a probabilistic-based design approach with game theoretical ...
a probabilistic-based design approach with game theoretical ...
a probabilistic-based design approach with game theoretical ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
A PROBABILISTIC-BASED DESIGN APPROACH WITH GAME<br />
THEORETICAL REPRESENTATIONS OF THE ENTERPRISE<br />
DESIGN PROCESS<br />
A Thesis<br />
Presented to<br />
the Academic Faculty<br />
By<br />
Gabriel Hernandez<br />
In Partial Fulfillment<br />
of the Requirements for the Degree of<br />
Master of Science in Mechanical Engineering<br />
Georgia Institute of Technology<br />
June 1998
A PROBABILISTIC-BASED DESIGN APPROACH<br />
WITH GAME THEORETICAL REPRESENTATIONS<br />
OF THE ENTERPRISE DESIGN PROCESS<br />
Approved:<br />
Farrokh Mistree, Chairman<br />
Professor<br />
Mechanical Engineering<br />
Janet K. Allen<br />
Senior Research Scientist<br />
Mechanical Engineering<br />
Soumen Ghosh<br />
Associate Professor<br />
School of Management<br />
David W. Rosen<br />
Assistant Professor<br />
Mechanical Engineering<br />
Mark L. Spearman<br />
Associate Professor<br />
Industrial and Systems Engineering<br />
Date Approved<br />
ii
DEDICATION<br />
iii<br />
A mi abuelo, Jose M. Perez Bravo
ACKNOWLEDGMENTS<br />
Only passions, great passions, can elevate the soul to great things<br />
iv<br />
Denis Diderot<br />
Since I was child I have been taught that the labor we accomplish shapes our souls. It<br />
must be taken <strong>with</strong> pride and passion because it is an essential feature of who we are. This<br />
thesis is a result of that belief, but I know it is only the first step, and probably the easiest one.<br />
I have also been taught that no accomplishment is possible in loneliness, but only <strong>with</strong> the<br />
help, inspiration and support of other people, such as the ones that I have been blessed to meet<br />
during this time. Specifically, I sincerely thank my advisors Dr. Farrokh Mistree and Dr. Janet<br />
K. Allen for their extraordinary guidance and patience. I also appreciate the insightful<br />
recommendations of my other members of the committee, Dr. David Rosen, Dr. Mark L.<br />
Spearman, and Dr. Soumen Ghosh.<br />
My beholden thanks to my family for the unconditional love and teaching that have driven<br />
me through life.
I am extremely grateful to Mr. Antonio Moratinos for his long-standing and sincere<br />
friendship to me and my family. Without his support I would not be writing these lines today.<br />
I give heartfelt thanks to the friends who have accompanied me during this time in Mexico<br />
and Europe, and whose constant communication has helped me to face bad times I have<br />
encountered: Javier Romo and his family, Celia Sánchez, Margarita Aviña, and Luis A. Calvillo<br />
and his family. I also thank those who built me a second home in Atlanta: Iván Zeron, Nicola<br />
Giberti, Guillermo and Raquel Rodríguez, and Eileen Lai. I apologize to all the good friends that<br />
the brevity of space does not allow me to name. My deepest gratitude to all of you.<br />
I thank the members of the Systems Realization Laboratory for their welcome and<br />
friendship. I give special thanks to Jesse Peplinski and Tim Simpson, whose vision and feats<br />
have greatly inspired me. They are the kind of people that make a difference in the world.<br />
I gratefully acknowledge the help of Dr. Eduardo Bascarán, Director of The Engineering<br />
Center of Carrier Mexico, for providing assistance for the realization of the present research. I<br />
hope this work is just the beginning of a long-term friendship.<br />
I thank Dr. George A. Hazelrigg, Director of Design and Integration Engineering of the<br />
National Science Foundation, for his valuable commentary on this thesis.<br />
I recognize and thank the Consejo Nacional de Ciencia y Tecnología of Mexico,<br />
CONACYT, for the economical support that made possible this endeavor. I shall work hard to<br />
pay back my country this assistance. Financial support from NSF grant DM1-96-12327 is<br />
v
gratefully acknowledged. The Systems Realization Laboratory of the Georgia Institute of<br />
Technology has underwritten the cost of computer time.<br />
vi
TABLE OF CONTENTS<br />
DEDICATION iii<br />
ACKNOWLEDGMENTS iv<br />
TABLE OF CONTENTS vii<br />
LIST OF TABLES xiii<br />
LIST OF FIGURES xvi<br />
NOMENCLATURE xx<br />
SUMMARY xxiii<br />
CHAPTER 1<br />
FOUNDATIONS FOR DESIGNING WITHIN AN ENTERPRISE DESIGN<br />
PARADIGM 1<br />
1.1 MOTIVATION 2<br />
1.2 INTELLECTUAL FOUNDATIONS 4<br />
1.2.1 Systems Thinking in Design and Manufacturing 4<br />
1.2.2 Design as a Science of the Artificial 6<br />
1.3 FRAME OF REFERENCE: DECISION-BASED DESIGN 12<br />
vii
1.4 A DECISION-BASED ENTERPRISE DESIGN 12<br />
1.4.1 The Concept of Enterprise Design 12<br />
1.4.2 A Decision-Based Method for Enterprise Design 14<br />
1.5 RESEARCH QUESTIONS AND HYPOTHESES 20<br />
1.6 ORGANIZATION OF THE THESIS 25<br />
CHAPTER 2<br />
ENTERPRISE DESIGN AS AN EXTENSIVE GAME 29<br />
2.1 OVERVIEW OF ENTERPRISE MODELING: ARCHITECTURES,<br />
TOOLS AND METHODS 30<br />
2.1.1 Modeling Architectures 30<br />
2.1.2 Modeling Tools 32<br />
2.1.3 Modeling Methods 33<br />
2.1.4 Enterprise Models and Enterprise Design 35<br />
2.2 A GAME THEORETICAL REPRESENTATION OF THE ENTERPRISE<br />
DESIGN PROCESS 37<br />
2.2.1 Decisions in a Manufacturing Enterprise 37<br />
2.2.2 Decisions and Game Theory 39<br />
2.2.3 Fundamentals of Game Theory in Design 41<br />
2.3 ENTERPRISE DESIGN AS AN EXTENSIVE GAME 47<br />
viii
2.3.1 Extensive Form Games 47<br />
2.3.2 Enterprise Design as an Extensive Game 52<br />
2.4 RELEVANCE OF MODELING AN ENTERPRISE DESIGN PROCESS<br />
CHAPTER 3<br />
AS A GAME 59<br />
FORMULATION AND SOLUTION OF ENTERPRISE DESIGN PROBLEMS 60<br />
3.1 ENTERPRISE INTEGRATION AS COORDINATION OF DECISIONS:<br />
IMPLICIT AND EXPLICIT COORDINATION 61<br />
3.2 GAME THEORY AS A FOUNDATION FOR COORDINATION OF<br />
DESIGN DECISIONS 68<br />
3.3 IMPLICIT COORDINATION THROUGH A COMMON PARADIGM<br />
FOR DECISION SUPPORT 71<br />
3.4 EXPLICIT COORDINATIO THROUGH A UNIFIED DECISION<br />
MODEL: THE COMPROMISE DSP 73<br />
3.5 FORMULATION OF DESIGN PROBLEMS IN ENTERPRISE DESIGN 81<br />
3.5.1 Approximation in Engineering Design 81<br />
3.5.2 The Cooperative Formulation using Compromise DSPs 84<br />
3.5.3 The Non-cooperative Formulation using Compromise DSPs 88<br />
3.6 A PROBABILISTIC-BASED DESIGN APPROACH FOR ENHANCING<br />
ix
COORDINATION WITH FLEXIBLE SOLUTIONS 94<br />
3.6.1 Why a Probabilistic-Based Design Approach? 94<br />
3.6.2 Preference Functions for Design Requirements 96<br />
3.6.3 Ranged Design Solutions 98<br />
3.6.4 Design Preference Index 100<br />
3.6.5 Constraint Feasibility Evaluation for Ranged Solutions 101<br />
3.6.6 A Probabilistic-Based Multiobjective Design Model 105<br />
3.7 TYING IT ALL TOGETHER: A MATHEMATICAL FRAMEWORK<br />
CHAPTER 4<br />
FOR ENHANCING COORDINATION IN ENTERPRISE DESIGN 109<br />
CASE STUDY: INTEGRATED PRODUCT AND PROCESS DESIGN OF<br />
AN ABSORBER/EVAPORATOR MODULE FOR A FAMILY OF<br />
ABSORPTION CHILLERS 111<br />
4.1 INTRODUCTION TO CASE STUDY 115<br />
4.1.1 Motivation for Case Study: Benefits of an Integrated Product and<br />
Process Design in Make-to-Order Systems 115<br />
4.1.2 A Methodology for Modeling Production Systems for Integrated<br />
Product and Process Design 117<br />
4.2 DESCRIPTION OF THE DESIGN PROBLEM AND ANCILLARY<br />
x
INFORMATION 120<br />
4.2.1 Absorption Refrigeration 120<br />
4.2.2 Objective and Ancillary Information 124<br />
4.3 BASELINE FORMULATION AND SOLUTION 128<br />
4.3.1 Formulation of Decision for Product Design 130<br />
4.3.2 Formulation of Decision for Production System Design 135<br />
4.3.3 Results of Baseline Approach 146<br />
4.4 COOPERATIVE FORMULATION AND SOLUTION 152<br />
4.4.1 Task 1: Expand Scope of Decision 154<br />
4.4.2 Task 2: Determine Input Needed for Analysis and Identify<br />
Modeling Techniques 154<br />
4.4.3 Task 3: Define Boundaries of Decision 156<br />
4.4.4 Results of Cooperative Model 162<br />
4.5 NON-COOPERATIVE FORMULATION AND SOLUTION 163<br />
4.5.1 Task 1: Expand Scope of Decision 165<br />
4.5.2 Task 2: Determine Input Needed for Analysis and Identify<br />
Modeling Techniques 166<br />
4.5.3 Task 3: Define Boundaries of Decision 169<br />
4.5.4 Results of Non-Cooperative Model 178<br />
4.6 NON-COOPERATIVE FORMULATION AND SOLUTION WITH<br />
xi
PROBABILISTIC DESIGN MODEL 180<br />
4.7 COMPARISON OF SOLUTIONS 189<br />
4.6 CLOSURE OF CASE STUDY AND OPEN ISSUES 198<br />
CHAPTER 5<br />
4.6.1 Review of Case Study 198<br />
4.6.2 Open Issues 199<br />
CLOSURE 201<br />
5.1 CRITICAL REVIEW OF THE THESIS: DISCUSSION OF RESEARCH<br />
QUESTIONS AND SHORTCOMINGS 202<br />
5.2 CONTRIBUTIONS AND RELEVANCE OF THE THESIS 208<br />
5.3 RECOMMENDATIONS FOR FUTURE WORK 210<br />
APPENDIX A<br />
THE PARAMETRIC DECOMPOSITION APPROACH 212<br />
APPENDIX B<br />
FORTRAN FILES 220<br />
xii
B.1 SAMPLE OF FORTRAN FILES FOR OptdesX 221<br />
B.2 SUBROUTINES TO FIND THE MINIMUM NUMBER OF TUBES<br />
GIVEN LENGTH AND TUBES SELECTION 246<br />
B.3 RESPONSE SURFACE NETWORK MODEL 251<br />
REFERENCES 260<br />
xiii
LIST OF TABLES<br />
Table 1.1 Research Questions and Hypotheses 25<br />
Table 2.1 Strategy, Tactics and Control Decisions 38<br />
Table 2.2 Contents of Static and Extensive-Form Games 49<br />
Table 3.1 Explicit and Implicit Coordination Regarding Processes Underlying<br />
Coordination 64<br />
Table 4.1 Estimated Distribution of Demand 124<br />
Table 4.2 Alternatives for Selection of the Evaporator Tube 131<br />
Table 4.3 Alternatives for Selection of the Absorber Tube 131<br />
Table 4.4 Parameters for Absorber-Evaporator Module Synthesis 132<br />
Table 4.5 Material Cost Goal for each Refrigeration Capacity 132<br />
Table 4.6 Constraints Values for Product Design 132<br />
Table 4.7 Cost of Raw Material at each Subassembly Station 133<br />
Table 4.8 Parameters for Production System Design 140<br />
Table 4.9 Goals and Constraint Values for Production System Design 140<br />
Table 4.10 Results for Each Capacity using Simulated Annealing 146<br />
Table 4.11 Results for Each Capacity using a Clustering-<strong>based</strong> Algorithm 147<br />
xiv
Table 4.12 Results for Each Capacity using FALP 148<br />
Table 4.13 Results for Minimum Production Cost and Cycle Time 150<br />
Table 4.14 Baseline Solution for Production System Design 151<br />
Table 4.15 Cycle Times Estimated <strong>with</strong> Response Surface Network Model and<br />
Discrete Event Simulation 151<br />
Table 4.16 Cycle Times for 20 Replicates of Simulation 152<br />
Table 4.17 Implementation of Cooperative Formulation and Solution 153<br />
Table 4.18 Initial Value of Design Variables for Optimization Process 162<br />
Table 4.19 Results of Cooperative Model 162<br />
Table 4.20 Results of Cooperative Model for Each Capacity 163<br />
Table 4.21 ANOVA Table for MfgC Regression 174<br />
Table 4.22 Summary of Fit for MfgC 174<br />
Table 4.23 ANOVA Table for CT Regression 175<br />
Table 4.24 Summary of Fit for CT 175<br />
Table 4.25 Initial Value of Design Variables for Optimization Process 178<br />
Table 4.26 Results of Product Design: MC and “Predicted” Values of MfgC,<br />
COST and CT 178<br />
Table 4.27 Results of Non-Cooperative Model for Each Capacity 179<br />
Table 4.28 Production System Design Results <strong>with</strong> Non-Cooperative Model 179<br />
xv
Table 4.29 Results of Product Design: MC and Predicted Values of MfgC ,<br />
COST and CT 188<br />
Table 4.30 Results of Non-Cooperative Probabilistic Model for Each Capacity 188<br />
Table 4.31 Production System Design <strong>with</strong> Probabilistic Non-Cooperative Model 188<br />
Table 4.32 Safety Stock of Evaporator and Absorber Tubes in Number of Tubes 194<br />
xvi
LIST OF FIGURES<br />
Figure 1.1 A Hybrid Paradigm for Decision Support 15<br />
Figure 1.2 A Decision-Based Method for Enterprise Design 17<br />
Figure 1.3 Implementation Strategy 26<br />
Figure 2.1 Various Formulations in Optimization Theory 42<br />
Figure 2.2 Example of an Extensive Game 51<br />
Figure 2.3 Enterprise Design as a Game 58<br />
Figure 3.1 Implicit and Explicit Coordination of Decisions 63<br />
Figure 3.2 Integrating Implicit and Explicit Coordination 65<br />
Figure 3.3 A Framework for Enterprise Design 67<br />
Figure 3.4 A Single Objective Optimization Problem and the Multigoal<br />
Compromise DSP 76<br />
Figure 3.5 Mathematical Form of a Compromise DSP 77<br />
Figure 3.6 Full Cooperation: Pareto Solutions 85<br />
Figure 3.7 Approximate-Cooperation using Compromise DSPs 86<br />
Figure 3.8 Solution of a Two-Players Approximate-Cooperative Model 87<br />
Figure 3.9 Conceptual Outline of RRS Construction 89<br />
Figure 3.10 Non-Cooperative Compromise DSPs 90<br />
xvii
Figure 3.11 Solution of a Two-Players Non-Cooperative Model 90<br />
xviii
Figure 3.12 Leader/Follower Solution Process 92<br />
Figure 3.13 Compromise DSPs for a Two-Players Leader/Follower Game 93<br />
Figure 3.14 Combining Cooperative Protocols in Enterprise Design 92<br />
Figure 3.15 Coordination of Decisions <strong>with</strong> Ranged Solutions 95<br />
Figure 3.16 A Flexible Preference Function 97<br />
Figure 3.17 A Comparison of Preference Structure 98<br />
Figure 3.18 Mapping between Design Variables and Design Performance 99<br />
Figure 3.19 Design Preference Index 101<br />
Figure 3.20 Statistical Constraint Feasibility 104<br />
Figure 3.21 The Probabilistic-Based Multobjective Design Model 108<br />
Figure 4.1 Hierarchical Objectives in a Manufacturing Organization 116<br />
Figure 4.2 Modeling the Production System as a Network of Response Surfaces 119<br />
Figure 4.3 Absorption Refrigeration Cycle 121<br />
Figure 4.4 Layout of Absorption Chillers 123<br />
Figure 4.5 Estimated Distribution of Demand 124<br />
Figure 4.6 Production line of the Absorber-Evaporator Module 125<br />
Figure 4.7 Roadmap to Case Study 127<br />
Figure 4.8 Agents Involved in the Design Process 129<br />
Figure 4.9 Baseline Compromise DSP for Product Designers 134<br />
Figure 4.10 Compromise DSP for Building a Payoff Matrix for Production<br />
xix
System Design 144<br />
Figure 4.11 Baseline Compromise DSP for Production System Design 145<br />
Figure 4.12 Comparison of the Value of the Final Deviation Function Z for each<br />
Capacity using San, GLOBAL and FALP 149<br />
Figure 4.13 Implementation of Design Strategy for Cooperative Model 159<br />
Figure 4.14 Compromise DSP for Cooperative Model 161<br />
Figure 4.15 Modified Deviation Function Z mp 164<br />
Figure 4.16 Stackelberg Protocol between Product Design and Production System<br />
Design 166<br />
Figure 4.17 Implementation of Design Strategy for Non-Cooperative Model 171<br />
Figure 4.18 Fit of MfgC Regression 173<br />
Figure 4.19 Fit of CT Regression 176<br />
Figure 4.20 Stackelberg Compromise DSP for Product Design 177<br />
Figure 4.21 Preference Functions for COST and CT 181<br />
Figure 4.22 Stackelberg Probabilistic Formulation for Product Design 186<br />
Figure 4.23 Stackelberg Probabilistic Formulation for Production System Design 187<br />
Figure 4.24 Deviation from Material Cost Target for Each Capacity 190<br />
Figure 4.25 Average Material Cost of the Various Solutions 190<br />
Figure 4.26 Average Production Cost of the Various Solutions 191<br />
Figure 4.27 Average Total Cost of the Various Solutions 192<br />
xx
Figure 4.28 Average Production Time of the Various Solutions 192<br />
Figure 4.29 Deviation from Targets of Total Cost and Production Time of the<br />
Various Solutions 193<br />
Figure 4.30 Sensitivity of Cooperative Solution to Changes in the Length 195<br />
Figure 4.31 Sensitivity of Probabilistic Solution to Changes in the Length 195<br />
xxi
a An action in a <strong>game</strong><br />
A Set of feasible actions in a <strong>game</strong><br />
B Design region <strong>with</strong> a local minimum<br />
CAD Computer Aided Design<br />
CCD Central Composite Design<br />
CE Concurrent Engineering<br />
CIM Computer Integrated Manufacturing<br />
NOMENCLATURE<br />
d Deviation variable in a Compromise DSP<br />
DFM Design for Manufacturability<br />
DBD Decision Based Design<br />
DOE Design of Experiments<br />
DPI Design Preference Index<br />
DSP Decision Support Problem<br />
f Cost (objective) function<br />
F Function or model<br />
g Equality vector constraint<br />
h Inequality vector constraint, history of a <strong>game</strong><br />
xxii
k Stage of a <strong>game</strong><br />
xxiii
K Total number of stages in a <strong>game</strong><br />
N Number of players<br />
P Preference function<br />
Po<br />
Payoff matrix<br />
r Dimensions of a cost function<br />
RRS Rational Reaction Set<br />
s A pure strategy of a player in a <strong>game</strong><br />
S Design space set<br />
SC Supercriterion<br />
T Target values<br />
W Weight of an objective in an Archimedean formulation<br />
x Design variable<br />
X Design vector<br />
y Level of performance of a system<br />
ˆ<br />
z Normalized deviation function of a Compromise DSP<br />
Z Deviation function of a Compromise DSP<br />
Subscripts<br />
p Pareto solution<br />
n Nash solution<br />
xxiv
Superscripts<br />
* Optimum value<br />
m Dimensions of a <strong>design</strong> space S<br />
mp modified Pareto solution<br />
w Worst value<br />
xxv
SUMMARY<br />
xxvi<br />
Enterprise <strong>design</strong> is embodied by the integration of product <strong>design</strong>, manufacturing process<br />
<strong>design</strong> and organization <strong>design</strong> to enhance competitiveness by achieving manufacturing<br />
enterprises that are <strong>design</strong>ed in an integrated and internally consistent manner. However, due to<br />
the complexity and size of an enterprise, it is likely that its <strong>design</strong> must be accomplished by a<br />
group of functional departments that deal <strong>with</strong> different aspects of its activity. Thus, it is<br />
necessary to provide a rigorous framework that enhances an efficacious collaboration of the<br />
various decision-makers that comprise an enterprise. In this thesis such a framework is built<br />
<strong>with</strong> the following elements:<br />
?? A representation of the enterprise activity as an extensive <strong>game</strong>.<br />
?? An enterprise <strong>design</strong> method <strong>based</strong> on a common model for decision support.<br />
?? A multiobjective decision model, the Compromise DSP.<br />
?? A <strong>probabilistic</strong> <strong>design</strong> model to attain flexible and robust decisions and handle uncertainty<br />
and imprecision.<br />
The merging of these elements provides a consistent mathematical framework for coordination<br />
of decisions <strong>based</strong> on <strong>design</strong> formulations, which fosters its implementation in a computer-<br />
supported environment. This framework is exercised in a case study, namely, the integrated
xxvii<br />
product and process <strong>design</strong> of an absorber/evaporator module for a family of absorption<br />
chillers.
