and the economic worth of customers: What
we know and what we need to learn. Journal of the Academy of Marketing
Science 28(1), 67–85 (2000)
Chapter 9
Optimization of Customer Satisfaction
Using an Improved Simulation Annealing
9.1 Introduction
Chapters 3 and 4 discussed using the fuzzy AHP approaches to determine the im-
portance weights of customer requirements of new product designs. Chapters 5 to
8 discussed using fuzzy and evolutionary methods to generate models which
represent relationships between customer requirements and the design attributes of
new products. Based on the models and the importance weights for customer re-
quirements, the optimization problems for maximizing overall customer satisfac-
tion for the new products can be formulated. However, nonlinearity exists between
customer requirements and design attributes of new products. Therefore, these op-
timization problems have multiple optima arising from local optima, and cannot
be handled by classical optimization methods such as gradient-based methods.
This chapter discusses a computational intelligence optimization method, namely
simulated annealing (SA), to solve these multi-optima problems for new product
design.
SA is a point-based stochastic optimization method, which explores the search
space through iterations from an initial solution to the optimum (Cerny 1985,
Kirkpatrick 1983). It has been widely used in solving many difficult optimization
problems (Laarhoven 1992, Laarhoven1987). Each iteration employs a neighbour-
hood function to generate a candidate solution by a randomized perturbation on a
current solution. The design of neighbourhood functions plays an important role in
developing an effective simulated annealing approach. Moreover, the searching
mechanism of SA has a very good convergent property (Locatelli 2000). Howev-
er, it can be noted in previous research (Aydin et al. 2004, Lin et al. 1995, Ruiz-
Torres 1997) that while SA can find good or reasonable solutions, in many cases it
cannot search for a global optimum.
In this chapter, enhanced simulated annealing algorithms (Ho et al. 2006, Chan
et al. 2010) are developed based on a neighbourhood function which is integrated
with orthogonal design for maximizing customer satisfaction in new product de-
sign. Potential solutions regarding customer satisfaction are generated based on
the combinations of the orthogonal array. In these neighbourhood functions, or-
thogonal arrays exploit the neighbourhood of a current solution by analyzing the
main effects of design attributes (Ho et al. 2006) and the effects of interactions
164 9 Optimization of Customer Satisfaction
between design attributes (Chan et al. 2010). The neighbourhood function, which
uses the orthogonal array for exploiting solution spaces, is called an orthogonal ar-
ray-based neighbourhood function (ONF). The effectiveness of using the en-
hanced simulated annealing algorithms for product design is demonstrated using
the optimization of an emulsified dynamite packing-machine design.
9.2 Development of Neighbourhood Function

Based on Orthogonal Experimental Design
for Product Design Purposes
This section discusses how the neighhourhood functions used in simulated anneal-
ing can be enhanced by integrating them with orthogonal arrays for maximizing
customer satisfaction with product design. Section 9.2.1 shows the orthogonal ar-
ray based neighbourhood function which is developed by (Ho et al. 2004a, b,
2006, Shu et al. 2004) and Section 9.2.2 shows an improved version developed by
(Chan et al. 2010).
9.2.1 Orthogonal Array Based Neighbourhood Function (ONF)

Ho et al. (Ho et al. 2004a,b ,2006, Shu et al. 2004) proposed an orthogonal array-
based neighbourhood function (ONF) which is intended to generate a candidate
solution by using the combinations of the orthogonal array, L2 N +1 (3 N ) . It can be
used on simulated annealing for maximizing customer satisfaction by optimizing
the design attributes of a new product design. ONF generates a candidate solution
Q=ONF(P1) from P1, where P1 = (S1 ,..., S m ) is the current solution. First, two
temporary solutions, P2 and P3 are generated by perturbing P1 as follows:
P2 = (S 11 ,..., S m1 ) and P3 = (S12 ,..., S m2 ) , (9.1)
where S i1 = S i + S i and S i2 = S i − S i , i=1,...,m. All S i are generated based on the

Cauchy-Lorentz probability distribution (Szu 1987). Consequently, Q is produced
by a combination of design attributes represented in P1, P2, and P3.
In ONF, P1, P2 and P3 are considered as level 2, level 1 and level 3 of an expe-
rimental design. To assign the m design attributes from P1, P2 and P3 into the N
parameters in L2 N +1 (3 N ) , the m design attributes are divided into N non-
overlapping groups. Each non-overlapping group of design attributes is considered
as a parameter in L2 N +1 (3 N ) . The number of design attributes in each group is de-
termined by a generated random integer li, i=1,2,…N, where
∑l
N
i
=m (9.2)
i =1
9.2 Development of Neighbourhood Function 165
ONF then uses the t-th combination of L2 N +1 (3 N ) to compute yt correspond-

ing to the t-th experiment with t=1,...2N+1, and computes all main effects
M with j=1,2,..,N and k=1,2,3 based on (9.1).
jk
N
If an orthogonal array L2N+1(3 ) with 3 levels is considered, for j=1,2,…N and k
= 1,2,3, the main effect Mjk of group j for level k is defined as:
2N + 1
M
jk
= ∑ yt × Ft (9.3)
t =1
where yt denotes an objective function value of the combination corresponding to
experiment t, and
⎧ 1 if level of parameter j of experiment t is k .
F =⎨
t ⎩0 if level of parameter j of experiment t is not k .
To maximize the customer satisfaction with product design, the best level Best(j)
of the j-th parameter is denoted as:
Best ( j ) = arg(max(M j1 , M j 2 , M j 3 )) (9.4)
where 'arg(max(..))' is a function that returns the index of the maximum value of
customer satisfaction. For example, if the value of Mj2 is the largest among the
values of Mjk where k=1,2 and 3, then Best(j)=2.
The algorithm of the ONF, Q=ONF(P1), is shown below:
Algorithm Q=ONF(P1)
Step 1) Generate P2 and P3 with P1 = (S1 ,..., S m ) based on (9.1).
Step 2) Divide P1, P2 and P3 into N groups of design attributes based on (9.2).
Step 3) Represent levels 2, 1 and 3 of the j-th parameter of L2 N +1 (3 N ) by the
j-th group of P1, P2 and P3 respectively.
Step 4) Compute yt based on the t-th combination of L2 N +1 (3 N ) as the t-th ex-
periment.
Step 5) Compute the main effect Mjk where j=1,...,N and k=1,2,3 based on
(9.3).
Step 6) Determine the best level Best(j) on the j-th parameter based on (9.4)
to maximize customer satisfaction with product design.
Step 7) The candidate solution Q is produced by combining the best levels of
parameters.
ONF uses the analysis of main effects to determine the optimal levels of parame-
ters which is the simplest approach to analyze experimental results (Box et al.
1978, Phadke 1987). However, it is quite common that an interaction effect exists
between two design attributes in new product design (Davidor 1991). To solve the
optimization problems where low interaction effects exist between parameters,
ONF could work properly. However, if strong interaction effects exist between pa-
rameters in optimization problems, the optimal combination based on ONF may
not be reproducible.
9.2.2 An Improved Orthogonal Array Based Neighbourhood

Function
An improved version of orthogonal array-based neighbourhood function (IONF),
which integrates interaction effects between parameters, is presented. In the ONF,
the children are produced by considering only the greatest main effects of parame-
ters. But, in IONF, the children are produced by considering both the main effects
of parameters and the interaction effects between parameters. Interaction plots
(Phadke 1987) are adopted to investigate the magnitudes of interaction effects be-
tween parameters.
In the IONF, an interaction matrix MIij is generated to estimate the magnitudes
of interaction effects between parameters i and j, where i, j=1,2,..N. It can be ex-
pressed as:
MI ij = (I ij (m, n ); for 1 ≤ m, n ≤ 3)3×3 (9.5)
where the numbers of rows and columns of the interaction matrix MIij are both
equal to the number of levels of L2N+1(3N) which is 3. The elements of MIij,
Iij (m, n) , which represents the average fitness of the i th parameter with level m
and j th parameter with level n, are defined as:
N
∑ f ⋅⎢
⎡ In the pth combination of L
⎢ 2N + 1 ( )
3 N , the level of ⎤
⎥
⎥
p
p =1 th th
I ij ( m, n ) = ⎣⎢ the i parameter is m and the j parameter is n ⎦⎥
( )
(9.6)
N ⎡⎢ In the p combination of L 3N , the level of ⎤
th
2N + 1 ⎥
∑ ⎢ ⎥
p = 1 ⎢ the ith parameter is m and the j th parameter is n
⎣ ⎦⎥
where fp denotes an objective function value of the combination corresponding to

the p-th row of L2N+1(3N), 1 ≤ m, n ≤ 3 and
[condition ] = ⎧⎨
1 if the statement inside the bracket is true.
⎩0 otherwise.
Then, interaction plots are used to investigate the magnitude of interaction effects
th
between parameters i and j. The r line of the interaction plot is defined as:
Line ij (r) = (I ij (1, r ), I ij (2, r ), I ij (3, r )) where r = 1, 2, 3. (9.7)

9.2 Development of Neighbourhood Function 167
Line with crosses Lineij(3)

Fitness
value Iij(3,2) Iij(3,3)
Iij(1,1)
Iij(1,2)
Lineij(1)
Iij(2,1)
Iij(1,3)
Iij(2,2) Iij(2,3)
Iij(3,1)
Lineij(2)
Level 1 Level 2 Level 3
Fig. 9.1 Strong interaction effect exists between parameters i and j
Figure 9.1 shows that the lines cross, indicating the existence of strong interac-
tions. If strong interaction effects do not exist in all parameter pairs, only the main
effects of parameters need to be studied. The candidate solution Q of the IONF is
generated by the combination of the parameters with the greatest main effects
based on (9.4) to maximize customer satisfaction for product design. In this case,
the algorithm of IONF is identical to the one of ONF. However, if strong interac-
tion effects exist in any one of the parameter pairs, the candidate solution Q is first
generated by the best level combinations of the orthogonal array L2N+1(3N) with the
optimal yt. For those parameters without strong interaction effects between each
other, the level combinations in Q are replaced by the parameters with the largest
main effects based on (9.4).
The algorithm of the IONF, Q=IONF(P1), is given as follows:
Algorithm Q=IONF(P1)
Step 1: Step 1 to Step 5 of ONF
Step 6: Construct the interaction matrices MI ij by (9.5), where i,
j=1,2,…N with i ≠ j .
Step 7: Construct the interaction plot for MI ij where i, j=1,2,…N with
i≠ j.
Step 8: Check whether the parameters i and j have a strong interaction
effect of each other, where i, j=1,2,…N with i ≠ j .
Step 9: If strong interaction effect exists in any one of the parameter
pairs, goto Step 10, otherwise goto Step 13.
Step 10: Form the candidate solution Q by the combination of the
L2 N +1 (3 N ) with the optimal yt.
Step 11: For the parameter pair with no strong interaction effect, the
level combinations in Q are replaced by the level combinations
with the largest main effects based on (9.4).
Step 12: Output Q as the resulting solution of IONF. Then goto step 15.
Step 13: Determine the best level Best(j) on the j-th parameter based on
(9.4) to maximize customer satisfaction for product design.
Step 14: The candidate solution Q is produced by the combinations of
best levels of parameters.
Step 15: Terminate the algorithm.
9.3 A Case Study: Emulsified Dynamite Packing Machine

In weapons manufacturing, dynamite is handled in powder form which can easily
explode during transportation or storage. Nowadays, dynamite is first emulsified
into a liquid state which is safer to handle. Usually, a machine is required to per-
form the emulsification of dynamite which is commonly called an emulsified dy-
namite packing machine. In this section, optimization of an emulsified dynamite
packing machine design based on the IOSA is illustrated. An optimization prob-
lem of the machine design formulated by (Chan et al. 2009) was adopted in this
research which contains the following engineering requirements (i.e. X=x1, x2,…,
x7) and customer requirements (i.e. Y=y1, y2,…, y4) below and summarized as
Figure 9.2.
x1 x2 x3 x4 x5 x6 x7
Emulsified dynamite packing

machine
y1 y2 y3 y4
Fig. 9.2 Illustration of the emulsified dynamite packing machine
x1 – precision of molding of clip

x2 – precision of dynamite packing
x3 – control force of dynamite packing
x4 – efficiency of dynamite packing
x5 – hardness of pressing hammer
x6 – noise of cam power transmission
x7 – height of machine bed
9.3 A Case Study: Emulsified Dynamite Packing Machine 169
y1 –quality of packing dynamite

y2 – efficiency of packing dynamite
y3 – packing noise
y4 – rigidity of machine
Details of the optimization problem (Chan et al. 2009) are shown below.
osc = (0.46 y1r + 0.28 y 2r + 0.16 y 3r + 0.10 y 4r ) r
1
(9.8)
subject to:
y1 = 5.98 x1 − 0.33x 2 − 1.37 x 3 + 0.88 (9.9)
y 2 = 2.45 x3 + 0.96 x 4 + 1.25 x 5 + 0.54 (9.10)
y 3 = 4.20 x 6 + 1.00 (9.11)
y 4 = 4.00 x 7 + 1.25 (9.12)
(f 1
− 1.3) + ( f 2 − 1) ≤ 0.5
2 2 2
(9.13)
f 1 = 0.464 y1 + 0.449 y 2 − 0.217 y 3 + 0.166 y 4 (9.14)
f 2 = −0.030 y1 − 0.100 y 2 − 0.508 y 3 − 0.695 y 4 (9.15)
50 + 20 x1 + 25 x 2 + 10 x 3 + 15 x 4 + 5 x 5 + 30 x 6 + 8 x 7 ≤ 100 (9.16)
1 ≤ y i ≤ 5, i = 1,...,4 (9.17)
0 ≤ x j ≤ 1, j = 1,...,7 (9.18)
where
1. xi (i=1,2,…7) is the level of attainment of Xi; yi (i=1,2,…4) is the value of
customer satisfaction of Yi;
2. (9.8) is the objective function of deriving overall customer satisfaction
(OCS);
3. (9.9) to (9.12) are the models of functional relationship between customer
requirement Yi, i=1,…,4 and design attributes, Xj, j=1,…,7;
4. (9.13) to (9.15) are the constraints of product positioning;
5. (9.16) is the cost constraint that is subject to a budget with the fixed cost and
the cost incurred for achieving each design attribute, Xj, j=1,…,7;
6. (9.17) and (9.18) are the ranges of values of the customer satisfactions and
levels of attainment of the design attributes respectively.
The IOSA was used to determine the optimal setting of levels of attainment of the
design attributes x1, x2, x3, x4, x5, x6 and x7, by maximizing the overall customer
satisfaction. The algorithm was implemented by using Matlab, in which a candi-
date solution is represented as:
s = [x1 , x 2 , x3 , x 4 , x5 , x6 , x 7 ] (9.19)
The objective function used in the algorithm is defined by maximizing the optimi-
zation function (9.20) subject to the constraints. It is defined as:
⎛ ⎧ min (− f b (s ),− f c (s )) if f b (s ) < 0 or f c (s ) < 0 ⎞
f (s ) = max⎜⎜ ⎨ ⎟ (9.20)
⎝ ⎩ f a (s ) otherwise ⎟
⎠
where f a (s ) = (0.46 y1r + 0.28 y 2r + 0.16 y 3r + 0.10 y 4r ) r ;

1
(9.21)
[
f b (s ) = 0.5 2 − ( f 1 − 1.3) + ( f 2 − 1) ;
2 2
] (9.22)
f c (s ) = 100 − (50 − 20 x1 − 25 x 2 − 10 x 3 − 15 x 4 − 5 x 5 − 30 x 6 − 8 x 7 ) ; (9.23)
⎧1 if y i < 1 and 5 < y i

∑ (y
4
f d (s ) = − 3) ⋅ R i ; R i = ⎨
2
(9.24)
⎩0 if 1 ≤ y i ≤ 5
i
i =1
y1, y2, y3 and y4 can be found in (9.9), (9.10), (9.11) and (9.12) respec-
tively; f1 and f2 can be found in (9.14) and (9.15) respectively. (9.22) is
formulated to handle the constraints (9.13)-(9.15). (9.23) and (9.24) are
formulated to handle the constraints (9.16) and (9.17) respectively. The
constraint (9.18) can be dealt with by setting the ranges of solutions in
the simulated annealing.
By referring to previous research (Ho et al. 2004a), the parameters used in OSA
and IOSA are T0=50, I0=5 and CR=0.95. The orthogonal array used in OSA and
IOSA is also the same as the one used in (Ho et al. 2004a), which is L2N+1(3N),
where N = 3⎣ 3 (log (2 p +1)⎦
)
− 1 / 2 and p is the number of variables of the optimi-
zation problem. Since only one functional evaluation is implemented in SNF, a
functional evaluation is required in each iteration of SSA. However, 2N+1 func-
tional evaluations are implemented in the ONF and IONF. Therefore, 2N+1 func-
tional evaluations are required in each iteration of the OSA and IOSA. To make
the number of functional evaluations of all the algorithms the same, the number of
iterations used in SSA is larger than those used in both the OSA and IOSA. There-
fore, the value of CR used in SSA is larger than the one used in the OSA and
IOSA. The parameters of SSA used are T0=50, I0=5 and CR=0.99. The pre-defined
number of functional evaluations used in all the algorithms is 10 000.
The variable r used in (9.24) was set to 1, 2, 3, 4 and 5. With a higher value of
r, interaction effects between design attributes are higher. Since the simulated an-
nealing algorithm is stochastic one, different solutions are obtained from different
runs. Therefore, 30 test runs were performed in order to obtain the two statistics,
namely mean and variance of overall customer satisfaction with respect to the val-
ues of r. After assessing the interaction effects between the design attributes as
shown in the house of quality for the machine design (Chan et al. 2009), they were
assessed as ‘low to medium’ and hence r value was set as 2.
Maximization of ocs is a multi-objective optimization which aims to optimize
the degrees of satisfaction of the four customer requirements, y1, y2, y3 and y4.
After running the IOSA for 30 times, the optimal solutions s corresponding to r=2
in terms of x1, x2, x3, x4, x5, x6 and x7 were found and are given in Table 9.1.
9.3 A Case Study: Emulsified Dynamite Packing Machine 171
Table 9.1 Optimal design attribute settings for the Emulsified dynamite packing machine
design
Precision of molding x1 0.8735

of clip
Precision of dyna- x2 0.0002
mite packing
Control force of dy- x3 0.0002
namite packing
Efficiency of dyna- x4 0.5479
mite packing
Hardness of press- x5 0.9913
ing hammer
Noise of cam power x6 0.3211
transmission
Height of machine x7 0.2076
bed
To further evaluate the effectiveness of the IOSA in solving real-world optimi-

zation problems compared with standard simulated annealing algorithms, SSA,
were also used to solve the same optimization problem using the same parameter
setting and number of test runs. To evaluate the performance of the three
neighbourhood functions (SNF, ONF and IONF), the simulated annealing algo-
rithm used by (Ho et al. 2006, Shu et al. 2004) was employed in this research
which is shown below.
Algorithm of simulated annealing