CHAPTER 1<br />
FOUNDATIONS FOR DEVELOPING A FRAMEWORK FOR ENTERPRISE<br />
DESIGN<br />
Enterprise Design is a term coined to characterize the integration of the processes of<br />
product <strong>design</strong>, manufacturing process <strong>design</strong> and organization <strong>design</strong> to enhance the<br />
competitiveness of manufacturing organizations by achieving enterprises that are <strong>design</strong>ed in a<br />
unified, integrated and internally consistent manner. In this work it is asserted that a practical<br />
and efficacious integration can be achieved by formulating and solving appropriately the<br />
individuals’ <strong>design</strong> problems under a unified decision-support framework. In this first chapter,<br />
the foundations for developing such a framework are examined, along <strong>with</strong> its motivations and<br />
research questions to be explored. The core notion is to envision <strong>design</strong> and manufacturing as<br />
disciplines that can be scientifically carried out <strong>with</strong> a decision-<strong>based</strong> perspective.<br />
1
1.1 MOTIVATION<br />
In recent years there has been a steady rise in customers’ demands. Customers expect variety,<br />
low prices, high quality, and responsive service. Due to these requirements, manufacturing<br />
organizations have to more efficaciously coordinate their activities to maintain their<br />
competitiveness. The notion of integrating the processes of product <strong>design</strong>, manufacturing<br />
process <strong>design</strong>, and organization <strong>design</strong> has been characterized <strong>with</strong> the term of “enterprise<br />
<strong>design</strong>” (Peplinski, 1997). It is postulated that such integration fosters competitiveness by<br />
achieving manufacturing enterprises that are <strong>design</strong>ed in an internally consistent manner.<br />
However, due to the overwhelming complexity and size of the problem of <strong>design</strong>ing an<br />
enterprise in its entirety, it is likely that the problem must be accomplished by a group of<br />
functional departments that deal <strong>with</strong> different aspects of the enterprise activity. In other words,<br />
an enterprise is a system whose scale and complexity necessitates decomposition into a set of<br />
smaller interdependent activities, undertaken by distinct groups making individual decisions.<br />
In this thesis a perspective in which this decomposition is not only necessary, but also desirable,<br />
is embraced, as voiced by Little (Little, 1992):<br />
Building complete models of a system and setting all the control variables misses the<br />
major opportunities for system improvement that are possible by finding new ways to<br />
empower the people on the front lines of the organization by giving them information,<br />
training and tools, <strong>with</strong> which to improve their own performance.<br />
Herbert Simon (Simon, 1996, p. 41) also supports this belief:<br />
2
As a single decision may be influenced by a large number of facts and criteria of choice,<br />
some fraction of these premises may be specified by superiors <strong>with</strong>out implying complete<br />
centralization. Organizations can localize and minimize information demands by<br />
decentralizing decisions. Matters of fact can be determined wherever the most skill and<br />
information is located to determine them.<br />
Organizations operating in a complete centralized way would exceed the limits of human<br />
procedural rationality and lose many of the advantages attainable from the use of hierarchical<br />
authority and employee empowerment. But, if decentralization of decisions is desirable, how<br />
can an effective integration of the enterprise activities be achieved? Peplinski and Mistree<br />
(1997) offer an answer to this dilemma:<br />
The alternative <strong>approach</strong> of enterprise <strong>design</strong> is that of promoting all of the domaindependent<br />
system modeling schemes across the enterprise and then integrating all of these<br />
varied models into a decision <strong>based</strong> <strong>design</strong>... The key shift in thinking is from visualizing a<br />
<strong>design</strong> process as initiated by a single <strong>design</strong>er to recognizing the actual enterprise <strong>design</strong><br />
process as the combined actions of the empowered group of decision makers that<br />
comprise the organization.<br />
The essential feature of this enterprise <strong>design</strong> vision is that enterprise <strong>design</strong> should not be<br />
viewed as an attempt to <strong>design</strong> products, manufacturing processes, and organizations all at once<br />
in a huge <strong>design</strong> effort. Instead, it is asserted that enterprise <strong>design</strong> should be <strong>based</strong> on an<br />
<strong>approach</strong> in which support is provided for identifying interdependencies, and tools are provided<br />
for modeling across boundaries.<br />
Decomposing a problem, in this case enterprise <strong>design</strong>, into smaller problems is a<br />
common <strong>approach</strong> in <strong>design</strong>. However, the decomposition of the enterprise in functional<br />
groups, <strong>with</strong> a variety of characteristics, needs, requirements and constraints, poses the problem<br />
3
of coordinating the activities of the various groups, as each one typically resorts to optimize its<br />
own subsystem. Although this <strong>approach</strong> is advantageous locally, the individually motivated<br />
decisions are usually not advantageous for the system as a whole. Even more, the objectives of<br />
the various <strong>design</strong> processes are frequently in conflict <strong>with</strong> each other. Thereupon, it is<br />
necessary to establish an effective cooperative framework for enterprise <strong>design</strong>, and to<br />
formulate and solve the resulting local <strong>design</strong> problems <strong>with</strong> an appropriate <strong>approach</strong>, consistent<br />
<strong>with</strong> the aforementioned framework. According to the preceding discussion, the objective of<br />
this work is twofold:<br />
1. To establish a mathematically supported cooperative framework that enhances a<br />
practical, effective and efficient integration of the enterprise <strong>design</strong> processes.<br />
2. To provide an appropriate <strong>approach</strong> for the formulation and solution of <strong>design</strong><br />
problems, consistent <strong>with</strong> such a cooperative framework.<br />
In working out this objective, the focus is on manufacturing enterprises that produce discrete<br />
products. Such systems are normal in mechanical, electrical and electronics industries.<br />
However, the framework that is constructed in this thesis is generic and can be readily applied<br />
to such systems.<br />
1.2 INTELLECTUAL FOUNDATIONS<br />
1.2.1 Systems Thinking in Design and Manufacturing<br />
4
Designers, engineers and production managers carry a heavy responsibility since their ideas,<br />
knowledge and skills determine in a decisive way the competitiveness of their enterprises. This<br />
responsibility places the onus on them to carry out their activities in a more rigorous way.<br />
Manufacturing enterprises cannot afford anymore a trial-and-error <strong>approach</strong> to <strong>design</strong> of<br />
products, its processes and organization.<br />
The traditional <strong>approach</strong> to <strong>design</strong> and planning has its roots in the reductionist method:<br />
the work is decomposed into its distinct subsystems and then each subsystem is “optimized.”<br />
But alternatively, <strong>design</strong> of subsystems can be viewed much more in terms of the interactions<br />
between them and in the light of the performance of the system as a whole. The main reason<br />
that motivates adopting such a systems thinking is that a system is usually more than the sum of<br />
its constitutive parts, because the parts are dynamically interrelated and interdependent. In<br />
other words, in complex systems it is both the organization of components and their physical<br />
properties that largely determine behavior. Hence, the major opportunities for improvements<br />
of the manufacturing enterprises lie at the interfaces (e.g., between sales and manufacturing, or<br />
between product development and manufacturing). The major benchmarks behind a systems-<br />
thinking mindset are (Wright, 1989):<br />
?? Goal seeking. A system’s inner actions result in some position of dynamic equilibrium<br />
where activities can be directed to goal attainment.<br />
5
?? Holism. A system is an inseparable entity. Hence, each system should be studied as a<br />
whole, i.e., holistically. Holism views a whole entity <strong>with</strong> all its parts in natural motion and<br />
does not force divisions of the indivisible.<br />
?? Hierarchy. A system has subparts or subsystems that are organized in a hierarchy, from<br />
those of major scope down to those subsets of minimal influence on goal attainment.<br />
?? Equifinality. An open system can reach its goals in a number of ways. This flexibility on the<br />
means to goals is made possible because, unlike a closed system <strong>with</strong> a pre-set alternative,<br />
open systems can change inputs and some can also change goals. Hence, the tendency<br />
toward inertia can be thwarted by open-systems when they can be managed.<br />
With such a systems-thinking perspective, subsystems are <strong>design</strong>ed <strong>with</strong> emphasis in the overall<br />
system performance, and the process of <strong>design</strong> is forced into a total systems context.<br />
1.2.2 Design as a Science of the Artificial<br />
Increasing manufacturing competitiveness is demanding organizations to develop products and<br />
their manufacturing processes more efficaciously. Designers should not anymore rely solely in<br />
experience and intuition. Such experience and intuition will always be desirable, and probably<br />
necessary, but <strong>design</strong>ers should be more systematic and support their activity <strong>with</strong> rigorous<br />
analysis in order to be effective in the new marketplace. Systematization and rigor brings<br />
<strong>design</strong> into the realm of science.<br />
6
It should be apparent that natural science and <strong>design</strong> are essentially different, the<br />
fundamental difference being that the former is concerned <strong>with</strong> analysis of phenomena, whereas<br />
<strong>design</strong> is concerned <strong>with</strong> the synthesis of artificial systems to function. However, if the general<br />
structure of <strong>design</strong> as a discipline and most of its elements can be understood, and it can be<br />
carried out <strong>with</strong>in a systematic methodology and supported <strong>with</strong> rigorous analysis and synthesis,<br />
<strong>design</strong> can be envisioned as a science of the artificial (Simon, 1996).<br />
Ordinary systems of logic serve natural sciences well because the concern of standard<br />
logic is <strong>with</strong> declarative statements, suited for assertions about the world and far for inferences<br />
from those assertions. On the other hand, <strong>design</strong> is concerned <strong>with</strong> how things ought to be.<br />
Even when there have been developed a number of constructions of modal logic for handling<br />
“shoulds”, and “oughts” of various kinds, none of these systems has been sufficiently developed<br />
to be adequate to handle the requirements of <strong>design</strong>. However, as affirmed by Herbert Simon<br />
(Simon, 1996, p. 115), “the requirements of <strong>design</strong> can be met fully by a modest<br />
adaptation of ordinary declarative logic. Thus a special logic of imperatives is<br />
unnecessary.”<br />
There exists a considerable area of <strong>design</strong> practice where standards on rigor in inference<br />
are high, in particular the so-called optimization methods. Since the optimization problem is a<br />
standard mathematical problem --to find an admissible combination of parameters that maximize<br />
a utility function subject to constraints-- it is evident that the logic used to deduce the answer is<br />
the standard logic of the predicate calculus on which mathematics rests. This formalism avoids<br />
7
making use of a special logic of imperatives by dealing <strong>with</strong> sets of possible worlds: we simply<br />
ask what values the command variables would have to maximize the utility function while<br />
meeting the stated constraints, and we conclude that these are the values the command variables<br />
should have. Hence, a scientific <strong>design</strong> practice should incorporate (Simon, 1996):<br />
1. A utility theory as a logical framework for rational choice among alternatives, and tools for<br />
actually making the computations.<br />
2. The body of techniques for actually deducing which of the available alternatives is the<br />
optimum, which include mathematical programming, simulation, queuing theory, and control<br />
theory, to name a few.<br />
3. Algorithms (that can be of a heuristic nature) to search for alternatives.<br />
4. A theory of structure and <strong>design</strong> organization.<br />
5. Representation of <strong>design</strong> problems.<br />
The subject of computational techniques need not be limited to optimization.<br />
Traditional engineering <strong>design</strong> methods make use of “figures of merit” to compare between<br />
<strong>design</strong>s in terms of “better” and “worse” <strong>with</strong>out actually providing a judgment of “best.” In<br />
<strong>design</strong>, we usually satisfy because we are unable to optimize, as the set of all alternatives cannot<br />
be always evaluated due to computational capacity, nor we can recognize the best alternative,<br />
even if we are fortunate enough to generate it. This is so because in real-world systems there is<br />
always present a certain amount of incompleteness and uncertainty in the information available.<br />
Herbert Simon coined the term “satisficing” to describe those solutions that are “good enough”<br />
8
ather than optimal. By using this term to describe the results of a “scientific” <strong>design</strong> effort it is<br />
recognized that in building a science of <strong>design</strong> there are segments of such an effort that can be<br />
organized <strong>with</strong> a systematic and formal theory, while other segments have to be more pragmatic,<br />
more empirical. This is the fundamental tenet of a decision-<strong>based</strong> <strong>design</strong> paradigm, which is<br />
described in the next section.<br />
1.3 FRAME OF REFERENCE: DECISION-BASED DESIGN<br />
Ideally, engineering <strong>design</strong> involves two steps: (1) determine all possible <strong>design</strong> options,<br />
and (2) choose the best one. Simple as this may sound, these two steps are extremely difficult<br />
or impossible to perform for all but the most simple systems. Some of the difficulties<br />
encountered are (Hazelrigg, 1996):<br />
?? For most systems, the range of possible <strong>design</strong> options is virtually limitless.<br />
?? It is not possible to know exactly how a particular <strong>design</strong> will perform until it is built. Thus,<br />
<strong>design</strong> is a matter of decision making under conditions of uncertainty and risk. Engineers<br />
frequently use models to assist in predicting the behavior of a <strong>design</strong>. But models, which are<br />
abstractions of reality, are never precise and never predict the future <strong>with</strong> certainty.<br />
Moreover, models cannot be fully validated until after the <strong>design</strong> has been decided upon and<br />
the system built.<br />
9
?? In order to compare and rank order <strong>design</strong> options, it is necessary to have a valid measure<br />
of value. It is not a trivial task to identify a valid value measure, particularly under conditions<br />
of uncertainty and risk.<br />
?? The dimensionality of a typical <strong>design</strong> is so large that it is computationally infeasible to<br />
search all possible options for the best one.<br />
As a consequence of these difficulties, engineering <strong>design</strong> can never be reduced to a prescriptive<br />
procedure exclusive of human input. Hazelrigg (1996) asserts that human input will always be<br />
needed for the following:<br />
?? assessment of probabilities (or their equivalent) describing random events,<br />
?? determination of an appropriate value measure, and<br />
?? judgment on which options to include in consideration of alternative <strong>design</strong>, and which to<br />
neglect.<br />
Thus, <strong>design</strong> is a decision-making process, rather than a problem solving process. “Decision-<br />
<strong>based</strong> <strong>design</strong>” is a term coined to emphasize such a perspective from which to develop<br />
methods for <strong>design</strong>.<br />
In decision-<strong>based</strong> <strong>design</strong> it is asserted that the principal role of a <strong>design</strong>er is to make<br />
decisions (Mistree, et al., 1990). Decisions help bridge the gap between an idea and reality.<br />
In general, decisions are characterized by information from many sources and disciplines and<br />
may have wide-ranging repercussions. In decision-<strong>based</strong> <strong>design</strong> they represent a unit of<br />
communication that has both domain-dependent and domain-independent features. This notion<br />
10
is consistent <strong>with</strong> the definition of a decision as, “a choice from among a set, either closed or<br />
open, of options, an irrevocable allocation of resources” (Hazelrigg, 1996). Such a<br />
conception of <strong>design</strong> brings the rationality of decision analysis to the field of engineering <strong>design</strong>.<br />
The discipline of decision analysis is <strong>based</strong> on a normative <strong>approach</strong> in which decision<br />
making has three main elements (Hazelrigg, 1996): identification of options, determination of<br />
expectations of each option, and an expression of values. The resulting rule is, “the preferred<br />
decision is the option whose expectation has the highest value.” An expectation comprises the<br />
range of possible outcomes of a decision paired <strong>with</strong> probabilities of occurrence. Determination<br />
of expectations on each option is the realm of modeling. Such expectations are practically<br />
always <strong>probabilistic</strong> because expectations relate to the future, and it is very difficult to predict<br />
the future <strong>with</strong> precision and certainty.<br />
There is a difference between solving problems and making decisions. The solution of a<br />
problem depends only on the statement of the problem, the data or boundary conditions, laws<br />
of nature and axioms of mathematics. Decisions however, involve options, expectations and<br />
values. And, whereas problems can be solved in the absence of options and values, decisions<br />
cannot be made <strong>with</strong>out their consideration. Thus, decision making cannot be done in the<br />
absence of human values, which are an expression of human needs and desires. There is no<br />
clear connection between the laws of nature and human values, and therefore, there are no<br />
known physical laws that govern the determination of values. The purpose of values in decision<br />
making is to enable the rank ordering of alternatives. It is convenient to use a quantitative<br />
11
function to automate the process of rank ordering options because it is too cumbersome to<br />
make an assessment of relative merits of every <strong>design</strong> alternative versus every other <strong>design</strong><br />
alternative. Such a comparison is generally too complex to make it accurately and consistently<br />
<strong>with</strong>out the use of a mathematical value model, particularly in the presence of uncertainty.<br />
Therefore, in decision-<strong>based</strong> <strong>design</strong>, a <strong>design</strong> decision is a decision that is intended to maximize<br />
value.<br />
By adopting decision-<strong>based</strong> <strong>design</strong> as the frame of reference for achieving enterprise<br />
integration it is possible to develop a comprehensive and practical <strong>approach</strong> for enterprise<br />
<strong>design</strong>, <strong>based</strong> on a common paradigm for decision support. This is discussed in Section 1.4.<br />
1.4 A DECISION-BASED ENTERPRISE DESIGN<br />
1.4.1 The Concept of Enterprise Design<br />
Given the focus of this work in manufacturing enterprises, a fitting description of an enterprise is<br />
offered by Gale and Eldred (1996):<br />
An enterprise is a system of business objects that are in functional symbiosis. Symbiosis<br />
means that business objects work together to accomplish mutually beneficial goals. The<br />
business objects of the enterprise include people, machines, buildings, processes, events,<br />
and information that combine to produce the products of the enterprise.<br />
In other words, an enterprise is the sum of its people and the jobs and tasks they perform, its<br />
products, processes and facilities, its information systems and flows, and so on. Clearly all of<br />
these aspects come into being at some time, and then remain dynamic, changing over time.<br />
12
Hence, as affirmed by Peplinski (1997), “an enterprise as a whole is <strong>design</strong>ed, either implicitly<br />
or explicitly.”<br />
In practice, because of the complexity of dealing <strong>with</strong> enterprises in their entirety, the<br />
<strong>design</strong> of each aspect is treated separately by functional departments such as research and<br />
development, product <strong>design</strong>, production, distribution, human resources, etc. However, in this<br />
thesis all of these aspects are grouped in three major different <strong>design</strong> processes: product <strong>design</strong>,<br />
manufacturing process <strong>design</strong>, and organization <strong>design</strong>. In this work, these processes are<br />
defined as follows:<br />
?? Product <strong>design</strong> is the conception of a physical artifact to fulfill a set of requirements.<br />
?? Manufacturing process <strong>design</strong> includes microprocess <strong>design</strong> and macroprocess <strong>design</strong>.<br />
Microprocess <strong>design</strong> is the field of process engineering, which deals <strong>with</strong> the specification of<br />
equipment, methods, fixtures and process conditions of individual operations for the<br />
manufacture of a product. On the other hand, macroprocess <strong>design</strong> is the scope of<br />
operations management, which can be defined as the <strong>design</strong> of the production system that<br />
create the firm’s products, and includes facility location and layout, and the planning,<br />
scheduling and operation of manufacturing facilities.<br />
?? Organization <strong>design</strong> refers to the <strong>design</strong> of the organization itself, through the setting of<br />
variables under management control. These variables include the organization structure<br />
(boundaries, levels), technologies (both product and process), people (hiring, training), tasks<br />
(work <strong>design</strong>), reward systems and information systems.<br />
13
Enterprise <strong>design</strong> is embodied by the integration of these three <strong>design</strong> processes<br />
(Peplinski, 1997). It is clear that enterprise <strong>design</strong> encompasses nearly all aspects of a<br />
manufacturing organization. Designing a manufacturing enterprise is a huge task, and clearly it is<br />
beyond the scope of one single person. It is necessary to carry out the enterprise activity<br />
through decomposition of tasks and responsibilities, assigned to various individuals. Almost all<br />
of these tasks involve making decisions (which usually require problem solving) rather than<br />
merely solving problems. Consequently, the enterprise activity becomes a decision-making<br />
process, in which such decisions are made in face of particular aspects of the enterprise<br />
requirements.<br />
The view of an enterprise activity as a decision-making process suggests pursuing<br />
enterprise integration through coordination of decisions. There are two options to coordinate<br />
decisions: either a centralized and prescriptive process can be specified to enforce coordination,<br />
or an empowered decision-<strong>based</strong> <strong>approach</strong> in which support is provided for identifying<br />
interdependencies and tools are provided for modeling across boundaries. Within this last<br />
vision, enterprise <strong>design</strong> becomes (Peplinski, and Mistree, 1997):<br />
?? identifying and assessing the enterprise-wide effects of each local decision,<br />
?? identifying and assessing the effects of other external decisions on the local decision at hand,<br />
and<br />
?? making each decision to satisfy as many of the enterprise-wide goals as possible, while being<br />
as robust as possible to external decisions that are beyond the decision-maker’s control.<br />
14
Based on this perspective of empowerment through decision-making, a method for enterprise<br />
<strong>design</strong> has been proposed by Peplinski and Mistree (1997), <strong>based</strong> on a common model for<br />
decision support. This method is examined in the next section.<br />
1.4.2 A Decision-Based Method for Enterprise Design<br />
In Figure 1.1 a model for decision support spanning operations research, engineering <strong>design</strong>,<br />
and systems theory, proposed by Peplinski and Mistree (1997), is illustrated. In this model for<br />
decision support, a decision is represented in terms of a set of <strong>design</strong> variables ? X ? and a set of<br />
n goals and constraints captured by target values {T1, T2, ..., Tn} and models {F1(X), F2(X), ...,<br />
Fn(X)}. Regions of interest are specified for each <strong>design</strong> variable a priori, and these regions<br />
define the realm of potential solutions for the decision. Goals and constraints represent the<br />
motives for making the decision, and may encompass measures of system cost, performance,<br />
quality, and so on. The target values Ti represent the “aspiration space” or what ideally will be<br />
achieved <strong>with</strong> the decision, and the models Fi(X) quantify to what extent these aspirations are<br />
achieved by the actual values of the <strong>design</strong> variables. Analysis is then the process of computing<br />
the values of Fi(X) for given values of {X}, and synthesis is the process of using these computed<br />
values to find the best settings for the <strong>design</strong> values. A specific value is identified from the<br />
region of interest for each <strong>design</strong> variable, and this set of values represents the solution that<br />
comes closest to achieving all of the goal targets while not violating any constraint.<br />
15
Synthesis<br />
Analysis<br />
Objectives & Constraints<br />
(Requirements)<br />
T 1 T 2 T n<br />
F 1(x)<br />
Models<br />
F 2 (x) F n (x)<br />
Design Variables<br />
{x}<br />
Figure 1.1 A Hybrid Paradigm for Decision Support (Peplinski, 1997)<br />
This representation builds upon the strengths of operations research, engineering <strong>design</strong><br />
and systems theory. It follows the operations research structure of decision variables,<br />
constraints, and an objective function. It is built upon the virtues of engineering <strong>design</strong> as it<br />
capitalizes on and integrates the basic activities in engineering <strong>design</strong>, namely modeling, analysis<br />
and synthesis, where modeling encompass the universe of techniques that exist to predict system<br />
performance or behavior, analysis represents the process of performance prediction, and<br />
synthesis represents the prescriptive process of finding the values of the <strong>design</strong> variables. This<br />
model for decision support has the following characteristics:<br />
?? It is domain-independent.<br />
?? The partitioning of activities into modeling, analysis and synthesis fosters the creation of<br />
methods and tools for decision support.<br />
?? It sets the stage for concrete and specific definitions of interdependence and integration in<br />
terms of decisions.<br />
Synthesi<br />
s<br />
16
Based on this model for decision support, Peplinski and Mistree proposed an enterprise <strong>design</strong><br />
method, which is described in broad strokes as a series of tasks in Figure 1.2. These tasks<br />
correspond to the development, formulation and solution of an enterprise <strong>design</strong> decision<br />
according to the decision support model. The method is represented as a relatively straight<br />
sequence of tasks, but in actuality iteration will likely occur <strong>with</strong>in and across tasks. Each of the<br />
tasks in Figure 1.2 are expanded into more detailed diagrams of tasks, information flows, and<br />
decisions, but due to space constraints these diagrams are not given here.<br />
In Figure 1.2, a <strong>design</strong>er begins <strong>with</strong> a baseline decision, represented in terms of <strong>design</strong><br />
variables, goals and constraints, and models for analysis. In product <strong>design</strong> the variables may<br />
represent system parameters or characteristics, while in organization <strong>design</strong> the variables may<br />
represent different policies or courses of action. This decision is likely geared toward local<br />
measures of merit and may not capture its enterprise-wide implications.<br />
17
Baseline Decision<br />
Task 1 Expand Scope of Decision<br />
Task 2a Identify Modeling Techniques<br />
Task 2b Determine Input Needed for<br />
Analysis<br />
Task 3 Define Boundaries of Decision<br />
Task 4 Transform and Integrate Models<br />
Task 5 Generate Potential Solutions<br />
Recommendation<br />
Figure 1.2 A Decision-Based Method for Enterprise Design<br />
(Peplinski, 1997)<br />
18
Task 1 of the method is employed to explore the potential impact of the baseline<br />
decision. A wide range of enterprise-wide effects are generated by a process of ideation<br />
(brainstorming is a specific example of this) both <strong>with</strong> the <strong>design</strong> variables and <strong>with</strong> an<br />
awareness of overall customer requirements. From this process, additional goals are identified<br />
to expand the scope of the decision.<br />
During Tasks 2a and 2b modeling schemes are identified to quantify the relationships<br />
between the baseline <strong>design</strong> variables and the additional goals from Task 1. These modeling<br />
schemes are selected from the diverse range of existing system modeling techniques in the<br />
enterprise (which include finite element analysis, system dynamics, discrete event simulation,<br />
stochastic models, regression models, physical prototypes, to name a few); in this fashion<br />
integration is fostered. The two tasks are represented separately to imply a process of iteration;<br />
modeling techniques are identified both by examining the goals and by examining the <strong>design</strong><br />
variables. Iteration may be required to resolve inconsistencies. At the conclusion of these tasks<br />
a modeling scheme is selected for each goal and constraint in the decision. These modeling<br />
schemes are tentative and require further hardening. Also, these modeling schemes may require<br />
additional input information, represented by additional clouds. The additional goals and input<br />
information from Tasks 2a and 2b may overlap <strong>with</strong> other decisions made by different people or<br />
at different times throughout the enterprise. Thus the interdependencies between decisions<br />
become explicit. Because of empowerment it not intended to gather all decisions under some<br />
centralized control, and because there are limitations on the size of any one decision, so at some<br />
19
point the boundaries of the decision must be drawn. This is represented in Task 3. These<br />
boundaries may be forced to slice through interdependencies, so they may be resolved by<br />
formulating a decision that is robust to external issues. The information that remains inside the<br />
decision boundary is then classified into constraints, <strong>design</strong> variables and noise factors.<br />
As Task 4 is initiated, all of the elements of the decision are potentially in place. Modeling<br />
techniques have been identified for each goal and constraint, and the input needed for each<br />