Begin
i=1
Randomly generate a candidate solution s
Repeat
t=T0; I=I0
Repeat:-
Q = neighbour_function(s)
if f(Q)< f(s) then replace s with Q
else
generate a random number r
if exp(-(f(s)-f(Q))/t)>r then replace s with Q
endif
endif
t=t*CR
until t<I*CR
i=i+1
Until i=( pre-defined number of function evaluations)
End.
Table 9.2 shows the means and standard deviations of the runs with different val-
ues of r. Here the means of the overall customer satisfaction based on the IOSA
for any value of r are larger than the ones based on the OSA and SSA, except r=1.
It is because there is no interaction between design attributes of (9.5) while r=1.
Therefore, the overall mean of customer satisfaction based on the IOSA is nearly
identical to that based on the OSA. Regarding the standard deviation, it can be
found from Table 9.1 that the average standard deviations of the IOSA is smaller
than the ones based on the OSA and SSA.
Table 9.2 Results of validation tests based on the case of design optimization
r= 1 2 3 4 5
SSA mean 4.1223 3.5952 4.0196 4.0130 3.9232
SSA std 0.1223×10-3 0.4756 0.1509×10-0 0.8520×10-1 0.1190×10-1
OSA mean 4.2372 4.0089 4.1333 4.1246 4.0010
OSA std 0.9903×10-6 0.4792 0.9413×10-1 0.1658×10-1 0.3740×10-1
IOSA mean 4.2373 4.3098 4.2372 4.2368 4.2373
IOSA std 0. 2679×10-5 0.4793 0.6072×10-1 0.1391×10-1 0.1264×10-1
std – standard deviation
The t-test was used to evaluate whether the IOSA is significantly better than the
standard simulated annealing algorithm in this optimization problem. The t-values
between the IOSA and the standard simulated annealing algorithms are shown in
Figure 9.3, from which it can be seen that all the t-values are higher than 1.675
except the t-value for OSA-IOSA while r=1. Based on the normal distribution ta-
ble, if the t-value is higher than 1.675, the difference of performance between two
algorithms is significant with a confidence level of 95.3%. Therefore, it can be
concluded that the performance of the IOSA is significantly better than the SSA
and OSA. As explained before, while r=1, interaction effects do not exist in (9.7),
there is no significant difference in performance between the OSA and IOSA. The
significance of the difference increases as r increases. It can be explained that with
a larger value of r, interaction effects between variables become stronger.
On the other hand, computational times of generating solutions based on the
IOSA were also compared with those based on the OSA and SSA. Figure 9.4
shows the computational times of the three algorithms with respect to different
values of r. Execution of the algorithms is based on a Pentium 4 PC with 2.26
MHz. It can be seen from the figure that the times taken to search for solutions
based on the IOSA are less than those based on the OSA and SSA for all values of
r, except r=1. When r=1, the time taken based on the IOSA is identical to that
based on the OSA, but is still less than that based on the SSA.
9.4 Conclusion 173
t-values between the algorithms
100
10
T-values
SSA - IOSA
OSA - IOSA
1
0.1
1 2 3 4 5
r -values
Fig. 9.3 t-values between the algorithms
Computational times on the algorithms

14
13
12
Computational time (second)
11
10
6 SSA
OSA
5
IOSA
4
1 1.5 2 2.5 3 3.5 4 4.5 5
Value of r
Fig. 9.4 Computational time on solving the emulsified dynamite packing-machine design
problem
9.4 Conclusion
In this chapter, an improved orthogonal array-based neighbourhood function
(IONF) for simulated annealing which considers both main effects of design
attributes and interaction effects between design attributes is described. The
proposed IONF is to make up for the deficiency of the existing orthogonal array-
based neighbourhood functions (ONF), which do not consider the effects of the in-
teraction between design attributes. The IOSA was successfully applied to solve
the design optimization problem of emulsified dynamite packing machines in

which interaction effects exist between the design attributes. Results indicate that
IOSA outperforms the other two simulated annealing algorithms in terms of quali-
ty and robustness of solutions. T-tests were also conducted. Results of the tests in-
dicate that the IOSA outperforms the other two algorithms significantly. Besides
this, the times taken to search for solutions based on the IOSA are the smallest
compared with those based on the other two algorithms.
In the following Chapter 10, we will discuss an innovative computational intel-
ligence method which integrates the mechanisms of orthogonal array and evolu-
tionary algorithms. This resulting innovative computational intelligence method
will be used to determine optimal design attribute settings, in order to maximize
overall customer satisfaction with new product design.
References
Aydin, M.E., Fogarty, T.C.: A distributed evolutionary simulated annealing for combina-
torial optimisation problems. Journal of Heuristics 10(3), 269–292 (2004)
Box, G.E.P., Hunter, W.G., Hunter, J.S.: Statistics for Experimenters. John Wiley (1978)
Bryne, D.M., Taguchi, S.: The Taguchi approach to parameter design. ASQC Quality Con-
gress Transaction, 168 (1986)
Cerny, V.: Thermodynamical approach to the travelling salesman problem: an efficient si-
mulation algorithm. Journal of Optimization Theory and Applications 45, 41–51 (1985)
Chatterjee, S., Carrera, C., Lynch, L.A.: Genetic algorithms and travelling salesman prob-
lems. European Journal of Operational Research 93, 490–510 (1995)
Chan, K.Y., Kwong, C.K., Luo, X.G.: Improved orthogonal array based simulated anneal-
ing for design optimization. Expert Systems with Applications 36, 7379–7389 (2009)
Davidor, Y.: Epistasis variance: a viewpoint on GA-hardness. In: Rawlins, G.J.E. (ed.)
Foundations of Genetic Algorithms. Morgan Kaufmann, San Mateo (1991)
Fogel, D.B.: An introduction to simulated evolutionary optimization. IEEE Transactions of
Neural Networks 5(1), 3–14 (1994)
Gong, G., Liu, Y., Qian, M.: An adaptive simulated annealing algorithm. Stochastic
Processes and their Applications 94, 95–103 (2001)
Ho, S.J., Ho, S.Y., Shu, L.S.: OSA: Orthogonal simulated annealing algorithm and its ap-
plication to designing mixed H2=H ∞ Optimal Controllers. IEEE Transactions on Sys-
tems. Man and Cybernetics – Part A: Systems and Humans 34(5), 588–600 (2004a)
Ho, S.Y., Ho, S.J., Lin, Y.K., Chu, W.C.C.: An orthogonal simulated annealing algorithm
for large floorplanning problems. IEEE Transactions on Very Large Scale Integration
(VLSI) Systems 12(8), 874–876 (2004b)
Ho, S.Y., Shu, L.S., Chen, J.H.: Intelligent evolutionary algorithms for large parameter op-
timization problems. IEEE Transactions on Evolutionary Computation 8(6), 522–541
(2004c)
Ho, S.J., Shu, L.S., Ho, S.Y.: Optimizing fuzzy neural networks for tuning PID controllers
using an orthogonal simulated annealing algorithm OSA. IEEE Transactions on Fuzzy
Systems 14(3), 421–434 (2006)
Kim, J.D., Choi, M.S.: Stochastic approach to experimental analysis of cylindrical lapping
process. International Journal of Machines Tools Manufacturing 35(1), 51–59 (1995)
References 175
Kirkpatrick, S., Gelatt, J., Vecchi, M.P.: Optimization by simulated annealing. Science 220,
671–680 (1983)
Kratica, J., Tosic, D., Filipovic, V., Ljubic, I.: Solving the simple plant location problem by
genetic algorithm. RAIRO Operations Research 35, 127–142 (2001)
Kwong, C.K., Chen, Y., Chan, K.Y.: Integrating perceptual product mapping with QFD for
new product design (working paper)
Lin, C.K.Y., Haley, K.B., Sparks, C.: A comparative study of both standard and adaptive
versions of threshold accepting and simulated annealing algorithms in three scheduling
problems. European Journal of Operational Research 83, 330–346 (1995)
Lin, Y.H., Tyan, Y.Y., Chang, T.P., Chang, C.Y.: An assessment of optimal mixture for
concrete made with recycled concrete aggregates. Cement and Concrete Research 34,
1373–1380 (2004)
Locatelli, M.: Simulated annealing algorithms for continuous global optimization: conver-
gence conditions. Journal of Optimization Theory and Applications 104(1), 121–133
(2000)
Mohan, N.S., Ramachandra, A., Kulkarni, S.M.: Influence of process parameters on cutting
force and torque during drilling of glass fiber polyester reinforced composites. Compo-
site Structures 71, 407–413 (2005)
Moilanen, A.: Simulated evolutionary optimization and local search: introduction and ap-
plication to tree search. Cladistics 17, 12–15 (2001)
Phadke, M.S.: Quality engineering using robust design. Prentic Hall, New York (1987)
Reeves, C.R., Wright, C.C.: An experimental design perspective on genetic algorithms.
Foundation of Genetic Algorithms 3, 7–22 (1995)
Reeves, C.R., Wright, C.C.: Epistasis in Genetic Algorithms: An Experimental Design
Perspective. In: Proceedings of the 6th International Conference on Genetic Algorithms,
pp. 217–224 (1995)
Reeves, C.R.: Predictive measures for problem difficulty. In: Proceedings of the 1999 Con-
gress on Evolutionary Computation, vol. 1, pp. 736–742 (1999)
Ruiz-Torres, A.J., Enscore, E.E., Barton, R.R.: Simulated annealing heuristics for the aver-
age flow-time and the number of Tardy jobs bi-criteria identical parallel machine Prob-
lem. Computers Industry Engineering 33, 257–260 (1997)
Shu, L.S., Ho, S.Y., Ho, S.J.: A novel orthogonal simulated annealing algorithm for optimi-
zation of electromagnetic problems. IEEE Transactions on Magnetics 40(4), 1791–1795
(2004)
Szu, H., Hartley, R.: Fast simulated annealing. Physical Letters 122, 157–162 (1987)
Szu, H.: Nonconvex optimization by fast simulated annealing. Proceedings of the
IEEE 75(11), 1538–1540 (1987)
Tsallis, C., Stariolo, D.A.: Generalized simulated annealing. Physica A 233, 395–406
(1996)
Unal, R., Stanley, D.O., Joyner, C.R.: Propulsion system design optimization using the Ta-
guchi Method. IEEE Transactions on Engineering Management 40(3), 315–322 (1993)
Van Laarhoven, P.J.M., Aarts, E.H., Lenstra, J.K.: Job shop scheduling by simulated an-
nealing. Operations Research 40(1), 113–125 (1992)
Van Laarhoven, P.J.M., Aarts, E.H.L.: Simulated Annealing: Theory and Applications. D.
Reidel Publishing Co. (1987)
Vaessens, R.J.M., Aarts, E.H.L., Lenstra, J.K.: A local search template. In: Proceedings of
Parallel Problem-Solving from Nature, pp. 65–74 (1992)
Whitley, D., Mathias, K., Rana, S., Dzubera, J.: Building better test function. In: Proceed-
ings of the 6th International Conference on Genetic Algorithms, pp. 239–246 (1995)
Wong, S.Y.W.: Hybrid simulated annealing/genetic algorithm approach to short term hy-
dro-thermal scheduling with multiple thermal plants. Electric Power Energy Systems 23,
565–575 (2001)
Yao, X., Lin, Y., Lin, G.: Evolutionary programming made faster. IEEE Transactions on
Evolutionary Computation 3(2), 82–102 (1999)
Yao, X.: Simulated annealing with extended neighbourhood. International Journal of Com-
puter Mathematics 40, 169–189 (1991)
Yao, X.: Comparison of different neighbourhood sizes in simulated annealing. In: Proceed-
ings of 4th Australian Conference on Neural Networks, pp. 216–219 (1993)
Yin, G.Z., Jillie, D.W.: Orthogonal design for process optimization and its application in
plasma etching. In: Bendell, A., Disney, J., Pridmore, W.A. (eds.) Taguchi Methods:
Applications in World Industry, pp. 181–198. IFS Publications/Springer-Verlag (1989)
Chapter 10
An Enhanced Genetic Algorithm Integrated
with Orthogonal Design
10.1 Introduction
Chapter 9 introduced an innovative computational intelligence method based on

simulated annealing, to perform optimization of new products. In this chapter, we
introduce another computational intelligence method known as evolutionary algo-
rithms to perform optimization of new products.
These evolutionary algorithms are increasingly being used in a wide range of
practical applications especially where the problem involves a non-differentiable
cost function or where the cost function is hard to quantify mathematically [Bonis-
sone et al. 2006, Dimopoulos and Zalzala 2000]. However, one of the main draw-
backs of evolutionary algorithms is their tendency to converge before reaching an
acceptable solution to challenging problems where the dimensions are high and
there are numerous local optima [KrishnaKumar et al. 1995]. To overcome this
problem, recent research [Ho et al.1999, Ho et al. 2004a, Leung and Wang 2001,
Tsai et al. 2004] has shown that optimization using evolutionary algorithms for
solving parametric problems with high dimension and multi-optima can be en-
hanced by embedding the approach of orthogonal design in the crossover.
Research indicates that the orthogonal array-based crossovers outperform the ex-
isting crossovers in solving travelling salesman problems [Ho et al. 1999], poly-
gonal approximation [Ho and Chen 2003], solving multimedia multicast routing
problems [Zhang and Leung 1999], searching Pareto-optimal solutions [Leung and
Wang 2000], development of fuzzy classifiers [Ho et al. 2004a], structure-
specified mixed H2/H∞ controller design [Ho et al. 2005], solving multi-objective
combinatorial optimization problems [Ho et al. 2004b], solving mesh optimization
problems for surface approximation [Huang and Ho 2003], and process design of
fluid dispensing [Kwong et al. 2006].
In this chapter, three versions of orthogonal array-based crossovers are intro-
duced to maximize customer satisfaction with product design by optimizing the
design attribute settings of a new product.
The first version of orthogonal array-based crossover [Leung and Wang 2000,
Leung and Wang 2001, Zhang and Leung 1999] is called the ‘orthogonal cross-
over’ (OC), where chromosomes are produced by exploring alleles in parents
178 10 An Enhanced Genetic Algorithm Integrated with Orthogonal Design
based on combinations of an orthogonal array. The two top chromosomes with

best fitness among all the chromosomes produced are selected as the two children
as the outcome of the OC. However, this approach considers only a limited num-
ber of combinations in the orthogonal array rather than taking all the combinations
as in a full factorial design. This may not be applicable for optimizing the design
attributes of a new product, as the optimal combination of design attributes might
not be included in the combinations of the orthogonal array.
The second version of orthogonal array-based crossovers [Ho et al. 1999, Ho
and Chen 2003, Ho et al. 2004 a,b, Ho et al. 2005, Huang and Ho 2003, Kwong et
al. 2006], is called the ‘main effect crossover’ (MC), because the combinations
excluded in the orthogonal array can be considered by analyzing the main effects
of design attributes in a new product. Children are formed from the best combina-
tions of genes with the best main effect in design attributes of a new product.
Thus, all combinations are considered, as is the case in full factorial design. This
is more promising for optimizing design attributes in new products than OC is on
its own. MC allows us to approximate the main effect on each gene, but it ignores
linkages in the form of interaction between design attributes in a new product. If
strong interaction exists in localized features of the search space, misleading re-
sults may be obtained [Reeves 1999].
The third version of orthogonal array-based crossover [Chan et al. 2010] is
called an ‘interaction crossover’ (IC), which considers the interactions between
the design attributes of a new product. It employs the interaction plot approach
[Phadke 1987], that has been commonly used to analyze the interactions between
parameters in industrial systems [Kim and Choi 1995, Lin et al. 2004, Mohan et
al. 2005, Unal et al. 1993], to analyze the interaction between genes. From the in-
teraction plot, a clear picture of the effects of interaction between genes can be ob-
tained. In the crossover operator, the children can be produced by considering both
the main effects of design attributes and the interaction between design attributes.
The effectiveness of the three versions of orthogonal array-based crossovers is
evaluated by using a case study of a car door design. Customer satisfaction with a
car door is maximized by optimizing its design attribute settings. The result is then
compared with the results obtained by standard crossover and the previously de-
veloped computational method [Bai and Kwong 2003].
10.2 Orthogonal Array Based Crossovers

The following sub-sections 10.2.1, 10.2.2 and 10.2.3 discuss the operations of the
three versions of orthogonal array based crossovers, orthogonal crossover OC
[Le‘ung and Wang 2000, Leung and Wang 2001, Zhang and Leung 1999], main
effect crossover MC [Ho et al. 1999, Ho and Chen 2003, Ho et al. 2004 a,b, Ho et
al. 2005, Huang and Ho 2003, Kwong et al. 2006] and IC [Chan et al. 2010] re-
spectively to maximize customer satisfaction with product design.
10.2 Orthogonal Array Based Crossovers 179
10.2.1 Orthogonal Crossover (OC)

In OC, an orthogonal array is integrated into the classical crossover operator so
that two parents can be used to generate a small but representative set of sampling
points, to be children based on the orthogonal array.
The chromosomes used in OC are in a real-coded representation, where the al-
leles in genes are real numbers. In a way similar to the classical crossover, two
parents P1 = ( p1,1 , p1,2 ,..., p1,l ) and P2 = ( p2,1 , p2,2 ,..., p2,l ) are selected randomly
from the population, where l is the number of design attributes in a new product.
Q is the number of levels, M is the number of rows and N is the number of col-
umns of the orthogonal array LM ( Q N ) respectively, while LM (Q N ) is integrated
in OC. Then the genes in P1 and P2 are quantified into Q levels such that the
difference between any two successive levels is the same. The i th level is de-
noted as Level ( i ) = ( βi ,1 , βi,2 ,... βi ,l ) , where i=1,2,…,Q, and βi , j is defined as:
⎧min ( p1, j , p 2 , j ), for i = 1 and 1 ≤ j ≤ l.

⎪
⎪
⎪⎪ ⎛ p1, j − p 2 , j ⎞
βi , j = ⎨min ( p1, j , p 2 , j ) + (i − 1) ⋅ ⎜ ⎟, for 2 ≤ i ≤ Q − 1 and 1 ≤ j ≤ l (10.1)
⎜ Q −1 ⎟
⎪ ⎝ ⎠
⎪
⎪
⎩⎪max ( p1, j , p 2 , j ), for i = Q and 1 ≤ j ≤ l.
After quantifying P1 and P2 , the Q levels are sampled as M potential offspring

based on the combinations of the M rows of parameter levels in the orthogonal ar-
ray LM ( Q N ) . Specifically, N-1 integers k1 , k 2 ,…, k N −1 are generated randomly
such that 1< k1 < k2 <…< k N −1 <l. Then N vectors are created such that:
( )
⎧ f 1 = f 11 , f 21 ,..., f k11 = (1,2,..., k1 ),
⎪
⎪ where the number of elements inside f 1 is k1 .
( )
⎪ f 2 = f 12 , f 22 ,..., f k22 − k1 +1 = (k1 + 1, k1 + 2,..., k 2 ),
⎪
⎪ where the number of elements inside f 2 is k 2 − k1 + 1.
⎨ (10.2)
⎪ : : :
⎪: : :
⎪ 1
(
⎪ f N = f N , f N ,..., f N
2 l − k N −1 +1
)
= (k N −1 + 1, k N −1 + 2,..., l ),
⎪⎩ where the number of elements inside f N is l − k N −1 + 1.
For i = 1,2,…M, the i th offspring oi is produced as:

oi = ( β a (i ), f 1 , β a ( i ), f 1 ,..., β a ( i ), f 1 , β a 2 , βa 2 ,...
1 1 1 2 1 k1 2 ( i ), f 2 ( i ), f
1 2
..., β a 2 ,..., β a N , βa N ,..., β a N ) (10.3)

2 ( i ), f N ( i ), f1 N ( i ), f 2 N ( i ), f l − k
k 2 − k1 +1 N −1 +1
row of the orthogonal array LM (Q N ) is de-

th
where the combination of the i
[
noted as a(i) = a1 (i ), a2 (i ),..., a N (i ) . The fitness ci of each offspring ]
(i.e. : oi for i = 1,2,..., M ) is evaluated according to the fitness function; i.e.
ci = fun ( oi ) , where fun ( ) denotes as the fitness function. Then the two off-
spring with the best fitness among M offspring are selected to be the two children
of the OC.
Algorithm 10.1: Orthogonal crossover (OC)
Detailed steps of the OC are as follows:
Algorithm 10.1: Orthogonal crossover (OC)
Step 1: Select two parents P1 and P2 from the population randomly.
Step 2: Quantize P1 and P2 based on (10.1), and produce
Level ( i ) = ( βi ,1 , βi ,2 ,... βi ,l ) , where i=1,2,…,Q.
Step 3: Randomly generate N-1 integers k1 , k2 ,…, k N −1 , such
that1< k1 < k2 <…< k N −1 <l in which l is the number of variables in
i
the chromosomes. Then create N vectors f based on (10.2),
where i=1,2,…,N.
Step 4: Apply LM Q ( ) to produce the M potential offspring o
N
i based
on (10.3), where i=1,2,…,M.
Step 5: Evaluate the fitness of the M potential offspring based on the fit-
ness function fun( ) .
Step 6: Select the two offspring with the best fitness among M potential
ones to be the two children of the OC.
th
For i = 1,2,…,M, the i th offspring oi is produced based on the i combination
( ) . Therefore M offspring are produced, meaning
of the orthogonal array LM Q
N
that M combinations are explored by the orthogonal array L (Q ) . However,

N
M
the total number of combinations of N genes with Q levels are N Q . In OC, only
M combinations are considered; thus N Q − M combinations are not explored.
This may not be applicable to all the product design problems as the best
combination may not be included in the orthogonal array. This is the potential
limitation of OC.
An example as shown in Figure 10.1 is used to explain the limitation in which
the orthogonal array L4 (2 3 ) [Taguchi and S. Konishi 1987] is used to sample the
genes from the two parents, P and P'. Each of the two parents is divided into three
genes, where ( p1 , p 2 , p3 ) and ( p1 ' , p 2 ' , p 3 ') are defined as the three genes of P
and P' respectively. These three genes from the parents are sampled based on the
four combinations of parameter levels in L4 (2 3 ) . Four potential offspring, O1 ,
O2 , O3 , and O4 are produced as shown in Figure 10.1. The best two offspring
among the four are selected to be the children of OC.
Sample the genes

O1 = ( p1, p2, p3 )
based on L4(2^3)
P= ( p1, p2, p3 )
O2 = ( p1, p2', p3' ) Select the best two
offspring to be the
O3 = ( p1', p2, p3' ) children
P'= ( p1', p2', p3' )
O4 = ( p1', p2', p3 )
Fig. 10.1 The orthogonal array L4(23) is used to sample the genes from P and P' for OC
Figure 10.2 illustrates the combinations in a full factorial design of 3 parame-

ters with 2 levels. It can be seen from Figure 10.2 that the total number of
combinations of 3 genes with 2 parameter levels is 8 (i.e. 23). However, only 4
combinations (the black points) are considered by the orthogonal array L4 (2 3 ) ,
and the other 4 combinations (the grey points) are not explored by L4 (2 3 ) . There-
fore, this approach may not be applicable to parametrical problems since the opti-
mal combination may be one of the combinations of the grey points.
Third gene
(p1', p2, p3')

p3'
(p1, p2', p3')

(p1, p2, p3)
p3
First gene
p1 p1'
p2
(p1', p2', p3)
p2'
Second gene
Fig. 10.2 Combinations of the orthogonal array L4 (2 3 )
10.2.2 Main Effect Crossover (MC)

The major steps of the main effect crossover (MC) [Ho et al. 1999] are similar to
those of the OC. In the OC, after evaluating the fitness of the M offspring
(oi for i = 1,2,...M ) the two top offspring among the offspring from the orthogo-
nal array are selected to be the two children, after evaluating the fitness of the M
offspring (oi for i = 1,2 ,...M ) . In MC, the fitness values of the M offspring are
analyzed further by considering the main effects in genes, and the children are
produced by taking the genes with the best levels.
The main effect of the jth gene with level k is defined as:
N
M jk = ∑ c i ⋅ Η ij (10.4)
i =1
where
⎧1, if a j (i ) = k ,
Η ij = ⎨
⎩0 otherwise,
ci is the fitness value of the i th offspring oi with i=1,2,…,M, i.e. ci = fun(oi ) ,

and aj(i) is the element of the jth column and the ith row of the orthogonal array
LM(QN). In other word, ci represents the fitness value of the genes formed by the ith
combination of the orthogonal array LM(QN).
The first child is formed from the best combinations with the best level on each
gene. If M j1 > M j 2 , the level 1 of the j th gene is better than the level 2 to
maximize customer satisfaction with new products. The best level Best(j) of the
j th gene is denoted as:
( )
Best ( j ) = arg max (M jk ) , where j = 1,2,..., N
k =1, 2 ,...,Q
(10.5)
where ' arg(max(..))' is a function that returns the indices of the maximum value of
the matrix.
The second child is identical to the first child except that the gene with the
lowest main effect difference at the other level is chosen, where the main effect
difference ( MED ) j on the j th gene is denoted as:
MED j = max (M jk ) − k =min (M jk ), where j = 1,2,..., N . (10.6)

k =1,2,..., Q
1,2,..., Q
Note that the main effect reveals the individual effect of a gene, thus the most ef-
fective gene has the largest main effect difference.
The detailed steps of the main effect crossover (MC) are as follows:
Algorithm 10.2: Main effect crossover (MC)
Step 1-Step 5: Step 1 to Step 5 are identical to Step 1 to Step 5 of Algo-

rithm 10.1.
Step 6: Based on (10.4), evaluate the main effect M jk of the j th gene
with level k, where j=1,2,…,N and k=1,2,…,Q.
Step 7: Determine the best level Best(j) of the j th gene based on (10.5),
where j=1,2,…,N.
Step 8: The first child is formed from the best level of each gene.
( )
Step 9: Determine the main effect difference MED j on the j th gene
based on (10.6), where j=1,2,…,N.
Step 10: The second child is identical to the first child except the gene
with the lowest main effect difference adopts the other level.
It should be emphasized that the analysis of the main effect is the simplest ap-
proach to data analysis [Box and Hunter 1978, Phadke 1987]. However, it is
common for two of the genes to interact and yield a result that is more dependent
upon the interaction between the two genes than on the main effects of either indi-
vidual gene [Davidor 1991]. Further analysis, which gives insights into interac-
tions and main effects inside the chromosomes in GAs, has been done [Reeves
1999]. Their central idea is to perform an 'analysis of variance (ANOVA)',
whereby the variability of the fitness values of the chromosomes (measured by
sums of squared deviations from mean fitness, and denoted by SS) is partitioned
into main effects and interactions.
Total SS = SS of main effects + SS of interactions
Therefore, a major weakness of MC is that it does not adequately deal with the po-
tential interactions between genes. If a chromosome exhibits very low interaction
between the genes, it could probably be processed efficiently by MC. Otherwise,
the predicted optimal combination may not be reproducible if strong interaction
exists between the genes.
Furthermore, the empirical results [Chan et al. 2003] show that MC outper-
forms OC on the parametrical problems where all variables are linearly indepen-
dent of each other. However, no significant improvement can be found on MC
over OC on the parametrical problems where the variables interact with each oth-
er. Therefore, it seems that MC cannot work well on parametric problems in
which variables interact with each other. In the following section, the improved
version of MC, which integrates the information of interactions between genes, is
proposed.
10.3 Interaction Crossover (IC)
The steps of the proposed new orthogonal array-based crossover, namely IC, are
similar to the ones in MC. In MC, the children are produced by considering only
the best main effects in genes. In IC, the children are produced by considering
both the main effects in genes and the interactions between genes. The approach
of the interaction plot [Phadke 1987], which is commonly used to analyze the
magnitudes of interaction between parameters in industrial systems [Lin et al.
2004, Mohan et al. 2005, Unal et al. 1993], is applied to IC. From the interaction
plot, a clear picture of the magnitudes of interactions between genes can be
indicated.
In IC, an interaction matrix MI ij is prepared in order to estimate the magni-
tudes of interaction between genes i and j, where 1 ≤ i, j ≤ N . It can be expressed
as:
MI ij = (I ij (m, n ); for 1 ≤ m, n ≤ Q )Q×Q (10.7)
where Q is the number of rows and columns of the interaction matrix MI ij . The
elements in MI ij , I ij (m, n ) , which represent the average fitness of the i th gene

with level m and j th gene with level n, are defined as:
⎡ ⎤
N the level of the p th offspring of the i th gene is m ⎥
∑ f ⋅ ⎢⎢ ⎥
p
p =1 th
⎣⎢and the j gene is n ⎦⎥
I ij (m, n ) = (10.8)
N ⎡ the level of the p th offspring of the i th gene is m ⎤
∑ ⎢⎢ ⎥
⎥
p = 1 ⎢and the j gene is n
th
⎣ ⎦⎥
where 1 ≤ m, n ≤ Q and
[condition ] = ⎧⎨
1 if the statement inside the bracket is true.
⎩0 otherwise.
Then the approach of interaction plot [Phadke 1987] is used to indicate the magni-
th
tude of interaction between genes i and j. The r line of the interaction plot is
defined as:
Line ij (r) = (I ij (1, r ), I ij (2, r ),..., I ij (Q , r )); for 1 ≤ r ≤ Q. (10.9)

10.3 Interaction Crossover (IC) 185
Parallel lines Non-parallel lines

Fitness Fitness
value value
Lineij (1) Lineij(1)
Lineij (2) Lineij(2)
Lineij (Q) Lineij(Q)
1 2 3 .......... Q Level 1 2 3 .......... Q Level
Fig. 10.3(a) No interaction exists between Fig. 10.3(b) Interaction exists between
genes i and j genes i and j
Lines with crosses

Fitness
value
Lineij (1)
Lineij(2)
Lineij (Q)
1 2 3 .......... Q Level
Fig. 10.3(c) Strong interaction exists between genes i and j
In the interaction plot, the Q points in each line represented by equation (10.9)
are put onto a two-dimension plot. The magnitude of interaction can be deter-
mined by the interaction plot. If the lines on the interaction plot (as shown in Fig-
ure 10.3(a)) are parallel, no interaction exists between genes i and j. If the lines on
the interaction plots are nonparallel (as shown in Figure 10.3(b)), interaction oc-
curs. If the lines cross (as shown in Figure 10.3(c)), strong interaction occurs. The
actual amount of interaction between genes i and j can be determined by the num-
ber of intersections on the interaction plot.
If strong interaction does not exist in any of the gene pairs, then the main ef-
fects on genes can be separated out. The first child is formed by the combination
of the genes with the best main effects based on (10.5). However, if strong interac-
tion does exist in any one of the gene pairs, the first child is formed in two parts:
The first part is the genes which do not carry any strong interaction between each
other and the second part that in which the genes carry strong interaction between
each other. In the first part, the level combination is formed by the genes with the
best main effects based on (10.5). For the second part, the level combination of the
genes, which gives the best fitness value, is chosen. Assume that strong interaction
exists between genes i and j. The best level combination of genes i and j is
given by:
⎡⎣ Best ( i ) , Best ( j ) ⎤⎦ = arg ( max

m , n =1,2,...Q
( I ( m, n ) ) )
ij (10.10)
where i, j =1,2,…N but i ≠ j . 'arg(max(…))' is a function that returns the indices

of the maximum value of the matrix.
If strong interaction exists both between gene i and j and between gene j and k,
and the estimated interaction between i and j is larger than the one between j and
k, then the gene pair of i and j will be selected, and the best level combination of
gene i and j are given by equation (10.10). Otherwise, the gene pair of j and k
are selected, and the best level combination of gene j and k are given by
equation (10.10).
The second child is identical to the first child except that the gene with the low-
est main effect difference in the other level is chosen. The main effect difference
of the genes can be found by equation (10.6).
The detailed steps of IC are as follows:
Algorithm 10.3: Interaction crossover (IC)
Step 1- Step 6: Step 1 to Step 6 are identical to Step 1 to Step 6 of Algo-
rithm 10.2.
Step 7: Construct the interaction matrix MI ij by (10.8), where
i,j=1,2,…N with i ≠ j .
Step 8: Construct the interaction plot for MI ij by using the lines given
by (10.9), where i, j=1,2,…N with i ≠ j .
Step 9: Identify whether or not the genes i and j have a strong interaction,
by checking whether any intersection exists on the interaction
plot, where i, j=1,2,…N with i ≠ j .
Step 10: The first child is formed in two parts. The first part is formed by
the genes without carrying any strong interaction based on
(10.5). The second part is formed by the genes that carry strong
interaction based on (10.10).
Step 11: The second child is formed by performing Step 9 and Step 10 in
Algorithm 2.
10.4 A Case Study: Car Door Design

In this section, the case study of the optimization of a car door design [Kim et al.
2000] was used to validate the effectiveness of ICEA. In the car door design, a
fuzzy optimization model was developed which contains the following engineer-
ing requirements (i.e. X=x1, x2,…, x6) and customer requirements (i.e. Y=y1, y2,…,
y5) as illustrated in Figure 10.4.
10.4 A Case Study: Car Door Design 187
x3 x4 x5
x2
x1 x6
Car door design
y1 y5
y2 y3 y4
Fig. 10.4 Car door design
x1 – energy to close the door y1 – easy to close from outside

x2 – check force on level y2 – stays open on a hill
ground y3 – rain leakage
x3 – check force on 10% slope y4 – road noise
x4 – door seal resistance y5 – cost
x5 – road noise reduction
x6 – water resistance
A fuzzy optimization model for the car door design is formulated as shown below:
Maximize λ
subject to
λ ≤ μ yi ( X ), i = 1, 2, ..., 5 (10.11)
λ ≤ μ f i ( X , Y ), i = 1, 2, ..., 5
λ ≤ μ g j ( X , Y ), j = 1, 2, ..., 6
λ ≤ μc (X )
where
Y=(y1,y2,…,y5);
X=(x1,x2,…,x6);
λ (0 ≤ λ ≤ 1) represents the overall value of membership functions, or overall
degree of satisfaction with performance characteristics achieved at a design X;
membership function μ yi ( X ) can be represented as:
⎧0 if y i ( X ) ≤ y imin
⎪
μ yi ( X ) = ⎨τ ( X ) if y imin ≤ y i ( X ) ≤ y imax (10.12)
⎪1 if y i ( X ) ≥ y imax
⎩
with the linear or non-linear fuzzy function τ ( X ) , and y imin and yimax
represent the lower and upper bounds of aspirations with respect to y i
respectively.
the membership function of the fuzzy relationship constraints respectively
are μ fi ( X , Y ) , μ g j ( X , Y ) , where yi=fi(x1,…,x6)and xj=gj(x1,..,xj-1,xj+1,…,x6)
with i=1,2…,5 and j=1,2,…6. The membership functions of a fuzzy con-
straint “ AX = b ”[52] can be represented as:
⎧0 if AX ≤ b − d or AX ≥ b + d
⎪
⎪ AX − b (10.13)
μ ( X ) = ⎨1 − if b − d < AX < b + d
⎪ d
⎪1 if AX = b
⎩
with the row vector A, the constant b and a chosen constant of admissible
violations of the constraint d.
The membership of the cost constraint μ c ( X ) can be represented in the
following form:
⎧1 if CX < c
⎪ CX − c
⎪
μ c ( X ) = ⎨1 − if c ≤ CX ≤ c + t (10.14)
⎪ t
⎪⎩0 if CX > c + t
where t is a pre-specified, non-negative tolerance level to the cost c. By solving

the above fuzzy optimization model, an optimal target value setting of the engi-
neering requirements can be obtained. A detailed description of the formulation of
the fuzzy optimization model is outside the scope of this book. For details, the
readers can refer to the previous work [Bai and Kwong 2003]. This is a non-
separable problem since interactions between both engineering requirements
(i.e. X=x1, x2,…, x6) and customer requirements (i.e. Y=y1, y2,…, y5) are
unavoidable.
A toolbox for the classical evolutionary algorithm coded in Matlab [Chipper-
field et al. 1994, 1995] was employed to investigate the performance of the or-
thogonal array-based crossovers (i.e.: OC, MC and IC), which were embedded in
the classical evolutionary algorithm. We set up and carried out the experimental
work regarding the following settings and configurations that can be classified
into two types: 1) orthogonal array based evolutionary algorithm embedded with
orthogonal array-based crossover, and 2) standard evolutionary algorithm embed-
ded with standard crossover:
1) The three versions of orthogonal array based crossovers (i.e. OC, MC and IC)
embedded in the above classical evolutionary algorithm [Chipperfield et al.
1994, 1995] have been tested. They are called orthogonal array-based evolu-
tionary algorithms:
• The first version is the orthogonal array-based evolutionary algo-
rithm (OCEA). The basic process of OCEA is identical to the clas-
sical evolutionary algorithm except that the crossover utilizes the
orthogonal crossover operator (OC) as discussed in Section 10.2.1.
• The second version is the orthogonal array-based evolutionary algo-
rithm (MCEA). The basic process of MCEA is identical to the clas-
sical evolutionary algorithm except the crossover utilizes the main
effect crossover operator (MC) as discussed in Section 10.2.2.
• The third version is orthogonal array-based evolutionary algorithm
(ICEA). The basic process of ICEA is identical to the classical evo-
lutionary algorithm except the crossover utilizes the interaction
crossover operator (IC) as discussed in Section 10.2.3.
( )
An orthogonal array L9 34 [Taguchi and S. Konishi 1987] has been used in the
three orthogonal array based crossover operators (i.e. OC, MC and IC) in all three
orthogonal array based evolutionary algorithms (i.e. OCEA, MCEA and ICEA).
2) Two standard evolutionary algorithms (SEAs) have been tested.
– The first version is the standard evolutionary algorithm (SEA1).
The basic process of SEA1 is identical to that of the classical evolu-
tionary algorithm [Chipperfield et al. 1994, 1995]. The standard
three-point crossover is used in SEA1 because three crossover
points are produced by the three orthogonal array-based crossovers
( )
(i.e. OC, MC and IC) with L9 34 . To unite the number of cross-
over points, three crossover points are used in the crossover opera-
tor in SEA1.
– The second version is the standard evolutionary algorithm two
(SEA2). The basic process of SEA2 is identical to that of the classi-
cal evolutionary algorithm [Chipperfield et al. 1994, 1995] except
for the crossover.
In the orthogonal array-based crossovers (i.e.: OC, MC and IC),
two parents are selected randomly from the population. Then nine
potential offspring are produced based on the combinations of the
( )
orthogonal array L9 34 . In OC, the two resulting children are pro-
duced by selecting two best potential offspring from among the
nine. In MC, the two children are produced by analyzing the main
effects of the genes of the nine offspring. In IC, the two children are
produced by analyzing both the main effects of the genes and the
interactions between the genes of the nine offspring. Therefore, ex-

tra selective pressure is created by the three orthogonal array-based
crossovers (i.e.: OC, MC and IC).
To investigate how the extra selective pressure influences the
performance of orthogonal array-based evolutionary algorithms, a
crossover operator with a parent tournament selection of nine is
used in SEA2. In the crossover operator, nine chromosomes are se-
lected randomly from the population. Then the standard three-point
crossover is performed on the two chromosomes with the best fit-
ness among the nine selected chromosomes, and two children are
generated for the next generation.
The following parameter values and scheme in the five evolutionary algorithms
(i.e. OCEA, MCEA, ICEA, SEA1 and SEA2) have also been adopted. The pre-
defined number of function evaluations in all algorithms is the same, which was
set as 80 000, which is the same as the one used in [Bai and Kwong 2003] for
solving the problem. The real coded representation was used in all algorithms. The
parameters of the crossover rate and mutation rate were kept constant and their
values were taken from [Leung and Wang 2001].
For the mutation rate, 1/11 is used where the number of design attributes for the
car door design is 11. The mutation operator of Gaussian perturbation of individu-
al variables was used in all algorithms. For the crossover rate, 0.1 was used in the
three orthogonal array-based evolutionary algorithms (i.e.: OCEA, MCEA and
ICEA) and 1.0 was used in the two SEAs (i.e.: SEA1 and SEA2).
The value of the crossover rate used in orthogonal array based evolutionary al-
gorithms is smaller than the one used in SEAs, since the orthogonal array based
( )
crossover operators (i.e.: OC, MC and IC) are using L9 34 to produce nine po-
tential offspring. With this value of crossover rate, the three orthogonal array-
based evolutionary algorithms can generate a reasonable number of potential
offspring in each generation. A population size of 100 and selective pressure of
1.5 are used in all algorithms, where 1.0 is the minimum selective pressure and 2
is the maximum selective pressure, thus the middle selective pressure 1.5 was used
[Baker 1985, 1987].
The evolutionary algorithms, SEA1, SEA2, OCEA, MCEA and ICEA, were
used to solve the optimization problem of determining the target values of the car
door design. This is modelled in (10.11). These algorithms are coded in Matlab.
To conduct a more comprehensive comparison, a genetic algorithm was developed
which integrated with a gradient search operator proposed by [Bai and Kwong
2003] .This algorithm was recoded in Matlab again to solve the problem. We call
Bai and Kwong’s algorithm BEA in this paper. In these evolutionary algorithms,
the population consisted of a set of real coded chromosomes in which 11 variables
are in each chromosome. The t-th chromosome in all evolutionary algorithms is
represented as:
Z (t ) = ⎡⎣ X ( t ) , Y ( t ) ⎤⎦ ,
where X (t ) = [ x1 (t ), x 2 (t ), x 3 (t ), x 4 (t ), x5 (t ), x 6 (t )] and Y (t ) = [ y1 (t ), y 2 (t ), y 3 (t ),
y 4 (t ), y 5 (t )] ; t=1,2,…,Popsize and Popsize is the total number of chromosomes in
the population. The t-th chromosome Z (t ) in the evolutionary algorithms is eva-
luated by the following fitness function the aim of which is to optimize the cost
function (10.11):
fitness (Z (t ) ) = fitness ( X (t ), Y (t ))
{ }
= min μ yi ( X (t )), μ fi ( X (t ), Y (t )), μ gi ( X (t ), Y (t )), μ c ( X (t ))
where i=1,2,…,5; j=1,2,…,6.