model is captured by the sets of <strong>design</strong> variables, noise factors, and constants that have been<br />
identified. The issues that remain are whether the models are efficient enough, and whether<br />
models are in a format that allows integration into a decision formulation. Peplinski and Mistree<br />
(1997) suggest the use of statistical metamodeling techniques during Task 4 to resolve these<br />
issues.<br />
During Task 5 an enterprise <strong>design</strong> decision is formulated mathematically and solved.<br />
“Solution” occurs through an in-depth process of exercising the formulation - changing<br />
parameters, testing assumptions, and exploring “what-if” scenarios - thereby generating sets of<br />
potential solutions for consideration. In the end one solution is selected, one that satisfies as<br />
many of the enterprise-wide goals as possible and one that is as robust as possible to the effects<br />
of external decisions outside the <strong>design</strong>er’s control. This solution is the recommendation<br />
submitted for implementation. This method is discussed in detail in Peplinski (1997).<br />
20
1.5 RESEARCH QUESTIONS AND HYPOTHESES<br />
As indicated in Section 1.1, the objective of this work is to establish a mathematical framework<br />
for enterprise <strong>design</strong> <strong>with</strong>in a decision-<strong>based</strong> mindset. There are various issues related to such<br />
an aim. The first is the need of modeling the enterprise <strong>design</strong> process itself. Accordingly, a<br />
first research question is:<br />
Research question 1: How can an enterprise <strong>design</strong> process be modeled<br />
mathematically in the context of decision-<strong>based</strong> <strong>design</strong>?<br />
The fundamental tenet behind this question is that the enterprise <strong>design</strong> processes, namely,<br />
product process <strong>design</strong>, manufacturing process <strong>design</strong>, and organization <strong>design</strong>, are decision-<br />
making processes carried out by several employees who have to coordinate their decisions.<br />
Mathematical analysis of strategic behavior, where a decision-maker’s action depends on<br />
decisions by others is typical in <strong>game</strong> theory. Thus, <strong>game</strong> theory can provide the underpinning<br />
for modeling mathematically an enterprise <strong>design</strong> process. Because all the decisions of an<br />
enterprise cannot be made simultaneously, the enterprise <strong>design</strong> process has an important<br />
dynamic structure. The <strong>game</strong>s that are used to model and analyze such dynamic situations are<br />
known as <strong>game</strong>s in extensive form. Therefore, the enterprise <strong>design</strong> process can be abstracted<br />
as an extensive <strong>game</strong>. This notion is posed as Hypothesis 1:<br />
Hypothesis 1: An enterprise <strong>design</strong> process can be mathematically modeled as an<br />
extensive <strong>game</strong> in a decision <strong>based</strong> context.<br />
21
Application of <strong>game</strong> theory classically has had two goals. A first goal is a descriptive goal<br />
of understanding why the players behave as they do which includes describing the results of a<br />
certain <strong>game</strong> structure. A second goal is a more practical one of being able to advise the<br />
players of the <strong>game</strong> the best strategy to take. In enterprise <strong>design</strong>, the later goal can be<br />
expressed as how to appropriately formulate and make local decisions, i.e., decisions carried<br />
out by an individual agent of the organization, while enhancing the convergence of the global<br />
enterprise <strong>design</strong> process to a common satisfying solution. Here global is used to characterize<br />
the activity carried out by all the decision-makers that comprise an enterprise. This last concern<br />
is expressed as a research question as follows:<br />
Research question 2: How can enterprise decisions be coordinated through a <strong>design</strong><br />
formulation <strong>based</strong> on a <strong>game</strong> <strong>theoretical</strong> representation of the enterprise <strong>design</strong><br />
process?<br />
Game theory is divided into two branches: cooperative and non-cooperative <strong>game</strong> theory.<br />
Essentially, in non-cooperative <strong>game</strong> theory the unit of analysis is the individual participant in the<br />
<strong>game</strong> who is concerned <strong>with</strong> doing as well for himself/herself as possible subject to clearly<br />
defined rules and possibilities. If individuals happen to undertake behavior that in common<br />
parlance would be described as “cooperative,” then this is done because such behavior is in the<br />
best interests of each individual. In comparison, in cooperative <strong>game</strong> theory the unit of analysis<br />
is most often the group or the coalition; when a <strong>game</strong> is specified, part of the specification is<br />
22
what each group or coalition of players can achieve <strong>with</strong>out reference to how the coalition<br />
would affect a particular outcome or result.<br />
At first, it seems desirable to coordinate enterprise decisions using a cooperative<br />
formulation. However, there are two major drawbacks in trying to attain such solutions: first, a<br />
cooperative formulation is difficult to implement since it requires a single model to quantify all the<br />
relevant aspects that might affect the payoff of each player. Second, some enterprise decisions<br />
involve quite different time scope and have to be made at different times. It is not realistic to<br />
suppose that a <strong>design</strong>er can incorporate all the various aspects of an enterprise in solving a local<br />
<strong>design</strong> problem. Therefore, the focus on this thesis is in the coordination of decisions using non-<br />
cooperative <strong>game</strong> theory. This perception is enunciated in Hypothesis 2.1:<br />
Hypothesis 2.1: Enterprise decisions can be effectively coordinated <strong>with</strong> a non-<br />
cooperative formulation of the individual <strong>design</strong> problems.<br />
A fundamental assumption in coordinating decisions through a <strong>game</strong> <strong>theoretical</strong><br />
formulation, either cooperative or non-cooperative, is that the decision-makers that comprise<br />
the enterprise behave rationally, that is, their decisions are unbiased results of their analyses and<br />
estimations.<br />
Multiple criteria and multiple objectives, sometimes in conflict <strong>with</strong> each other, usually<br />
characterize the resulting <strong>design</strong> formulations. Therefore, a multi-objective <strong>design</strong> model is<br />
necessary to embody the <strong>design</strong> formulation. In this thesis the Compromise Decision Support<br />
Problem is used as such a model. The Compromise DSP is a multi-objective decision model<br />
23
which is a hybrid formulation <strong>based</strong> on Mathematical Programming and Goal Programming<br />
(Mistree et al., 1993). It is used to determine the values of <strong>design</strong> variables to satisfy a set of<br />
constraints and to achieve as closely as possible a set of conflicting goals. This assertion is<br />
captured in Hypothesis 2.2:<br />
Hypothesis 2.2: The Compromise DSP is an adequate decision model to embody <strong>game</strong><br />
<strong>theoretical</strong> <strong>design</strong> formulations <strong>with</strong>in an enterprise <strong>design</strong> framework.<br />
The embodiment of the aforementioned protocols using the Compromise DSP is discussed in<br />
Chapter 3.<br />
When considering an enterprise <strong>design</strong> problem, a <strong>design</strong>er has to deal <strong>with</strong> uncertain and<br />
incomplete information from other <strong>design</strong>ers. This is so because of the use of non-local<br />
information approximations, the need to forecast information, and the unavoidable randomness<br />
of the manufacturing environment (e.g., demand fluctuations). Therefore, the <strong>design</strong> problem is<br />
no longer choosing the right course of action but to find a way of calculating, very<br />
approximately, a good course of action. With this shift, the <strong>design</strong> solution becomes estimation<br />
under uncertainty. The real goal of a local action, <strong>with</strong>in a systems-thinking perspective, is to<br />
establish a satisficing local solution that is also satisficing to the other members of the system.<br />
The question is what are satisficing conditions for other members of the system?<br />
The use of approximations of non-local relevant information helps setting expectations of<br />
other members of the organization (assuming rational behavior). Therefore, it is possible to<br />
provide them <strong>with</strong> conditions that are reasonably close to their expectations. However, because<br />
24
each one wants to optimize their own performance, it is expected (furthermore, it is desirable)<br />
that there will be a certain number of negotiation and iteration. One way to improve<br />
convergence of the process is to provide each <strong>design</strong>er <strong>with</strong> the flexibility to find a satisficing<br />
solution in a region instead of trying to meet a single point. One <strong>design</strong> <strong>approach</strong> that is<br />
consistent <strong>with</strong> this notion is the <strong>probabilistic</strong>-<strong>based</strong> <strong>design</strong> model of Chen and Yuan (1997). In<br />
this work, it is suggested the use of Chen’s and Yuan’s <strong>probabilistic</strong>-<strong>based</strong> model in formulating<br />
and solving Compromise DSPs as an effective means to enhance coordination of decisions.<br />
This is expressed in the following research question and its corresponding hypothesis:<br />
Research question 3: How can the <strong>design</strong> problems be solved to enhance the<br />
coordination of enterprise <strong>design</strong> decisions?<br />
Hypothesis 3: A <strong>probabilistic</strong>-<strong>based</strong> <strong>design</strong> <strong>approach</strong> enhances the coordination of<br />
enterprise <strong>design</strong> decisions.<br />
In this <strong>probabilistic</strong> <strong>approach</strong>, the distribution of performance is collected through<br />
statistical analysis and the <strong>design</strong> parameters are expressed by <strong>probabilistic</strong> distributions. There<br />
are two main reasons behind Hypothesis 3:<br />
?? As argued in Section 1.3, when dealing <strong>with</strong> uncertain and incomplete information, only an<br />
expected value (i.e., a <strong>probabilistic</strong> solution) is a valid measure to rank alternatives.<br />
?? By providing <strong>design</strong> solutions as ranges, instead of single point values, the possibility to meet<br />
a common satisfying solution is enhanced.<br />
The <strong>probabilistic</strong>-<strong>based</strong> <strong>design</strong> model is explained in Chapter 3.<br />
25
1.6 ORGANIZATION OF THE THESIS<br />
In this chapter a series of research questions and their respective hypothesis are elaborated.<br />
These questions and hypotheses are summarized in Table 1.1.<br />
Table 1.1 Research Questions and Hypothesis<br />
RESEARCH QUESTIONS AND HYPOTHESES<br />
Research question 1<br />
How can an enterprise <strong>design</strong> process be modeled mathematically in the context of decision<strong>based</strong><br />
<strong>design</strong>?<br />
?? Hypothesis 1: An enterprise <strong>design</strong> process can be mathematically modeled as an<br />
extensive <strong>game</strong> in a decision-<strong>based</strong> <strong>design</strong> context.<br />
Research question 2<br />
How can enterprise decisions be coordinated through a <strong>design</strong> formulation <strong>based</strong> on a <strong>game</strong><br />
<strong>theoretical</strong> representation of the enterprise <strong>design</strong> process?<br />
?? Hypothesis 2.1: Enterprise decisions can be effectively coordinated <strong>with</strong> a noncooperative<br />
formulation of the individual <strong>design</strong> problems.<br />
?? Hypothesis 2.2: The Compromise DSP is an adequate decision model to embody <strong>game</strong><br />
<strong>theoretical</strong> <strong>design</strong> formulations <strong>with</strong>in an enterprise <strong>design</strong> framework.<br />
Research question 3<br />
How can the <strong>design</strong> problems be solved to enhance the coordination of enterprise <strong>design</strong><br />
decisions?<br />
?? Hypothesis 3: A <strong>probabilistic</strong>-<strong>based</strong> <strong>design</strong> <strong>approach</strong> enhances the coordination of<br />
enterprise <strong>design</strong> decisions.<br />
26
These hypotheses embody the framework for enterprise <strong>design</strong> that is posed as the<br />
objective of this work in Section 1.1. A scheme of the strategy followed in developing this<br />
framework is illustrated in Figure 1.3.<br />
A common<br />
paradigm for<br />
decision<br />
support<br />
The<br />
Compromise<br />
DSP<br />
Enterprise Design as an Extensive Game<br />
Game Theory<br />
Decision-Based Design<br />
Design as a Scientific Discipline<br />
Systems Thinking<br />
Closure<br />
Case Study<br />
A<br />
Probabilistic<br />
Design<br />
Model<br />
Figure 1.3 Implementation Strategy<br />
Chapter 5<br />
Chapter 4<br />
Chapter 3<br />
Chapter 2<br />
Chapter 1<br />
In this chapter the foundations of a framework for enterprise <strong>design</strong> are established,<br />
namely, systems thinking, decision-<strong>based</strong> <strong>design</strong>, and <strong>design</strong> as a scientific discipline. In<br />
Chapter 2 the enterprise <strong>design</strong> process is abstracted as an extensive <strong>game</strong>. The abstraction is<br />
27
ased on the fact that an enterprise is composed of decision-makers who have to negotiate the<br />
value of <strong>design</strong> parameters to find mutually satisfying solutions. An overview of existing<br />
enterprise modeling tools and methods is presented in Section 2.1. In Section 2.2 the use of<br />
<strong>game</strong> theory to model and support <strong>design</strong> processes is examined. Based on such analysis, the<br />
extensive-form <strong>game</strong> is discussed as a means of modeling the enterprise <strong>design</strong> process in<br />
Section 2.3. The implications of the resulting model are discussed in Section 2.4.<br />
In Chapter 3 the formulation of <strong>design</strong> problems is discussed, <strong>based</strong> on the perception of<br />
enterprise integration as coordination of decisions. In Section 3.1, two different coordination<br />
mechanisms, namely implicit coordination and explicit coordination, are examined. A way to<br />
merge them and capitalize their different advantages <strong>with</strong> a decision-<strong>based</strong> perspective using a<br />
common model of decision support and the Compromise DSP is presented in Sections 3.2 and<br />
3.3. Section 3.4 contains a discussion of how coordination <strong>based</strong> on the use of compromise<br />
DSPs can be translated into specific <strong>design</strong> formulations, either cooperative or non-cooperative,<br />
as well as methods to implement both. In Section 3.6 a <strong>probabilistic</strong> <strong>approach</strong> for solving the<br />
resulting compromise DSPs is proposed, ideally facilitating the cooperative process.<br />
In Chapter 4, the framework developed in the first three chapters is exercised in a case<br />
study: the integrated product and process <strong>design</strong> of an absorber/evaporator module for a family<br />
of absorption chillers. Therefore, the <strong>approach</strong> is applied to both product <strong>design</strong> and process<br />
<strong>design</strong>. The synthesis of the system is performed <strong>with</strong> four different <strong>approach</strong>es:<br />
1. The various individuals involved in the <strong>design</strong> process acting in isolation (Section 4.3).<br />
28
2. The individuals coordinating their decisions through a cooperative formulation of the <strong>design</strong><br />
problem (Section 4.4).<br />
3. The individuals coordinating their decisions through a non-cooperative formulation of the<br />
<strong>design</strong> problem and using a non-<strong>probabilistic</strong> <strong>design</strong> model (Section 4.5).<br />
4. The individuals coordinating their decisions through a non-cooperative formulation of the<br />
<strong>design</strong> problem and using a <strong>probabilistic</strong> <strong>design</strong> model (Section 4.6).<br />
In Section 4.7 the results obtained <strong>with</strong> the four <strong>approach</strong>es are compared and discussed, and<br />
in Section 4.8 the case study is summarized and open issues are examined.<br />
Finally, in Chapter 5 the thesis is surveyed, its relevance for the engineering and <strong>design</strong><br />
practice is discussed and possible directions of future work are identified.<br />
29
CHAPTER 2<br />
ENTERPRISE DESIGN AS AN EXTENSIVE GAME<br />
In this chapter, the first element of the enterprise <strong>design</strong> framework proposed in Chapter<br />
1 is examined, namely, the abstraction of an enterprise <strong>design</strong> process as an extensive <strong>game</strong>.<br />
The motivation for using a <strong>game</strong> <strong>theoretical</strong> representation of the enterprise is to capture the<br />
dynamics of interaction between decision makers who are trying to coordinate their activities<br />
through the appropriate formulation of their decisions. First, a review of previous enterprise<br />
modeling representations, along <strong>with</strong> a discussion on their use as a decision support tool, is<br />
presented. Then, the use of <strong>game</strong> theory in engineering <strong>design</strong> is reviewed in order to introduce<br />
the basic concepts that are used to abstract an enterprise <strong>design</strong> process as a <strong>game</strong>. The core<br />
of this chapter is Section 2.3 where it is addressed how an enterprise <strong>design</strong> process can be<br />
mathematically modeled as an extensive <strong>game</strong>. Attention is paid to a particular type of solution<br />
of extensive <strong>game</strong>s, Stackelberg equilibrium. The assumptions and definitions used in<br />
abstracting <strong>design</strong> processes as <strong>game</strong>s are examined. Finally, the chapter ends <strong>with</strong> a<br />
discussion of the relevance of modeling an enterprise activity as a <strong>game</strong>.<br />
29
2.1 OVERVIEW OF ENTERPRISE MODELING: ARCHITECTURES,<br />
TOOLS AND METHODS<br />
An imperative for achieving a framework for enterprise <strong>design</strong> and integration is enterprise<br />
modeling. Designers need models of the enterprise to allow them to “tune” the formulation of<br />
their individual decisions <strong>with</strong> the performance of the system.<br />
Most of the efforts in enterprise modeling have been undertaken in Computer Integrated<br />
Manufacturing (CIM), and they have primarily focused on creating architectures, modeling tools<br />
and modeling methods (Kateel, et al., 1996). This section presents a brief overview of some<br />
selected architectures, tools and methods to exemplify the enterprise models that are being<br />
currently used in CIM.<br />
2.1.1 Modeling Architectures<br />
In CIM, an architecture is defined as a “body of rules that define those system features which<br />
directly affect the manufacturing environment into which the system is placed. These features<br />
include system configuration, component locations, interfaces between the system and its<br />
environment, and modes of operation” (O’Sullivan, 1994). Several enterprise modeling<br />
architectures have been proposed in the literature, and surveys can be found in (Kateel, et al.,<br />
1996) and (O’Sullivan, 1994). A brief overview of selected architectures follows:<br />
CIM-OSA: This CIM-Open Systems Architecture has three levels of model derivation:<br />
requirements definition, <strong>design</strong> specification and implementation description. It also has four<br />
30
views (functional, informational, resource, and organizational) and three levels of model<br />
instantiation: generic, partial, and particular (Jorysz and Vernadat, 1990).<br />
NBS: This architecture, developed by the National Bureau of Standards, is <strong>based</strong> on the<br />
concept of hierarchical control. The NBS architecture was developed to consist of five levels of<br />
levels of hierarchy: facility, shop, cell, work station, and machine. The decomposition is <strong>based</strong><br />
on procedures, functions, and rules (O’Sullivan, 1994).<br />
ICAM: The Air Force’s Integrated Computer Aided Manufacturing project offers a suite<br />
of IDEF (ICAM DEFinition) modeling methods, from functional modeling (IDEF0), information<br />
modeling (IDEF1), data modeling (IDEF3), object oriented <strong>design</strong> (IDEF4), and ontology<br />
capture (IDEF5), among others (O’Sullivan, 1994).<br />
CAM-I: The Computer Aided Manufacturing-International architecture is a general<br />
representation of manufacturing enterprises. The functional decomposition method used to<br />
create this architecture is prescribed to indicate the details such as company policies and<br />
procedures, organizational structure and standards (Doumeingts, et al., 1995).<br />
IMPACS: The IMPACS architecture uses IDEF0, Data Flow Diagrams (DFD),<br />
“Graphe a Resultatis er Activities Intercies” (GRAI) Grids and nets, IDEF1x and group<br />
technology. IMPACS outlines a cellular architecture. Software modules such as dispatcher,<br />
scheduler, mover, producer and monitor control the production cells. These software modules<br />
are <strong>design</strong>ed to be compatible <strong>with</strong> each other even if different vendors develop them.<br />
O’Sullivan (1994) states that the IMPACS architecture has been widely accepted among<br />
31
manufacturing software vendors as a useful and practical interpretation of the production<br />
management system.<br />
2.1.2 Modeling Tools<br />
CIM models are built using the modeling tools prescribed by modeling methodologies. These<br />
tools refer to techniques used for diagrammatically representing functions or activities. Some of<br />
the most widely known modeling tools are:<br />
IDEF Modeling Tools: IDEF0, IDEF1, IDEF1x, and IDEF2 are modeling tools<br />
developed by the ICAM project of the US Air Force. These four main modeling tools,<br />
complement each other to provide functional, informational and dynamic models of the system.<br />
a) IDEF0 Models: IDEF0 is a comprehensive and expressive functional modeling<br />
language capable of graphically representing a wide variety of business, manufacturing<br />
and other types of enterprise operations to any level of detail. The basic construction<br />
block of IDEF0 is a function block linked to other function blocks by inputs, outputs,<br />
mechanisms and controls. Links between the blocks may be either physical objects<br />
(material flow) or information flow. These models have three important features:<br />
“context” which indicates the position that the subject model takes up in the systems<br />
hierarchy, “viewpoint’ which refers to the perspective which the model adopts, and the<br />
“purpose” which indicates the reason for existence of the model (Colquhoun, et al.,<br />
1993).<br />
32
) IDEF1, and IDEF1x Models: IDEF1 is a technique for modeling information<br />
requirements of a function, in terms of the structure of information. IDEF1x is an<br />
extension of IDEF1 and deals <strong>with</strong> the flow of information (Pandya, 1995).<br />
c) IDEF2 Models: IDEF2 models have the capability to describe as well analyze a<br />
system. However, IDEF2 is an unsupported simulation language. Hence, other<br />
simulation models such as ARENA (Drevna and Kasales, 1994), or PROMODEL<br />
(Baird and Leavy, 1994) are more commonly used.<br />
Structured System Analysis (SSA): SSA is a data flow <strong>approach</strong> to systems <strong>design</strong>. It<br />
is also quite effective in representing the flow of physical entities. This technique prompts the<br />
user to think in terms of what to accomplish rather than how. It is more detailed and software<br />
oriented than IDEF0 and is the most preferred modeling tool for data flows (Gane and Sarson,<br />
1979).<br />
GRAI Grids and GRAI Nets: GRAI grids and nets are the tools developed by the<br />
GRAI Laboratories of France to model decision-flow processes in manufacturing environments.<br />
In the GRAI grid, various activities are modeled <strong>with</strong> respect to the decisions and information<br />
flows between them while the GRAI net delineates the decision-making process itself.<br />
2.1.3 Modeling Methods<br />
33
Modeling methods in CIM are guidelines to combine the modeling tools described in the<br />
previous section to model a particular system. Some of the methods used for modeling<br />
enterprise systems include IDEF, SSADM, SADT, GIM and CIM-OSA cube:<br />
IDEF Method: This methodology prescribes the integration of IDEF0, IDEF1 and<br />
IDEF2 and models to describe functional, informational and dynamic models of an enterprise.<br />
Structured Systems Analysis and Design Methodology (SSADM): SSADM is a<br />
procedural framework developed specifically for use in system development projects. It uses<br />
three modeling tools: data flow diagrams, logical data structures, and entity life histories to<br />
provide function, data, and event views of the systems (Pandya, 1995).<br />
Structured Analysis and Design Technique (SADT): A system analysis and <strong>design</strong><br />
methodology to represent the structure of the system diagrammatically. This method makes use<br />
of a number of tools such as activity diagrams, data diagrams, node lists and data dictionaries to<br />
represent the structure of the system. The analysts are required to consider activity or function<br />
and data views of the system being modeled, thereby encouraging the creation of integrated<br />
enterprises.<br />
CIM-OSA Method: Defines an integrated methodology to <strong>design</strong>, implement, operate,<br />
and maintain an enterprise. The CIM-OSA method deploys many well-accepted ideas and<br />
principles to <strong>design</strong> a CIM enterprise.<br />
Object-Oriented Approach: This methodology proposed by Kim, et al. (1993) consists<br />
of an analysis phase and a <strong>design</strong> phase. In the analysis phase, functional decomposition is<br />
34
employed to define the information flow among the manufacturing functions and their<br />
infrastructures. Decomposed functions are represented by functional diagrams, which in turn<br />
are transformed into an object-oriented information model (an object oriented database<br />
management system).<br />
2.1.4 Enterprise Modeling and Enterprise Design<br />
The enterprise modeling schemes discussed so far are integrated modeling schemes, i.e., they<br />
are intended to be applied to model an enterprise at nearly any level of abstraction. Enterprise<br />
integration in CIM is thus pursued by constructing and applying an integrated modeling scheme<br />
or by integrating existing models. The argument in favor of the former <strong>approach</strong> (a single<br />
model) is that local models built <strong>with</strong> the big picture in mind do not search for local optima,<br />
rather they contribute to the system effectiveness. On the other hand, the argument in favor of<br />
the second <strong>approach</strong> (of many small linked models) is that a unified model that captures all<br />
aspects and variables of an enterprise is extremely cumbersome for practical use.<br />
In order to determine how to model an enterprise it is necessary to consider what is the<br />
model for and how is it going to be used. In a framework for decision-<strong>based</strong> enterprise <strong>design</strong>,<br />
an enterprise model is needed to support and improve <strong>design</strong> decisions. This need motivates the<br />
following research question:<br />
35
How can an enterprise <strong>design</strong> process be modeled mathematically in a decision-<strong>based</strong><br />
<strong>design</strong> context?<br />
The hypothesis posed in Section 1.5 to answer this question is that an enterprise <strong>design</strong><br />
process, in the context of decision-<strong>based</strong> <strong>design</strong>, can be mathematically modeled using <strong>game</strong>-<br />
<strong>theoretical</strong> constructs. Emphasis is given to the need of a mathematical model because<br />
mathematical models support rational choice among <strong>design</strong> alternatives. Without mathematical<br />
models, the comparison of alternatives and the tradeoff between competing objectives lacks the<br />
rigor that is demanded in decision-<strong>based</strong> <strong>design</strong>. This <strong>game</strong> <strong>theoretical</strong> representation of the<br />
enterprise is the first floor of the enterprise <strong>design</strong> framework that is being constructed in this<br />
thesis.<br />
Before proceeding <strong>with</strong> a discussion on how to model the enterprise activity using <strong>game</strong><br />
theory, it is worth examining why the CIM architectures previously described do not to fulfill the<br />
need that originated the preceding research question. For a decision-<strong>based</strong> enterprise <strong>design</strong>,<br />
we require a representation of the enterprise that provides insight in how the individuals or<br />
agents of an enterprise should coordinate their decisions in carrying out their respective tasks.<br />
Even if a single model of an enterprise were affordable, computationally and economically, to<br />
support any <strong>design</strong> decision <strong>with</strong>in an enterprise (which is very unlikely), there would still be the<br />
problem that such a model would not provide any insight in how to formulate those decisions,<br />
which in a decision-<strong>based</strong> <strong>design</strong> paradigm have to be formulated, at least the important ones.<br />
On the other hand, mathematically modeling a system’s behavior, where one decision-maker’s<br />
36
action depends on decisions by others, is well established in <strong>game</strong> theory. Examples of modeling<br />
a <strong>design</strong> process using <strong>game</strong> <strong>theoretical</strong> representations can be found in (Rao, et al., 1997),<br />
(Lewis and Mistree, 1997) and (Lewis, 1997).<br />
2.2 A GAME THEORETICAL REPRESENTATION OF THE<br />
ENTERPRISE DESIGN PROCESS<br />
2.2.1 Decisions in a Manufacturing Enterprise<br />
Manufacturing decisions differ greatly <strong>with</strong> regard to the length of time over which their<br />
consequences persist. This makes it essential to use different planning horizons in the<br />
decision-making process. For example, the decision to construct a new plant influences<br />
operations for years; thus, we must forecast these effects years into the future in order to make<br />
a reasonable decision. On the other hand, the effect of selecting a particular part to work on a<br />
particular workstation affects the system only <strong>with</strong>in hours or minutes. Within a given company,<br />
longer time horizons are generally used at the corporate office, which is responsible for long-<br />
range business planning, than at the plant where day-to-day execution decisions are made. In<br />
37
Table 2.1 various manufacturing decisions that are made over long, intermediate, and short<br />
planning horizons are listed according to Hopp and Spearman (1996).<br />
In general, long-range decisions address strategy, by considering such questions as<br />
what to make, how to make it, how to finance it, how to sell it, where to make it, where to get<br />
materials, and general principles for operating the system. Intermediate-range decisions address<br />
tactics for determining what to work on, who will work on it, what actions will be taken to<br />
maintain the equipment, and so on. These tactical decisions must be made <strong>with</strong>in the physical<br />
and logical constraints established by the strategic long-range decisions. Finally, short-range<br />
decisions address control by moving materials and workers, adjusting processes and<br />
equipment, and taking whatever actions are required to ensure that the system continues to<br />
function towards its goal. Therefore, when trying to integrate the various activities of the<br />