Since all evolutionary algorithms are stochastic algorithms, different solutions
can be obtained with different runs. The better the evolutionary algorithm, the
larger is the mean and the smaller is the variance of overall customer satisfaction
obtained in different runs. Therefore, 100 testing runs were performed to collect
the two statistics of the means and variances of overall customer satisfaction.
These are detailed in Figures 10.5 and 10.6 together with the six algorithms. It can
be found from Figure 10.5 that ICEA achieves the largest mean of overall custom-
er satisfaction among the six algorithms, and also from Figure 10.6, that the stan-
dard deviation of overall customer satisfaction with ICEA is the smallest one.
Therefore, ICEA can yield the best and most robust solutions compared with those
of the other five evolutionary algorithms.
Solution qualities of the evolutionary algorithms

Mean of overall custom er
0.78
0.76
satisfactions
0.74
0.72
0.7
0.68
0.66
BEA SEA1 SEA2 OCEA MCEA ICEA
Evolutionary algorithms
Fig. 10.5 Means of overall customer satisfaction of runs found by the evolutionary
algorithms
Std. of overall customer Robustness of the evolutionary algorithms
0.16
0.14
satisfactions
0.12
0.1
0.08
0.06
0.04
0.02
0
Evolutionary algorithms
Fig. 10.6 Standard deviations of overall customer satisfaction of runs found by the evolu-
tionary algorithms
The t-test is then used to evaluate the significance level at which the ICEA is bet-
ter than the other evolutionary algorithms in this validation test. The t-values be-
tween ICEA and the other evolutionary algorithms are shown in Figure 10.7, which
shows that all t-values are higher than 2.15. Based on the normal distribution table,
if the t-value is higher than 2.15, one has a significant difference at a 98% confi-
dence level. Since the mean of ICEA is below than that of the other algorithms, one
can conclude that the performance of ICEA is significantly better than those of the
other five evolutionary algorithms with 98% confidence in solving this problem.
T-values between the two evolutionary algorithms
6
5
4
t-values
3
2
1
0
BEA-ICEA SEA1-ICEA SEA2-ICEA OCEA-ICEA MCEA-ICEA
Evolutionary algorithm pairs
Fig. 10.7 T-values between ICEA to other evolutionary algorithms

After performing the validation test, the convergence plots of all evolutionary
algorithms averaged over the 100 runs are shown in Figure 10.8. The figure shows
the progress of the evolutionary algorithms through the searches. It can be ob-
served clearly from the figures that in general the convergence speeds of the or-
thogonal array-based evolutionary algorithms, OCEA, MCEA and ICEA, are in
general faster than the other three evolutionary algorithms BEA, SEA1 and SEA2.
Finally, it is also obvious that ICEA can produce better solutions than those of the
other five evolutionary algorithms.
The convergence plots of the algorithms

0.8
0.7
0.6
ICEA
MCEA
0.5 OCEA
fitness value
SEA2
0.4
BEA
0.3
0.2
SEA1
0.1
0
0 1 2 3 4 5 6 7 8
number of evaluations 4
x 10
Fig. 10.8 Convergence curves of the evolutionary algorithm for solving the problem of de-
termining the target values in car door design
However, solely from the convergence curves, it is difficult to ascertain the

computational effort the algorithms needed in order to reach the acceptable solu-
tions. In [Bai and Kwong 2003], it has already been demonstrated that BEA can
produce the most acceptable solutions for this problem. Table 10.1 shows the
computational times used (in seconds) on all algorithms, that can reach the accept-
able solutions found by BEA. It can also be found from Table 10.1 that ICEA can
reach acceptable solutions in the shortest computational time compared with the
other five algorithms. It also shows that ICEA used less than half the computa-
tional effort to reach the acceptable solutions than did the BEA [3], even through
the number of operations used in IC is larger than the other orthogonal array based
evolutionary algorithms OCEA and MCEA, and the standard evolutionary algo-
rithms SEA1 and SEA2.
Table 10.1 Computational time (in seconds) used by the algorithms (i.e. BEA, SEA1,
SEA2, OCEA, MCEA and ICEA) until the acceptable solution reached

Computational 61.2400 44.69 41.61 27.846 23.11 22.21
time taken to
reach the ac-
ceptable solu-
tion
Recall that the steps of the algorithms are similar, except that different opera-
tors are used. In BEA, the gradient search operator is used. In both SEA1 and
SEA2, both three-point crossovers, one suppressed with normal selective pressure
and one suppressed with high selective pressure, are used. In OCEA, OC is used.
In MCEA, MC is used. In ICEA, IC is used. These results indicate that IC
can help the evolutionary algorithm to give the best mean solution quality and
more robust solutions with the shortest computational time compared with the
other algorithms.
10.5 Conclusion
This chapter presented three versions of orthogonal array-based crossover which is
intended to enhance the effectiveness of classical crossover when exploring solu-
tions for product design. The effectiveness of the orthogonal array-based crossov-
er is evaluated by addressing the design problem regarding the design of a car
door. Results show that the three versions of orthogonal array-based crossovers
outperform the classical crossover in solving this product design problem. The
most current version of orthogonal array-based crossover (IC) was found to yield
better results in terms of quality and stability compared with those based on the
evolutionary algorithms embedded with the other two versions of orthogonal ar-
ray-based crossover (OC and MC), which ignore interactions between design
attributes. Referring to the statistical results of the t-test, it can be confirmed that
IC significantly outperforms the other algorithms involved in the validation test.
Also, IC can provide acceptable solutions with faster convergence speeds and
smaller computational effort compared with the other algorithms that were tested.
After the optimal design attributes of a new product have been specified, the
next step is to manufacture the new product through the production line. To ensure
that the specified design attributes of the product can be realized, manufacturers
need to control variability at each of the many processing steps in the production
line. Also, all of the variables that control the desired quality of the design
attributes of a new product need to be understood and optimized if tight control is
to be maintained. The following two chapters will discuss the use of computation-
al intelligence methods to ensure that the desired quality of new products with
lower production cost can be achieved in practice.
References 195
References
Box, G.E.P., Hunter, W.G., Hunter, J.S.: Statistics for Experimenters. John Wiley (1978)
Bai, H., Kwong, C.K.: Inexact genetic algorithm approach to target values setting of engi-
neering requirements in QFD. International Journal of Production Research 41(16),
3861–3881 (2003)
Baker, J.E.: Adaptive selection methods for genetic algorithms. In: Proceedings of the First
International Conference on Genetic Algorithms, pp. 101–111 (1985)
Baker, J.E.: Reducing bias and inefficiency in the selection algorithm. In: Proceedings of
the Second International Conference on Genetic Algorithms, pp. 14–21 (1987)
Bonissone, P.P., Subbu, R., Eklund, N., Kiehl, T.R.: Evolutionary algorithms + domain
knowledge = real-world evolutionary computation. IEEE Transactions on Evolutionary
Computation 10(3), 256–280 (2006)
Chan, K.Y., Emin Aydin, M., Fogarty, T.C.: A Taguchi method-based crossover operator
for the parametrical problems. In: Proceedings of the IEEE International Congress on
Evolutionary Computation, pp. 971–977 (2003)
Chan, K.Y., Kwong, C.K., Jiang, H., Aydin, M.E., Fogarty, T.C.: A new orthogonal array
application to car door design. Expert Systems Applications 37(5), 3853–3862 (2010)
Chipperfield, A.J., Fleming, P.J., Fonseca, C.M.: Genetic Algorithm Tools for Control Sys-
tems Engineering. In: Proceedings of Adaptive Computing in Engineering Design and
Control, pp. 128–133 (1994)
Chipperfield, A.J., Fleming, P.J.: The MATLAB genetic algorithm toolbox. In: Proceedings
of the IEE Colloquium on Applied Control Techniques using MATLAB, pp. 10/1–10/4
(1995)
Cvetkovic, D., Muhlenbein, H.: The optimal population size for uniform crossover and
truncation selection, in Technical Report GMD-AS-TR-94-11, St Augustine, Germany
(1994)
Davidor, Y.: Epistasis variance: a viewpoint on GA-hardness. In: Rawlins, G.J.E. (ed.)
Foundations of Genetic Algorithms. Morgan Kaufmann, San Mateo (1991)
Dimopoulos, C., Zalzala, A.M.S.: Recent developments in evolutionary computation for
manufacturing optimization: problems, solutions, and comparisons. IEEE Transactions
on Evolutionary Computation 4(2), 93–113 (2000)
Davision, E.J.: Benchmark problems for control system design. International Federation of
Automatic Control (May 1990)
Ho, S.Y., Shu, L.S., Chen, H.M.: Intelligent genetic algorithm with a new intelligent cros-
sover using orthogonal arrays. In: Proceedings of the Genetic and Evolutionary Compu-
tation Conference, vol. 1, pp. 289–296 (1999)
Ho, S.Y., Shu, L.S., Chen, J.H.: Intelligent evolutionary algorithms for large parameter op-
timization problems. IEEE Transactions on Evolutionary Computation 8(6), 522–541
(2004)
Ho, S.Y., Chen, H.M., Ho, S.J., Chen, T.K.: Design of accurate classifiers with a compact
fuzzy-rule base using an evolutionary scatter partition of feature space. IEEE Transac-
tions on Systems, Man and Cybernetics –Part B: Cybernetics 34(2), 1031–1044 (2004)
Ho, S.Y., Chen, J.H., Huang, M.H.: Inheritable genetic algorithm for bi-objective 0/1 com-
binatorial optimization problems and it applications. IEEE Transactions on Systems,
Man and Cybernetics –Part B: Cybernetics 34(1), 609–620 (2004)
Ho, S.J., Ho, S.Y., Hung, M.H., Shu, L.S., Huang, H.L.: Designing structure-specified
mixed H2/H¥ optimal controllers using an intelligent genetic algorithm IGA. IEEE
Transactions on Control Systems Technology 13(6), 1119–1124 (2005)
Ho, S.Y., Chen, H.M.: A GA-based systematic reasoning approach for solving traveling sa-
lesman problems using an orthogonal array crossover. In: Proceeding of the Fourth In-
ternational Conference on High Performance Computing in the Asia Pacific Region,
vol. 2, pp. 659–663 (2000)
Ho, S.Y., Chen, H.M.: An efficient evolutionary algorithm for accurate polygonal approxi-
mation. Pattern Recognition 34, 2305–2317 (2003)
Huang, H.L., Ho, S.Y.: Mesh optimization for surface approximation using an efficient
coarse-to-fine evolutionary algorithm. Pattern Recognition 36, 1065–1081 (2003)
KrishnaKumar, K., Narayanaswamy, S., Garg, S.: Solving large parameter optimization
problems using a genetic algorithm with stochastic coding. In: Winter, G., Périaux, J.,
Galán, M., Cuesta, P. (eds.) Genetic Algorithms in Engineering and Computer Science.
Wiley, New York (1995)
Kwong, C.K., Chan, K.Y., Aydin, M.E., Fogarty, T.C.: An orthogonal array based genetic
algorithm for developing neural network based process models of fluid dispensing. In-
ternational Journal of Production Research 44(12), 4815–4836 (2006)
Khuri, A.I., Cornell, J.A.: Response Surfaces Design and Analysis. Marcel Dekker, Inc.,
New York (1996)
Kim, J.D., Choi, M.S.: Stochastic approach to experimental analysis of cylindrical lapping
process. International Journal of Machines Tools Manufacturing 35(1), 51–59 (1995)
Kim, K., Moskowitz, H., Dhingra, A., Evans, G.: Fuzzy multicriteria models for quality
Leung, Y.W., Wang, Y.: Multiobjective programming using uniform design and genetic al-
gorithm. IEEE Transactions on Systems, Man, and Cybernetics – Part C: Applications
and Reviews 30(3), 293–304 (2000)
Leung, Y.W., Wang, Y.: An orthogonal genetic algorithm with quantization for global nu-
merical optimization. IEEE Transactions on Evolutionary Computation 5(1), 41–53
(2001)
Lin, Y.H., Tyan, Y.Y., Chang, T.P., Chang, C.Y.: An assessment of optimal mixture for
concrete made with recycled concrete aggregates. Cement and Concrete Research 34,
1373–1380 (2004)
Mohan, N.S., Ramachandra, A., Kulkarni, S.M.: Influence of process parameters on cutting
force and torque during drilling of glass fiber polyester reinforced composites. Compo-
site Structures 71, 407–413 (2005)
Montgomery, D.C.: Design and Analysis of Experiments. John Wiley and Sons, Inc., New
York (1997)
Muhlenbein, H.: How genetic algorithms really work - Part I: Mutation and hill climbing.
In: Proceedings of the 2nd International Conference on Parallel Problem Solving from
Nature, pp. 15–25 (1992)
Phadke, M.S.: Quality engineering using robust design. Prentice Hall, New York (1987)
Reeves, C.R.: Predictive measures for problem difficulty. In: Proceedings of the 1999 Con-
gress on Evolutionary Computation, vol. 1, pp. 736–742 (1999)
Taguchi, G., Konishi, S.: Orthogonal Arrays and Linear Graphs. American Supplier Insti-
tute, Dearborn (1987)
References 197
Tsai, J.T., Liu, T.K., Chou, J.H.: Hybrid Taguchi-genetic algorithm for global numerical
optimization. IEEE Transactions on Evolutionary Computation 8(4), 365–377 (2004)
Unal, R., Stanley, D.O., Joyner, C.R.: Propulsion system design optimization using the Ta-
guchi Method. IEEE Transactions on Engineering Management 40(3), 315–322 (1993)
Whitley, D.: The genitor algorithm and selective pressure: why rank-based allocation of re-
productive trials is best. In: Proceedings of the Third International Conference on Genet-
ic Algorithms, pp. 116–121 (1989)
Whitley, D., Mathias, K., Rana, S., Dzubera, J.: Building better test function. In: Proceed-
ings of the 6th International Conference on Genetic Algorithms, pp. 239–246 (1995)
Yao, X., Lin, Y., Lin, G.: Evolutionary programming made faster. IEEE Transactions on
Evolutionary Computation 3(2), 82–102 (1999)
Zhang, Q., Leung, Y.W.: An orthogonal genetic algorithm for multimedia multicast
routing. IEEE Transactions on Evolutionary Computation 3(1), 53–62 (1999)
Zimmermann, H.J.: Fuzzy Set Theory and Its Applications, 3rd edn. Kluwer, Boston (1996)
Chapter 11
A Nonlinear Fuzzy Regression for Developing
Manufacturing Process Models
11.1 Introduction
It is well recognized that manufacturing concerns need to be considered in product
design stage such that quality of manufactured products can be improved and their
production cost can be reduced. To address these concerns, one common method
is to develop manufacturing process models that relate the quality requirements of
a new product to the variables of manufacturing processes. Based on the models,
proper settings of process parameters and the predicted quality of new products
can be obtained in the product design stage.
Fuzzy regression has demonstrated its ability to model manufacturing processes
where the processes have fuzziness and the number of experimental data sets for
modeling them is limited. However, previous studies yield only fuzzy linear re-
gression based process models which do not address variables or higher order
terms. In fact, it is widely recognized that in manufacturing processes, there are
often interactions among variables or higher order terms. In this chapter, a ge-
netic programming based fuzzy regression, namely GP-FR, is proposed for model-
ing manufacturing processes. The proposed method uses the general outcome of
GP to construct models whose structure of which is based on a tree representation,
which could carry interaction and higher order terms. Then, a fuzzy linear regres-
sion algorithm is used to estimate the contributions and the fuzziness of each
branch of the tree, so as to determine the fuzzy parameters of genetic program-
ming based fuzzy regression model.
The effectiveness of the GP-FR for process modeling is evaluated, by applying
it to the modeling of a solder paste dispensing process. Results are compared with
those based on statistical regression and fuzzy linear regression. Of the three mod-
els, it was found that the GP-FR can achieve better goodness-of-fit. Also, the pre-
diction accuracy of the model developed using the GP-FR is better than that
achieved by the other two methods.
The rest of this chapter is organized as follows. Section 11.2 defines and dis-
cusses the formulation of nonlinear fuzzy regression and discusses how genetic
programming can be used to generate nonlinear fuzzy regression models. In Sec-
tion 11.3, a case study of solder paste dispensing process is used to illustrate the
effectiveness of the proposed nonlinear fuzzy regression approach. Section 11.4
provides the conclusion.
200 11 A Nonlinear Fuzzy Regression for Developing Manufacturing Process Models
11.2 Nonlinear Fuzzy Regression

The general form of the fuzzy regression models, which involves interactions be-
tween variables and higher order terms, can be represented as:
y = f0 + ∑ fi ( xi ) + ∑∑ fij ( xi , x j ) + ... + f1,2,... N ( x1 , x2 ,...xN )

N N N
(11.1)
i =1 i =1 j =1
in which f0 is a fuzzy bias term and fi ( xi ) , fij ( xi , x j ) , …represent a univariate
fuzzy component, and a bivariate fuzzy component, … respectively (Friedman
1991). A higher order high-dimensional Kolmogorov-Gabor polynomial (Gabor
et al. 1961) is one of the forms of (11.1), which can be written as:
N N N N N d
y = fNR ( x ) = A0 + ∑ Ai1 xi1 + ∑∑ Ai1i2 xi1 xi2 + ...∑ ...∑ Ai1 ...id ∏ x j (11.2)
i1 =1 i1 =1 i2 =1 i1 =1 id =1 j =1
where A0 = (α 0 , c0 ) , A1 = (α1 , c1 ) , A2 = (α 2 , c2 ) , …

A N = (α N , cN ) , A11 = (α11 , c11 ) , A12 = (α12 , c12 ) ,… A NN = (α NN , cNN ) ,…
A N ... N = (α N ... N , cN ... N ) .
The fuzzy regression model (11.2) can be rewritten as:
y = A ' x ' + A ' x ' + A ' x ' ... A ' x '
0 0 1 1 2 2 N NR N NR (11.3)
or y = ( c '0 ,α '0 ) x '0 + ( c '1 ,α '1 ) x '1 + ( c '2 ,α '2 ) x '2 + ... c 'N NR ,α 'N NR x 'N NR( ) (11.4)
where 1+NNR is the number of terms of (11.1), (11.2) and (11.3); A '0 = A0 ,
A '1 = A1 , A '2 = A2 ,… A 'N NR = A N ... N ; x '0 = 1 , x '1 = x1 , x '2 = x2 , …
x 'N NR = x1 ⋅ x2 ⋅ ...xd ; and A '0 = ( c0 , α 0 ) , A '1 = ( c '1 ,α '1 ) , …
( )
A 'N NR = c 'N NR ,α 'N NR . A 'i and x 'i are called the fuzzy parameters and the trans-
formed variables respectively, where i=0,1,2,…NNR.
The vectors of the fuzzy parameters are defined as:
( ) ( (
A ' = A '0 , A '1 ,... A 'N NR = ( c '0 , α '0 ) , ( c '1 ,α '1 ) ,... c 'N NR ,α 'N NR )) , (11.5)
c ' = ( c ' , c ' ,...c ' ) ,

0 1 N NR (11.6)
and α ' = (α ' ,α ' ,...α ' ) .
0 1 N NR (11.7)
The vector of the transformed variables is defined as:

(
x ' = x '0 , x '1 , x '2 ,...x 'N NR . ) (11.8)
Using the vectors of the fuzzy parameter and the vector of transformed variables,
(11.3) can be rewritten as:
y = A '⋅ x 'T (11.9)
11.2 Nonlinear Fuzzy Regression 201
Figure 11.1 shows a fuzzy regression model which contains all samples within the
nonlinear polynomial intervals.
α 'T x '
α 'T x '+ c 'T x '
α 'T x '− c 'T x '

x
'
Fig. 11.1 The nonlinear fuzzy regression
Since some terms in (11.3) may be redundant, prudent selection of significant

terms or orders is advisable if a more parsimonious and adequate model is desired.
In this chapter, the genetic programming based fuzzy regression (GP-FR) ap-
proach, is proposed and descriptions of it are given. The pseudocode for the genet-
ic programming based fuzzy regression (GP-FR) is shown below.
t=0
Initialize Ω(t)=[θ1(t), θ2(t),… θPOP(t)]
Assign fuzzy parameters to all θi(t)
// Ω(t) is the population of the t-th generation.
// θi(t) is the i-th individual of Ω(t).
Evaluate all θi(t) according to a fitness function
Parent Selection Ω(t+1)
Crossover Ω(t+1)
Mutation Ω(t+1)
Determine fuzzy parameters in all θi(t+1) by using
Tanaka’s fuzzy regression
Evaluate all θi(t+1)
Ω(t)= Ω(t+1)
t=t+1
}
The GP-FR starts by first creating a random initial population Ω(t) with POP
individuals θi(t), while t=0. Each individual θi(t) is in a form of tree structure, that
can be used to represent the structure of the fuzzy regression model as defined in
(11.2). Then the fuzzy parameters are assigned to each individual θi(t) by applying
Tanaka and Watada’s (1988) fuzzy regression. All individuals are evaluated
according to a defined fitness function which is aimed at evaluating the goodness-
of-fit of the fuzzy regression model. The parent selection process uses the good-
ness-of-fit of each individual to determine the selection of potential individuals for
performing crossover or mutation. Finally, new individuals with the determined
fuzzy parameters are evaluated using the fitness function in order to create a new
population Ω(t+1). The process continues until the pre-defined termination condi-
tion is fulfilled. Major aspects of applying the GP-FR on modeling the functional
relationships are described below:
11.2.1 Model Representation

In the GP-FR, one of the most popular methods of representing structures is by us-
ing hierarchical trees which are composed of functions F and terminals T (Koza
1992). The fuzzy regression model (11.2) contains only the three arithmetic opera-
tions, +, - and *, thus F is represented as F = {+, -, *}. The set of terminals T = {x,
p } contains the variable set x={ x1, x2, … xN} of the fuzzy regression model and
{
the fuzzy parameter set p = p 0 , p1 , p 2 ,..., p N NS } of the fuzzy regression model,
where n is the number of variables and NNS is the number of terms of the fuzzy re-
gression model. A potential solution is depicted as a labeled tree with ordered
branches. In the tree, operations from the function set F are used as internal nodes,
and arguments from the terminal set T are used as terminal nodes. For example, a
hierarchical tree has the following formulation:
(x1*x1) - (x2*x2) + (x1*x2*x4)
which is equivalent to:

x12 – x22 + x1 x2 x4
{
The fuzzy parameters set p = p 0 , p1 , p 2 ,..., p N NS } can be obtained after determin-
ing the structure of the function from the tree. The number of fuzzy parameters of
the fuzzy regression model is 4. Therefore, the completed fuzzy regression model
can be represented as follows:
p 0 + p1 ·x12 – p 2 ·x22 + p 3 ·x1 x2 x4,
It can also be represented by:
p 0 + p1 ·x’1 – p 2 ·x’2 + p 3 ·x’3,
where x’1= x12, x’2= x22 and x’3= x1·x2·x4.
11.2 Nonlinear Fuzzy Regression 203
In this chapter, the fuzzy parameters, p 0 , p1 , … p N NS , are determined accord-

ing to Tanaka and Watada’s (1988) fuzzy regression. The following linear pro-
gramming problem is formulated for the fuzzy regression problem with reference
to Tanaka et al. (1982). By solving the linear programming problem, the fuzzy pa-
rameters, p 0 , p1 , … p N NS , can be determined.
N NR
⎛ M
⎞
Minimize J = ∑ ⎜c 'j ∑ x ' (i ) ⎟⎠
j (11.10)
j =0 ⎝ i =1
where M is the number of data sets, and x ' j ( i ) is the j-th transformed variable of
the fuzzy polynomial model of the i-th data set, subject to:
N NR N NR
∑α '
j =0
j x ' j ( i ) + (1 − h)∑ c ' j x ' j ( i ) ≥ y ( i )
j =0
(11.11)
N NR N NR
∑α '
j =0
j x ' j ( i ) + (1 − h)∑ c ' j x ' j ( i ) ≤ y ( i )
j =0
(11.12)
c ' j ≥ 0, α ' j ∈ R, x0 ' ( i ) = 1 for all i, (11.13)

0 ≤ h ≤ 1, i = 1, 2, " M , j = 0, 1, 2, " N NR . (11.14)
J in (11.10) is the total fuzziness of the regression model. The value of h in

(11.11) and (11.12) is between 0 and 1. h refers to the degree to which the fuzzy
linear model fits the given data sets, and is subjectively chosen by decision mak-
ers. Constraints (11.11) and (11.12) impose the restriction that the observation of
the i-th data set y (i ) has at least h degree of belonging to y ( i ) as
μ y (i ) ( y ( i ) ) ≥ h (i = 1, 2, ", M ) . Therefore, the objective of solving the linear pro-
gramming problem (11.10-11.14) is to determine the fuzzy nonlinear parameters
A ' j = ( c ' j ,α ' j ) such that the total vagueness J is minimized subject to
μ y (i ) ( y ( i ) ) ≥ h (i = 1, 2, " , M ) .
11.2.2 Fitness Function

GP-FR evaluates the goodness-of-fit of each individual by using a fitness function,
which is based on the mean absolute error (MAE), and can reflect the differences
between the predicted values of the model and the actual values of the data sets.
The MAE of the j-th individual can be calculated based on (11.15).
1 M y ( k ) − Fj ( x ( k ) )
MAE j = 100% ×
M
∑k =1 y(k )
, (11.15)
where Fj is the fuzzy regression model represented by the j-th individual,

( y ( k ) , x ( k )) = ( y ( k ) , ( x ( k ) , x ( k ) ,...x ( k )) )
1 2 N is the k-th training data set, N is
the number of variables of the training data set, and M is the number of training
data sets used for developing the fuzzy regression model.
(11.15) is commonly known as an indicator of training errors in a model. It re-
flects how well the model can fit the training data sets. However, a model may
contain a lot of unnecessary and complex terms. A complex over-parameterized
model with a large number of parametrical terms reduces the transparency and
ease of interpretation of the model. To prevent the GP-FR from generating models
which are too complex, a fitness function is designed to balance the tradeoff be-
tween the reduction of complexity and model accuracy. In this research, penalty
terms are introduced into the fitness function of the GP-FR (McKay et al. 1997),
and the fitness of the j-th individual is denoted as:
1 − MAE j
fitness j =
(1 + exp ( c ( L − c )))
(11.16)
1 j 2
where fitnessj is the fitness value, Lj is the number of nodes of the j-th individual,
and c1 and c2 are both penalty terms.

Like other evolutionary algorithms, the two main evolutionary operators are cros-
sover and mutation. The crossover operation produces a pair of offspring that in-
herit characteristics from both parents by selecting a random node in each of the
hierarchical tree structures of the parents and exchanging the associated sub-
expressions of the hierarchical tree structures. Because of the dynamic representa-
tion used in GP-FR, typically the parents are different in size, shape and content.
The process of mapping the genotype onto the phenotype does not correspond to a
one-to-one relationship. Therefore, the resulting offspring can be expressed by
more than one different tree structure and some diversification of the population is
allowed. Mutation is performed by randomly selecting a node that is an internal or
terminal node, and by replacing the associated sub-expression with a randomly
generated sub-expression.

After the operations of crossover and mutation, individuals from the current popu-
lation with relatively better fitness are selected to serve as parents for the next
generation. The approach of a roulette-wheel, which is one of the most commonly
used methods for selecting individuals to perform reproduction operations in evo-
lutionary algorithms (Goldberg 1989), is used for the selection of individuals.
11.3 Validation of Genetic Programming 205
Regarding the j th individual, its fitness is assigned a value, fitness j , and the se-
lection probability value, prob j , is defined as:
fitness j
prob j = POP
(11.17)
∑ fitness j
j =1
where POP is the population size of the GP-FR. Equation (11.17) shows that
the individual with the greatest fitness value has a higher probability of being
selected.
After the selection, the population evolves and improves iteratively until a
stopping condition is met. In genetic programming, there are several stopping
conditions. In this research, the stopping criterion is met when the number of gen-
erations is equal to a pre-defined number of generations. Otherwise, the GP-FR
goes on to the next evolutionary iteration.
11.3 Va lidatio n of Genetic Programming
11.3 Validation of Genetic Programming Based Fuzzy

Regression Approach to Modeling Manufacturing
Processes
11.3 Va lidatio n of Genetic Programming
In the manufacture of electronics, solder dispensing machines are used to automat-

ically place a certain amount of solder paste on a printed circuit board. The solder
paste spots must be positioned correctly and must contain the specified amount of
solder paste such that when a surface mounted IC component is placed onto the
solder paste, all of the IC’s leads are aligned correctly with each solder spot.
The motion of the dispensing machine has three axes. The x-axis and y-axis are
used to place the solder paste dispenser on the desired location, and the z-axis is
used to position the tip of the solder paste dispernser at the desired height above
the board. The dispensing machine works as follows:
A series of x, y and z coordinates describing the locations of where the solder
paste is to be deposited is entered into the system’s computer program. For each of
these locations two programmable process parameters are specified: 1) the amount
of solder paste, which is governed by the amount of time that the pump is en-
gaged, and 2) the dwell time, which is the length of time the dispensing system
remains over the location after the pump has been disengaged. A schematic dia-
gram of a solder paste dispensing system is shown in Figure 11.2. The continuous-
ly running motor is connected to a clutch. The output of the clutch drives a screw
pump. The amount of time, that the clutch is engaged, determines the amount of
solder paste deposited, which is called the ‘shot size’. The solder paste exits
through the interchangeable needle. The different solder pastes come prepackaged
in tubular containers, which are inserted in the receptacle adjacent to the motor
and clutch structure.
Motor
Air Gear box and

Clutch
Support
Solder
Frame
Screw Pump
Needle
Solder paste
Fig. 11.2 Solder paste dispensing system
In the process of solder paste dispensing, the key quality characteristic is the di-
ameter of the circular solder pads. The four significant operating parameters (fac-
tors) for the solder paste dispensing process to be studied are: pressure, needle inner
diameter, shot size and dwell time which are represented by x1, x2, x3 and x4 respec-
tively. In the experimental plan, each factor has two levels. Table 11.1 shows the
setting of each level of the factors. Table 11.2 shows the experimental results.
Table 11.1 Factors and their levels
Factor levels
Factors Level 1 (0) Level 2 (1)
x1, Pressure (bar) 0.276 0.827
x2, Needle inner diameter 0.041 0.584
(mm)
x3, Short size (ms) 150 500
x4, Dwell time (ms) 0 500
For model development whether using GP-FR, statistic regression or fuzzy li-
near regression, the four operating parameters x1, x2, x3 and x4 need to be norma-
lized to [0,1], and their resulting values are shown in Table 11.1. Evaluation of the
effectiveness of the models can be carried out by investigating the mean of train-
ing errors and the variance of training errors, which are defined as Re and Rv re-
spectively as follows:
1 M y ( i ) − P ( x1 ( i ) , x2 ( i ) , x3 ( i ) , x4 ( i ) )
Re =
M
∑
i =1 y (i )
× 100% (11.18)
and
1 M ⎡ y ( i ) − P ( x1 ( i ) , x2 ( i ) , x3 ( i ) , x4 ( i ) ) ⎤
Rv = ∑ ⎢
M − 1 i =1 ⎢ y (i )
× 100% − Re ⎥ (11.19)
⎥⎦
⎣
In (11.18) and (11.19), M is the number of experimental data sets; y ( i ) is the i-th
measured value of the solder spot diameter; x1 ( i ) , x2 ( i ) , x3 ( i ) and x4 ( i ) are
parameter values of the i-th experimental data set; and
P ( x1 ( i ) , x2 ( i ) , x3 ( i ) , x4 ( i ) ) is the predicted value for the i-th experiment.
Table 11.2 24 factorial experimental plan and results
Normalization Solder spot

diameters (mm)
j-th experimental x1(j) x2(j) x3(j) x4(j) y(j)
data set
1 0 0 0 0 1.1176
2 1 0 0 0 1.1176
3 0 1 0 0 1.2954
4 1 1 0 0 1.2192
5 0 0 1 0 2.1082
6 1 0 1 0 2.1336
7 0 1 1 0 2.2098
8 1 1 1 0 1.9558
9 0 0 0 1 1.3970
10 1 0 0 1 1.3716
11 0 1 0 1 1.2700
12 1 1 0 1 1.1430
13 0 0 1 1 2.1590
14 1 0 1 1 2.1336
15 0 1 1 1 2.1590
16 1 1 1 1 1.9304
Using the 16 experimental data sets and their results shown in Table 11.2, the
proposed GP-FR was implemented using Matlab to relate solder paste diameter to
the operating parameters x1, x2, x3 and x4. The GP parameters are set as shown
with reference to (Madar et al. 2005): population size = 50; maximum number of
evaluated individuals=5000; generation gap=0.9; crossover rate = 0.5; mutation
rate = 0.5; probability of changing terminal via non-terminal = 0.25; penalty fac-
tors with c1=0.5 and c2=30. Since GP-FR is a stochastic method, different results
will be obtained from different runs. To evaluate its overall performance, 30 runs
on the GP-FR were carried out, and the mean of the 30 runs was calculated.
The model with the smallest mean training error among the 30 runs is shown
below as:
y = (1.1887,0.6985 ) + ( 0.1473,0.0058 ) ⋅ x2 + ( 0.8763,0.0071) ⋅ x3
(11.20)
+ ( 0.1397,0.0005 ) ⋅ x4 − ( 0.1854,0.0008) ⋅ x1 ⋅ x2 − ( 0.2032,0.0003) ⋅ x2 ⋅ x4
Re and Rv were found to be 3.2580% and 0.1285 respectively. Using the same
experimental data sets as shown in Table 2, the following statistical regression
model was determined.
y = 1.2929 − 0.0914 ⋅ x1 − 0.0483 ⋅ x2 + 0.8458 ⋅ x3 + 0.0483 ⋅ x4 (11.21)
The R2 value of the model (11.21) is 96.7%. Re is 4.4575% and Rv is 0.1551.

Based on the same set of data, the following fuzzy linear regression model was al-
so determined.
y = (1.2649,1.4605 ) + ( −0.0711,0.0152 ) ⋅ x1 + ( −0.0457,0.0152 ) ⋅ x2
(11.22)
+ ( 0.8433,0.0051) ⋅ x3 + ( 0.0584,0.0025) ⋅ x4
Re and Rv were found to be 4.9874% and 0.1418 respectively. Table 11.3 summa-
rizes all the Re and Rv of the three methods, the proposed GP-FR, statistical re-
gression and fuzzy regression. From Table 11.3, it can be seen that both the Re
and Rv of the proposed GP-FR are smaller than those of the statistical regression
and fuzzy linear regression. This indicates that the proposed GP-FR can fit the da-
ta sets with the smallest mean of errors and the smallest variance of errors.
Table 11.3 Comparisons of the modeling results
GP-FR Statistical Fuzzy regression

regression
Mean of training 3.2580 5.0291 4.9874
errors Re (%)
Variance of 0.1285 0.1551 0.1418
training errors
Rv
To further validate the modeling performance of the GP-FR, four data sets were
randomly selected from the 16 data sets, as shown in Table 11.3, as testing data
sets and the remaining 12 data sets were used to develop a GP-FR model, a statis-
tical regression model and a fuzzy linear regression model. Their prediction errors
were calculated. The validations were repeated 12 times. We ran the GP-FR 30
times in each validation test and the mean of the 30 runs was calculated. Table 5
summarizes the prediction errors of the three methods. From the table, it can be
seen that GP-FR yields the smallest mean of prediction errors and variance of pre-
diction errors. Results of the 12 validation tests are shown in Figure 11.4 from
which it can be seen that the prediction errors of GP-FR, for all the tests, are the
smallest.
Table 11.4 Prediction errors of the three methods
Validations Testing data set GP-FR Statistical Fuzzy linear

regression regression
1 1 13 15 16 7.7065 8.2668 9.5705
2 1 5 6 12 8.7606 8.9494 9.7422
3 6 12 15 16 8.9340 9.6433 9.6136
4 3 5 10 13 6.4408 6.8848 6.3889
5 4 6 13 16 6.4891 7.0259 7.2404
6 2 8 15 16 6.5076 7.8022 7.9411
7 1 10 11 14 6.7059 12.3621 9.8317
8 1 2 4 15 6.6578 15.0754 14.6359
9 8 12 14 16 9.2822 11.9544 12.0100
10 1 7 11 14 6.4753 9.0530 8.0198
11 4 7 13 16 5.5274 7.4860 6.4493
12 4 6 15 16 5.8662 6.4916 6.6645
Mean of prediction errors 7.1128 9.2496 9.0090
Variance of prediction errors 1.5586 6.0657 6.8675
Prediction errors
Prediction errors (%)
20
15
10
5
0
1 2 3 4 5 6 7 8 9 10 11 12
Testing order
Statistical regression Fuzzy linear regression GP-FR
Fig. 11.4 Prediction errors for each testing order using statistical regression, fuzzy linear
regression and GP-FR
Based on the fuzzy polynomial model (11.20), which relates the solder paste
diameter to the manufacturing process parameters, pressure, needle inner diame-
ter, short size and dwell time, the corresponding manufacturing process parame-
ters can be determined with respect to the specified solder paste diameter. For
example, the engineer needs to set the solder paste diameter at 0.4mm, as shown
in Figure 11.5. Then, the pressure is required to be set based on the following:
⎛ ⎞
⎟⎟ ( (1.1887, 0.6985 ) + ( 0.1473, 0.0058 ) ⋅ x2 + ( 0.8763, 0.0071) ⋅ x3
1
x1 = ⎜
⎜ ( 0.1854, 0.0008 ) ⋅ x
⎝ 2 ⎠
+ ( 0.1397, 0.0005) ⋅ x4 − 0.4 − ( 0.2032, 0.0003) ⋅ x2 ⋅ x4 ) (11.20)
Fig. 11.5 Diameter of the solder paste
11.4 Conclusion
The existing fuzzy regression approaches are not able to produce a model that in-
corporates interaction terms and higher order terms which are important in model-
ing manufacturing processes. In this chapter, a genetic programming based fuzzy
regression (GP-FR) approach is proposed for modeling manufacturing processes,
whereby models can be developed with proper interaction terms and higher order
terms. The proposed GP-FR uses the general outcomes of GP to construct a model
using a tree representation structure. Then, a fuzzy regression algorithm is used to
estimate the contributions and the fuzziness of each branch of the tree so as to de-
termine the fuzzy parameters of each term of the model.
References 211
To validate the proposed GP-FR approach to modeling manufacturing