enterprise it is necessary to formulate decisions that involve quite different scope and time<br />
horizon, which complicates their integration into joint formulations.<br />
Table 2.1 Strategy, Tactics, and Control Decisions<br />
(Hopp and Spearman, 1996)<br />
Time Horizon Length Representative Decisions<br />
Long-term (strategy) Year-decades Financial decisions<br />
Marketing strategies<br />
Product <strong>design</strong>s<br />
Process technology decisions<br />
Capacity decisions<br />
Facility locations<br />
Supplier contracts<br />
38
Personnel development programs<br />
Plant control policies<br />
Quality assurance policies<br />
Intermediate-term (tactics) Week-year Work scheduling<br />
Staffing assignments<br />
Preventive maintenance<br />
Sales promotions<br />
Purchasing decisions<br />
Short-term (control) Hour-week Material flow control<br />
Worker assignments<br />
Machine setup decisions<br />
Process control<br />
Quality compliance decisions<br />
Emergency equipment repairs<br />
Due to the different horizon and scope of decisions, manufacturing enterprises are<br />
organized in a hierarchical structure, <strong>with</strong> the lowest level dealing <strong>with</strong> issues that are either<br />
limited in scope or where, because of their frequency, a programmed response is appropriate.<br />
At the middle level of the hierarchy, there is greater freedom for making decisions but still <strong>with</strong>in<br />
constraints of authority posed by the higher levels. At high levels in the hierarchy, the decision is<br />
typically one that affects the whole system. Clearly, this hierarchical organization of the<br />
enterprise requires defining what is the range of authority of its agents, and mechanisms to<br />
coordinate and resolve their conflicts at any level. Game theory can provide a mathematical<br />
representation of an organization that facilitates the coordination of individuals. This is discussed<br />
in the next section.<br />
39
2.2.2 Decisions and Game Theory<br />
Rousseau, in his Discourse on the Origin and Basis of Equality Among Men, comments<br />
(quoted from Fudenburg and Tirole, 1993, p. 3):<br />
If a group of hunters set out to take a stag, they are fully aware that they would all have<br />
to remain faithfully at their posts in order to succeed; but if a hare happens to pass near<br />
one of them, there can be no doubt that he pursued it <strong>with</strong>out qualm, and that once he<br />
had caught his pray, he cared very little whether or not he had make his companions<br />
miss theirs.<br />
To make the situation described by Rousseau into a <strong>game</strong>, suppose that there are only two<br />
hunters, and that they must decide simultaneously whether to hunt for stag or for hare. If both<br />
hunt for stag, they will catch one stag and share it equally. If both hunt for hare, they each will<br />
catch one hare. If one hunts for hare while the other tries to take a stag, the former will catch a<br />
hare and the latter will catch nothing. Each hunter prefers half a stag to one hare. This is a<br />
simple example of a <strong>game</strong>. The hunters are the players. Each player has the choice between<br />
two strategies: hunt stag and hunt hare. The payoff to their choice is the prey. How should the<br />
hunters act? What is the expected outcome of the <strong>game</strong>? These are the issues of interest in<br />
<strong>game</strong> theory.<br />
In a general sense, a “<strong>game</strong>” is a set of rules completely specifying a competition, including<br />
the permissible actions of, and information available to each participant, the criteria for<br />
termination of the competition, and the distribution of payoffs (Websters, 1984). From a<br />
systems perspective, a “<strong>game</strong>” is defined as follows (Myerson, 1991):<br />
40
A <strong>game</strong> consists of multiple decision-makers who each control a specified subset of<br />
system variables and who each seek to minimize their own cost functions subject to their<br />
individual constraints.<br />
Game theory is the study of the strategic interactions of such <strong>game</strong>s (Fudenburg and Tirole,<br />
1995).<br />
Now consider a more practical case than the aforementioned hunting <strong>game</strong>. Consider the<br />
problem of a manufacturer -let’s call it P- that must choose between a high and a low price for<br />
the seasonal batch of products. In making this choice P wants to think about what the prices of<br />
other similar and substitute products are likely to be. P could simply define its pricing policy for<br />
some given exogenous beliefs about the competitor’s prices, but it seems more satisfactory to<br />
try to predict these prices from some knowledge of the industry. In particular, P knows that the<br />
other firms choose their own prices on the basis of their own predictions of the market<br />
environment, including P’s prices. The <strong>game</strong>-theoretic way for P to use this knowledge is to<br />
build a model of the behavior of each individual competitor, and to look for a solution that forms<br />
an equilibrium of the model. This relatively simple example of pricing is typical in <strong>game</strong> theory,<br />
which has been extensively used in business administration, economics and disciplines alike, but<br />
<strong>game</strong> theory has recently been applied to other fields such as engineering <strong>design</strong>.<br />
2.2.3 Fundamentals of Game Theory in Design<br />
41
Traditionally, “optimal” <strong>design</strong> of parametric systems is generally referred to the concept of a<br />
minimum or maximum. However, there are many other optimization concepts defined for<br />
parametric systems if we allow optimal <strong>design</strong> to include multiple criteria optimization <strong>with</strong><br />
multiple optimizers. Such an extended view, more in line <strong>with</strong> a systems perspective, brings<br />
optimality in parametric systems into the realm of <strong>game</strong> theory along <strong>with</strong> its numerous<br />
optimization concepts.<br />
Generally during a <strong>design</strong> process, the variables in a system are constrained by a system<br />
of equality and inequality equations of the form (Vincent, 1983):<br />
42<br />
g(X) = 0 (2.1)<br />
h(X) ? 0 (2.2)<br />
where X is a vector variable and g and h are vector functions, that is ? x ,..., x ?<br />
? ? g ( X ),..., g ( X ) ? , and h( X ) ? h1(<br />
X ),..., hq(<br />
X ) ?<br />
g( X ) 1<br />
n<br />
X 1<br />
? ,<br />
? . The equations given by (2.1)<br />
and (2.2) may be thought of as defining the system under study. In mathematical terms the<br />
“system” is a constraint set defined by<br />
? X g(<br />
X ) ? 0 and h(<br />
X ? 0?<br />
m X ? S ?<br />
)<br />
(2.3)<br />
where S m <strong>design</strong>ates an m-dimensional space of real numbers. Thus, X is the set of all points in<br />
the m-dimensional variable space that satisfy all the constraints. For each choice of X, one or<br />
more costs (or objectives) may be defined by means of a cost function (or objective function)<br />
f(X):<br />
? f ( X ),..., f ( X ) ?<br />
f ( X ) ? 1<br />
r<br />
(2.4)<br />
m
When r = 1 we have a scalar-valued cost function. For r >1 we have either a vector-valued<br />
cost criterion or a <strong>game</strong>, depending whether one or more <strong>design</strong>ers are involved in selecting<br />
the point X and the relation of the various costs to the various players. This classification of<br />
optimality problems is illustrated in Figure 2.1.<br />
N = 1<br />
N 1<br />
?<br />
Number of Objectives<br />
r = 1 r 1<br />
?<br />
Scalar<br />
Optimization<br />
Problem<br />
Scalar<br />
Games<br />
Vector<br />
Optimization<br />
Problem<br />
Vector<br />
Games<br />
Figure 2.1 Various Formulations in Optimization Theory<br />
(Mesterton-Gibbons, 1992)<br />
Typical optimization processes focus on the upper-left quadrant, namely scalar<br />
optimization problems <strong>with</strong> one objective and one decision-maker. With only one <strong>design</strong>er<br />
selecting <strong>design</strong> parameters we have either a problem in scalar optimization or vector<br />
optimization. In the former case, a point X ? is a minimum point if and only if there does not<br />
exist a point such that f(X)
y replacing X? S m by X? B m ? S m , where B m is a region centered at X ? and B m ? S m is the set<br />
of points common to B m and S m .<br />
For vector costs a point Xp? S m is said to be a Pareto-minimum if and only if there does<br />
not exist a point X? S m such that fi(X) ? f(Xp) for all i ? ? 1,...,r?<br />
and fj(X) ? f(Xp) for at least<br />
one j ? 1,...,<br />
r?<br />
? (Vincent, 1983). If Xp is Pareto-minimal then for any other X at least one of<br />
the costs that compose the vector cost must increase or all costs remain the same. Under this<br />
definition it is usually possible for many points to have the Pareto-minimum property. Indeed for<br />
most problems in engineering <strong>design</strong>, the Pareto-minimal points are not isolated. Vincent<br />
(1983) provides a thorough discussion of the mathematical conditions for achieving Pareto-<br />
minimum points. Here, it is enough to note that one <strong>design</strong>er choosing parameters for a vector-<br />
valued optimization problem should try to restrict choices to the Pareto-minimal set. This set<br />
may be obtained by a number of methods such as the mini-max method (Osyczka, 1984)<br />
multicriteria dynammic programming (Rosenman and Gero, 1983), and scalarization methods<br />
such as compromise programming (Goicoechea, et al., 1982), goal programming (Ignizio,<br />
1985), the global criterion method (Rao, 1984), the ratios method (Metwalli, et al., 1984), to<br />
name some. Vanderpooten (1989) and Hwang (1980) provide a survey of various of these<br />
methods.<br />
Now suppose that a <strong>design</strong> project has been broken down into a number of subprojects<br />
each <strong>with</strong> a set of <strong>design</strong> objectives and each <strong>with</strong> a <strong>design</strong> team in charge of optimizing those<br />
objectives. Since communication is generally possible between the subgroups and since each<br />
44
<strong>design</strong>er must consider the possible choices made by other <strong>design</strong> teams, it is envisioned that<br />
there will be many tentative <strong>design</strong>s which will be passed back and forth between the various<br />
<strong>design</strong> teams before some ultimate <strong>design</strong> is finally obtained. Operations researchers and<br />
economists, in studying competitive systems, have developed theories for <strong>game</strong>s that are<br />
applicable to this kind of situation. This is the most general case of the various optimization<br />
formulations that are shown in Figure 2.1.<br />
Two <strong>game</strong>-<strong>theoretical</strong> models are of relevance for this thesis to describe the interaction of<br />
players: the cooperative <strong>game</strong> model <strong>based</strong> on the concept of Pareto equilibrium, and the<br />
non-cooperative model <strong>based</strong> on the concept of Nash equilibrium. In cooperative <strong>game</strong>s,<br />
each player compromises his/her objective in order to allocate the resources in such a way that<br />
the performance of the system of players be as optimal as possible. On the other hand, in non-<br />
cooperative <strong>game</strong>s each player selects a share of resources <strong>with</strong> the view of optimizing his/her<br />
performance. These models are described next using a simple 2-player terminology:<br />
?? Pareto or cooperative model:<br />
In this model both players have knowledge of the other players’ information and they work<br />
together to find a Pareto solution. If the players cooperate, they can be expected to obtain<br />
better solutions than when they do not. A pair (x1p, x2p) is Pareto optimal if no other pair (x1,<br />
x2) exists such that:<br />
45
f 1 (x 1, x 2 ) ? f 1 (x 1 p , x 2p ) & f 2 (x 1 , x 2 ) ? f 2 (x 1p , x 2p ) (2.5)<br />
In the cooperative solution the group of players allocate resources <strong>with</strong> the intent that all<br />
players be at an optimal point as far as possible. The group must decide how to distribute<br />
resources so that a gain for one player does not result in an unacceptable loss for other players.<br />
One method is to distribute the resources such that all players are as far from their worst cases<br />
as possible. To find such a solution it is necessary to formulate a supercriterion as a measure<br />
of compromise. This supercriterion, a function of the objectives, is essentially a penalty function<br />
that penalizes solutions <strong>with</strong> objectives that are too close to predefined worst cases (Rao,<br />
1991). Implementing cooperative <strong>game</strong> theory requires a two-step algorithm. First, the set of<br />
Pareto solutions is generated. Then the Pareto solution that has the most optimal supercriterion<br />
is selected as the best <strong>design</strong>. This method is difficult to implement in automated routine. Rao<br />
and Freiheit (1991) suggest a method to perform both tasks simultaneously by combining the<br />
Pareto optimal solution generation and the supercriterion into one objective by subtracting the<br />
supercriterion from the Pareto optimal objective. A single-objective optimization algorithm is<br />
used to minimize this new objective. This formulation change the characteristic of the<br />
optimization problem such that it is neither truly minimizing the Pareto objective nor is it truly<br />
maximizing the supercriterion, but according to Rao and Freihet (1991), “from an engineering<br />
point of view, the computational savings warrant the acceptance of a solution which deviates<br />
slightly from a true Pareto optimum.” They support their assert <strong>with</strong> numerical examples. The<br />
algorithm to implement this <strong>approach</strong> is presented in Chapter 3.<br />
46
?? Nash or non-cooperative model:<br />
In this model players make decisions <strong>based</strong> on prediction of others players’ strategy. A<br />
strategy pair (x1n, x2n) is a Nash solution if<br />
where<br />
f1( x1<br />
n , x2<br />
n ) min f1(<br />
x1<br />
, x2<br />
n ), x1<br />
? X1<br />
47<br />
? (2.6)<br />
f 2(<br />
x1<br />
n , x 2n<br />
) ? min f 2(<br />
x1n<br />
, x 2 ), x 2 ? X 2<br />
(2.7)<br />
? x1n<br />
? X 1 : f1(<br />
x1<br />
, x2<br />
) ? min f1(<br />
x1<br />
, x 2 ) ?<br />
x1 n ( x2<br />
) RRS 1(<br />
x1<br />
) ?<br />
n<br />
x1<br />
? X 1<br />
? (2.8)<br />
? x 2n<br />
? X 2 : f 2(<br />
x1<br />
, x 2 ) ? min f 2(<br />
x1<br />
, x2<br />
) ? x2<br />
X 2<br />
x 2 n ( x1<br />
) RRS 2(<br />
x1<br />
) ?<br />
n<br />
?<br />
? (2.9)<br />
are called the rational reaction sets (RRS) of the two players. The RRS is a set of solutions<br />
that a player constructs <strong>based</strong> on public information from other players.<br />
Non-cooperative <strong>game</strong> theory is <strong>based</strong> on the assumption that each player is looking<br />
out for his/her own interests. Each players selects a share of the resources only <strong>with</strong> the view of<br />
optimizing his/her own objective. The players then bargain <strong>with</strong> each other until equilibrium is<br />
reached. The resultant solution is known as Nash equilibrium, and no player may improve his<br />
objective by deviating from it as long as the other players maintain their choices. The solution<br />
can be dependent on the order in which the resources are allocated and the order in which the<br />
players choose the resources when more than one equilibrium point exists.<br />
In complex problems, finding exact expressions for the RRS is difficult since it involves<br />
finding functions relating the independent variables of one player as a function of the independent<br />
variables of the other players. In this case, the RRS have to be approximated.
So far, the fundamental concepts on <strong>game</strong> theory required to build a <strong>game</strong>-<strong>theoretical</strong><br />
representation of the enterprise <strong>design</strong> process have been introduced. In the next section these<br />
fundamentals are applied in building such a model.<br />
2.3 ENTERPRISE DESIGN AS AN EXTENSIVE GAME<br />
2.3.1 Extensive-Form Games<br />
In the examples presented in Section 2.2, namely, the hunters’ <strong>game</strong> and the oligopoly-pricing<br />
problem, the players choose actions simultaneously. However, in the <strong>design</strong> of an enterprise,<br />
i.e., its product, its manufacturing process and its organization, all decisions cannot be made<br />
simultaneously. A more general form of non-cooperative <strong>game</strong>s, the so-called <strong>game</strong>s in<br />
extensive form, can be used to model such a dynamic situation. Attention is given in this work<br />
to a particular kind of extensive-form <strong>game</strong>s known as multi-stages <strong>game</strong>s <strong>with</strong> observed<br />
actions.<br />
A multi-stage <strong>game</strong> <strong>with</strong> observed actions are <strong>game</strong>s in which (1) all players (including<br />
“Nature”) knew the actions chosen at all previous stages 0, 1, 2, ..., k-1 when choosing their<br />
actions at stage k, and (2) all players move “simultaneously” in each stage (Fudenburg and<br />
Tirole, 1995). Players move simultaneously in stage k if each player chooses an action at stage<br />
k <strong>with</strong>out knowing the stage-k action of any other player. However, “simultaneous” moves do<br />
not exclude <strong>game</strong>s where players move in alternation, as there is the possibility that some of the<br />
players have the one-element choice “do nothing.”<br />
An extensive form of a <strong>game</strong> contains the following information:<br />
48
1. The set of players.<br />
2. Number of stages and order of moves, i.e., who moves when.<br />
3. The player’s payoffs as a function of the moves that were made.<br />
4. What the players’ choices are when they move.<br />
5. What each player knows when making choices.<br />
6. Probability distributions over any exogenous events.<br />
The difference in the content of information of extensive <strong>game</strong>s and static <strong>game</strong>s is shown in<br />
Table 2.2. The fundamental difference is that in extensive <strong>game</strong>s strategies correspond to<br />
contingent plans, whereas in static <strong>game</strong>s strategies correspond to uncontingent actions. That<br />
is, in a <strong>game</strong> where sequential moves are to be performed (as in chess), one action is taken in<br />
contingency of possible future actions, whereas in a static or one-stage <strong>game</strong> (like flipping once<br />
a coin) there is no need to develop a strategy that considers future actions.<br />
Table 2.2 Contents of Static- and Extensive-Form Games<br />
STATIC GAME EXTENSIVE GAME<br />
Set of players<br />
? i?<br />
i ? 1,<br />
2 N<br />
I ? ,...,<br />
One stage <strong>game</strong><br />
K=1<br />
Set of players<br />
? i?<br />
i ? 1,<br />
2 N<br />
I ? ,...,<br />
Number of stages K ? 1<br />
and order of Moves<br />
Players’ payoff as function of <strong>design</strong> variables Players’ payoff as function of <strong>design</strong> variables and<br />
history<br />
49
k<br />
f i (X )<br />
( X , H )<br />
Players’ choices<br />
A ?<br />
i<br />
S<br />
i<br />
f k ? ? 1,<br />
2,...,<br />
K?<br />
i<br />
Players’ choices<br />
k<br />
A ? A ( H ) ? S k ? ? 1,<br />
2,...,<br />
K?<br />
Players’ information when making a choice Players’ information when making a choice<br />
Probability distributions over exogenous events Probability distributions over exogenous events<br />
In a stage k of a multi-stage <strong>game</strong>, all players i ? ? choose simultaneously actions from<br />
choice sets Ai(h k ) (some of the choice sets may be “do nothing”) where h k denotes the<br />
information sets or “history” of the <strong>game</strong> at the stage k of the <strong>game</strong>. Note that Ai is a subset of<br />
the <strong>design</strong> space of the player, Si. At the end of each stage, all players observe the action<br />
profile of the stage.<br />
Let a k ? a1 k , a2 k ,..., an k ? ? be the stage-k action profile (the actions performed by the n<br />
players at stage k), where ai k is the value of the <strong>design</strong> vector Xi that is controlled by player i at<br />
stage k. Let also h 0 ? ? be the history at the start of the play. At the beginning of stage 1,<br />
players know history h 1 , which can be identified <strong>with</strong> a 0 given that h 0 is trivial. In general, the<br />
available actions of the ith player in stage k depend on what has happened previously,<br />
1 k?1<br />
k<br />
? a , a ,..., a ? f ? ?<br />
a ?<br />
i<br />
i<br />
50<br />
k 0 ? f<br />
h<br />
(2.10)<br />
In this setting, a pure strategy for player i is simply a plan of how to play in each stage k for<br />
possible history h k .
Let K be the total number of stages in the <strong>game</strong>, and let H k denote the set of all stage-k<br />
histories as follows:<br />
A ( H<br />
i<br />
k<br />
51<br />
k<br />
) ? ? A ( h )<br />
(2.11)<br />
hk<br />
? H k<br />
A pure strategy si for player i is a sequence of maps ? ? K k<br />
of player i’s feasible actions Ai(H k ):<br />
i<br />
si k 0<br />
? , where each s k k<br />
i maps H to the set<br />
si k : H k ? Ai(H k ) (2.12)<br />
Hence, the stage-0 actions are a 0 =s 0 (h 0 ), the stage-1 actions are a 1 =s 1 (a 0 ), the stage-2 actions<br />
are a 2 =s 2 (a 0 ,a 1 ), and so on. This is called the path of the strategy profile. Since the terminal<br />
histories represent an entire sequence of play, we can represent each player i’s payoff as a<br />
K ?1 r<br />
function f : H ? R .<br />
i<br />
To illustrate this, consider a <strong>game</strong> <strong>with</strong> two stages and two players, depicted as a <strong>game</strong><br />
tree in Figure 2.2 where qi represents the selection of player i. Both Player 1 and Player 2 have<br />
as choice options the set of actions {Ø,3,4,6} (Ø is the option “do nothing”) and Player 2 acts<br />
only after Player 1 has acted. Thus,<br />
A1(<br />
H<br />
A ( H<br />
1<br />
1<br />
2<br />
) ?<br />
) ?<br />
1<br />
? 3,<br />
4,<br />
6?<br />
A2(<br />
H ) ? ? ? ?<br />
2<br />
? ? A2<br />
( H ) ? ? 3,<br />
4,<br />
6?<br />
The payoff vectors are displayed next to the corresponding terminal nodes. The information<br />
that players have when choosing their actions is usually represented using the information set
h ? H that partitions the nodes of the tree; that is, every node is in exactly one information set h<br />
and the action set at h is denoted as A(h). Finally, the probability distributions over exogenous<br />
events are represented as moves by “Nature” that have some probability distribution. For<br />
example, in the previous <strong>game</strong>, if Player 1 were “Nature”, a probability distribution on its<br />
actions could be a vector [p(3)=0.3, p(4)=0.2, p(6)=0.5].<br />
q2=3<br />
q2=4<br />
q1=3<br />
q 2=6<br />
q2=3<br />
q 1=4<br />
Player1<br />
q2=4<br />
q 2=6<br />
q1=6<br />
q 2=3<br />
q2=4<br />
q2=6<br />
(18,18) (15,20) (9,18) (20,15) (16,16) (8,12) (18,9) (12,8) (0,0)<br />
Payoff of player<br />
1<br />
Player 2<br />
Figure 2.2 Example of an Extensive Game<br />
2.3.2 Enterprise Design as an Extensive Game<br />
Payoff of<br />
player 2<br />
The mathematics of extensive-form <strong>game</strong>s, described in the previous section, can be readily<br />
applied to model an enterprise <strong>design</strong> process. As a simple example, consider a<br />
Leader/Follower relationship between two <strong>design</strong> teams, where the <strong>design</strong> teams are abstracted<br />
52
as players in a <strong>game</strong>. The actions of the two players are to set the value of the <strong>design</strong> variables<br />
X1 for player 1 and X2 for player 2.<br />
Player 1, the leader, chooses the value of X1 first, and player 2 observes X1 before<br />
choosing X2. Since player 2 observes player 1’s choice before choosing X2, player 2 can<br />
condition his choice of X2 on the observed level of X1. Since player 1 moves first, she cannot<br />
condition her output on player 2’s. Thus, it is natural to expect that player 2’s strategy in this<br />
<strong>game</strong> be a map of the form:<br />
s2: X1? X2<br />
Given this strategy, the outcome is a vector X=[X1, s2(X1)] <strong>with</strong> payoffs f1(X) and f2(X).<br />
In the Stackelberg equilibrium, which is the expected equilibrium in this <strong>game</strong>, player 2’s<br />
strategy is to choose, for each X1 the level of X2 that solves min f2(X1, X2), so that X2 is equal to<br />
the reaction function or rational reaction set RRS2(X1). Given that player 1 expects the<br />
strategy of player 2 to be RRS2(X1) , her choice of X1 should be the solution to min f1(X1,<br />
RRS2(X1)). Thus, a formulation of a strategy for player 1, the leader, would be as follows:<br />
minimize f 1 ( X1,<br />
X2<br />
), ( X1,<br />
X2<br />
) S1<br />
? S2<br />
? (2.13)<br />
satisfying ? RRS ( X ) ? X ( X )<br />
(2.14)<br />
X2 2 1 2n<br />
1<br />
Why is this Stackelberg equilibrium the expected one? An informal answer is that, as<br />
observed by Fudenburg and Tirole (1995), this is the only “credible” equilibrium if both players<br />
act “rationally.” A formal answer is that the Stackelberg equilibrium is consistent <strong>with</strong> the<br />
notion of backward induction, so called because the idea is to start by solving the optimal<br />
53
choice of the last mover for each possible situation he might face, and then work backward to<br />
compute the optimal choice for the player before.<br />
The ideas of credibility and backward induction have been generalized <strong>with</strong> the concept of<br />
sub<strong>game</strong>-perfect equilibrium (see Fudenburg, and Tirole, 1995), which extends the idea of<br />
backward induction to extensive <strong>game</strong>s where players move simultaneously in several periods.<br />
In this case, the backward-induction algorithm is not applicable because there are several “last<br />
movers” and each of them must know the moves of the others to compute the optimal choice.<br />
As an illustrative example of how a two-players <strong>game</strong> can be used in an actual <strong>design</strong><br />
scenario consider the collaboration between a stress engineer and a manufacturing engineer in<br />
the <strong>design</strong> and production of a wing section of an aircraft who have to negotiate a solution. The<br />
stress engineer and the manufacturing engineer interact using a CAD drawing (the public<br />
information). The stress engineer needs a solution that transmits loads well through the<br />
structure, and the manufacturing engineer needs a structure that is easy to fabricate, using, for<br />
example, an automatic riveting machine. The criteria used by each specialist are private in that<br />
they are complex and concerned <strong>with</strong> their particular technologies. In this example, the stress<br />
engineer is concerned <strong>with</strong> arrangements such that the loads carried in the elements are well<br />
formed, in that internal load is transmitted throughout the structure under a given external load<br />
specification. The description of the stress engineer concerns loads, stresses, transmission, and<br />
techniques for finding them. He/she works privately <strong>with</strong> a model to calculate load patterns that<br />
must match the load transmission properties of the sized geometry. On the other hand, the<br />
54
manufacturing engineer is concerned <strong>with</strong> a <strong>design</strong> that can be produced <strong>with</strong> the machines<br />
currently available using techniques and tooling currently in use. For example, he/she would like<br />
to keep rivet spacing constant, or at least to a small number of different rivet spacings, in order<br />
to limit tooling set-up cost.<br />
Aircraft <strong>design</strong> is one example of a large-scale <strong>design</strong> of a system, where there are<br />
many specialists interacting, each <strong>with</strong> their own technology and language. For example, there<br />
are aerodynamicists who use surface models and flow-field equations; there are maintainability<br />
engineers concerned <strong>with</strong> access, disassembly, and replacement; there are hydraulic engineers,<br />
and thermodynamic experts, to name just a few. They have to collaborate and negotiate to<br />
produce a <strong>design</strong> that satisfies all of them. This means that each agent has a justification of the<br />
<strong>design</strong> that he or she is satisfied <strong>with</strong>.<br />
Clearly, a common and single model would not be practical to support each <strong>design</strong><br />
decision of the aircraft development. Thus, its <strong>design</strong> has to be partitioned into a series of<br />
problems handled by disciplinary teams, and the overall <strong>design</strong> proceeds by the cooperation<br />
and negotiation of individuals, teams and/or departments, each one <strong>with</strong> private information.<br />
Moreover, each agent may use several different models or partial models for different purposes.<br />
Because the criteria used by each agent is private to them in that they are complex and<br />
concerned <strong>with</strong> their particular goals and technologies, each agent has limited ability to<br />
understand other agents’ motivations and decisions. Therefore, the <strong>design</strong> proceeds by<br />
successive coordination of decisions.<br />
55
Aircraft <strong>design</strong> is also an example of a decision making process in which each agent has<br />
public and private information during the <strong>design</strong> process. In order to take the aircraft <strong>design</strong><br />
process into an extensive-form <strong>game</strong> consider how the six points that define an extensive <strong>game</strong><br />
can be readily applied:<br />
1. The set of players. They would vary depending on the level of hierarchy of the organization<br />
and the extent of integration that is considered. At a high level of hierarchy the players<br />
could be functional departments such as product engineering, manufacturing, marketing,<br />
purchasing, etc. At a lower level they could be disciplinary teams <strong>with</strong>in a department and<br />
some “external” players to include relevant external factors over which the team does not<br />
have control.<br />
2. The order of moves. Who moves when depends on the policies of the organization.<br />
Usually there is always an established plan for product development that is used to specify<br />
when and how do different agents participate in the <strong>design</strong> process.<br />
3. The players’ payoffs as a function of the moves (or history of the <strong>design</strong> process). This is<br />
straightforward if there are clear goals defined in the beginning of each stage of the <strong>design</strong><br />
process (the payoff could be the deviation from a previously specified target).<br />
4. The players’ choices when they move. This is a subtle point when considering <strong>design</strong><br />
processes, specially at early stages, because the alternatives may not be known in advance<br />
(for example, new technologies might be developed during the process).<br />
56
5. Player knowledge when making choices. This includes private and public information.<br />
Public information is any information that is common to all the <strong>design</strong> teams, such as a CAD<br />
model which serves as baseline model for all the players involved in the process between<br />
updating meetings.<br />
6. Probability distributions over exogenous events. Here, exogenous events may refer to<br />
movements by “Nature” such as market demand, interest rates, suppliers’ deliveries, etc.<br />
It should be apparent that an enterprise <strong>design</strong> process can be abstracted as a multi-stage<br />
<strong>game</strong>, in which each of the <strong>design</strong> groups is a player who strives to act in the interest of the<br />
organization, while at the same time tries to optimize his/her standing (e.g., minimize production<br />
cost or maximize reliability).<br />
The relevant assumptions and definitions that are used in this thesis to model an enterprise<br />
<strong>design</strong> process as a <strong>game</strong> are:<br />
Assumption 2.1: Decisions are supported <strong>with</strong> mathematical models.<br />
Assumption 2.2: No agent of an enterprise has sufficient competence and resources to solve an<br />