processes, it has been applied to the modeling of the solder paste dispensing
process, and has been compared with the other commonly used explicit modeling
methods, statistical regression and fuzzy linear regression. The result shows that
the smallest number of training errors can be achieved by GP-FR. This indicates
that GP-FR is more capable of fitting the data sets than are the other two tested
methods. Also, a comparison of the validation results shows that the smallest
number of prediction errors and errors in variance can be achieved by GP-FR. The
achievement of better results can be explained by the introduction of interaction
terms in GP-FR, although the two most commonly used methods ignore these.
In Chapter 12, there follows a discussion of a rule based system that can extract
relevant information from experimental data which is collected from the manufac-
turing process. This knowledge is essential to the engineers if they are to improve
the quality of manufactured products
References
Babets, K., Geskin, E.S.: Application of fuzzy logic for modeling of water jet depainting.
Machining Science and Technology 4(1), 81–100 (2000)
Bates, D.M., Watts, D.G.: Nonlinear Regression Analysis and Its Applications. Wiley, New
York (1998)
Chang, Y.H.O., Ayyub, B.M.: Fuzzy regression methods – a comparative assessment.
Chen, D.X.: Modeling and off-line control of fluid dispensing for electronics packaging.
PhD thesis, University of Saskatchewan (2002)
Chen, Y., Tang, J., Fung, R.Y.K., Ren, Z.: Fuzzy regression based mathematical program-
ming for QFD. International Journal of Production Research 42(5), 3583–3604 (2004)
Chiang, H.H., Hieber, C.A., Wang, K.K.: A unified simulation of the filling and postfilling
stages in injection molding, Part 1: formulation. Polymer Engineering and Science 31,
116–124 (1991)
Friedman, J.H.: Multivariate adaptive regression splines. The Annals of Statistics 19(1), 1–
141 (1991)
Gabor, D., Wildes, W., Woodcock, R.: A Universal non-linear filter, predictor and simula-
tor which optimizes itself by a learning process. Proceedings of IEE 108B, 422–438
(1961)
son-Wesley (1989)
Gray, G.J., Murray-Smith, D.J., Li, Y., Sharman, K.C.: Nonlinear model structure identifi-
cation using genetic programming and a block diagram oriented simulated tool. Elec-
tronic Letters 32, 1422–1424 (1996)
Han, R., Shi, L., Gupta, M.: Three-dimensional simulation of microchip encapsulation
process. Polymer Engineering and Science 40(3), 776–785 (2000)
Holland, J.H.: Adaptation in Natural and Artificial Systems. Michigan Press (1975)
Ip, C.K.W., Kwong, C.K., Bai, H., Tsim, Y.C.: The process modeling of epoxy dispensing
for microchip encapsulation using fuzzy linear regression with fuzzy intervals. Interna-
tional Journal of Advanced Manufacturing Technology 22, 417–423 (2003)
Ip, K.W., Kwong, C.K., Wong, Y.W.: Fuzzy regression approach to modeling transfer
moulding for microchip encapsulation. Journal of Materials Processing Technology 140,
147–151 (2003)
Kang, S.Y., Xie, H., Lee, Y.C.: Physical and fuzzy logic modeling of a flip-chip thermo-
compression bonding process. Journal of Electronic Packaging 115, 63–70 (1993)
Khalil, H.K.: Nonlinear Systems. Prentice-Hall (2002)
(1994)
Kwong, C.K., Bai, H.: Fuzzy Regression Approach to Process Modeling and Optimization
of Epoxy Dispensing. International Journal of Production Research 43(12), 2359–2375
(2005)
Lai, Y.J., Chang, S.I.: A fuzzy approach for Multiresponse optimization: an off-line quality
engineering problem. Fuzzy Sets and Systems 63, 117–129 (1994)
676 (2000)
Li, H.X., Tso, S.K., Deng, H.: A concept approach to integrate design and control for the
epoxy dispensing process. International Journal of Advanced Manufacturing Technolo-
gy 17, 677–682 (2001)
Li, H.L., Chou, T., Chou, C.P.: Optimization of resistance spot welding process using Ta-
guchi method and a neural network. Experimental Techniques 31(5), 30–36 (2007)
(2005)
McKay, B., Willis, M.J., Barton, G.W.: Steady-state modeling of chemical processes using
genetic programming. Computers and Chemical Engineering 21(9), 981–996 (1997)
Schaiable, B., Lee, Y.C.: Fuzzy logic based regression models for electronics manufactur-
ing applications. Advances in Electronic Packaging 1, 147–155 (1997)
Simpson, P.K.: Artificial neural systems. Pergamon Press, New York (1989)
Tong, K.W., Kwong, C.K., Yu, K.M.: Intelligent process design system for the transfer
moulding of electronic packages. International Journal of Production Research 42(10),
1911–1931 (2004)
Willis, M.J., Hiden, H., Hinchliffe, M., McKay, B., Barton, G.W.: Systems modeling using
genetic programming. Computers and Chemical Engineering 21, 1161–1166 (1997)
Xie, H., Lee, Y.C.: Process optimization using a fuzzy logic response surface method. IEEE
Transactions on Components, Packaging and Manufacturing Technology – Part A 17(2),
202–210 (1994)
Zimmermann, H.J.: Fuzzy sets theory and its applications. Kluwer, Boston (1985)
Chapter 12
Rule Extraction from Experimental Data
for Manufacturing Process Design
12.1 Introduction
In order to study a manufacturing process, it is common for engineers to conduct a

large number of experiments and generate experimental data sets. Experimental
data sets must first be processed and/or analyzed in order to extract patterns, use-
ful information or knowledge. The development of effective and efficient methods
for deriving knowledge from these data is important as the knowledge extracted
from the data not only has to have a high predictive accuracy, but also needs to be
understood by users [Fayyad et al. 1996, Freitas 1997, Freitas 1999]. Rule induc-
tion is one of the common forms of data mining [Langlery and Simon 1995]. It is
a method for discovering a set of “IF THEN” rules that can be used for converting
uninformative data into either a knowledge base for decision support or an easily
understood description of the system behavior so that knowledge that humans
can understand can be explored. Moreover, it is able to search for all possible in-
teresting patterns from data sets.
In this chapter, a knowledge discovery system based on a genetic algorithm
(GA) for mining rules from a number of experimental data sets for a manufactur-
ing process is proposed. The fluid dispensing process, which is a very common
process used for die-bonding and microchip encapsulation in electronics packag-
ing, is analyzed using this method to demonstrate the approach. Currently, engi-
neers determine the process parameters for selecting the settings in fluid dispens-
ing, by using their experience and intuitive judgments, which is usually a very
time consuming process. With the use of the rules generated by the knowledge
discovery system, it is hoped that the time needed for identifying proper process
parameter settings can be significantly reduced.
The organization of this chapter is as follows: Section 12.2 introduces the fluid
dispensing for microchip encapsulation in electronic packaging. Section 12.3
presents the operations of the proposed GA based knowledge discovery system for
rule mining. In Section 12.4, validation of the rules generated by the GA-based
knowledge discovery system is performed with the aid of the developed computa-
tional system [Kwong et al. 2009]. Numerical results and discussion are also giv-
en. Section 12.5 concludes the chapter.
214 12 Rule Extraction from Experimental Data for Manufacturing Process Design
12.2 Fluid Dispensing for Microchip Encapsulation

In fluid dispensing processes of microchip encapsulation, normally, silicon chips
are covered with an epoxy encapsulant using an X-Y numerically controlled dis-
pensing system that delivers the epoxy encapsulant through a needle. The material
is commonly dispensed in a pattern, working from the center outwards. An epoxy
dam around the die site and second wire bond points can be made to contain the
flow of material and this produces a more uniform looking part as shown in Figure
12.1. Fluid dispensing is a highly nonlinear process and creates a highly coupled
multi-variable system that involves complex inter-relationships between the epoxy
properties, process conditions, needle design parameters and overall encapsulation
quality. In semi-conductor manufacturing, trial-and-error is still a common me-
thod used to identify appropriate process parameter settings. However, this me-
thod involves a long process development time and optimum encapsulation quality
may not be obtained. A detailed description of fluid dispensing can be found in
[Gilleo 2004].
Bond Wire Encapsulation
Epoxy Dam
IC Chip Substrate
Fig. 12.1 Encapsulation of COB packages
In order to determine the optimal process condition of fluid dispensing, it is

necessary to understand the process’ behavior. Three significant process parame-
ters and their normal operating ranges were identified as follows:
• The compressed air pressure (1 bar to 4 bar), x1

• The height between the substrate and the needle (250 to 2000 steps of a
stepping motor), x2
• The pump motor speed (400 rpm to 1000 rpm), x3 .
Two quality characteristics were studied in this research which are the encapsula-
tion weight (mg), y , and the encapsulation thickness (mm), z. 96 experiments
were carried out based on a full factorial design with 4 levels in compressed air
pressure ( x1 ), 4 levels in the height between the substrate and the needle ( x2 ) and
6 levels in pump motor speed ( x3 ).
12.3 GA-Based Rule Discovery System 215
12.3 GA-Based Rule Discovery System

In this section, a genetic algorithm GA-based knowledge discovery system of the
fluid dispensing process for microchip encapsulation is described, and this is used
to generate rules from the experimental data sets. First, an experimental data set,
involving process parameters and measures of encapsulation, is collected by carry-
ing out experiments on the fluid dispensing process. Then, a knowledge discovery
system that consists of a conjunction of encapsulation requirements and the rules
consequently recommended for searching domains of process parameters, is de-
veloped by the genetic algorithm. Based on the GA-based rule discovery system,
informative rules involving a small searching domain of process parameters can
be recommended with respect to the required encapsulation. The rules generated
can be represented as follows.
if y = yw and z = zw then R1l ≤ x1 ≤ R1u and R2l ≤ x2 ≤ R2u and R3l ≤ x3 ≤ R3u
where yw is the required encapsulation weight; z w is the required encapsulation
thickness; R1l ≤ x1 ≤ R1u is the range of setting of the process parameter x1 ;
R2l ≤ x2 ≤ R2u is the range of setting of the process parameter x2 ; R3l ≤ x3 ≤ R3u is
the range of setting of the process parameter x3 . All the ranges are recommended
by the GA based knowledge discovery system. With a set of training data samples,
Figure 12.2 shows a schematic diagram of the GA-based knowledge discovery sys-
tem. Details of the GA-based knowledge discovery system are described below.
Fig. 12.2 GA-based knowledge discovery system

12.3.1 Generation of Random Strings

The first step in the GA-based knowledge discovery system is to randomly gener-
ate a population of strings which represent the ranges of the process parameters.
The strings can be expressed as ⎡⎣ R1l , R1u , R2l , R2u , R3l , R3u ⎤⎦ , where Ril and Riu
are the lower and upper ranges of the i-th process parameter xi with i=1, 2 and 3
respectively.
Real and binary encoding are two commonly used approaches for string repre-
sentation in GAs. In binary encoding representation, strings need to be encoded to
real values for fitness evaluation and they need to be decoded again for reproduc-
tion operations. However, in real encoding representation, there is no need for
string encoding and decoding. The omission of encoding and decoding can help to
reduce the computational time. Since the ranges of process parameters are all real
values, real encoding is chosen.
12.3.2 Fitness Evaluation

The fitness function of the GA-based knowledge discovery system is used to eva-
luate how well a rule fits the data samples of the epoxy dispensing process. Due to
the limited number of data sets, the required conditions of encapsulation weight
y w and thickness z w are covered by the ranges Ywl ≤ yw ≤ Ywu and
Z wl ≤ zw ≤ Z wu defined by the following rule:
if Ywl ≤ y = yw ≤ Ywu and Zlw ≤ z = zw ≤ Zuw then

R1l ≤ x1 ≤ R1u and R2l ≤ x2 ≤ R2u and R3l ≤ x3 ≤ R3u
where the ranges Ywl ≤ yw ≤ Ywu and Z wl ≤ zw ≤ Z wu covers 10% of the whole op-
erating ranges of the encapsulation weight and encapsulation thickness respective-
ly; and R1l , R1u , R2l , R2u R3l and R3u are the values of the string as discussed in
Section 12.3.1 and they determine the fitness of a rule.
Rules need to be evaluated during the training process in order to establish
points of reference for the GA-based knowledge discovery system. The fitness
function considers the data sets as: correctly classified, still to be classified, and
wrongly classified. In the GA-based rule discovery system, the fitness function
(12.1), which was suggested by Carvalho and Freitas [2000] is used. The fitness
function evaluates the predictive accuracy of a rule based on both true positive
rate and true negative rate that considerably mitigates some pitfalls associated
with the problems of overfitting and lack of balance,
Fitness = true _ positive _ rate × true _ negative _ rate (12.1)

( no. of TP )
where true _ positive _ rate = (12.2)
( no. of TP ) + ( no. of FN )
( no. of TN )
and true _ negative _ rate = (12.3)
( no. of TN ) + ( no. of FP )
with
• TP means True Positive which refers to the data sets covered by the rule
correctly classified;
• FP means False Positive which refers to the data sets covered by the rule
wrongly classified;
• TN means True Negatives which refers to the data sets not covered by
the rule but differing from the training target class;
• FN means False Negatives which refers to the data sets not covered by
the rule but matching the training target class.
With higher numbers of TP and TN, and lower numbers of FP and FN, a better
rule is generated. For a comprehensive discussion about rule-quality measures, the
reader can refer to [Hand 2001].
The following shows a rule generated by the GA knowledge discovery system:
if 65.2 ≤ y = 67 ≤ 68.1 and 0.55 ≤ z = 0.59 ≤ 0.62 then

(12.4)
1 ≤ x1 ≤ 2 and 50 ≤ x2 ≤ 600 and 250 ≤ x3 ≤ 400
where y = 67 and z = 0.59 are the required values of the encapsulation weight
and encapsulation thickness respectively; R1l ( = 1) , R1u ( = 2 ) , R2l ( = 50 ) ,
R2u ( = 600 ) , R3l ( = 250 ) and R3u ( = 400 ) are the values from the string of the GA
based knowledge discovery system. To evaluate the fitness of the rule, the 4 train-
ing data sets as shown in Table 12.1 are used,
Classifications of the training data sets are shown in the last column of Table 1.
• The 1-st data set is classified as FN class, since y=70.1 is not within the
range, 65.2<y<68.1, and also both x1 = 0.8 and x3 = 200 are not within the
ranges, 1 ≤ x1 ≤ 2 and 250 ≤ x3 ≤ 400 . This means the sample is not cov-
ered by the rule, but matches the rule.
• The 2-nd data set is classified as FP class, as y=64.3 and z=0.51 are not
within the ranges 65.2<y<68.1 and 0.55<z<0.62 respectively, but all
x1=1.2, x2=400 and x3=350 are within the ranges, 1<x1<2, 50<x2<600 and
250<x3<400. Therefore the data set is not covered by the rule, but is
wrongly classified as belonging to the target class.
• The 3-rd data set is classified as TP class, since y=66.9 and z=0.57 are all
within the ranges 65.2<y<68.1 and 0.55<z<0.62 respectively, and also all
x1=1.8, x2=350 and x3=300 are within the ranges 1<x1<2, 50<x2<600 and
250<x3<400 respectively. Therefore, the data set is covered by the rule
and is correctly classified.
• The 4-th data set is classified as TN class, since y=65.5 and z=0.61 are all
within the ranges 65.2<y<68.1 and 0.55<z<0.62 respectively, and both
x2=40 and x3=220 are not within the ranges 50<x2<600 and 250<x3<400
respectively. This means the data set is not covered by the rule but differs
from the target class.
In this example, the number of data sets in all FN, FP, TP and TN classes is 1.
Thus based on the fitness function (12.1), the fitness of rule (12.4) can be
calculated as:
Fitness = true _ positive _ rate × true _ negative _ rate

TP TN 1 1
= × = × = 0.25
TP + FN TP + FP 1 + 1 1 + 1
Table 12.1 Training data sets for rule (12.4)
Data sets y z x1 x2 x3 Class

1-st 70.1 0.53 0.8 100 200 FN
2-nd 64.3 0.51 1.2 400 350 FP
3-rd 66.9 0.57 1.8 350 300 TP
4-th 65.5 0.61 1.6 40 220 TN

The population evolves and improves in each generation until a stopping condition
is met. In genetic algorithms, there are quite a few stopping conditions. In this re-
search, the stopping criterion is fulfilled when the number of generations is equal
to a pre-defined number of generations or one of the solutions in the population of
the genetic algorithm achieves a full fitness score of 1. Otherwise, the GA-based
knowledge discovery system performs the selection operation for the next evolu-
tionary generation.
For the selection of strings, the approach of the roulette-wheel is used, which is
one of the most common selection methods for selecting strings to perform repro-
duction operations [Goldberg 1989]. This is unlike other selection approaches
such as rankbased selection [Whitley 1989], tournament selection [Goldberg and
Deb 1991], where their selective pressures need to be controlled by adjusting their
inbuilt parameters. The selection of strings produced by the roulette-wheel selec-
tion algorithm is completely based on the fitness of the strings. Therefore, it can
provide a zero bias to strings in the population.
This selection method imitates the roulette-wheel game, where the thrown dice
would most probably end up by being in the slot with the largest area. Following
this, one can conclude that the string with the largest fitness value is most likely to
be chosen because it has the largest slot size. The fitness value of the j th string in a
population is fit j . The fitness values are used to calculate the probability of selec-
tion, prob j , to the j th string. The probability of selection prob j is defined as:
fit j
prob j = Popsize
(12.5)
∑j =1
fit j
where Popsize is the population size of the GA based knowledge discovery

system.

Discrete Crossover Operation [Muhlenbein and Voosen 1993] is the most com-
mon crossover operation and is performed by exchanging variable values between
parent strings. However, it can generate only corners of the hypercube defined by
the parent strings. Furthermore, experimental results have indicated that the com-
bination of biases is far from optimal and has undesirable side-effects on the
exploratory power of crossover [Eshelmann 1991]. Another common crossover
operation for real encoding representation is Intermediate Crossover [Muhlenbein
and Voosen 1993]. It is capable of producing any point within a hypercube which
is larger than that defined by the parent strings. Therefore, it can be adapted to
sustain a higher explorative search in the searching domain than when using Dis-
crete Crossover Operation.
In the development of the GA-based knowledge discovery system, intermediate
crossover, which can produce a new string around and between the variables of
the two selected parent strings, is used. Referring to the representation of the ge-
netic algorithm, a new string ⎡⎣ R1l , R1u , R2l , R2u , R3l , R3u ⎤⎦ is produced according to
the following rule:
⎡⎣ R1l , R1u , R2l , R2u , R3l , R3u ⎤⎦ = ⎡⎣ 1R1l , 1R1u , 1R2l , 1R2u , 1R3l , 1R3u ⎤⎦ +
(12.6)
{
α ⎣⎡ 1R1l , 1R1u , 1R2l , 1R2 u , 1R3l , 1R3u ⎦⎤ − ⎣⎡ 2 R1l , 2 R1u , 2 R2l , 2 R2u , 2 R3l , 2 R3u ⎦⎤ }
where α is a scaling factor chosen uniformly at random
over an interval [ −0.25, 1.25] , and ⎡⎣ 1R1l , 1R1u , 1R2l , 1R2u , 1R3l , 1R3u ⎤⎦ and
⎡⎣ R , R , R2 , R2 , R3 , R ⎤⎦ are the two selected parent strings. Ranges of
2
1
l 2
1
u 2 l 2 u 2 l 2
3
u
process parameters in the new string are the result of combining the values of the
parent strings according to (12.6) with a scaling factor α chosen for each range
of process parameter. In geometric terms, intermediate crossover is capable of
producing new parameter values within a slightly larger hypercube than that de-
fined by the parent strings, but these values are constrained by a range of scaling
factor α .
Mutation is carried out by randomly changing one or more values of a selected

string between the operating ranges of process parameters. During mutation, the
value of each range of process parameter in a rule has a finite probability of
changing. Therefore, the probability of searching within the operating ranges of
process parameters is never zero. This prevents complete loss of genetic material
through selection and elimination.
The mutation operator of Gaussian perturbation [Muhlenbein and Voosen
1993] of individual variables was used in the GA based knowledge discovery sys-
tem. For example, the variable R j is selected for mutation. After performing the
mutation, its value becomes:
R j ' = R j + MutMx × D j × δ (12.7)
where MutMx = +1 or -1 with equal probability; D j = 0.5 × operating range of

the j-th process parameter; δ = a value in the range [0,1] for shrinking the muta-
tion range based on Gaussian perturbation.
After being generated, the newly produced strings are put into the old popula-
tion to generate a new population. This can be done by replacing the least fit
strings in the old population with the newly produced strings. Such replacement
can also be done by randomly replacing the strings in the old population with the
newly produced strings. This chapter presents a random reinsertion approach.
12.3.5 Rule Induction

Ninety-six experiments were carried out using a full factorial design with 4 levels
in compressed air pressure ( x1 ), 6 levels in pump motor speed ( x2 ) and 4 levels in
the height between the substrate and the needle ( x3 ). 88 out of the 96 experimen-
tal data sets were used to train the GA based knowledge discovery system, and the
remaining 8 experimental data sets were used for system validation.
The GA-based knowledge discovery system was implemented using Matlab
programming software. The parameter settings, Crossover rate = 0.8 and Mutation
rate = 1 / n , where n is the number of variables of the string, suggested by
[Schaffer et al. 1989] were adopted. Since the number of variables of the string is
6 (i.e. n=6), mutation rate was set at 1 / 6 . The number of generations and popula-
tion size were set at 500 and 100 respectively.
If the required encapsulation weight y is 50 mg and the required encapsulation
thickness z is 0.5 mm, Figure 12.3 shows a rule recommended by the GA based
knowledge discovery system. From Figure 12.3, it can be found that the numbers
of TP = 6, FP = 1, FN = 4, TN =77, and the fitness value of the recommended
rule is 0.5923. From the rule, more specified ranges of parameter settings can be
obtained.
12.4 Results Verification 221
Fig. 12.3 The user interface of the GA based knowledge discovery system
12.4 Results Verification