entire enterprise <strong>design</strong> problem. In other words, even if an individual involved in a <strong>design</strong><br />
problem strives consciously to solve it considering all the aspects and variables of the system,<br />
he/she will do so in a way that reflects cognitive and computational limitations.<br />
Assumption 2.3: The common link among each agent is the company for which they all work.<br />
They all act respecting the social rules established by the organization.<br />
57
Assumption 2.4: The <strong>design</strong> variables of different agents do not overlap. The local control of<br />
each agent is exclusive. That is, if one agent controls X1? S1 and another agent controls X2? S2<br />
then S S ? ? .<br />
1 ? 2<br />
Given these assumptions, the following definitions are used to formalize an enterprise<br />
<strong>design</strong> process as a <strong>game</strong>:<br />
Definition 2.1: In enterprise <strong>design</strong>, a player in a <strong>game</strong> is the decision maker or agent, which is<br />
embodied by a <strong>design</strong>er, <strong>design</strong> team or department, and their associated analysis and synthesis<br />
<strong>design</strong> tools.<br />
Definition 2.2: A strategy of an agent is embodied by a <strong>design</strong> formulation.<br />
Definition 2.3: A payoff of an agent is the value of the motivating function at a given move.<br />
Definition 2.4: Public information is knowledge that is accessible to any agent of an<br />
enterprise.<br />
Definition 2.6: A stage of the <strong>game</strong> is demarcated by public events when one or more agents<br />
of an enterprise make <strong>design</strong> decisions public.<br />
The equivalency between <strong>design</strong> and <strong>game</strong>s <strong>based</strong> on the discussion carried out in this<br />
chapter is shown in Figure 2.3.<br />
58
Figure 2.3 Enterprise Design as a Game<br />
In the following section, the relevance of a <strong>game</strong> <strong>theoretical</strong> representation of the<br />
enterprise for enterprise <strong>design</strong> is discussed.<br />
2.4 RELEVANCE OF MODELING AN ENTERPRISE DESIGN PROCESS AS<br />
A GAME<br />
Enterprise Design<br />
Enterprise Design Process<br />
Agents of the Enterprise<br />
Design Formulations<br />
Value of Cost Function<br />
Game<br />
Extensive Game<br />
Players<br />
Players’ Strategy<br />
Players’ Payoff<br />
The first and perhaps most important contribution of a <strong>game</strong> <strong>theoretical</strong> representation of the<br />
enterprise lies in the framing of issues. The language of extensive form <strong>game</strong>s permits us to ask<br />
questions about the dynamics of interaction between decision makers; and the language of other<br />
parts of non-cooperative <strong>game</strong> theory are well suited to pose questions about cooperative<br />
interactions in which parties have proprietary information. Indeed, a <strong>game</strong> <strong>theoretical</strong> model<br />
focuses attention on issues of dynamics and proprietary information, framing the issues so that<br />
the “physics” of interaction -who does what, when, <strong>with</strong> what information- come to the fore. If<br />
59
the mechanics of interaction between decision-makers matter, as they do when trying to<br />
integrate enterprise activities, then simply focusing attention and framing cooperation in this way<br />
is a contribution. It has to be noted that a <strong>game</strong> <strong>theoretical</strong> representation does not exclude the<br />
use of other enterprise models, such as CIM models. Moreover, CIM models could provide<br />
knowledge on dependencies between the agents of an enterprise, thus facilitating the<br />
construction of a <strong>game</strong> <strong>theoretical</strong> model useful for individual decision making. In Chapter 3 is<br />
examined how agents can formulate and solve problems using a <strong>game</strong> <strong>theoretical</strong> representation<br />
of an enterprise <strong>design</strong> process.<br />
60
CHAPTER 3<br />
FORMULATION AND SOLUTION OF ENTERPRISE DESIGN PROBLEMS<br />
In Chapter 2 it is presented a <strong>game</strong> <strong>theoretical</strong> representation of the enterprise <strong>design</strong><br />
process. This representation provides the underpinning of a framework for enterprise <strong>design</strong><br />
that is <strong>based</strong> on three key elements:<br />
?? An enterprise <strong>design</strong> method <strong>based</strong> on a common paradigm for decision support.<br />
?? A multi-objective decision model, the Compromise DSP.<br />
?? A <strong>probabilistic</strong> <strong>design</strong> model to attain flexible and robust decisions and handle uncertainty<br />
and imprecision<br />
In this chapter, these elements are described and integrated into a consistent construction for<br />
supporting coordination of decisions through <strong>design</strong> formulations, which is the essential feature<br />
of a decision-<strong>based</strong> perspective for enterprise integration.<br />
60
3.1 ENTERPRISE INTEGRATION AS COORDINATION OF<br />
DECISIONS: IMPLICIT AND EXPLICIT COORDINATION<br />
Peplinski (1997) defines organization <strong>design</strong> as “the <strong>design</strong> of the organization itself through the<br />
setting of variables under management control.” These variables include the organization<br />
structure, technologies, people, tasks, reward systems, information systems, etc. These issues<br />
are largely influenced by human behavior, which makes organization behavior only partially<br />
amenable to mathematical modeling and analysis. However, there are basic patterns of human<br />
behavior that can be recognized and used for improving the <strong>design</strong> of an organization. For<br />
example, it is clear that individuals act according to their motivations and rewards. Therefore,<br />
improving the quality of individual decisions does not necessarily improve the performance of<br />
the system unless the coordination of such decisions is improved as well. Such coordination is<br />
typically pursued through team meetings, notices, centralized decision making or information<br />
sharing (Lizotte and Chaib-draa, 1997; Browning, 1997).<br />
In the team-meetings <strong>approach</strong>, individuals from different disciplines or groups reason on<br />
each other’s decisions as a team. Notices are recipes describing a decision sent by a<br />
responsible party to all the groups of the enterprise that might be affected. In the centralized<br />
<strong>approach</strong>, all the <strong>design</strong> decisions are gathered in one place and one decision- maker reviews<br />
them to coordinate all tasks and resources and solve conflicts. These three <strong>approach</strong>es are<br />
informal ways to pursue enterprise integration, and do not enhance the quality of individual<br />
decision making.<br />
61
In the information-sharing <strong>approach</strong> enterprise integration is pursued by constructing and<br />
applying integrated architectures that encompass all the aspects of an enterprise such that the<br />
agents that comprise it have more complete information to make decisions. In Section 2.1 it is<br />
argued that the effectiveness of this <strong>approach</strong> is limited because no individual can handle such<br />
amount of information efficaciously. This <strong>approach</strong> has also the inconvenient that it does not<br />
provide any insight in how to formulate and make decisions consistently in an integrated<br />
enterprise environment.<br />
In the decision-<strong>based</strong> perspective embraced in this work, enterprise integration is pursued<br />
by coordinating decisions, i.e., managing dependencies between the various agents through the<br />
appropriate formulation and solution of their decisions. There are two kinds of coordination,<br />
namely, implicit and explicit coordination. In implicit coordination the agents of an enterprise<br />
establish social rules that they agree to respect in order to act in a coordinated manner. A<br />
simple example of an implicit coordination mechanism, in the context of the integration of<br />
product <strong>design</strong> and manufacturing, are DFM guidelines (e.g., Boothroyd, 1989), which<br />
represent a set of rules that a <strong>design</strong>er should follow to avoid potential manufacturing problems.<br />
Explicit coordination, on the other hand, asks agents to make explicit choices to ensure<br />
coordination. In order to make such explicit choices, they have to identify potential interactions<br />
between their activities and those of other agents. An example of this kind of coordination are<br />
the intelligent CAD systems (e.g., Lu and Subramanyam, 1988), which utilize the state-of-the<br />
art computer technology to explicitly anticipate manufacturing problems of a <strong>design</strong>.<br />
62
IMPLICIT COORDINATION<br />
Coordination achieved<br />
by establishing social<br />
rules that agents should respect<br />
Coordination<br />
achieved by<br />
identifying<br />
interactions<br />
<strong>with</strong> other<br />
agents<br />
EXPLICIT COORDINATION<br />
Figure 3.1 Implicit and Explicit Coordination of Decisions<br />
In Table 3.1 it is shown a comparison of implicit and explicit coordination according to<br />
Lizotte and Chaib-draa (1997). Implicit coordination is generally easier and less costly to<br />
<strong>design</strong> and implement than explicit coordination because two of the processes required for<br />
achieving coordination are reduced to a minimum. The group decision-making process is<br />
determined in advance using conventions called “social rules.” However, implicit coordination<br />
reduces the autonomy of agents because they should only consider alternatives that respect such<br />
social rules. Explicit coordination, on the other hand, increases an agent’s autonomy since the<br />
agent possesses a margin of authority that is not limited to alternatives respecting social rules.<br />
Nevertheless, increasing decision-makers’ autonomy forces them to deal <strong>with</strong> incomplete and<br />
uncertain information of other agents of the enterprise. An organization should be able to<br />
63
alance the advantages of both kinds of coordination to achieve an effective integration of its<br />
activities.<br />
Table 3.1 Explicit and Implicit Coordination Regarding Processes Underlying<br />
Coordination (Lizotte and Chaib-draa, 1997)<br />
Processes Implicit Coordination Explicit Coordination<br />
1. Coordination Less costly to <strong>design</strong> and more<br />
effective, but can reduce agent’s<br />
autonomy and it is domain<br />
dependent<br />
2. Group decision making Static, i.e., the decision is<br />
determined in advance using<br />
conventions<br />
3. Communication Generally done by exchanging signs<br />
and signals<br />
4. Perception of<br />
common<br />
objects<br />
Perceiving same agents and<br />
objects (sharing data files, etc.)<br />
More costly to <strong>design</strong> and less<br />
effective but can improve agent’s<br />
autonomy<br />
64<br />
Dynamic, i.e., implies evaluating<br />
alternatives and making a common<br />
decision between agents by<br />
authority (decision power)<br />
Generally done by exchanging<br />
alternatives, evaluations and<br />
choices<br />
Knowing part of agents’ intentions,<br />
goals and plans<br />
For example, <strong>with</strong>in explicit coordination, in order to determine the effect of setting the<br />
value of variable x as a or b, a <strong>design</strong>er would have to specify the states and actions of all other<br />
players as a function of x. An alternative <strong>approach</strong> would be to determine the possible<br />
strategies of each player independently, and to make the transition function from one stage to<br />
the next <strong>probabilistic</strong>. In this last <strong>approach</strong>, for example, if player 1 decides to set x = a, the<br />
result would be a <strong>design</strong> <strong>with</strong> an expected payoff that depends on other players’ behavior. In<br />
the first <strong>approach</strong>, complete information about other players must be supplied and the transition
function produces a specific prediction that supports a decision. In the second <strong>approach</strong>, the<br />
transition function produces a highly uncertain prediction.<br />
Instead of adopting either extreme view, a more practical <strong>approach</strong> is to use the social<br />
rules of an implicit-coordination mechanism that constrain actions of agents to “frame” explicit<br />
coordination. These rules should specify which of the actions that are in general available are in<br />
fact allowed in a given stage. They do so by a predicate over that stage. The transition function<br />
can now take as an argument not only the previous history of the process and the possible<br />
actions taken by the agents, but also the system constraints in force, such that it produces a<br />
better estimate of possible points of equilibrium between agents. Generally speaking, a<br />
prediction <strong>based</strong> on constraints will be more general than a prediction that is <strong>based</strong> on the<br />
precise states and actions of all other players, and more specific than a prediction <strong>based</strong> on pure<br />
<strong>probabilistic</strong> estimates. This idea is illustrated in Figure 3.2.<br />
Enterprise Design Space<br />
Constraints<br />
set by social<br />
rules<br />
Points of<br />
Equilibrium<br />
Figure 3.2 Integrating Implicit and Explicit Coordination<br />
65
The goal of a coordination-mechanism <strong>design</strong> is to strike a good balance between<br />
implicit and explicit coordination, allowing freedom to the individual decision-makers on the one<br />
hand and ensuring a harmonious cooperative behavior among them on the other. Furthermore,<br />
the integrative mechanism should be systematic to take advantage of a computer-supported<br />
environment. A systematic coordination mechanism, that capitalizes the advantages of both<br />
kinds of coordination, and is mathematically supported, can be established by promoting a<br />
framework for enterprise <strong>design</strong> <strong>based</strong> on the following key elements:<br />
?? A representation of the enterprise activity as an extensive <strong>game</strong>.<br />
?? An enterprise <strong>design</strong> method <strong>based</strong> on a common paradigm for decision support.<br />
?? A multi-objective decision model, the Compromise DSP.<br />
?? A <strong>probabilistic</strong> <strong>design</strong> model to attain flexible and robust decisions and handle uncertainty<br />
and imprecision.<br />
The merging of these elements provides a consistent mathematical framework for coordination<br />
of decisions <strong>based</strong> on <strong>design</strong> formulations, which is the objective of this work. This framework<br />
is illustrated in Figure 3.3, where it is shown how this framework can support the enterprise<br />
<strong>design</strong> method of Peplinski and Mistree (1997), described in Section 1.4.<br />
Is in Task 1, when a decision-maker expands the scope of a baseline decision to address<br />
needs, goals and constraints of other agents of the enterprise, that a representation of the <strong>design</strong><br />
process as a <strong>game</strong> is brought up. Depending on the number of objectives that are considered,<br />
the <strong>game</strong> can be either a scalar <strong>game</strong> or a vector <strong>game</strong>.<br />
66
Enterprise Design Method<br />
Baseline Decision<br />
Task 1<br />
Expand Scope<br />
of Decision<br />
Task 2<br />
Determine Input<br />
Needed for<br />
Analysis<br />
Identify Modeling<br />
Techniques<br />
Task 3<br />
Define Boundaries<br />
of Decision<br />
Task 4<br />
Transform and<br />
Integrate Models<br />
Task 5<br />
Generate Potential<br />
Solutions<br />
Recommendation<br />
Enterprise Design as an Extensive Game<br />
Identify Players<br />
of the Game<br />
Define Order of Moves<br />
Define Players’ Choices<br />
Identify Public and<br />
Private Information<br />
Probabilities on<br />
Exogenous Events<br />
Define Players’ Payoffs<br />
Define Strategies<br />
Given<br />
Find<br />
Satisfy<br />
Minimize<br />
Compromise DSP<br />
Figure 3.3 A Framework for Enterprise Design<br />
Scalar<br />
Games<br />
Vector<br />
Games<br />
Probabilistic<br />
Design Model<br />
During Tasks 2 and 3, where it is determined the input needed for analysis and modeling<br />
techniques, and when the boundaries of decision are set, is when the basic information required<br />
to define a strategy in this <strong>game</strong> is identified, i.e., the order of moves, players’ choices,<br />
information available, probabilities on exogenous events, and the players’ payoffs (see Section<br />
2.3.1). With this information it is possible to embody a strategy, using a compromise DSP, as<br />
67
explained in Section 3.5. Finally, and depending on the characteristics of the problem, the<br />
compromise DSP can be formulated as a <strong>probabilistic</strong> decision model, in order to obtain<br />
interval solutions, and to capture uncertainty and imprecision. This framework is described in<br />
detail in the rest of this chapter.<br />
3.2 GAME THEORY AS A FOUNDATION FOR COORDINATION OF DESIGN<br />
DECISIONS<br />
In Section 3.1 it is argued that an organization must adopt a coordination mechanism that<br />
balances the advantages of implicit and explicit coordination. Given the abstraction of an<br />
enterprise activity as a <strong>game</strong>, as discussed in Chapter 2, it should not be difficult to consider the<br />
social rules of an implicit coordination mechanism analog to the rules of a <strong>game</strong>. During an<br />
enterprise <strong>design</strong> process, abstracted as an extensive-<strong>game</strong>, the actions of a player are taken<br />
from a set of alternatives that might depend on the previous history of the <strong>game</strong>. The problem in<br />
defining the strategies of players is that the state in which each player ends up after taking a<br />
particular action at a particular state depends also on actions and states of other players. Thus,<br />
in principle, the transition of the entire system can be thought of a mapping from the states and<br />
actions of all players to the new states of all players. In other words, the agents or players in<br />
this <strong>game</strong> have to think ahead and devise a strategy considering other agents’ goals, intentions<br />
and motivations.<br />
A player in a <strong>game</strong> that acts rationally seeks the best solution for his/her own performance<br />
while respecting the rules of the <strong>game</strong>. In doing so he/she has to assume that all other players<br />
68
will do so as well. These rules on the one hand constrain her autonomy and possible choices,<br />
but on the other hand guarantee certain behaviors on the part of other agents, so that she can<br />
formulate a specific <strong>design</strong> strategy. If all players involved in the <strong>game</strong> act in such a rational<br />
way, the expected outcome is a point of equilibrium (if it exists), either a Pareto solution or a<br />
Nash solution, depending on the cooperative protocol. Thus, the <strong>design</strong> of the organization, i.e.,<br />
the <strong>design</strong> of its structure, communication and information systems, rewards systems, etc., has<br />
to be carried out in such a way that motivates and facilitates agents to achieve this equilibrium.<br />
Adopting <strong>game</strong> theory as a foundation for coordination of agents has important<br />
implications. A Pareto or full cooperative solution requires centralization of decisions, which in<br />
turn requires a model of the enterprise to quantify all the important effects of a decision. A non-<br />
cooperative protocol, on the other hand, empowers employees by decentralizing decisions and<br />
does not require the large information demands for decision making that are necessary in<br />
attaining a Pareto solution. This discussion has settle the foundation for Hypothesis 2.1:<br />
Hypothesis 2.1: Enterprise decisions can be effectively coordinated using a non-<br />
cooperative formulation of the individual <strong>design</strong> problems.<br />
However, it must be kept in mind that, given the same conditions, a cooperative solution is<br />
usually better than a non-cooperative one.<br />
According to the preceding discussion, the rules of an enterprise <strong>design</strong> <strong>game</strong> should<br />
define at least, but are not restricted to:<br />
69
1. Who are the players and which is their respective authority (organization structure).<br />
2. How players interact (information, communication and negotiation systems).<br />
3. Individual and systems goals which in turn define the payoffs of players (a reward system).<br />
If these rules are clear it is easier for individuals to formulate strategies of action and<br />
coordinate their decisions. Note that these “rules” already exist in almost every manufacturing<br />
organization. The difference is that now these rules should be settle in such a way that they<br />
promote the achievement of a particular kind of equilibrium.<br />
In either a Pareto or a Nash solution, each player perceives the solution as optimal, or at<br />
least “satisficing,” in the sense that they cannot do better by deviating from such solution. This is<br />
a natural perspective to coordinate agents because it fosters the <strong>design</strong> of an organization <strong>with</strong><br />
controls that are consistent <strong>with</strong>, not contrary to, the behavior of individuals, who generally act<br />
according to their own interests. This perspective has also the advantage that brings the insight,<br />
constructs and tools of <strong>game</strong> theory to improve the individual decision-making.<br />
How is one to derive appropriate rules for this <strong>game</strong>, i.e., to <strong>design</strong> the coordination<br />
mechanism? One <strong>approach</strong> could be to handcraft laws for each organization given certain<br />
existing conditions. An alternative <strong>approach</strong> could be to derive a general theory of social laws<br />
(see Shoham and Tennenholtz, 1992) and derive from it appropriate laws for individual<br />
organizations. These issues are beyond the scope of this thesis, but certainly <strong>game</strong> theory can<br />
70
provide insightful contributions to organization <strong>design</strong> in general, and the <strong>design</strong> of coordination<br />
mechanisms for enterprise integration in specific.<br />
In Sections 3.3 and 3.4 it is shown how a <strong>game</strong>-<strong>theoretical</strong> coordination mechanism can<br />
be implemented in practice by using a common model for decision support across the enterprise<br />
and a multi-objective decision model, the Compromise DSP.<br />
3.3 IMPLICIT COORDINATION THROUGH A COMMON PARADIGM FOR<br />
DECISION SUPPORT<br />
In the previous section it is mentioned that the basic elements that rule the behavior of the agents<br />
of an organization are:<br />
?? The structure of the organization, i.e., to settle which individuals and responses are<br />
associated to each level of decision.<br />
?? Information, communication and negotiation systems, i.e., the way in which information is to<br />
be filtered and passed to other agents at the same or different level of authority.<br />
?? Individual and systems goals.<br />
In order to pursue a decision-<strong>based</strong> enterprise <strong>design</strong>, it must be added to the former<br />
requirements another element, namely, “the way in which decisions are formulated.” This<br />
element is important because it fosters enterprise integration through the creation and<br />
implementation of a common paradigm for decision support and a unified decision model,<br />
thus promoting coordination between agents by providing a common language for<br />
71
communication and negotiation. These common paradigm for decision support and decision<br />
model should be mathematically supported, because mathematics provides some of the basic<br />
elements of a scientific <strong>design</strong> practice, discussed in Section 1.2.2, such as:<br />
?? A valid framework for rational choice among alternatives (therefore, to solve conflicts) and<br />
tools for actually making the computations.<br />
?? A body of techniques for actually deducing which of the available alternatives is the<br />
optimum.<br />
?? Algorithms to search for alternatives and perform tradeoffs.<br />
As mentioned in Section 1.3.1, Peplinski (1997) proposes a model for decision support in<br />
enterprise <strong>design</strong> that fulfills this aim. In such a model, a decision is represented in terms of:<br />
?? A set of <strong>design</strong> variables {X}<br />
?? A set of goals and constraints and their targets Ti, and<br />
?? Models Fi(X) that quantify the relationships between the variables and objectives.<br />
This model sets the stage for concrete and specific definitions of interdependence and<br />
integration in terms of decisions. Based on this model, Peplinski (1997) propose an enterprise<br />
<strong>design</strong> method, through which diverse issues across product <strong>design</strong>, manufacturing process<br />
<strong>design</strong> and organization <strong>design</strong> can be integrated (see Section 1.4). This enterprise <strong>design</strong><br />
method establishes a specific set of steps in making decisions, thus predicates over each agent<br />
by constraining the decision making <strong>approach</strong>. In other words, it asks individuals to follow a<br />
72
formal and rigorous <strong>approach</strong> in making decisions, thus guaranteeing at some extent the rational<br />
outcome that is assumed in <strong>game</strong> theory.<br />
In the following section it is examined the implementation of an explicit coordination<br />
mechanism <strong>based</strong> on the use of a common decision model across the enterprise, the<br />
Compromise DSP, which is the next element of the framework for enterprise <strong>design</strong> that is<br />
being constructed.<br />
3.4 EXPLICIT COORDINATION THROUGH A UNIFIED DECISION MODEL:<br />
THE COMPROMISE DSP<br />
As mentioned in Section 3.1, in explicit coordination the agents of an organization have to be<br />
aware of the system-wide effects of their decisions and make choices <strong>based</strong> on such awareness.<br />
Various <strong>approach</strong>es for explicit coordination have been considered in recent years in the<br />
context of a concurrent product and manufacturing process development, such as <strong>design</strong> <strong>with</strong><br />
features (Dixon, et al., 1988), enhanced CAD systems (Lu and Subramanyam, 1988), and<br />
multi-objective <strong>design</strong> evaluations (Hwang, et al., 1980). Agrell (1994) examines the<br />
advantages of each one and concludes that a multi-objective <strong>design</strong> evaluation <strong>approach</strong> is the<br />
best way to deal <strong>with</strong> an explicitly integrated <strong>design</strong> problem. In this thesis the Compromise<br />
DSP is used as a rigorous and practical tool to solve multi-objective <strong>design</strong> problems.<br />
The Compromise DSP is a multi-objective decision model which is a hybrid formulation<br />
<strong>based</strong> on Mathematical Programming and Goal Programming Technique (Mistree and Muster,<br />
73
1988). The Compromise DSP is used to determine the values of <strong>design</strong> variables to satisfy a<br />
set of constraints and to achieve as closely as possible a set of conflicting goals. The<br />
Compromise DSP is used to model such decisions since it is capable of handling constraints,<br />
goals, and multiple objectives. In particular, the Compromise DSP offers the following<br />
capabilities:<br />
• Accurately represent single-objective or multi-objectives.<br />
• Use either preemptive or Archimedean formulation to prioritize objectives.<br />
• Have hard constraints or soft constraints (goals).<br />
• Generate feasible solutions more frequently.<br />
• Quickly generate results for several different weighting schemes.<br />
The Compromise DSP has been successfully used in <strong>design</strong>ing aircrafts (Lewis and Mistree,<br />
1997), thermal energy systems (Bascaran, et al., 1987) damage tolerant structural systems<br />
(Shupe et al., 1984), and material composite <strong>design</strong> (Karandikar and Mistree, 1990), to name<br />
some applications. A compromise DSP is stated in terms of the following system descriptors:<br />
?? Variables: system variables and deviation variables<br />
?? Constraints: system constraints and goal constraints<br />
?? Bounds: on system and deviation variables<br />
?? Objective: in terms of deviation variables only<br />
The resulting compromise DSP is stated in words as follows:<br />
74
Given<br />
An alternative that is to be improved through modification.<br />
Assumptions used to model the domain of interest.<br />
The system parameters.<br />
The goals for the <strong>design</strong>.<br />
Find<br />
The values of the independent system variables that describe the attributes of an artifact.<br />
The values of the deviation variables that indicate the extent to which the goals are achieved.<br />
Satisfy<br />
The system constraints that must be satisfied for the solution to be feasible.<br />
The system goals that must achieve a specified target value as much as possible.<br />
The upper and lower bounds on the system variables.<br />
Minimize<br />
The objective function Z, which is a measure of the deviation of the system performance<br />
from that implied by the set of goals and their associated priority levels or relative weights.<br />
The difference between the compromise DSP and a traditional single objective<br />
formulation for a two dimensional problem is illustrated in Figure 3.4. In the case of the single<br />
objective formulation, shown in Figure 3.4a, the objective is a function of the system variables.<br />
The space representing all feasible solutions (the feasible <strong>design</strong> space) is surrounded by the<br />
system constraints and bounds of the problem. The objective is to maximize the value of Z, and<br />
as shown in the figure, the solution is at vertex A.<br />
In the compromise DSP, the set of system constraints and bounds again define the<br />
feasible <strong>design</strong> space, whereas the set of system goals defines the aspiration space (see Figure<br />
3.4b). For feasibility, the system constraints and bounds must again be satisfied whereas the<br />
system goals are to be achieved to the extent possible. The solution to this problem represents<br />
a tradeoff between that which is desired (as modeled by the aspiration space) and that can be<br />
achieved (as modeled by the <strong>design</strong> space).<br />
75
X 2<br />
Objective<br />
Function<br />
Z = W1 A1(X) + W2 A2(X) + W3 A3(X)<br />
Feasible<br />
Design<br />
Space<br />
A<br />
Bounds<br />
System constraints<br />
(a)<br />
Direction of increasing Z<br />
X 1<br />
X 2<br />
Feasible<br />
Design<br />
Space<br />
A1(X) + d1 - - d1 + = G1<br />
A<br />
Bounds<br />
System constraints<br />
System goals<br />
(b)<br />
Aspiration<br />
Space<br />
X 1<br />
76<br />
A2(X) + d2 - - d2 + = G2<br />
A3(X) + d3- - d3 + = G3<br />
Deviation<br />
Function<br />
Z = W3 (d1 -+ d1 +) + W3 (d2 -+ d2 +) + W3 (d3 -+ d3 +)<br />
Figure 3.4 A Single Objective Optimization Problem and the Multi-goal<br />
Compromise DSP (Mistree, et al., 1993)<br />
The solution for the compromise DSP shown in Figure 3.4b is at vertex A. This is the<br />
same solution as that obtained for the problem illustrated in Figure 3.6a except that in the<br />
compromise DSP case (<strong>with</strong> the aspiration space modeled), the best possible solution can be<br />
identified. This solution is referred to as a satisficing solution since it is a feasible point that<br />
achieves the system goals to the extent that is possible. Given a solution, it is left to a<br />
<strong>design</strong>er to accept this solution or to explore the problem further by modifying the aspirations<br />
and/or the feasible <strong>design</strong> space and re-solving. The values of the deviation variables provide a<br />
measure for assessing the degree to which each of the goals has not been achieved.<br />
The system descriptors, namely, system and deviation variables, system constraints,<br />
system goals, bounds and the deviation function are described in (Mistree, et al. 1993) and are<br />
therefore not repeated here. The mathematical form of the compromise DSP is shown in Figure<br />
3.5.