To validate the effectiveness of the rules generated by the GA-based knowledge
discovery system, the computational system for fluid dispensing developed by
Kwong et al. [2009] was employed.
Given operating ranges of process parameters ( x1 , x 2 , x3 ), and the required
encapsulation weight β and thickness γ, the computational system determines the
setting of the three process parameters, compressed air pressure ( x1 ), pump motor
speed ( x 2 ), and the distance between the substrate and needle ( x3 ), based on the
requirements of encapsulation weight β and thickness γ. The system consists of a
neural network (NN) based prediction model, and a GA based optimization unit as
shown in Figure 12.4. Here we call it a pure computational system.
In the GA-based optimization unit, the following objective function is used:
⎛ y−β z −γ ⎞
Objective Function: Min ⎜ λ1 + λ2 ⎟⎟ (12.8)
⎜ β γ
⎝ ⎠
subject to: 1 ≤ x1 ≤ 4 , 400 ≤ x2 ≤ 1000 , 250 ≤ x3 ≤ 2000 ,

where λ1 and λ2 are the weights of the two quality characteristics, encapsulation
weight and encapsulation thickness respectively.
To validate the effectiveness of the GA-based knowledge discovery system, the
system was integrated with the pure computation system developed by Kwong et
al. [2009] as shown in Figure 12.5. In the enhanced computational system, rec-
ommended ranges of parameter settings are generated by the GA-based know-
ledge discovery system and input to the GA-based optimization unit. It is hoped
that the parameter settings recommended by the enhanced computational system
will produce better results in terms of the two quality characteristics than the pure
computational system.
Pure computational system
NN based pre-
Process parameters
lation y and z
diction model
Predicted encapsu-
x1, x2 and x3
Recommended
process
parameters of x1,
Operating para- x2 and x3
meter ranges of GA based op-
x1, x2 and x3 timization unit
Required encap-
sulation β
and γ Relative errors
(y-β)/ β and
(z- γ)/ γ
Fig. 12.4 Pure computational system

Pure computational
Required encapsula-
system
tion β and γ
Predicted encapsulation β and γ

NN based pre-
Process parameters
diction model
GA based
knowledge Recom-
x1, x2 and x3
discovery mended
system process pa-
rameters of
x1, x2 and x3
Recommended
parameter GA based opti-
ranges of x1, x2 mization unit
and x3
Relative er-
rors (y-β)/ β
and (z- γ)/ γ
Fig. 12.5 Enhanced computational system
To validate the GA-based knowledge discovery system, eight validation tests

were carried out. First, eight sets of required encapsulation weights and thick-
nesses as shown in Table 12.2 were input to the GA-based knowledge discovery
system.
Table12.2 Eight sets of required encapsulation weights and thicknesses
Weight Thickness
y z
1 72.3 0.58
2 43.2 0.48
3 87.4 0.67
4 37.2 0.46
5 75.1 0.62
6 59.3 0.57
7 62.4 0.53
8 53.1 0.53
The corresponding eight rules were generated as shown below:

1 IF y=72.3 and z =0.58
THEN 1.6124<x1<2.4414 AND 403<x2<1682.7 AND
493.99<x3<687.3
2 IF y=43.2 and z =0.48
THEN 2.5476<x1<3.268 AND 250<x2<2000 AND
850<x3<1000
3 IF y=87.4 and z =0.67
THEN 1.0561<x1<2.718 AND 644.34<x2<1862.7 AND
400<x3<541.85
4 IF y=37.2 and z =0.46
THEN 1<x1<1.4294 AND 1665.5<x2<2000 AND
819.61<x3<1000
5 IF y=75.1 and z =0.62
THEN 1<x1<2.5259 AND 960.94<x2<1497.2 AND
400<x3<537.3
6 IF y=59.3 and z =0.57
THEN 1.8869<x1<2.8044 AND 250<x2<465.58 AND
548.05<x3<745.13
7 IF y=62.4 and z =0.53
THEN 1.2048<x1<3.9475 AND 250<x2<897.88 AND
773.09<x3<834.33
8 IF y=53.1 and z =0.53
THEN 1.103<x1<2.3456 AND 250<x2<2000 AND
764.92<x3<849.78
These recommended ranges of parameters settings were then input to the GA-
based optimization unit in order to reduce the searching space. Because both the
pure computational system and the enhanced computational system involve the
stochastic algorithm GA, 50 runs were carried out in the eight validations. Then
we evaluated the effectiveness and robustness of both systems by analyzing the
statistical results of the 50 runs.
To investigate the quality and the robustness of solutions found in both the pure
computational system and the enhanced computational system, the means and the
variances of the relative errors found in both systems for the 8 validations were
analyzed. Table 12.3 and Table 12.4 show the means and variances of the relative
errors of both systems respectively. It can be seen clearly from the tables that the
enhanced computational system yields better solutions in terms of mean errors and
variance of relative errors compared with the pure computational system in the 8
validations.
Table 12.3 Means of the relative errors
Mean errors of encapsulation Mean errors of encapsulation

weight thickness
Vali- Pure com- Enhanced Pure com- Enhanced
dation putational computational putational computational
system system system system
(%) (%) (%) (%)
1 0.1847 0.0289 0.2377 0.0534
2 0.1796 0.0132 0.0886 0.0147
3 2.9548 2.5588 1.3575 1.2588
4 1.2272 0.3691 0.5150 0.1594
5 0.5632 0.4533 0.2360 0.2031
6 0.1345 0.1338 0.084 0.0064
7 0.8581 0.5141 0.4547 0.3945
8 0.9710 0.7174 0.4358 0.3527
Table 12.4 Variances of the relative errors
Variances of errors of encap- Variances of errors of encap-

sulation weight sulation thickness
Vali- Pure com- Enhanced Pure com- Enhanced
dation putational computational putational computational
system system system system
(%) (%) (%) (%)
1 0.6589×10-5 0.0000×10-5 0.1246×10-5 0.0000×10-5
2 0.7143×10-5 0.0000×10-5 0.1351×10-5 0.0000×10-5
-4
3 0.3722×10 0.0010×10-4 0.3171×10 -4
0.0002×10-4
-5
4 0.9816×10 0.0016×10-5 0.3736×10 -5
0.0002×10-5
-3
5 0.2512×10 0.0004×10-3 0.2565×10 -3
0.0001×10-3
-5
6 0.8252×10 0.0009×10-5 0.3451×10 -4
0.0002×10-4
-4
7 0.1001×10 0.0396×10-4 0.0007×10 -4
0.0001×10-4
-4
8 0.2136×10 0.1283×10-4 0.0001×10 -4
0.0000×10-4
The t-test was used to determine the significance of differences between the
pure computational system and the enhanced computational system. Table 12.5
shows two sets of all t-values comparing the pure computational system and the
enhanced computational system for the validation tests for both encapsulation
weight and encapsulation thickness. It can be found that all the t-values are higher
than 2.15, which indicates that the significance is at 98% level of confidence.
Therefore the performance of the enhanced computational system is significantly
better than that of the pure computational system with 98% confidence, in terms of
prediction accuracy.
Table 12.5 t-values between pure computational system and enhanced computational sys-
tem for the relative errors of encapsulation weight and encapsulation thickness
Valida- T-values of encapsulation T-values of encapsulation

tion weight between pure and en- thickness between pure
hanced computational systems and enhanced computa-
tional systems
1 4.2918×102 1.1674×103
2 4.4024×102 4.4957×102
3 4.5836×102 1.2390×102
4 1.9351×103 1.3006×103
5 48.9922 14.5229
6 5.4459 93.3788
7 6.5079×102 1.5050×103
8 3.0667×102 5.8761×103
12.5 Conclusion
This chapter presents a GA-based knowledge discovery system which is intended
to generate rules from experimental data sets of the fluid dispensing process in
which three process parameters are involved: compressed air pressure, the height
between the substrate and the needle and pump motor speed, and two quality re-
quirements, encapsulation weight and thickness. Based on rules generated from
the GA-based knowledge discovery system, more specified ranges of process pa-
rameter settings can be obtained. Engineers could make use of the specified ranges
in order to more efficiently determine the appropriate setting of process parame-
ters for fluid dispensing compared with the time they spent in their conventional
practice. To validate the effectiveness of the rules generated from the GA-based
knowledge discovery system, the system was integrated with a computational sys-
tem for fluid dispensing. Eight validation tests were carried out. Results of the
tests indicate that the enhanced computational system can recommend process pa-
rameter settings which lead to smaller prediction errors and smaller variance of the
errors in comparison, with the standard GA method.
References
Bojarczuk, C.E., Lopes, H.K., Freitas, A.A.: A hybrid genetic algorithm/decision tree ap-
proach for coping with unbalanced classes. In: Proceedings of the 3rd International Con-
ference Practical Applications of Knowledge Discovery and Data Mining, pp. 61–70
(1999)
Carvalho, D.R., Freitas, A.A.: A hybrid decision tree/genetic algorithm for coping with the
problem of small disjoints in data mining. In: Proceeding of the Conference of Genetic
and Evolutionary Computation, pp. 1061–1068 (2000)
References 227
Chipperfield, A.J., Fleming, P.J.: The MATLAB genetic algorithm toolbox. In: Proceedings
of the IEE Colloquium on Applied Control Techniques using MATLAB, pp. 10/1–10/4
(1995)
Chiu, C., Hsu, P.L.: A constraint-based genetic algorithm approach for mining classifica-
tion rules. IEEE Transactions on Systems, Man, and Cybernetics –Part C: Applications
and Reviews 35(2), 205–220 (2005)
Cooper, G.F., Herskovits, E.: A Bayesian method for constructing Bayesian belief networks
from databases. In: Proceedings of the 7th Annul Conference of Uncertainty in Artificial
Intelligence, pp. 86–94 (1991)
Davis, L.: Handbook of Genetic Algorithms. Van Nostrand Reinhold, New York (1991)
Dehuri, S., Mall, R.: Predictive and comprehensible rule discovery using a multi-objective
genetic algorithm. Knowledge-Based Systems 19, 413–421 (2006)
Dzeroski, S.: Inductive logic programming and knowledge discovery in databases. In: Ad-
vances in Knowledge Discovery and Data Mining, pp. 117–152. AAAI Press, Menlo
Park (1996)
Elder, J., Pregibon, D.: A statistical perspective on knowledge discovery in database. In:
Advances in Knowledge Discovery and Data Mining, pp. 83–113. AAAI Press, Menlo
Park (1996)
Eshelmann, L.J.: The CHC Adaptive Algorithm: How to have safe search when engaging in
non-traditional genetic recombination. Foundations of Genetic Algorithms 1, 265–283
(1991)
Fayyad, U., Piatetsky-Shapiro, G., Smyth, P.: From data mining to knowledge discovery:
An overview. In: Advances in Knowledge Discovery and Data Mining, pp. 1–36 (1996)
Freitas, A.A.: A genetic programming framework for two data mining tasks: classification
and generalized rule induction. In: Proceedings of 2nd Annual Conference on Genetic
Programming, pp. 96–101 (1997)
Freitas, A.A.: On rule interestingness measures. Knowledge-Based Systems 12, 309–315
(1999)
Giles, C., Lee, C., Lawrence, S., Tsoi, A.C.: Rule inference for financial prediction using
recurrent neural networks. In: IEEE Proceedings of Conference on Computational Intel-
ligence for Financial Engineering, pp. 253–259 (1997)
Gilleo, K.: Area array packaging processes. McGraw-Hill, NewYork (2004)
son-Wesley, Reading (1989)
Goldberg, D.E., Deb, K.: A comparative analysis of selection schemes used in genetic algo-
rithms. In: Foundations of Genetic Algorithms, pp. 69–93 (1991)
Hand, D.J.: Principles of Data Mining. MIT Press (2001)
Hayashi, Y., Imura, A.: Fuzzy neural expert system with automated extraction of fuzzy if–
then rules from a trained neural network. In: The First International Symposium on Un-
certainty Modeling and Analysis, pp. 489–494 (1990)
Holland, J.H.: Adaptation in Natural and Artificial Systems. Univ. Michigan Press, Ann
Arbor (1975)
Kim, M.J., Han, I.: The discovery of experts’ decision rules from qualitative bankruptcy da-
ta using genetic algorithms. Expert Systems with Applications 25, 637–646 (2003)
Kolodner, J.: Case-Based Reasoning. Morgan Kaufman, San Francisco (1993)
Kwong, C.K., Chan, K.Y., Wong, H.: Empirical approach to modeling fluid dispensing for
electronic packaging. International Journal of Advanced Manufacturing Technology
(2007) (in Print)
Kwong, C.K., Chan, K.Y., Wong, H.: A computational system of process optimization of
fluid dispensing for electronic package (2007) (reviewing)
Langlery, P., Simon, H.: Application of machine learning and rule induction. Communica-
tion of ACM 38(11), 54–64 (1995)
Lee, R., Sikora, M.S.: A genetic algorithm based approach to flexible flow-line scheduling
with variable lot sizes. IEEE Transactions on Systems, Man, and Cybernetics – Part
B 27(1), 36–54 (1995)
Lin, C.T., Lee, C.S.G.: Neural network-based fuzzy logic control and decision system.
IEEE Transactions on Computer 12, 1320–1336 (1991)
Michalewicz, Z.: Genetic Algorithms + Data Structure = Evolution Programs. Springer,
Berlin (1994)
Muhlenbein, H., Voosen, D.S.: Predictive models for the breeder genetic algorithm: I. Con-
tinuous parameter optimization. Evolutionary Computation 1(1), 25–49 (1993)
Quinlan, J.: C4.5: Programs for machine learning. Morgan Kaufmann, San Francisco
(1992)
Schaffer, J., Caruana, R., Eshelman, L., Das, R.: A study of control parameters affecting
online performance of genetic algorithms for function optimization. In: Proceedings of
the 3rd International Conference on Genetic Algorithms, pp. 51–60 (1989)
Sikora, R.: Learning control strategies for a chemical process: A distributed approach. IEEE
Expert, 35–43 (1992)
Sikora, R., Shaw, M.: A double-layered learning approach to acquiring rules for classifica-
tion: Integrating genetic algorithms with similarity-based learning. ORSA Journal on
Computing 6(2), 174–187 (1994)
Srikant, R., Agrawal, R.: Mining sequential patterns: generalizations and performance im-
provements. In: Proceedings of the 5thInternational Conference on Extending Database
Technologies, pp. 3–17 (1996)
Whitley, D.: The GENITOR Algorithm and Selection Pressure: Why Rank-Based Alloca-
tion of Reproductive Trials is Best. In: Proceedings of the Third International Confe-
rence on Genetic Algorithms, pp. 116–121 (1989)
Chapter 13
Conclusion and Future Work
13.1 Conclusions
In terms of new product development, marketing personnel are usually concerned
with making the most of market opportunities by choosing the right price and un-
derstanding ‘consumer needs’, while engineering personnel may be concerned
only with ascertaining whether the engineering requirements can be met satisfac-
torily. Product designers are concerned with the product characteristics and ap-
pearance of the new product while manufacturing personnel are mainly concerned
with the manufacturing process design, quality of manufactured products, and
manufacturing time and cost. Therefore, they have different notions about the
drivers of success, the optimization variables, and the nature of constraints for new
product design. This book has presented and discussed several methodologies for
incorporating the concerns of marketing, engineering and manufacturing person-
nel into new product development.
Product design aims to address engineering concerns, manufacturing concerns
and marketing concerns regarding new product development, in order to maximize
overall customer satisfaction with new products. Various issues for product design
ranging from identification of customer requirements, ranking of importance of
customer requirements, determination of optimal design attribute settings for new
products, integration of marketing and engineering concerns for product design, to
affective product design have been discussed in Chapter 1.
However, the transformation of market opportunities in the marketing domain
into product development technologies in the engineering domain is a complex
procedure, as two very different domains are involved. Moreover, they have not
been well formalized. Computational intelligence technologies have been applied
to achieve these transformations, which fuse historical design information distri-
buted in space and time into coherent and understandable design knowledge.
Chapter 2 introduced and discussed the recent computational intelligence methods
applied to product design engineering. These methods offer modeling approaches
and optimization algorithms that can contribute greatly to design formalization
and automation.
For modeling methods, computational intelligence can search and summarize
the legacy data in a usable model, since the collection of huge amounts of custom-
er data by means of surveys is growing at an unprecedented rate. The resulting
230 13 Conclusion and Future Work
models will help designers to understand the relationship between customer re-
quirements and the design attributes of products. These modeling algorithms in-
cluding fuzzy systems and neural networks were introduced in Chapter 2. After
the model has been developed, we can formulate an optimization problem in order
to maximize the overall customer satisfaction. Due to the nonlinear nature of the
model, this optimization problem is multi-optima. Computational intelligence al-
gorithms including evolutionary algorithm, evolutionary programming, simulated
annealing and particle swarm optimization which are highly effective for solving
multi-optimization problems, were introduced to solve this optimization problem.
The literature shows that classical computational intelligence approaches can
provide a reasonable result regarding a particular product design problem. How-
ever, the result obtained might not be the best one possible. Therefore, the integra-
tion of appropriate techniques into a classical computational intelligence approach
is usually required. The resulting system is called the ‘hybrid computational intel-
ligence method’, which integrates statistical methods or the other computational
intelligence approaches. In general, the hybrid computational intelligence method
can enhance the effectiveness of the classical computational intelligence approach.
Chapters 3 to 12 discussed the applications of hybrid computational intelligence
methods to address different product design issues. These approaches are able to
obtain higher quality solutions than those obtained by the classical approaches re-
garding the particular product design problems. They can be sub-divided into the
following:
1) Determination of importance weights for customer which is summarized in
Section 13.1.1;
2) Development of customer satisfaction models which is presented in Section
13.1.2;
3) Optimization of overall customer satisfaction which is demonstrated in
Section 13.1.3;
4) Development of manufacturing process models for quality prediction of
products as discussed in Section 13.1.4.
13.1.1 Determination of Importance Weights for Customer

Requirements
In the past, the classical AHP was commonly used to determine the importance
weights for customer requirements which are crucial to product design. However,
linguistically expressed customer requirements always contain ambiguity and mul-
tiplicity of meanings and are often vague. Therefore, the classical AHP cannot
adequately capture customer requirements explicitly, or accurately determine the
importance weight of customer requirements.
In Chapter 3, we discussed a new version of AHP, namely fuzzy AHP inte-
grated with fuzzy systems, to determine importance weights of customer require-
ments (Kwong and Bai 2002) . The fuzzy AHP intends to overcome the limitation
of the traditional AHP that cannot address the vagueness and uncertainty, due to
human judgment being fuzzy. The fuzzy AHP is presented to determine the
13.1 Conclusions 231
importance weights of customer requirements. It is integrated with fuzzy scales

and is intended to overcome the limitation of the classical AHP that cannot ad-
dress the vagueness and uncertainty that is due to human judgment. The fuzzy
AHP aims to improve the imprecise ranking of customer requirements and is
based on the classical AHP. A case study of a bicycle splashguard was used to il-
lustrate the effectiveness of the fuzzy AHP.
In order to improve on the hitherto imprecise ranking of importance weights of
customer requirements inherited from the previous works based on the conven-
tional AHP and the fuzzy AHP discussed in Chapter 3, the enhanced fuzzy AHP
(Kwong and Bai 2003) converts the linguistic assessment of customer require-
ments to triangular fuzzy numbers, which are used to build the pairwise compari-
son matrix of the AHP. The enhanced fuzzy AHP uses the extent analysis method
and the principles of comparison of fuzzy numbers to derive weight vectors, in or-
der to improve the imprecise ranking of importance weights of customer require-
ments inherited from the previous works based on the conventional AHP and the
fuzzy AHP discussed in Chapter 3. The enhanced fuzzy AHP was discussed in
Chapter 4. Compared to the fuzzy AHP approach discussed in Chapter 3, the en-
hanced fuzzy AHP is simpler and easier to implement when prioritizing customer
requirements for product design. A case study of a hair dryer design was used to
illustrate the effectiveness of the enhanced fuzzy AHP.
13.1.2 Development of Customer Satisfaction Models