Given<br />
An alternative to be improved through modification.<br />
Assumptions used to model the domain of interest.<br />
The system parameters:<br />
n number of system variables<br />
p+q number of system constraints<br />
p equality constraints<br />
q inequality constraints<br />
m number of system goals<br />
gi(X) system constraint function:<br />
gi(X) = Ci(X) – Di(X)<br />
fk(di ) function of deviation variables to be minimized at priority level k<br />
for the preemptive case.<br />
Find<br />
Xi i = 1, …, n<br />
? ?<br />
i , di<br />
d i = 1, …, m<br />
Satisfy<br />
System constraints (linear, nonlinear)<br />
gi(X) = 0 ; i = 1, …, p<br />
gi(X) = 0 ; i = p+1, …, p+q<br />
System goals (linear, nonlinear)<br />
Bounds<br />
Ai(X) + d - i – d+ i = Gi ; i = 1, …, m<br />
X i min = Xi = X i max ; i = 1, …, n<br />
d - i , d+ i = 0 ; i = 1, …, m<br />
d - i . d + i = 0 ; i = 1, …, m<br />
Minimize<br />
Archimedean formulation<br />
? ?<br />
Z ? Wi(di ? di )<br />
?<br />
Figure 3.5 Mathematical Form of a Compromise DSP<br />
(Mistree, et al., 1993)<br />
77
In the compromise DSP, each goal, Ai, has two associated deviation variables<br />
?<br />
d i , which indicate the extent of the deviation from the target. The deviation variables,<br />
?<br />
i<br />
78<br />
?<br />
d i and<br />
?<br />
di and<br />
d , are both positive, and the product constraint, ? ? 0<br />
? ?<br />
d d , ensures that at least one of the<br />
deviation variables for a particular goal is always zero.<br />
Usually, goals are not equally important. To determine a solution on the basis of<br />
preference, the goals may be rank-ordered into priority levels. Customers rate certain product<br />
qualities higher than other qualities. Designers usually seek a solution that minimizes all<br />
unwanted deviations from the desired qualities. There are various methods for measuring the<br />
effectiveness of the minimization of these unwanted deviations. In this work it is used the<br />
Archimedean or weighted sum method, in which the deviation function Z is formulated as:<br />
Z ? W i(d i<br />
<strong>with</strong> Wi ? 0<br />
?<br />
?<br />
W i ? 1<br />
i<br />
? ? di<br />
The weighted sum method generates a Pareto solution, but is deficient in that it is difficult<br />
to determine which weights will produce the best compromise solution. A systematic<br />
redistribution of the weights on the objectives does not necessarily produce a continuous<br />
Pareto-optimal solution set. To overcome this problem Rao and Freiheit (1991) proposed a<br />
modified <strong>approach</strong> as follows (in terms of a compromise DSP deviation function Z):<br />
1. From a constant initial <strong>design</strong> vector, minimize each of the objectives separately, recording<br />
the values of the other objectives at each optimal <strong>design</strong> vector.<br />
? )<br />
i
Minimize Zi ? di ? ? di ? , i ? 1, 2, ...,m (3.1)<br />
subject to system constraints. For each objective function Zi an “optimal” solution Xi *<br />
is obtained.<br />
2. With the values obtained in the previous step, construct a pay-off matrix as:<br />
? Z1(<br />
X<br />
?<br />
? ? Z1(<br />
X<br />
P o ?<br />
?<br />
? Z1<br />
( X<br />
*<br />
1<br />
*<br />
2<br />
*<br />
m<br />
)<br />
)<br />
)<br />
Z<br />
Z<br />
Z<br />
2<br />
2<br />
2<br />
( X<br />
( X<br />
( X<br />
Normalize the objectives so that no objective due to its magnitude will be favored as<br />
follows:<br />
*<br />
1<br />
*<br />
2<br />
*<br />
m<br />
Z ˆ<br />
i ? Zi ? Zi *<br />
w<br />
Zi ? Z i<br />
)<br />
)<br />
)<br />
Z<br />
Z<br />
Z<br />
m<br />
m<br />
m<br />
( X<br />
( X<br />
( X<br />
*<br />
1<br />
*<br />
2<br />
*<br />
m<br />
) ?<br />
?<br />
) ?<br />
?<br />
?<br />
) ?<br />
79<br />
* (3.2)<br />
where Zi w is the worst value and Zi * is the optimum value of the ith objective.<br />
3. Formulate a super-criterion SC. A formulation for a super-criterion which has been used<br />
effectively (Rao and Feiheit, 1991) is:<br />
m<br />
SC ? [1 ? ˆ ? Z i ]<br />
(3.3)<br />
Note that due to the normalization, SC will always have a value between zero and one,<br />
which is the same magnitude of the normalized objectives.<br />
4. Formulate a Pareto optimal objective Z using the normalized objectives:<br />
i? 1<br />
Z ? W ˆ ? iZ i i ? 1, ...,m (3.4)
5. Since Z has to be minimized and S has to be maximized, a new objective Z mp that has to be<br />
minimized is constructed as:<br />
subject to the system constraints and<br />
80<br />
Z mp ? Z ? SC (3.5)<br />
Z ˆ<br />
i ? ? Zi mp ? ˆ Z i w (3.6)<br />
<strong>with</strong> the <strong>design</strong> vector X now including the priorities on objectives. This technique greatly<br />
simplifies the search for a “satisficing” solution since it is not necessary to construct the entire<br />
set of Pareto points. However, it must be remembered that it does so by changing the<br />
characteristic of the original problem.<br />
Karandikar and Mistree (1993) and Peplinski, et al. (1996) have already shown that the<br />
Compromise DSP is an adequate tool to support the integration of enterprise <strong>design</strong> processes.<br />
This method has proven to be efficient because of the expected reduction in the number of<br />
iterations and it is effective because it can deal <strong>with</strong> the complexity of the problem in terms of<br />
hierarchy and the existence of multiple objectives in <strong>design</strong>.<br />
The discussion carried out in this section provides the foundation for Hypothesis 2.2:<br />
Hypothesis 2.2: the Compromise DSP is an adequate decision model to embody <strong>game</strong><br />
<strong>theoretical</strong> <strong>design</strong> formulations <strong>with</strong>in an enterprise <strong>design</strong> framework.<br />
In the next section it is discussed how the common paradigm for decision support<br />
discussed in Section 3.4 facilitates the creation of compromise DSPs, thus establishing a<br />
consistent mechanism for coordination of decisions, which fosters its implementation in a
computer-supported environment, and which capitalize on both implicit and explicit<br />
coordination.<br />
3.5 FORMULATION OF DESIGN PROBLEMS IN ENTERPRISE<br />
DESIGN<br />
3.5.1 Approximation in Engineering Design<br />
Explicit coordination requires individuals to make explicit choices <strong>based</strong> on awareness of the<br />
system-wide effects of their decisions. To do so they have to be able to predict such effects. In<br />
“The Sciences of the Artificial” Herbert Simon comments (Simon, 1996):<br />
In real life business firm must choose product quality and the assortment of products it<br />
will manufacture. It often has to invent and <strong>design</strong> some of these products. It must<br />
schedule the factory to produce a profitable combination of them and devise marketing<br />
procedures and structures to sell them. So we proceed step by step from the simple<br />
caricature of the firm depicted in the textbooks, to the complexities of real firms in the<br />
real world of business. At each step towards realism, the problem gradually changes<br />
from choosing the right course of action (substantive rationally) to finding a way of<br />
calculating, very approximately, where a good course of action lies (procedural<br />
rationality).<br />
In a perfect world, approximation and procedural rationality would not be needed, but<br />
limited human beings as we are, we do not have the ability nor sufficient information to create<br />
“exact” models that capture the real behavior of natural or artificial systems, even the most<br />
simple ones. Therefore, we build models that approximate such behavior <strong>with</strong> different degrees<br />
of accuracy. Usually, the boundaries, elements and internal relationships of these models are<br />
selected to explore different parameters drawn form a single simple model and each <strong>based</strong> on<br />
81
different simplifying and independence assumptions depending on our needs and interests. In<br />
essence, a formulated <strong>design</strong> problem and its supporting models must be stated in sufficiently<br />
simple terms that finding a solution is a manageable process and, at the same time, they must be<br />
a close-enough approximation of the real-world that the solution is useful. In doing so, we have<br />
two choices when faced <strong>with</strong> complex real-world problems in <strong>design</strong> (Muster and Mistree,<br />
1988):<br />
1. To develop an algorithm, <strong>based</strong> on a relatively simple model, by means of which an exact<br />
optimal solution can be found, and provided the assumptions on which the model is <strong>based</strong><br />
can be satisfied exactly.<br />
2. To develop an approximate algorithm or heuristic <strong>based</strong> on a relatively complex model that<br />
represents the real world more closely than the simple model referred to earlier.<br />
When dealing <strong>with</strong> multidisciplinary problems, like in enterprise <strong>design</strong> problems, the motivations<br />
for approximation become even more critical because direct coupling of a <strong>design</strong> space search<br />
(DSS) code to several multidisciplinary analysis models may be impractical due to<br />
computational constraints. Consequently, most optimizations of complex engineering systems<br />
couple a DSS code to easy-to-calculate approximations of the objective and constraints. An<br />
“optimum” of the approximate problem is found; then, the approximation models are upgraded<br />
and the search is performed again on a smaller <strong>design</strong> space that surrounds the “optimum” point<br />
found on the previous step.<br />
82
Approximation of complex models <strong>with</strong>out simplifying assumptions may take many forms.<br />
Global reduced-basis approximations (e.g., Kirsch, 1991) are occasionally used in engineering<br />
system optimization. However, most often the approximations are either local, linear, or<br />
quadratic, <strong>based</strong> on the derivatives. Occasionally, intermediate variables or intermediate<br />
response quantities (e.g., Kodiyalam and Vanderplaats, 1989) are used to improve the<br />
accuracy of the approximation. Another method is a variable-complexity modeling (VCM)<br />
technique (e.g., Unger, et al., 1992) in which both simpler and more sophisticated models<br />
alternate during the optimization procedure. The sophisticated model provides a scale factor,<br />
which is periodically updated to correct the simpler model. Traditional derivative-<strong>based</strong><br />
approximations can be combined <strong>with</strong> global VCM approximations by using a derivative-<strong>based</strong><br />
linear approximation for the scale factor (Chang, et al., 1993). Toropov and Markine (1996)<br />
generalize this <strong>approach</strong>, constructing an approximation <strong>based</strong> on both the <strong>design</strong> variables and<br />
the prediction of the simpler model.<br />
A different kind of approximation is statistical-<strong>based</strong> approximation. Response surface<br />
techniques replace the objective and constraints functions <strong>with</strong> simple functions, often<br />
polynomials, which are fitted to data computed at a set of carefully selected <strong>design</strong> points<br />
(Chen, et al., 1996). These approximations may be regarded either as global or local<br />
approximations depending on the <strong>design</strong> space that they span. Other <strong>approach</strong>es of this nature<br />
are Neural Networks-<strong>based</strong> regression, kriging and inductive learning. Unfortunately, statistical-<br />
<strong>based</strong> approximations do not scale well to a large number of variables (Sobieszcanski-Sobieski<br />
83
and Haftka, 1997). Nevertheless, these techniques provide a convenient representation of data<br />
from one discipline to other disciplines and to the system. Because <strong>design</strong> points are pre-<br />
selected, rather than chosen by an optimization algorithm, planning and coordinating the solution<br />
process “off line” is possible.<br />
In summary, to permit practical solution of enterprise <strong>design</strong> problems, we require to<br />
approximate the behavior of systems <strong>with</strong> relatively simple models, but as Simon (1996) notes:<br />
“it is done at the cost of shaping and squeezing the real-world problem.”<br />
3.5.2 The Cooperative Formulation using Compromise DSPs<br />
Up to this point, it has been examined how <strong>game</strong> theory can provide a foundation for<br />
coordination of decisions. In Section 3.5.1 it is asserted that the Compromise DSP is an<br />
adequate decision model to embody <strong>game</strong> <strong>theoretical</strong> formulations of decisions. In this section<br />
it is explained how to formulate a <strong>design</strong> problem to achieve a Pareto or cooperative solution<br />
(defined mathematically in Section 2.2.3), using the Compromise DSP. The way in which this<br />
solution can be achieved is as follows:<br />
1. Combine the models of each player into one encompassing model.<br />
2. Solve the model using an appropriate technique.<br />
Here the strategy of a player is embodied by a <strong>design</strong> formulation, in this case using a<br />
compromise DSP. A full cooperative model (using compromise DSPs) is illustrated in Figure<br />
3.6. The compromise DSPs of Player 1 and Player 2 are combined into one compromise DSP<br />
84
and solved using the cumulative <strong>design</strong> variables, constraints, goals, and deviation variables of<br />
both players.<br />
Player 1<br />
Player 2<br />
Find<br />
x1, d +<br />
1 ,d1 -<br />
Satisfy<br />
constraints<br />
g 1(x1,s1,x2,s2 h 1 (x1 ,s1 ,x2 ,s 2 )=0.0<br />
goals<br />
f 1 (x1 ,s1 ,x2 ,s2 )+ d -<br />
1 -d +<br />
1 =1.0<br />
Minimize<br />
Z1 (x1 ,s1 ,x2 ,s2 )<br />
Find<br />
x2 , d +<br />
2 ,d -<br />
2<br />
Satisfy<br />
constraints<br />
g2(x1,s1,x2,s2 h 2(x1,s 1,x2,s2)=0.0 goals<br />
f -<br />
2(x1,s1,x2,s 2)+d 2 -d +<br />
2 =1.0<br />
Minimize<br />
Z2(x 1,s1,x2,s2) Find<br />
x1 ,x2 , d +<br />
1 ,d -<br />
1 , d +<br />
2 ,d2 -<br />
Satisfy<br />
constraints<br />
g 1(x1,s1,x2,s 2) 0.0<br />
h 1 (x1 ,s1 ,x2 ,s2 )=0.0<br />
g 2(x1,s1,x2,s 2) 0.0<br />
h 2(x1,s1,x2,s2)=0.0 goals<br />
f -<br />
1(x1,s1,x2,s2)+d 1 -d +<br />
1 =1.0<br />
f 2 (x1 ,s1 ,x2 ,s2 )+d -<br />
2 -d +<br />
2 =1.0<br />
Minimize<br />
Z(x1 ,s1 ,x2 ,s2 )<br />
Figure 3.6 Full Cooperation: Pareto Solutions (Lewis, 1997)<br />
Although conceptually obtaining a cooperative solution is simple and <strong>theoretical</strong>ly sound, it<br />
is difficult to implement since the models that the different agents utilize are usually complex<br />
analysis packages, and many times are solved at different points in a <strong>design</strong> process using<br />
approximate and/or incomplete information. Combining these models is typically impractical due<br />
to the sheer size of the models and analysis routines. Therefore, the definition of a cooperative<br />
solution in complex systems <strong>design</strong> has to be extended to “approximate cooperation” (Lewis,<br />
1997) where models remain separate but linked through approximations of the coupling<br />
variables which are needed by more than one agent. Approximate cooperation is achieved<br />
85
using approximations of the state variables, including constraints and goals needed from the<br />
other players, obtained using any suitable technique. However, approximation of every<br />
constraint, goal, and state variable is unrealistic. Hence, only those system variables that are<br />
largely affected by the local <strong>design</strong> variables are approximated. Because of the non-linearity of<br />
the resulting enterprise <strong>design</strong> problems it is difficult to know a priori which variables of other<br />
players are relevant to the local <strong>design</strong> decision. CIM models, screening experiments and<br />
sensitivity analysis can help a <strong>design</strong>er’s judgment to this aim. The approximation technique to<br />
be used depends on the problem; for example, when dealing only <strong>with</strong> continuum variables a<br />
derivative-<strong>based</strong> method, such as the Global Sensitivity Equations <strong>approach</strong> (see Lewis, 1997)<br />
is appropriate. Using the Compromise DSP and approximations of non-local information, the<br />
cooperative <strong>design</strong> problems are now formulated as “approximate-cooperative DSPs”<br />
(Lewis, and Mistree, 1997). This concept is illustrated in Figure 3.7.<br />
Player 1<br />
Given<br />
X2 Approximations of Player 2’s<br />
information<br />
Player 1’s Models<br />
Find<br />
X1 + -<br />
d1i , d1i Satisfy<br />
constraints<br />
goals<br />
Minimize<br />
Z1 = [ f1 , f2 ]<br />
Player 2<br />
Given<br />
X1 Approximations of Player 1’s<br />
information<br />
Player 2’s Models<br />
Find<br />
X2 + -<br />
d2i , d2i Satisfy<br />
constraints<br />
goals<br />
Minimize<br />
Z2 = [ f1 , f2 ]<br />
P1 P2<br />
Figure 3.7 Approximate-Cooperation using Compromise DSPs<br />
86
Note that due to the use of approximations the problem is not anymore a single problem,<br />
and iterations between the players’ solutions are required. If the heuristic proposed by Rao and<br />
Freihet (1991) previously described in Section 3.4 (or a similar one) is not used, in any single<br />
iteration each player has to generate an individual set of Pareto solutions, each one requiring<br />
solving a compromise DSP. From the individual’s set each player choose a solution and then<br />
convergence is checked. Clearly, the modified method is advantageous in order to reduce the<br />
number of compromise DSPs to be solved in an approximate-cooperative model. The<br />
algorithm for the solution of the approximate-cooperative <strong>design</strong> model is illustrated in Figure<br />
3.8 considering two players. The process starts <strong>with</strong> some initial values of their <strong>design</strong><br />
variables, say x10 and x20, and each player solves an approximate-cooperative DSP. Once the<br />
solution of each player is obtained, the values of the resulting <strong>design</strong> variables, x1 and x2, are<br />
compared, and if they differ they update the values of these variables and solve again their<br />
respective compromise DSPs until convergence is achieved.<br />
x 10<br />
x 20<br />
x 1<br />
x 2<br />
Approximate-Cooperative<br />
DSP for Player 1<br />
Approximate-Cooperative<br />
DSP for Player 2<br />
No<br />
Convergence?<br />
x 1<br />
x 2<br />
87
Figure 3.8 Solution of a Two-Players Approximate-Cooperative Model<br />
3.5.3 The Non-Cooperative Formulation using Compromise DSPs<br />
As discussed in Section 3.2, it might be more practical in enterprise <strong>design</strong> to pursue<br />
coordination of <strong>design</strong> decisions using a non-cooperative protocol than using a cooperative one.<br />
In this section it is shown how to achieve non-cooperative solutions using the Compromise<br />
DSP.<br />
Finding the rational reaction sets (RRS) of the players is the first step in formulating a<br />
strategy for a non-cooperative <strong>game</strong>. Similar to the cooperative case where complete<br />
cooperation is not realistic in the <strong>design</strong> of complex systems, finding exact mathematical RRS of<br />
other players is not always possible. Each player might use models that consist of multiple,<br />
nonlinear constraints and objectives that require nonlinear programming and/or heuristic<br />
techniques in order to solve for the independent variables. To develop a closed form equation<br />
for one or more variables as functions of other variables is difficult. Therefore, the RRS of<br />
players might have to be found by using approximation techniques. In this work the RRS are<br />
approximated using response surfaces.<br />
The steps followed in this work to construct the RRS of each player is as follows (adapted<br />
from Lewis, 1997):<br />
1. Use statistical software as the <strong>design</strong> of experiments (DOE) driver.<br />
88
1.a) Based on the characteristics of the input and response variables set up the appropriate<br />
experimental <strong>design</strong>.<br />
1.b) Set the variables of the players which are required constant in a compromise DSP<br />
2. Solve the individual compromise DSP using the modified Pareto solution described in<br />
Section 3.4.2.<br />
3. Send the values of the <strong>design</strong> variables obtained in the previous step to the DOE driver.<br />
4. Determine if the full experiment is finished.<br />
?? If not, continue by moving to the next experiment point and repeat Step 2.<br />
?? Otherwise, construct the required response surfaces.<br />
This algorithm is illustrated in Figure 3.9.<br />
sP1<br />
1<br />
P1’s Design Space<br />
xP1<br />
xP1, sP1<br />
xP1, sP1<br />
xP1, sP1<br />
xP1, sP1<br />
P2’s Compromise<br />
DSP<br />
Given<br />
Find<br />
Satisfy<br />
Minimize<br />
Given<br />
Find<br />
Satisfy<br />
Minimize<br />
Given<br />
Find<br />
Satisfy<br />
Minimize<br />
Given<br />
Find<br />
Satisfy<br />
Minimize<br />
xP2, sP2<br />
xP2, sP2<br />
xP2, sP2<br />
xP2, sP2<br />
Rational Reaction<br />
Set xP2, sP2 = f(xP1, sP1)<br />
y<br />
Response Surface<br />
Equations<br />
Figure 3.9 Conceptual Outline of RRS Construction (Lewis, 1997)<br />
2<br />
3<br />
89
In Figure 3.10 is shown the compromise DSPs of two players <strong>with</strong> a non-cooperative<br />
formulation. The given information is information required from the other player that is<br />
unknown. Therefore, a player’s solution must be found using unknown variables.<br />
Player I Player II<br />
Given<br />
unknown s II, x II<br />
Find<br />
x I, d i + ,di -<br />
Satisfy<br />
constraints<br />
g I(x I,s I,x II,s II<br />
h I (x I ,s I ,x II ,s II )=0.0<br />
goals<br />
f i(x I,s I,x II,s II)+d i - -d i + =1.0<br />
Minimize<br />
Z I (x I ,s I ,x II ,s II )<br />
Given<br />
unknown s I , x I<br />
Find<br />
x II , d i + ,di -<br />
Satisfy<br />
constraints<br />
g II (x I ,s I ,x II ,s II<br />
h II (x I ,s I ,x II ,s II )=0.0<br />
goals<br />
f i (x I ,s I ,x II ,s II )+d i - -d i + =1.0<br />
Minimize<br />
Z II(x I,s I,x II,s II)<br />
Figure 3.10 Non-cooperative Compromise DSPs (Lewis, 1997)<br />
The steps to solve a non-cooperative formulation are as follows (Lewis, 1997):<br />
1. Construct Rational Reaction Set of each player I, RRSi.<br />
2. Using an appropriate technique, find the intersection points of the RRS of the players.<br />
90<br />
X RRS ? RRS (i ? j)<br />
(3.7)<br />
n ? i<br />
j<br />
3. Determine which solution fall in the ranges of the <strong>design</strong> variables.<br />
These three steps are shown for a two-player case in Figure 3.11, where player 1 controls the<br />
value X1 and player 2 controls the value of X2.
Construct RRS 1<br />
Construct RRS 2<br />
Figure 3.11 Solution of a Two-Player Non-cooperative Model<br />
A particular case of non-cooperative <strong>game</strong>s that is used in this work is the Stackelberg or<br />
Leader/Follower <strong>game</strong>. This model arises in some dynamic situations (extensive <strong>game</strong>s) where<br />
one player dominates others. This is the case, for example, of a sequential <strong>design</strong> process<br />
where a product <strong>design</strong>er hands the <strong>design</strong> over the production department. In the Stackelberg<br />
formulation the leader may be able to use knowledge of the followers’ response to his<br />
advantage in minimizing his own deviation function. The followers may also benefit from having<br />
a leader in that they do not have to guess what the leader will do. This behavior of the follower<br />
is dictated by his strategy and embodied by his/her RRS, which describes how the follower will<br />
behave in response to any decision made by the leader (this is the principle of backward<br />
induction).<br />
1997):<br />
1 2<br />
The process for solving the Stackelberg formulation is as follows (adapted from Lewis,<br />
1. Construct the RRS of the followers<br />
Find X 1 and X 2 such that<br />
RRS 1(<br />
X 2)<br />
? RRS 2(<br />
X 1)<br />
2. Allowing the leader access to the followers’ RRS solve the leader’s compromise DSP.<br />
3. Solve the followers’ compromise DSPs<br />
3a) If one of the followers is the leader of the remaining players, solve a<br />
3<br />
Determine which solutions<br />
fall in the feasible <strong>design</strong> space<br />
of players<br />
91
leader/follower formulation for remaining players using steps 1 and 2.<br />
3b) Otherwise, solve the appropriate <strong>design</strong> formulation (cooperative, non-<br />
cooperative, or single player).<br />
These steps to solve a Stackelberg formulation are shown schematically in Figure 3.12, and the<br />
compromise DSPs of the leader and follower are shown in Figure 3.13, considering two<br />
players. The leader’s given information includes the RRS of the follower, while the follower’s<br />
given information includes the leader’s solution.<br />
1st. Leader<br />
Given<br />
Compromise DSP<br />
Construct RRS<br />
of the followers Compromise DSP<br />
Yes<br />
Leader among<br />
previous leader’s<br />
followers?<br />
Figure 3.12 Leader/Follower Solution Process<br />
F , sF = f(xL ,sL )<br />
Find<br />
xL , d +<br />
i ,di -<br />
Satisfy<br />
constraints<br />
g(xL ,sL ,xF ,sF h(xL ,sL ,xF ,sF )=0.0<br />
goals<br />
fi (xL ,sL ,xF ,sF )+d -<br />
i -di + =1.0<br />
Minimize<br />
ZL (xL ,sL ,xF ,sF )<br />
No<br />
Given<br />
x L, s L<br />
Find<br />
x F, d j + ,d j -<br />
Satisfy<br />
constraints<br />
g(x L,s L,x F,s F<br />
h(x L,s L,x F,s F)=0.0<br />
goals<br />
f j (x L ,s L ,x F ,s F )+d j - -d j + =1.0<br />
Minimize<br />
Z F (x L ,s L ,x F ,s F )<br />
92<br />
Solve Follower’s<br />
Compromise DSP<br />
Figure 3.13 Compromise DSPs for a Two-Players Leader/Follower Game
(Lewis, 1997)<br />
It is important to note that these cooperative protocols can be combined for coordination<br />
of decisions in an enterprise. For example, some small groups can use a cooperative protocol<br />
between them, and maintain a non-cooperative one <strong>with</strong> other groups. This idea, which is<br />
illustrated in Figure 3.14, is in line <strong>with</strong> a hierarchical organization of manufacturing enterprises.<br />
GROUP B<br />
Full Cooperative<br />
GROUP A<br />
Leader/Follower<br />
LEADER/FOLLOWER<br />
NON-COOPERATIV E<br />
GROUP C<br />
Approximate-Cooperative<br />
Figure 3.14 Combining Cooperative Protocols in Enterprise Design<br />
93
In the next section is described how to enhance this framework <strong>with</strong> a <strong>probabilistic</strong><br />
<strong>design</strong> <strong>approach</strong> such that the players can obtain flexible solutions, ideally facilitating<br />
coordination of decisions.<br />
3.6 A PROBABILISTIC-BASED DESIGN APPROACH FOR ENHANCING<br />
COORDINATION WITH FLEXIBLE SOLUTIONS<br />
One of the most critical issues in <strong>design</strong> is making early decisions on a sound basis. When<br />
addressing the interests of multiple groups and criteria in making a decision, a <strong>design</strong>er faces the<br />
additional problem of finding a solution that is satisfying to all those groups. However, the early<br />
stages of <strong>design</strong> are the most uncertain and obtaining precise information from other groups<br />
upon which support decisions can be made is usually impossible. Thus, in the early stages it is<br />
known that a solution that seems “optimal” at the present will probably have to be modified to<br />
meet the requirements of those other groups that will suggest changes <strong>based</strong> on their private<br />
information and concerns. Therefore, obtaining the “best” solution in an early stage has little<br />
meaning. Instead, it is better to attain solutions that are flexible, yet robust to future changes. In<br />
this section it is described a <strong>design</strong> <strong>approach</strong> that is intended to provide such a kind of solutions.<br />
3.6.1 Why a Probabilistic-Based Design Approach?<br />
94
Antonsson and Otto (1995) distinguish two kinds of uncertainty in <strong>design</strong>: uncertainty<br />
associated <strong>with</strong> uncontrolled variations (e.g., manufacturing variations, material property<br />
variations, demand fluctuations) and uncertainty in choosing among alternatives, that they called<br />
“imprecision.” An imprecise variable is a variable that may potentially assume any value <strong>with</strong>in a<br />
range because the <strong>design</strong>er does not have all the information required for deciding the final<br />
value. To represent an imprecise variable, x, a range of numbers can be used or, more<br />
completely, a range and a function defined on the range to describe the preference for particular<br />
values. In this way variables whose values are not known precisely can be formally<br />
represented.<br />
The need for a methodology to represent and manipulate uncertainty and imprecision have<br />
led to a series of <strong>approach</strong>es such as Bayesian methods (Vadde, et al., 1991), fuzzy <strong>design</strong><br />
methods (Zhou, et al., 1992), and utility theory (Thurston and Liu, 1991), to name some.<br />
Independently of the method chosen, the use of intervals encourages the passing of imprecise<br />
<strong>design</strong> information between agents (Antonsson and Otto, 1995), and permits the early release<br />
of <strong>design</strong> information from one group to others in advance of more precise information, thus<br />
ideally facilitating the integration of the enterprise <strong>design</strong> activities. It also facilitates<br />
coordination and group decision making because it is easier to find common satisfying solutions<br />
when the “satisficing” values of the <strong>design</strong> variables are not fixed points but regions. This notion<br />
is illustrated in Figure 3.15.<br />
95
Point Solutions<br />
X 1 ? A<br />
X 2 ? B<br />
* *<br />
Ranged Solutions<br />
X 1 ? A ? ? A<br />
* *<br />
X 2 ? B ? ? B<br />
Figure 3.15 Coordination of Decisions <strong>with</strong> Ranged Solutions<br />
In this work, the <strong>probabilistic</strong>-<strong>based</strong> <strong>design</strong> model of Chen and Yuan (1997) is used, because it<br />
provides a consistent scheme. The following sections are mostly summarized from Chen and<br />
Yuan (1997) and Parkinson, et al. (1993).<br />
3.6.2 Preference Functions for Design Requirements<br />
In conventional <strong>design</strong> optimization, <strong>design</strong> requirements have to be precisely satisfied under all<br />
circumstances. For example, a <strong>design</strong> requirement on weight, W, can be stated as “it is desired<br />
the weight to be less than 800 lbs.” This can be modeled by a constraint, W ? 800 lbs, plus<br />
maybe a <strong>design</strong> objective on minimizing W. The modeling is slightly different if using goal<br />
programming, where <strong>design</strong>ers’ preferences are specified by the performance target value and<br />
expressed using one of the following three notions: (1) smaller-the-beter (STB), (2) larger-the-<br />
better (LTB), or (3) center-the-better (CTB). The previous requirement on W could be<br />
modeled by a goal (STB) <strong>with</strong> the target value set at 800 lbs. The goal could be considered as<br />
96
either rigid (constraint) or soft (objective). In all circumstances, 800 lbs is picked as a precise<br />
number to distinguish <strong>design</strong>s that are desirable or undesirable. However, the target level of<br />
performance that will be judged as desirable may be uncertain, especially in the early stages of<br />
product development and when addressing the interests of various enterprise groups. Hence, it<br />
is hard for a <strong>design</strong>er to make such a distinct evaluation.<br />
In the previous example, <strong>design</strong>ers may not want to absolutely reject the <strong>design</strong> when the<br />
gross weight is slightly above 800 lbs. Instead, they would like to consider <strong>design</strong>s in the range<br />
of 750 to 850 lbs, <strong>with</strong> a varying degree of desirability. To capture a varying degree of<br />
preference on the levels of performance, a preference function might be included in a <strong>design</strong><br />
model. An example of a preference function is shown in Figure 3.16.<br />
P(y)<br />
1<br />
0<br />
750 800 850<br />
W (lbs)<br />
Figure 3.16 A Flexible Preference Function (Chen and Yuan, 1997)<br />
In general, a preference function P(y) is a function defining the relationships between the<br />
degree of desirability, P, and the level of performance, y. The function is usually defined in the<br />
range of 0 and 1, where a value of 1 represents full preference. Any <strong>design</strong> <strong>with</strong> desirability 1 is<br />
fully acceptable, whereas a <strong>design</strong> <strong>with</strong> a value of 0 is unacceptable. For the same problem<br />
97
mentioned earlier, a linear function may be used as the preference function for W. In Figure<br />
3.16 when W ? 800 lbs, a <strong>design</strong> is fully acceptable, and a <strong>design</strong> is unacceptable if W is<br />
greater than 850 lbs. In the range of 800 to 850 lbs the degree of desirability decreases from 1<br />
to 0. A comparison is made in Figure 3.17 to show that when the preference functions are used<br />
for expressing STB, LTB and CTB, they are more flexible, compared those in conventional<br />
optimizations. In the next sections is studied how to use this flexible specification of preference<br />
for ranged <strong>design</strong> solutions. From now on, it is assumed that the ranges of <strong>design</strong> variables are<br />
symmetrical <strong>with</strong> respect to their mean.<br />
STB<br />
LTB<br />
CTB<br />
Conventional<br />
Optimization<br />
P(y)<br />
1.0<br />
0<br />
P(y)<br />
1.0<br />
0<br />
P(y)<br />
1.0<br />
0<br />
y 0<br />
y 01<br />
y0<br />
y0<br />
y02<br />
y<br />
y<br />
y<br />
Proposed Model<br />
P(y)<br />
1.0<br />
0<br />
P(y)<br />
1.0<br />
0<br />
y 1<br />
y 0 y 1<br />
Figure 3.17 A Comparison of Preference Structure (Chen and Yuan, 1997)<br />
3.6.3 Ranged Design Solutions<br />
P(y)<br />
1.0<br />
0<br />
y 1<br />
y 01<br />
y 0<br />
y 0<br />
y 02<br />
y 2<br />
y<br />
y<br />
y<br />
98
To enhance coordination of decisions and integration since an early stage of the enterprise<br />
<strong>design</strong> process it is desirable to have a <strong>design</strong> model that considers ranges of solutions for<br />
<strong>design</strong> variables instead of point solutions. In this work, the uncertainty associated <strong>with</strong> such a<br />
ranged <strong>design</strong> solution is modeled by <strong>probabilistic</strong> distributions of <strong>design</strong> and performance<br />
variables. This is illustrated in Figure 3.18, where it is shown how variation in <strong>design</strong> variables,<br />
shown on the left side, causes a variation in the dependent performance function, as shown on<br />
the right side.<br />
x1<br />
x 2<br />
Figure 3.18 Mapping between Design Variables and Design Performance<br />
Monte Carlo simulations or other statistical techniques, such as DOE, can be used to<br />
generate the probability density function of the performance attribute. However, when the<br />
computation time is demanding, it is an accepted practice to assume the performance follows a<br />
normal distribution (Chen and Yuan, 1997). This practice follows from the central limit<br />
theorem, which states that if a variable depends on a large number of independently acting<br />
random causes or variables, then its distribution closely follows a normal distribution, even<br />
y<br />
99
though the independent causes or variables are not normally distributed. In such a case, the<br />
probability density function is as follows:<br />
100<br />
2<br />
1 ?<br />
?<br />
? ?<br />
? 1 ? y ? y ?<br />
f y ? exp ? ? ? ?<br />
(3.8)<br />
? 2?<br />
? 2 ? ? ?<br />
y ? ? ? y ? ?<br />
If this assumption holds, the standard deviation ? can be approximated as ? y / 3,<br />
where ? y is<br />
the half-width of the variable range. In this case, the probability of the performance falling <strong>with</strong>in<br />
the range y ? ? y is 99.8%.<br />
3.6.4 Design Preference Index<br />
To evaluate the goodness of a <strong>design</strong> when both the <strong>design</strong> performance and the preference<br />
level of performance vary <strong>with</strong>in ranges, Chen and Yuan suggest the use of a Design Preference<br />
Index (DPI). Mathematically, the DPI is defined as the expected preference function value of<br />
<strong>design</strong> performance <strong>with</strong>in the range of <strong>design</strong> solutions. This is estimated as follows for a<br />
continuous y:<br />
? ?<br />
y ? ? y<br />
?<br />
DPI ? E P( y) ? P( y) f ( y) dy<br />
y ? ? y<br />
(3.9)<br />
This function can be used as a measure to evaluate the goodness of <strong>design</strong>s <strong>with</strong> varying<br />
performance in successfully satisfying a ranged set of requirements.<br />
Three different <strong>design</strong>s <strong>with</strong> different values of DPI are shown in Figure 3.19. In <strong>design</strong><br />
(a) the preference function value P(y) remains <strong>with</strong>in the range y ? ? y , where y is the average
value of the performance measured and DPI=1. This means that this <strong>design</strong> can fully meet the<br />
<strong>design</strong> requirements under the variation of performance. However, DPI=1 cannot be achieved<br />
in some cases, as illustrated in <strong>design</strong> (b) and (c). In <strong>design</strong> (c) the whole range of performance<br />
falls totally outside the preferred range, which results in DPI=0. Hence, a <strong>design</strong> <strong>with</strong> a value of<br />
DPI=1 corresponds to the <strong>design</strong> <strong>with</strong> the highest expected value of performance <strong>with</strong> respect<br />
to the target acceptance function. Thereupon, maximizing DPI as closely as possible to 1<br />
becomes the <strong>design</strong> objective. To evaluate the DPI, a numerical method can be used to<br />
integrate the product of P(y) and f(y) <strong>with</strong>in the range of variation y ? ? y .<br />
P(y)<br />
1.0<br />
0<br />
(a) DPI = (b) 0
Equation (3.9) is used to measure the probability of <strong>design</strong>s being acceptable. The novelty in<br />
the model proposed by Chen and Yuan is in applying this concept to evaluate ranged solutions<br />
that are assumed to be picked from a range on a random basis.<br />
3.6.5 Constraint Feasibility Evaluation for Ranged Solutions<br />
When solving a constrained <strong>design</strong> problem where the <strong>design</strong> variables are expressed as ranges<br />
there are two option to evaluate feasibility: (1) to formulate the problem such that the constraint<br />
g ( x ) has to be rigidly satisfied, i.e., satisfied for the whole range of values of x, or (2) to<br />
formulate the constraint as to be satisfied <strong>with</strong>in a confidence interval using statistical analysis.<br />
For the former case, known as worst-case specification, the constraint is formulated as<br />
102<br />
g ( x,<br />
x ) b<br />
w<br />
? ?<br />
(3.10)<br />
where g ( x,<br />
x )<br />
w<br />
? represents the worst possible value of the constraint in the interval of<br />
solution, and b is the maximum value of the constraint function for the solution to be feasible. A<br />
constraint formulated as in Equation (3.10) is used to guarantee that the whole range of the<br />
<strong>design</strong> solution, x ? ? x , satisfies the constraint. If the worst value, g ( x,<br />
x ),<br />
w<br />
? is unknown, it<br />
can be approximated by two means: (1) evaluating the combinations of <strong>design</strong> solutions in the<br />
range <strong>based</strong> on Monte Carlo simulation and then choosing the worst case, or (2) estimating it<br />
from a first order Taylor series approximation as:<br />
? g<br />
? g ? ? ? x j<br />
(3.11)<br />
j ? x j
where ? g is the worst-case deviation of the constraint function:<br />
g w<br />
103<br />
? g ? ? g<br />
(3.12)<br />
Worst-case specification is almost always overly conservative. There are some<br />
conditions that must be treated as a worst-case, but generally it is unlikely that all variables will<br />
simultaneously occur in the worst possible combination. More reasonable, for such cases, is to<br />
consider variations in performance as random variables, and the low probability of a worst-case<br />
combination can be taken into account through statistical analysis. By allowing some fixed<br />
number of infeasible <strong>design</strong>s to occur, a <strong>design</strong>er can use larger ranges and/or back away from<br />
the “optimum” <strong>design</strong> some amount smaller than for a worst-case specification. In this case,<br />
assuming that all variables are independent and the fluctuations are small, the standard deviation<br />
of a constraint can be estimated as follows:<br />
?<br />
g<br />
?<br />
? ? g ?<br />
? ? ? x j ?<br />
j ? ? x j ?<br />
where ? g is the standard deviation of the constraint g.<br />
2<br />
(3.13)<br />
Because constraint functions are now described by <strong>probabilistic</strong> distributions, we can<br />
modify the conventional notion of constraint feasibility. If a constraint distribution straddles the<br />
right hand side, b in Figure 3.20, only a certain fraction of <strong>design</strong>s will be feasible, depending on<br />
the location of the distribution relative to b. Suppose that at the nominal value (no variation<br />
considered), the constraint function is equal to b, then the constraint is binding. Introducing<br />
variation, and if the induced constraint distribution is symmetric, for half of the possible <strong>design</strong>s
the constraint is violated, as shown in the right side of Figure 3.20. It is possible to control the<br />
number of infeasible <strong>design</strong>s by shifting the distribution of the constraint into the feasible region.<br />
How far the distribution is to be shifted depends on the particular distribution type, the values of<br />
distribution parameters (such as standard deviation) and which percentage of possible <strong>design</strong>s<br />
are allowed to be infeasible.<br />
y=b<br />
Infeasible <strong>design</strong>s<br />
<strong>with</strong> no shift<br />
y<br />
Shift<br />
Figure 3.20 Statistical Constraint Feasibility<br />
b<br />
Infeasible <strong>design</strong>s<br />
<strong>with</strong> shift<br />
Statistical specification of feasibility can be used to develop a method for achieving<br />
“feasibility robustness,” which refers to a feasibility that is characterized by a definable<br />
probability, set by a <strong>design</strong>er, to remain feasible relative to a nominal constraint, as it is<br />
subjected to variations in the variables (Parkinson, 1993). Feasibility robustness can be<br />
obtained by increasing the value of the constraint function during optimization by a suitable<br />
amount (for a constraint g ? b ). This has the effect of reducing the size of the feasible region<br />
<strong>with</strong> an accompanying degradation in the optimal value of the objective. In equation form,<br />
104
where ? g is the shift of the constraint.<br />
105<br />
g ? ? g ? b<br />
(3.14)<br />
For a worst-case analysis the shift is as given in Equation (3.11). For a statistical<br />
specification it is given as:<br />
? g ? k ? g<br />
(3.15)<br />
The value of k is selected to obtain a particular confidence of feasibility. For example, if a<br />
constraint is described by a normal distribution, then k=3 shifts the constraint distribution such<br />
that 99.8% of the possible <strong>design</strong>s are feasible for that constraint.<br />
The linear approximations, Equations (3.11) and (3.13), are accurate for normal<br />
distributions. When this assumption does not hold it is possible to extend the analysis by using<br />
second order Taylor series to approximate first to third moments of the distribution, i.e., the<br />
mean, the standard variation and the skewness, and then fitting a three-parameter gamma<br />
distribution to these three values (Lewis and Parkinson, 1998). This allows defining a whole<br />
family of distributions. When the skewness is zero, the gamma distribution closely matches a<br />
normal distribution. The analysis can be significantly more accurate when the functions involved<br />
are highly non-linear, but it can also become computationally expensive.<br />
3.6.6 A Probabilistic-Based Multiobjective Design Model<br />
So far, four basic notions that are used to obtain flexible <strong>design</strong> solutions have been described,<br />
namely:
(1) flexible preference functions to capture a varying degree of preference for a <strong>design</strong>,<br />
(2) evaluation of system behavior <strong>with</strong>in a range using probability density functions,<br />
(3) the combination of the former two concepts in an index, the DPI, to evaluate interval <strong>design</strong><br />
solutions, and<br />
(4) evaluation of feasibility for interval solutions.<br />
With these four elements it is now possible to formulate a compromise DSP for solving <strong>design</strong><br />
problems that involve interval <strong>design</strong> variables, i.e., problems whose solution is a set of ranged<br />
variables. This is the final block for constructing a mathematical framework for enterprise<br />
<strong>design</strong>, as proposed in Section 1.1.<br />
In general, the <strong>probabilistic</strong>-<strong>based</strong> <strong>design</strong> model, shown in Figure 3.21, is constructed by<br />
modifying the conventional compromise DSP (Section 3.4) to include <strong>probabilistic</strong><br />
representations of performance variations and the multiple goals on maximizing the <strong>design</strong><br />
metrics, DPIs, as close as possible to 1. The required changes are as follows (from Chen and<br />
Yuan, 1997):<br />
(1) In Section “Given” of the compromise DSP include, in addition to the given elements of a<br />
conventional compromise DSP, the following:<br />
?? Preference functions, Pi ( yi<br />
), for the ranged performance variables yi ? y ? ? y.<br />
106<br />
?? Probability density functions ( y ) of the functional relationship between <strong>design</strong><br />
fi i<br />
variables xj and performance, yi ? F(<br />
x j ) , obtained through probability analysis,<br />
Monte Carlo simulations, DOE or any other suitable technique.
107<br />
(2) In Section “Find,” substitute the <strong>design</strong> variables x j <strong>with</strong> the average value of the variable<br />
x j and the variation of each variable, ? xj, as <strong>design</strong> variables.<br />
(3) In Section “Satisfy”:<br />
(3.1) For the system constraints:<br />
??Include lower limiting values for the deviation of <strong>design</strong> variables, ? xi<br />
min , to ensure a<br />
minimum flexibility of the <strong>design</strong> solutions. Without this constraint, the range of a <strong>design</strong><br />
variable, ? xi, will otherwise tend to converge to zero.<br />
??Add a limiting value for the deviation of performance, ? yi<br />
max,<br />
as a constraint, in order to<br />
find “robust” <strong>design</strong> regions, i.e., ranged solutions for which the performance does not<br />
deviate more than a specified amount.<br />
??Formulate constraint feasibility, g k ( x j ), using either a worst-case specification (Equation<br />
3.10) or a statistical specification (Equation 3.15).<br />
(3.2) For the system goals:<br />
??Instead of the original form for defining goals in a conventional compromise DSP,<br />
? y ? d ? d ? G (where Gi is the value of the goal for variable yi), define the system<br />
i<br />
i<br />
?<br />
i<br />
i<br />
goals as to achieve a DPIi of 1 for each of the original objectives, as shown in Equation<br />
(3.12), where the DPIs are defined as in Equation (3.8).<br />
(3.3) For the bounds of the <strong>design</strong> variables:<br />
??Define bounds for both the average value of the <strong>design</strong> variables, x j , and their<br />
variation, ?<br />
x j .