When functional models, which represent the relationship between customer re-
quirements and design attributes of new products, are available, product develop-
ment teams can study the degree of satisfaction of various customer requirements
by adjusting the design attribute setting of the new product. The following ap-
proaches to developing such functional models have been discussed:
- Genetic programming
- Fuzzy regression based genetic programming
- Generalized fuzzy least-squares regression
- Neuro-fuzzy networks
The genetic programming approach has commonly been used to develop a func-
tional model which relates design attributes to customer requirements of new
products (Chan et al. 2009b). It is intended to overcome two main limitations of
the previous modeling approaches, statistical linear regressions and computational
intelligence methods namely: i) statistical linear regressions are not able to gener-
ate functional models with nonlinear or high order terms in relating design
attributes to customer requirements; and ii) computational intelligence methods,
such as neural networks or neural fuzzy networks, can generate functional models
only as black-boxes from which no explicit information can be derived. They are
not preferred by product designers who usually want to understand the explicit
relationships between variables regarding design attributes and customer require-
ments. Chapter 5 discussed the use of genetic programming to develop the func-
tional models in relating design attributes to customer requirements of new
products. A case study of a digital camera design showed that genetic program-
ming is more effective in developing functional models for customer requirements
than is the method of statistical linear regression. It also showed that the functional
models developed by genetic programming are explicit, and thus, product design-
ers prefer them rather than the implicit models developed by other computational
intelligence methods such as neural networks, neural fuzzy networks etc.
However, the coefficients generated by genetic programming are real numbers
which are not able to address human feeling in customer survey data which is
usually fuzzy in nature. Also, previous modelling methods can only address either
nonlinearity or fuzziness. To overcome the deficiencies of the existing modeling
approaches, a fuzzy regression-based genetic programming approach to generate
functional models was proposed by (Chan et al. 2010a). This fuzzy regression
based genetic programming was discussed in Chapter 6. A case study of the affec-
tive design of mobile phones was used to evaluate the effectiveness of the hybrid
fuzzy regression based genetic programming approach. Results showed that
the functional models developed by the hybrid fuzzy regression based genetic
programming approach can generate functional models with better generalization
capability than those generated by fuzzy regression methods and statistical
regression.
Further to the hybrid fuzzy regression based genetic programming approach, a
generalized fuzzy least-square regression method (Kwong et al. 2010) was devel-
oped to address two uncertainties when conducting customer surveys: fuzziness
and randomness. The generalized fuzzy least-squares regression can also be used
to develop models for the functional relationships based on fuzzy observations and
crisp observations, which involves both crisp type and fuzzy type observed data. It
overcomes the limitation of the approach discussed in Chapter 6 and the other ex-
isting fuzzy regression approaches that cannot simultaneously take into account
the two uncertainties of fuzziness and randomness. The mechanisms for genera-
lized fuzzy least-squares regression have been discussed in Chapter 7. A case
study of an emulsified dynamite packing machine design showed that the genera-
lized fuzzy least-squares regression outperforms classical fuzzy regression which
addresses uncertainty only, due to the fuzziness of customer survey data.
Apart from all these regression approaches, the neuro-fuzzy networks approach
has also been used to develop the functional models for product design. Generally,
the neuro-fuzzy network approach outperforms the regression approaches to de-
velop models for highly nonlinear systems. However, the models in neuro-fuzzy
network form are black-box in nature so no explicit information and models can
be obtained. Even if the models in a neuro-fuzzy network form are more accurate
than those in regression form, product development teams do not prefer to use
these models, as no explicit information can be extracted.
Kwong et al. (2009a) proposed an enhanced neuro-fuzzy network approach for
the development of customer satisfaction models, which represents the relation-
ship between customer requirements and design attributes of a new product. The
approach is intended to overcome the limitation of the existing neural-fuzzy mod-
eling approaches whereby no explicit model can be developed.
13.1 Conclusions 233
This approach also overcomes the limitation of the genetic programming ap-
proach (discussed in Chapter 5) which cannot address fuzziness of customer re-
quirements. It also overcomes the limitation of the fuzzy regression approaches
(discussed in Chapters 6 and 7) which cannot model the relationships between
customer requirements and design attributes with strong nonlinearity. The model
developed by this method consists of a set of fuzzy rules which relate customer
requirements to the design attributes of a new product. Then a rule extraction me-
thod is used to extract significant rules to identify the appropriate ranges of design
attributes, which are significant to customer satisfaction for a new product. Then,
an explicit customer satisfaction model can be developed by aggregrating the in-
ternal models of the extracted rules.. In Chapter 8, an example of a notebook com-
puter design was used to illustrate the methodology.
13.1.3 Optimization of Overall Customer Satisfaction

When the importance weights of customer requirements and the functional models
become available, the optimization problems for maximizing the overall customer
satisfaction for new products can be formulated. As nonlinearity exists between
customer requirements and design attributes, these optimization problems are mul-
ti-optimal problems, which cannot be solved properly by traditional optimization
methods such as gradient-based methods. Chapter 9 and Chapter 10 discussed the
computational intelligence optimization methods, namely simulated annealing
(SA) and evolutionary algorithms respectively, to solve these multi-optima prob-
lems of new product design. Enhancements of these two computational intelli-
gence algorithms were conducted by integrating them with experimental design
methods, namely orthogonal design. Two case studies regarding optimizations of
packing machine design and car door design were presented to demonstrate the ef-
fectiveness of the enhanced computational optimization methods. They showed
that the hybrid computational intelligence algorithms outperform the classical
computational intelligence algorithms on maximizing the overall customer
satisfaction.
13.1.4 Development of Manufacturing Process Models for

Quality Prediction of Products
Commonly, settings of some design attributes are used as the quality requirements
of manufactured products. To ensure that a product can be manufactured to
achieve the quality requirements, manufacturers need to control the variability fac-
tor in each of the many manufacturing processes. Also, all the variables that con-
trol the desired quality of a new product need to be understood and optimized in
order to maintain tight control. To do this, a manufacturing process model that re-
lates the quality requirements of the new product to the manufacturing process va-
riables needs to be developed. Regression approaches have commonly been used
to develop process models in order to ensure the quality of the products. However,
it has been noticed that statistical regression methods cannot address fuzziness
which comes from experimental data resulting from the types of instrumental
measures used in the manufacturing processes. Even if traditional fuzzy regression

could address the fuzziness of experimental data, it cannot address the nonlinearity
of the manufacturing processes. A hybrid algorithm has been proposed by Chan et
al. (2009b) to integrate the mechanisms of fuzzy regression and genetic program-
ming in order to develop models for manufacturing processes that will improve
the quality of the manufactured product (Chan et al. 2009b). The approach over-
comes the existing fuzzy regression approaches which cannot address the nonli-
nearity of the manufacturing processes. The mechanism of the approach was dis-
cussed in Chapter 11. The case study of a solder paste injection process showed
that more accurate models can be generated using the hybrid algorithm than those
generated using the traditional statistical regression and classical fuzzy regression
approaches.
Apart from all these regression approaches, the neuro-fuzzy network approach
has also been used to develop models for manufacturing processes. Based on the
neuro-fuzzy networks, both nonlinearity and fuzziness of the manufacturing
processes can be addressed. In general, the models in neuro-fuzzy network form
are more accurate than those in regression form. However, manufacturing engi-
neers do not prefer to use these neuro-fuzzy models, as no explicit information can
be indicated. Kwong et al. (2009b) proposed a knowledge extraction approach,
which aims to extract information in the form of rules from the neural fuzzy net-
works. This approach overcomes the limitation of the existing neural fuzzy net-
works that are black-box in nature and no information and explicit models can be
obtained. It allows engineers to understand the behaviors of the manufacturing
systems. After they have more knowledge on the manufacturing process, they can
enhance quality of the manufactured products more easily. The mechanism of the
approach was discussed in Chapter 12. A case study of the epoxy dispensing
process is used to demonstrate the operations and procedures of the rule extraction
approach.
13.2 Future Works

There are several possible extensions of the work presented in this book and we
discuss these below as future work arising from this book. Future work can be di-
vided into two domains:
1) Collection of customer survey data using web mining
2) Investigation of innovative computational intelligence approaches
13.2.1 Collection of Customer Survey Data Using Web Mining

Customer survey data can be collected via questionnaires or customer interviews.
However, a significant amount of time is usually required to collect the data, and
also the size of the collected data is usually not enough for designing an ideal new
product for the whole customer domain. With the advent of the Social Web, a
great number of customers constantly generate opinions about new products
through Weblogs, wikis, online forums/groups, video/audio/file sharing sites,
References 235
social networks, which are related to consumers. It has been reported that at least
1.5 million daily posts appeared on 70 million Weblogs in 2007. In addition, at
least 3.5 billion brand-related conversations were generated through various online
media every day. Future work could focus on developing a Web Data Collecting
model which would periodically gather customers’ opinions distributed across the
Web for buying new products. It would pre-process customers’ opinions so that
these could be retrieved in an efficient and scalable manner and be used for the
development of new products. It should also deal with data cleaning, which is an
essential pre-processing step in data mining. In particular, an opinion anti-spam is
crucial, as it can ensure the quality and accuracy of customers’ opinions in buying
new products. It should also be able to compress the massive volume of opinions
based on several dimensions of customer requirements of new products.
13.2.2 Investigation of Innovative Computational Intelligence

Approaches
In product design, it is essential to develop an accurate functional model which
can precisely explain the relationship between design attributes and customer re-
quirements of the new products. The following two further works could be under-
taken in order to enhance the generalization capabilities of the functional models,
which are developed based on computational intelligence approaches: i) Outliers,
which are abnormal data points in the customers’ survey data, always exist in col-
lections of survey data. The generalization capability of the functional models will
decrease if customers’ survey data including outliers is used for developing the
models. Effective approaches for removing outliers in customers’ survey data
which may lead to a misleading outcome generated by the functional models need
to be developed; ii) an over-fitted functional model is usually produced during the
training process where the number of components inside the functional model in-
creases to fit all training data. There is also a need to develop approaches to avoid
generating an over-fitted functional model which is over-trained by survey data
collected from customers.
References
Chan, K.Y., Kwong, C.K., Tsim, Y.C.: Improved orthogonal array based simulated anneal-
ing with interaction analysis between variables for design optimization. Expert Systems
with Applications 36, 7379–7389 (2009a)
Chan, K.Y., Kwong, C.K., Tsim, Y.C.: A fuzzy nonlinear regression based on genetic pro-
gramming to modeling manufacturing processes. International Journal of Production
Research 48(7), 1967–1982 (2009b)
Chan, K.Y., Kwong, C.K., Dillon, T.S., Fung, K.Y.: An intelligent fuzzy regression ap-
proach for affective product design that captures nonlinearity and fuzziness. Journal of
Engineering Design 22(3), 523–542 (2010a)
opment using genetic programming. Journal of Engineering Design 22(1), 601–613
(2010b)
Chan, K.Y., Kwong, C.K., Tsim, Y.C., Aydin, M.E., Fogarty, T.C.: A new orthogonal array
application to car door design. Expert Systems with Applications 37(5), 3853–3862
(2010c)
facturing 13, 367–377 (2002)
626 (2003)
Kwong, C.K., Wong, T.C., Chan, K.Y.: A methodology of generating customer satisfaction
models for new product development using a neuro-fuzzy approach. Expert Systems
with Applications 36(8), 11262–11270 (2009a)
Kwong, C.K., Chan, K.Y., Tsim, Y.C.: A genetic algorithm based knowledge discovery
system for the design of fluid dispensing processes for electronic packaging. Expert Sys-
tems with Applications 36, 3829–3838 (2009b)
approach to modelling functional relationships in QFD. Journal of Engineering De-
sign 21(5), 601–613 (2010)
Index
affective design 111 customer surveys 5, 8, 12

affective response 119 customer survey data 8, 13, 27, 43,
aggregating internal model 149 111, 129, 232, 234
Analytic Hierarchy Process 59, 79
analytical model 19 data mining 213, 235
analysis of residual 14 design attributes 3, 8, 15, 95, 111,
asymmetrical distribution 33 150, 163
asymmetric triangular fuzzy design specification 2, 22
coefficients 14, 33 design automation 25, 229
design formalization 25, 229
biological neurons 33 desired quality 19, 194
biological system 33 digital camera design 53, 102, 232
black-box model 53, 160 discovery system 213, 215
car door design 54, 186 empirical model 13

classification 66, 217 emulsified dynamite packing machine
cluster analysis 6 design 130, 164, 232
comparison-based approach 8 encapsulation 212, 213
competitive analysis 6 engineering requirement 2, 168, 186,
consistency index 85, 86 229
consistency ratio 85, 86 engineering characteristics 1, 59, 129,
control variability 19, 194 133
crisp observations 130, 232 engineering constraints 1, 18
cross validation 104, 124 engineering personnel 19, 21, 90
constrained programming model 132, epoxy dispensing 216
134, 141 epoxy encapsulant 214
convergence rate 42 error sum of squares 137
customer requirements evolutionary programming
identification 5, 9, 53 convergence 95
importance weights 53, 59, 65, 68 crossover 48, 55, 101
hierarchical structure 6, 60, 65 fitness function 50, 96
model 3, 11 hierarchical trees 98
optimization 163, 165, 168 mutation 47, 48, 98
ranking of importance 76, 79 representation 100
relative importance 7, 8, 61 roulette-wheel 101
weight vectors 79, 86, 89, 92 selection 101, 102
customer attribute 15, 150 evolutionary algorithm
customer needs 1, 5, 6 convergence 193
customers’ opinions 235 crossover 180, 182, 183
customer perception 133, 138, 145, fitness function 191
150 mutation 190
customer preference 12 experimental data 199, 207
238 Index
explicit information 54, 95, 231, 234 injection moulding 19, 20

explicit model 125 interaction 15, 21, 48, 67, 83, 103,
extent analysis method 53, 76, 79 163, 178
interval 30, 31
feature-based approach 8
feed-forward neural network 36 Kano model 16
first-order linear model 14 Kolmogorov-Gabor polynomial 112,
fluid dispensing 19, 177, 213 200
fuzzy AHP 56, 63, 80
fuzzy additions 62 linear programming 14, 21, 30, 115,
fuzzy distance 62 132
fuzzy multiplications 64 linear modeling 13, 14
fuzzy comparison matrices 65 linear terms 95, 98
fuzzy eigenvalues 64 linearity 111
fuzzy matrix 80 linguistic description 146, 148
fuzzy number 89
fuzzy regression 29, 111, 112, 198, major components 9, 10
204 manufacturing process 19, 21, 54, 199
generalized fuzzy least square manufacturing systems 54, 234
regression 129 manufacturers 19, 194
Peters’ fuzzy regression 30 marketing personnel 22, 229
Tanaka’s fuzzy regression 30 market opportunities 1, 229
fuzzy coefficient 33, 34, 111, 112 membership function 20, 28, 63, 82
fuzzy regression based genetic microchip encapsulation 212, 213
programming 111, 112 mobile phone design 119
functional model representation 115 multiple linear regression 14
fitness function 115 multilayer feed-forward neural
evolutionary operations 117 networks 34, 36
fuzzy rules 156, 157 multi-objective optimization 170
fuzzy rule-based approach 148, 149 multi-optima problems 163, 233
genetic programming new product opportunities 3, 9, 11, 17

convergence 101 new product design 17, 23, 90, 163,
crossover 100 174
fitness function 99 noisy processes 20
hierarchical trees 98 non-differentiable function 42, 177
mutation 100 nonlinear fuzzy regression 199, 200
representation 99 nonlinear modeling 13, 15
selection 101 nonlinear polynomial interval 201
gradient based methods 163 nonlinear separation problem 36
neural fuzzy networks 145
hierarchical structure 60, 65 weight 146, 150
higher order terms 21, 95, 107 parameters 147
hill climbing algorithm 44 neural networks 36
house of quality 139, 170 configurations of neural
hybrid computational intelligence 53 networks 34
hybrid evolutionary computation 95, growing neural network 39
111 gradient based learning methods
hybrid fuzzy regression 129 42
Hopfield network 37, 38
IC Chip 214 learning algorithms 40, 41
informative rule 215 pruning neural network 39, 40
Index 239
radial basis function network 38, QFD 13, 14, 77, 79, 129
39 quantitative information 7
self-organizing map 15, 37 quality requirement 19, 20, 199, 226,
stochastic learning methods 40, 42 233
transfer function 35
new product opportunities 3, 4, 9 randomness 54, 126, 129
notebook computer design 145, 233 randomized perturbation 163
nonlinearity 160, 163, 232 regression parameter 33
removable mountain bicycle
orthogonal array 163, 167 splashguard 65, 75
orthogonal array based crossovers rule discovery system 215, 216
178 rule induction 220
orthogonal crossover 178–180
main effect crossover 182 sentiment-based approach 8
interaction crossover 184 simulated annealing 26, 46
orthogonal design 163, 177 neighbourhood function 44, 163
optimization 163, 177 significant rules 145, 233
overall customer satisfaction 16, 163, social web 234
174 solder spot diameter 207
optimizing customer satisfaction 16, statistical regression 14, 20, 120 145
164, 177 statistical significance 7
overfitting in modeling 235 stochastic optimization 43, 163
soldering paste dispensing 205
partial differential equations 19
Particle Swarm Optimization 43, 51, tolerance 31, 34, 188
55, 57 triangular fuzzy coefficients 14, 33 131
penalty factor 124
perceptual map 9, 12 uncertainty
predictive accuracy 213, 216 fuzziness 14, 26, 30, 62, 81, 107,
prediction error 124, 126, 208 111, 126
probabilistic decision 45 randomness 54, 126, 129
production line 194
production cost 3, 22, 194 vagueness 31, 53
product designers 1, 75, 90
product feature 2, 8 Web 2.0 7, 8
product material 4, 5, 8, 43, 75 web mining 234
process parameter 199, 205, 210 Weblog 234

12 Computational Intelligence Techniques For New Product Design

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

12 Computational Intelligence Techniques For New Product Design

Uploaded by

Copyright:

Available Formats

Kit Yan Chan, C.K. Kwong, and Tharam S.

ISSN 1860-949X e-ISSN 1860-9503

c Springer-Verlag Berlin Heidelberg 2012

Printed on acid-free paper

• Chapter 5 presents a genetic programming method in order to develop a

• Chapter 12 discusses a rule-extraction-based genetic algorithm approach

1 Integrated Product Design .............................................................................1

2 Computational Intelligence Technologies for Product Design..................25

3 Determination of Importance of Customer Requirements Using

4 An Enhanced Fuzzy AHP Method with Extent Analysis for

5 Development of Product Design Models Using Classical

6 Development of Product Design Models Using Fuzzy

7 Generalized Fuzzy Least Square Regression for Generating

8 An Enhanced Neuro-fuzzy Approach for Generating Customer

9 Optimization of Customer Satisfaction Using an Improved

10 An Enhanced Genetic Algorithm Integrated with Orthogonal

11 A Nonlinear Fuzzy Regression for Developing Manufacturing

12 Rule Extraction from Experimental Data for Manufacturing

12.3.5 Rule Induction ..................................................................220

13 Conclusion and Future Work ..................................................................229

performance targets, features and costs. Engineering personnel might be concerned

Fig. 1.1 Product design model

Fig. 1.2 Design specification of the new product

In this chapter, a methodology is described, which aims to integrate marketing

Identification of new product

Modeling the functional relationships

Fig. 1.3 A methodology for defining design specifications of new products

1.2 Determination of Importance of Customer Requirements

Determination of design attribute set-

Appropriate materials for the casing

Fig. 1.4 Design of notebook computer case

Therefore, before designing a new product, it is critical to acknowledge and

Step 1: Identification of Customer Needs. Step 1 determines the custom-

CR1, CR2, CR3, CR4

Step 2 Relative importance Step 3 Competitive analysis

CR1>CR3> CR4> CR2 CR1>CR2> CR3> CR4

Step 4 Final importance

CR1>CR3> CR2> CR4

Fig. 1.5 Determination of importance of customer requirements

Step 2: Relative Importance Ratings Based on Customer Data. Following

Customers’ Web mining

Fig. 1.6 Collection of customer data

A commonly used approach for collecting this information is via customer

1.3 Identification of New Product Opportunities

Table 1.1 Benchmarks of the five major competitors

Step 2) Transforming Customer Requirements Into Two Major Components:

Table 1.2 Rotated factor loadings and communalities

Variable Factor 1 (f1) Factor 2 (f2) Communality

Table 1.3 Factor score coefficients

Variable Factor 1 (f1) Factor 2 (f2)

Step 3) Construction of Perceptual Map: Based on the above two models, a

1.4 Functional Modeling of the Relationships between

Fig. 1.7 Perceptual map of packing machines

DA1 DA2 DA3 DA4 DA5 DA6 DA7

Fig. 1.8 House of Quality for packing machine design

Linear modeling Nonlinear modeling

Fig. 1.9 Modeling methodologies for generating empirical models

1.4.1 Linear Modeling Methods

1.4.2 Nonlinear Modeling Methods

+ β 12 x1 ⋅ x2 + β 12 x1 ⋅ x3 ... + β n−1,n xn−1 ⋅ xn

1.5 Maximization of Overall Customer Satisfaction

The final task in product design is to maximize overall customer satisfaction

attributes of a new product. As illustrated in Figure 1.10, it depends on two fac-