The <strong>probabilistic</strong> <strong>design</strong> model, embodied in the compromise DSP shown in Figure 3.20, is a<br />
formal <strong>design</strong> model to achieve flexible, yet robust, <strong>design</strong> decisions which is domain-<br />
independent, thus appropriate for the coordination of decisions of the different enterprise <strong>design</strong><br />
processes. Particularly, the Design Preference index, DPI, provides a valid measure of value for<br />
evaluating imprecise <strong>design</strong>s, thus enabling rational comparison and choice between interval<br />
solutions, which is one of the fundamental requirements for <strong>design</strong>ing in a decision-<strong>based</strong> <strong>design</strong><br />
paradigm, as discussed in Section 1.3.<br />
Given<br />
System parameters:<br />
n number of <strong>design</strong> variables<br />
q number of system constraints<br />
m number of system goals<br />
gi (x) system constraint function<br />
F i(x) functional relationship between performance and <strong>design</strong> variables<br />
Pi (yi) preference functions of <strong>design</strong> requirements<br />
fi (yi) probability density functions of <strong>design</strong> performance<br />
Find<br />
System variables<br />
x i<br />
nominal value of <strong>design</strong> variables, i = 1, ..., n<br />
? xi range of <strong>design</strong> variables, i = 1, ..., n<br />
? ?<br />
d i , d i<br />
deviation variables, i = 1, ..., m<br />
Satisfy<br />
System constraints<br />
? xi ? ? ximin i = 1, ..., n<br />
? yi ? ? yimax i = 1, ..., m<br />
g ( x , x ) b<br />
w<br />
? ? (for worst-case) i = 1, ..., q (Eq. 3.10)<br />
g ? k ? g ? b (for statistical feasibility) i = 1, ..., q (Eq. 3.15)<br />
System goals<br />
? ?<br />
DPIi ? d i ? di<br />
= 1 i = 1, ..., m (Eq. 3.9)<br />
Bounds<br />
x L ? x ? x U i = 1, ..., n<br />
? xi L ? ? xi ? ? xi U i = 1, ..., n<br />
? ?<br />
d i , d i ? 0 i = 1, ..., m<br />
108
? ?<br />
d i ? di<br />
= 0<br />
Minimize<br />
i = 1, ..., m<br />
Deviation function in preemptive or archimedean formulation<br />
? ?<br />
Z = Z( d i , di<br />
) i = 1, …, m<br />
Figure 3.21 The Probabilistic-Based Multiobjective Design Model<br />
(Chen and Yuan, 1997)<br />
109
When early <strong>design</strong> decisions are expressed as ranges, <strong>design</strong>ers have more freedom in<br />
adapting their decisions to satisfy the needs and goals of other agents of the organization, which<br />
is the paramount objective of any coordination mechanism. This is the final block of the<br />
mathematical framework for enterprise <strong>design</strong> proposed in Section 1.1. In Section 3.7, all the<br />
elements that comprise the framework are tied together into a consistent structure for supporting<br />
coordination of decisions in enterprise <strong>design</strong>.<br />
3.7 TYING IT ALL TOGETHER: A MATHEMATICAL FRAMEWORK<br />
FOR ENTERPRISE DESIGN<br />
Up to this point, the four basic elements that comprise a mathematical framework, which is<br />
shown in Figure 3.3, for integration of the enterprise <strong>design</strong> processes have been presented.<br />
The fundamental tenet that bears this framework, rooted in a decision-<strong>based</strong> perspective, is that<br />
the integration of the enterprise achieved through the appropriate coordination of decisions is<br />
generally more efficacious than integrating those decisions into joint formulations. These four<br />
basic elements are, as mentioned in Section 3.1:<br />
?? A representation of the enterprise activity as an extensive <strong>game</strong>.<br />
?? A common paradigm for decision support.<br />
?? A multi-objective decision model, the Compromise DSP.<br />
?? A <strong>probabilistic</strong> <strong>design</strong> model to attain flexible and robust decisions and handle uncertainty<br />
and imprecision.<br />
110
The merging of these elements provides a consistent mathematical framework for coordination<br />
of decisions <strong>based</strong> on <strong>design</strong> formulations. The partitioning of elements of a decision into <strong>design</strong><br />
variables, models, requirements (goals and constraints) and activities (modeling, analysis and<br />
synthesis) of the unified decision support model facilitates the mathematical formulation of<br />
decisions as compromise DSPs. Game theory provides the mathematics to appropriately<br />
formulate those compromise DSPs in seeking specific kinds of equilibrium. Finally, these<br />
compromise DSPs can be formulated to achieve flexible solutions and to <strong>design</strong> <strong>with</strong> uncertain<br />
and incomplete information, using a <strong>probabilistic</strong> <strong>design</strong> model, thus facilitating the coordination<br />
of decisions and, therefore, the integration of the enterprise <strong>design</strong> processes.<br />
111<br />
Observe that each of these elements is “independent” of the others, including the<br />
enterprise <strong>design</strong> method. That is, each of them holds by itself as a tool/method for decision<br />
support. However, it is their merging that creates a practical and comprehensive framework for<br />
enterprise integration that is anchored on a solid mathematical foundation.<br />
Another characteristic of this framework is its “openness.” It does not exclude any tool or<br />
method that contributes to an improvement of individual or group decision-making. The better<br />
the individual decision-making, the easier is to find points of equilibrium. Thus, this framework<br />
contributes to the integration of the enterprise activities, while empowering the agents of an<br />
organization by decentralizing decisions and encouraging a better individual decision making.<br />
Finally, the systematic and mathematical nature of this framework fosters its<br />
implementation on a computer-supported environment. This framework is exercised in Chapter
4 in a case study, namely, the integrated product and process <strong>design</strong> of an absorber/evaporator<br />
module for a family of absorption chillers.<br />
112
CHAPTER 4<br />
CASE STUDY: INTEGRATED PRODUCT AND PROCESS DESIGN OF AN<br />
ABSORBER/EVAPORATOR MODULE FOR A FAMILY OF ABSORPTION<br />
CHILLERS<br />
In this chapter the framework for enterprise <strong>design</strong> constructed in Chapters 2 and 3 is<br />
exercised in the preliminary <strong>design</strong> of an absorber/evaporator module for a family of absorption<br />
chillers, which are produced in a make-to-order system.<br />
The <strong>approach</strong> is demonstrated step by step for this case study, in which the <strong>design</strong><br />
process is abstracted as an extensive <strong>game</strong> that involves nine players: eight product <strong>design</strong>ers<br />
who want to minimize cost of material and one production system <strong>design</strong>er who wants to<br />
minimize manufacturing cost and production time. Thus, the benefits of the proposed enterprise<br />
<strong>design</strong> framework for the coordination of enterprise <strong>design</strong> decisions are studied in this case<br />
study, while considering the issue of <strong>design</strong> strategies for family <strong>design</strong> accounting for production<br />
process.<br />
111
NOMENCLATURE FOR CASE STUDY<br />
A Absorber tube selection<br />
BC Burden (overhead) cost<br />
Ca<br />
Ce<br />
COST Total cost<br />
Cost of absorber tube<br />
Cost of evaporator tube<br />
CT Cycle time (total production time)<br />
D Expected demand per year<br />
DPI Design Preference Index<br />
E Evaporator tube selection<br />
F Replenishment frequency<br />
FALP Foraging-Adaptive Linear Programming algorithm<br />
GLOBAL Global optimization algorithm <strong>based</strong> on clustering method<br />
IC Inventory cost<br />
ir Annual interest rate<br />
k Proportional factor in burden cost estimation<br />
LC Labor cost<br />
LFC Leader/follower solution using a conventional <strong>design</strong> <strong>approach</strong><br />
LFP Leader/follower solution using a <strong>probabilistic</strong> <strong>design</strong> <strong>approach</strong><br />
LT Normalized tube length<br />
112
MC Material cost<br />
MfgC Production cost<br />
MSi Capacity at station Si<br />
Na<br />
Ne<br />
Number of tubes of the absorber<br />
Number of tubes of the evaporator<br />
NAT Normalized number of tubes of the absorber<br />
NET Normalized number of tubes of the evaporator<br />
n(r) expected number of backorders during a cycle<br />
Ops Number of operators<br />
PC Product cost<br />
Q Replenishment quantity<br />
r Reorder point<br />
SA Safety stock of absorber tubes<br />
SAi Subassembly station i<br />
SAn Simmulated Annealing Algorithm<br />
SE Safety stock of evaporator tubes<br />
Si Station i<br />
SCV Squared coefficient of variation<br />
TH Throughput<br />
UA Heat transfer coefficient multiplied by the area of heat exchange<br />
113
? P Pressure drop?<br />
?? Mean processing time<br />
?? Expected demand during lead time?<br />
? Standard deviation<br />
Subindices<br />
Abs Absorber<br />
AP Aggregated product<br />
Evap Evaporator<br />
T Individual absorber/evaporator <strong>design</strong><br />
114
4.1 INTRODUCTION TO CASE STUDY<br />
In Chapters 2 and 3 is presented a framework for enterprise <strong>design</strong> that is <strong>based</strong> on a <strong>game</strong><br />
<strong>theoretical</strong> representation of the enterprise <strong>design</strong> process. Following the implementation<br />
strategy presented in Section 1.6, the next step is to exercise this framework in a case example,<br />
namely the <strong>design</strong> of an absorber/evaporator module for a family of absorption chillers, which<br />
are produced in a make-to-order system.<br />
4.1.1 Motivation for Case Study: Benefits of an Integrated Product and<br />
Process Design in Make-to-Order Systems<br />
115<br />
Make-to-order systems are “systems in which jobs, both the physical material and the job<br />
order, arrive from outside the shop, and the process that generates the arrivals is independent of<br />
whatever is happening in the shop, and its associated input and output stores” (Buzacott and<br />
Shanthikumar, 1993). This kind of systems is commonly found in the production of complex<br />
equipment <strong>with</strong> low demand, such as ships, aircrafts, turbines, etc. These systems, particularly<br />
those that offer product variety, face a large amount of variability in manufacturing processes<br />
time, supply deliveries and demand, which in turn cause long lead-times and large work-in-<br />
process. How can product <strong>design</strong> improve the performance of make-to-order systems?<br />
A hierarchy of objectives that links the fundamental objective of a manufacturing<br />
organization to various supporting subordinate objectives according to Hopp and Spearman<br />
(1996) is illustrated in Figure 4.1. Clearly, there are conflicts. High inventory is desirable for
fast response, but low inventory is attractive for keeping total assets low to return on investment<br />
be high. High utilization is needed to keep unit costs down, and low utilization is needed for<br />
good responsiveness. More variability is needed for greater product variety, but less variability<br />
is needed to keep inventory low and throughput high. Ideally, these objectives should be<br />
analyzed concurrently in order to obtain an appropriate compromise between them, particularly<br />
between greater product customization and shorter production times, which are both critical in<br />
customers’ preference for this kind of products. The motivation for this case study is, therefore,<br />
to apply the framework for enterprise <strong>design</strong> and integration constructed in Chapters 2 and 3 to<br />
explore its benefit in the integrated product and process <strong>design</strong> of a make-to-order product.<br />
High<br />
Throughput<br />
Less<br />
Variability<br />
Low<br />
Costs<br />
Low Unit<br />
Costs<br />
High<br />
Utilization<br />
Low<br />
Inventory<br />
High<br />
Profitability<br />
Quality<br />
Product<br />
Short<br />
Cycle Time<br />
High<br />
Sales<br />
Fast<br />
Response<br />
Low<br />
Utilization<br />
High Customer<br />
Service<br />
High<br />
Inventory<br />
Many<br />
Products<br />
Figure 4.1 Hierarchical Objectives in a Manufacturing<br />
Organization (Hopp and Spearman, 1996)<br />
More<br />
Variability<br />
116
Before addressing the case study, the strategy followed in this work to model and analyze<br />
a production system for an integrated product and process <strong>design</strong> is presented in Section 4.1.2.<br />
4.1.2 A Methodology for Modeling Production Systems for Integrated<br />
Product and Process Design<br />
In order to assess the effect of product changes in the production system, it is necessary<br />
to have some model that relates such changes <strong>with</strong> the performance of the production line. In<br />
simulation-<strong>based</strong> manufacturing models, the <strong>probabilistic</strong> nature of many events, such as product<br />
rework or arrival variability, is represented by sampling from a distribution representing the<br />
pattern of occurrence of the event. Thus, to represent the typical behavior of a system it is<br />
necessary to run the simulation model for a sufficiently long time, so that all events can occur a<br />
reasonable number of times. Direct coupling of product and manufacturing processes models<br />
(such as FEM models) <strong>with</strong> a discrete event simulation of the production system is<br />
computationally too expensive and time consuming to support product <strong>design</strong> decisions.<br />
117<br />
To overcome this problem, Peplinski, et al., (1997) propose creating a “bank” of<br />
response surfaces, one for each machine or processing center. These response surfaces<br />
capture the relationship between a reduced number of local input variables and a small set of<br />
<strong>design</strong> requirements such as flowtime, scrap rate, etc. These response surfaces could then be<br />
“assembled” to represent the larger manufacturing system. They suggest two types of<br />
constraints to assemble the response surfaces: (1) the exit flow from each upstream machine
must match the arrival flow of each downstream, and (2) input flows to a processing center must<br />
not exceed its capacity. It is easy to see that a set of small response surface modules should<br />
increase a <strong>design</strong>er’s flexibility in quickly modeling different shop floor configurations.<br />
However, this <strong>approach</strong> fails because, in general, the variability of arrivals, and therefore of<br />
departures, from a station is not independent of other stations of the system.<br />
In this thesis a hybrid method is proposed, which combines the notion of a bank of<br />
response surfaces, as proposed by Peplinski and co-authors, <strong>with</strong> the parametric<br />
decomposition <strong>approach</strong> (described in Appendix A) to build models of manufacturing systems<br />
as response surface networks, which do not have the aforementioned limitation.<br />
118<br />
In the parametric decomposition <strong>approach</strong>, each station is characterized by two<br />
parameters: the mean (?? and the squared coefficient of variation (SCV) of the processing times.<br />
Therefore, only two parameters for each station are needed. Thus, it is not necessary to fit a<br />
response surface for the flowtime, yield, waiting time, machine utilization, etc. of each station as<br />
proposed by Peplinski and co-authors, and it is possible to capture interaction between stations.<br />
The proposed method of building models of the manufacturing system as response surface<br />
networks consists of four steps, as illustrated in Figure 4.2:<br />
1. Identify relevant factors and ranges for each process (machining, forging,<br />
welding, assembly, etc).
2. Design appropriate experiments for fitting the mean processing times and its<br />
SCV for each process as functions of the factors identified in Step 1, using the<br />
appropriate modeling technique.<br />
3. Fit the response surfaces.<br />
4. Assemble the response surfaces in a network using the parametric<br />
decomposition <strong>approach</strong>.<br />
1. Identify Factors and Ranges<br />
x<br />
Control<br />
Factors<br />
Noise<br />
Factors<br />
z Service<br />
Mean Time<br />
?<br />
Process<br />
y<br />
x2<br />
x1<br />
Welding<br />
y<br />
2 c<br />
SCV<br />
x2<br />
x1<br />
Cutting<br />
y<br />
2. Design Experiments<br />
The Parametric<br />
Decomposition Approach<br />
4. Assemble the Response Surfaces in a Network<br />
x2<br />
x2<br />
x1<br />
x1<br />
Machining Assembly<br />
MANUFACTURING SYSTEM<br />
MODEL<br />
y<br />
Models of Individual<br />
Processes<br />
3. Build Response Surface<br />
Models of Individual<br />
Processes<br />
Figure 4.2 Modeling the Production System as a Network of Response Surfaces<br />
y<br />
x2<br />
x1<br />
119
In Step 1, the initial exploration space is defined, i.e., the control and noise factors which<br />
mostly influence the mean process times and the SCVs are identified as well as their range of<br />
value. In the second and third steps, response surfaces of the mean processing times and SCVs<br />
as function of the control and noise variables are built and validated. Finally, in the fourth step,<br />
the different stations are synthesized in a network using the response surfaces obtained in the<br />
previous steps. It should be apparent that once a network of response surfaces is built, it is<br />
possible to quickly observe the effect of any change in a product or process in the system<br />
performance, thus allowing an efficient and systematic product-process exploration.<br />
Furthermore, any change in the network is straightforward and requires a minimum remodeling<br />
effort, which is not the case <strong>with</strong> simulation-<strong>based</strong> models. In the following section this method<br />
is applied in solving the case study.<br />
4.2 DESCRIPTION OF THE DESIGN PROBLEM AND ANCILLARY<br />
INFORMATION<br />
4.2.1 Absorption Refrigeration<br />
Today there are several manufacturers of large absorption machines, which are used when the<br />
operating and maintenance costs of electrically driven refrigeration equipment are not<br />
economical or are restricted due to environmental regulations. This product is a typical example<br />
of expensive and highly customized products <strong>with</strong> low production volumes which obliges<br />
120<br />
manufacturers to produce them in make-to-order systems. Manufacturers of absorption
equipment usually face large demand fluctuations, hard-to-forecast market status, complex<br />
supply logistics, large work-in-process levels, and considerable lead-times. This case study<br />
illustrates how the proposed enterprise <strong>design</strong> framework can help to improve the performance<br />
of such systems via an integrated product and process <strong>design</strong>.<br />
121<br />
Absorption chillers consists of four basic components: (1) the evaporator, (2) the<br />
absorber, (3) the generator, and (4) the condenser; as shown in the absorption cycle illustrated<br />
in Figure 4.3. Usually, the absorbent in conventional absorption cycles is Lithium Bromide<br />
(LiBr), a salt that has the ability to absorb water (the refrigerant).<br />
Heat<br />
Exchanger<br />
GENERATOR CONDENSER<br />
Strong Solution<br />
Weak Solution<br />
Steam<br />
ABSORBER EVAPORATOR<br />
Condensing<br />
Water<br />
Refrigerant Vapor<br />
Refrigerant Vapor<br />
Refrigerant<br />
Refrigerant<br />
Figure 4.3 Absorption Refrigeration Cycle<br />
Condensing<br />
Water<br />
Cooling<br />
load
A brief description of these components follows:<br />
GENERATOR. The generator is the equivalent of the compressor in a conventional<br />
compression refrigeration cycle. It is a vessel where heat is applied which causes the mixture of<br />
refrigerant and absorbent to boil and provides the motive force for the process. The refrigerant<br />
vapor (steam) and absorbent droplets (LiBr and water) leave the generator and enter a<br />
condenser.<br />
CONDENSER. The vapors pass from the separator to the condenser where cooling<br />
water from a cooling tower circulates through the condenser removing the heat added at the<br />
generator. The low temperature steam vapor is condensed into a liquid refrigerant where it<br />
passes through an orifice into the evaporator which operates under a vacuum. This condenser<br />
performs the same function as the condenser of a conventional compression refrigeration<br />
system.<br />
122<br />
EVAPORATOR. The evaporator serves the same function as in a conventional<br />
compression refrigeration system inasmuch as it absorbs the heat from the process load. The<br />
expansion of the refrigerant (water) <strong>with</strong>in a vacuum causes boiling and therefore absorbs heat<br />
from the chilled water coils located <strong>with</strong>in the evaporator. This chilled water is used to satisfy<br />
the process load.<br />
ABSORBER. The refrigerant (water) vapor will not condense under the vacuum<br />
required for evaporation. Some means of converting the vapor back into a liquid must take
place before the refrigerant is returned back to the generator to repeat the process. This is<br />
where the absorption cycle differs from the conventional compression refrigeration cycle. The<br />
LiBr solution, which was concentrated <strong>with</strong>in the low temperature generator when the<br />
refrigerant (water) was boiled off, passes from the low temperature generator through a heat<br />
exchanger to the absorber. Cooling coils <strong>with</strong>in the absorber absorb the heat load from the<br />
evaporator and, unavoidably, some of the residual heat from the concentrated solution, to be<br />
dissipated through the cooling tower. The concentrated LiBr solution absorbs the refrigerant<br />
vapor from the evaporator. This LiBr/water solution is now diluted and is pumped through the<br />
heat exchanger, to recover heat from the concentrated solution, before returning to the<br />
generator to repeat the process. When the refrigerant vapor is absorbed, a vacuum is created<br />
which allows expansion to take place between the condenser and the evaporator.<br />
123<br />
Usually, the absorber and evaporator are combined in a single shell, the<br />
absorber/evaporator module, illustrated in Figure 4.4, which is the focus of this case example.
Figure 4.4 Layout of Absorption Chillers<br />
(Carrier Mexico)<br />
4.2.2 Objective and Ancillary Information 1<br />
A manufacturer of absorption refrigeration equipment offers a family of chillers <strong>with</strong> the<br />
following refrigeration capacities: 600, 700, 800, 900, 1000, 1100, 1200 and 1300 tons. A<br />
market study has estimated the demand for the next years as approximately 80 units/year, <strong>with</strong> a<br />
distribution of such demand as shown in Table 4.1 and Figure 4.5.<br />
Table 4.1 - Estimated Distribution of Demand<br />
Tons, T 600 700 800 900 1000 1100 1200 1300<br />
%<br />
demand<br />
DT<br />
0.05<br />
0.25<br />
0.2<br />
0.15<br />
0.1<br />
0.05<br />
0<br />
0.15<br />
0.16<br />
0.05<br />
0.15<br />
0.12<br />
600 700 800 900 1000 1100 1200 1300<br />
Capacity (tons of refrigeration)<br />
1 The actual information of the product <strong>design</strong> and the production system has been disguised to protect the<br />
competitive position of the manufacturer.<br />
0.25<br />
124<br />
0.07
Figure 4.5 Estimated Distribution of Demand<br />
Prior to this work, the absorber-evaporator module has been uniquely <strong>design</strong>ed for each<br />
capacity to minimize the material cost (the traditional <strong>approach</strong> in <strong>design</strong>ing heat exchangers).<br />
This in turn has created a large variety of components. Because of this variety, low production<br />
volumes and large fluctuations in demand, it is not economical at the present to protect the line<br />
against disruptions, delays and randomness by carrying safety stock. It also creates difficulties<br />
for inventory management, material handling, storage and control of the floor shop. In this<br />
example, the concern is <strong>with</strong> the preliminary <strong>design</strong> of the absorber/evaporator module (to<br />
determine its length and number and type of tubes) for the family of chillers using the<br />
mathematical framework established in previous chapters and applying the <strong>design</strong> principles of<br />
standardization and robustness.<br />
Consider a production line of the absorber/evaporator module as shown in Figure 4.6.<br />
125
SA1<br />
SA2<br />
SA3<br />
SUB-ASSEMBLIES<br />
STATIONS<br />
Absorber and<br />
Evaporator Tubes<br />
S1 S2<br />
S3<br />
STATION 1 STATION 2 STATION 3<br />
Figure 4.6 Production Line of the Absorber/Evaporator Module<br />
126<br />
DEPART<br />
There are 3 main subassemblies produced in the subassembly stations SA1, SA2, and<br />
SA3, each <strong>with</strong> one server. These subassemblies are assembled in Station S1. After this<br />
station, the assemblies go to a second station, S2, and finally to Station S3, where the<br />
evaporator tubes and the absorber tubes are assembled. There are two working shifts of 8<br />
hours each and the system operates only during weekdays.<br />
The manufacturer wants to reduce the variety of tubes to one combination of tube type<br />
and length for the absorber and one for the evaporator (to reduce tube variety from 16 to 2).<br />
This will allow carrying safety stock to avoid disruptions at Station S3, and will facilitate material<br />
handling and storage, thus fostering a better control of the shop floor, shorter lead times, lower<br />
costs and better customer service. Simply stated, the objective is to determine one<br />
combination of length and tubes selection (material, diameter, thickness, number of fins<br />
per inch, etc.) for the absorber/evaporator modules.
The <strong>design</strong> of the absorber/evaporator module is carried out <strong>with</strong> four <strong>approach</strong>es. The<br />
objective is to show the application of the framework for enterprise <strong>design</strong> described in Chapter<br />
3 and test the hypothesis posed in Chapter 1. In Section 4.3 is described and implemented an<br />
<strong>approach</strong> in which each of the agents involved in the <strong>design</strong> process acts in isolation (which is<br />
called Baseline Solution). These agents are 8 product <strong>design</strong>ers, each responsible of the <strong>design</strong><br />
of an absorber/evaporator module for a particular refrigeration capacity, and one production<br />
system <strong>design</strong>er who is responsible of <strong>design</strong>ing a production system that can meet the expected<br />
demand <strong>with</strong> minima production cost and cycle time. The information and formulation of these<br />
baseline decisions are the input of the enterprise <strong>design</strong> method.<br />
Once the baseline decisions are formulated, the expansion of decisions into a <strong>game</strong>, in<br />
order to coordinate decisions between product <strong>design</strong> and production system <strong>design</strong>, is<br />
carried out <strong>with</strong> three different <strong>approach</strong>es, described in Sections 4.4 to 4.6. This plan of<br />
action is illustrated in Figure 4.7.<br />
127
Baseline Decisions<br />
Product Design<br />
Section 4.3<br />
Production System<br />
Design<br />
Reformulate Decisions to<br />
Integrate Product Design and<br />
Production System Design<br />
Using a Cooperative<br />
Game Model<br />
Section 4.4<br />
Section 4.5<br />
Using a Non-cooperative<br />
(Stackelberg) Game Model<br />
Reformulation of<br />
Design Model<br />
as a Probabilistic<br />
Design Model<br />
Section 4.6<br />
Figure 4.7 Roadmap to Case Study<br />
190000<br />
187500<br />
185000<br />
182500<br />
180000<br />
177500<br />
175000<br />
Comparison of<br />
Solutions<br />
Baseline Coop Conv-NC Prob-NC<br />
Section 4.7<br />
In Section 4.4, it is studied a solution (called Cooperative Solution) obtained <strong>with</strong> a<br />
cooperative protocol between the 9 agents. In Sections 4.5 is adopted a cooperative protocol<br />
between the 8 product <strong>design</strong>ers, and in Section 4.6 is adopted a non-cooperative protocol<br />
between them as a group and the production system <strong>design</strong>er. This last case is formulated using<br />
a “conventional” <strong>design</strong> <strong>approach</strong> in Section 4.5 (called Conventional Non-Cooperative<br />
Solution), and a <strong>probabilistic</strong> <strong>design</strong> <strong>approach</strong> in Section 4.6 (called Probabilistic Non-<br />
Cooperative Solution). Here “conventional” <strong>approach</strong> refers to a non-<strong>probabilistic</strong> one.<br />
4.3 BASELINE FORMULATION AND SOLUTION<br />
128
The objective of this section is threefold:<br />
1. To present information regarding the <strong>design</strong> problem that is used in Sections 4.4 to 4.6.<br />
2. To obtain a <strong>design</strong> solution for comparison <strong>with</strong> those solutions obtained by applying the<br />
enterprise <strong>design</strong> framework developed in Chapters 2 and 3, and<br />
3. To verify the accuracy of the manufacturing systems model and the effectiveness of the<br />
optimization algorithm that is used in obtaining subsequent solutions.<br />
129<br />
Assume that there are 8 <strong>design</strong>ers, each responsible of achieving a <strong>design</strong> of an<br />
absorber/evaporator module for a particular refrigeration capacity that minimizes the material<br />
cost. There is also a production system <strong>design</strong>er who, after the <strong>design</strong> of the family of products<br />
is obtained by the group of product <strong>design</strong>ers, has to <strong>design</strong> a production system that can meet<br />
the expected demand of the absorber/evaporator modules, while seeking minima production<br />
cost and cycle time. This situation is illustrated in Figure 4.8. In this section it is described a<br />
solution achieved when each of these <strong>design</strong>ers acts in isolation.
Goal<br />
Minimize Material Cost<br />
Product Designers<br />
600 700 800 900 1000 1100 1200 1300<br />
Capacities<br />
DESIGN VARIABLES<br />
DESIGN VARIABLES<br />
Manufacturing Process<br />
Designer<br />
Length, Number of Tubes, Tube Selection<br />
for each Capacity<br />
Goal<br />
Minimize Production Cost<br />
Minimize Production Time<br />
Capacity of Stations S1, S2 and S3<br />
Tubes Replenishment Frequency<br />
Figure 4.8 Agents Involved in the Design Process<br />
Nine baseline decisions are identified, one for each agent. In Section 4.3.1 is described<br />
the formulation of decisions for the product <strong>design</strong>ers, and in Section 4.3.2 for the production<br />
system <strong>design</strong>er. In Section 4.3.3 the solution obtained is presented, and in Section 4.3.4<br />
validation and verification issues are studied.<br />
130
4.3.1 Formulation of Decision for Product Design<br />
The variables that a product <strong>design</strong>er controls are the value of the product <strong>design</strong> variables for a<br />
particular capacity, i.e., the length of the tubes (LT), number of evaporator tubes (NeT), number<br />
of absorber tubes (NaT), and the selection of tubes for the evaporator and the absorber, ET and<br />
AT, where the subscript T is used to refer to individual products/<strong>design</strong>ers (T=1,...,8). Each<br />
product has to be carried out for the following standard specifications:<br />
Entering temperature of chilled water 54 o F<br />
Leaving temperature of chilled water 44 o F<br />
Flow rate of chilled water 2.4 GPM/ton<br />
Refrigerant fluid Fresh Water<br />
Saturation temperature of refrigerant 41 o F<br />
Flow rate of water in the absorber 4.5 GPM/ton<br />
Entering (absorber) water temperature 85 o F<br />
Solution fluid LiBr/H2O<br />
Entering solution concentration 62%<br />
Entering solution temperature 110 o F<br />
Entering solution flow rate 222.5 lb / (hr ton)<br />
Fouling factor for tube walls 0.00025 hrft 2 o F/BTU<br />
Fouling factor for shell walls 0.00025 hrft 2 o F/BTU<br />
131
4.2.<br />
The commercial alternatives for selection of tubes for the evaporator are shown in Table<br />
Table 4.2 Alternatives for Selection of the Evaporator Tubes<br />
Evaporator Tube E 1 2 3 4<br />
Outer Diameter (in) 0.625 0.75 0.875 1.0<br />
Wall Thickness (in) 0.025 0.025 0.025 0.025<br />
Fins/in 30 30 30 30<br />
Fin Height (in) 0.05 0.05 0.05 0.05<br />
Fin Thickness (in) 0.02 0.02 0.02 0.02<br />
Material copper copper copper copper<br />
Cost C e (dollars/in) 6.00 6.50 7.0 7.5<br />
Similarly, the options for selection of tubes for the absorber are shown in Table 4.3.<br />
Table 4.3 Alternatives for Selection of the Absorber Tubes<br />
Absorber Tube A 1 2 3 4<br />
Outer Diameter (in) 0.625 0.75 0.875 1.0<br />
Wall Thickness (in) 0.03 0.03 0.03 0.03<br />
Material copper copper copper copper<br />
Cost C a (dollars/in) 3.25 3.50 3.75 4.0<br />
The classification of parameters that define the synthesis of the products is shown in Table 4.4.<br />
The only goal is to achieve some target of material cost and the constraints include a minimum<br />
UA (the overall heat transfer coefficient U multiplied by the area of heat exchange A) in both the<br />
absorber and the evaporator, and a maximum pressure drop of the chilled water. These<br />
product cost goals and constraints values for each refrigeration capacity are shown in Tables<br />
132
4.5 and 4.6. The cost of raw material (<strong>with</strong>out considering tubes) for each station is given in<br />
Table 4.7.<br />
Table 4.4 Parameters for Absorber-Evaporator Module Synthesis<br />
Parameter Name Symbol Category Range Units<br />
1 Tube length L <strong>design</strong> variable 17.0 - 24.0 ft<br />
2 Evaporator tube<br />
selection<br />
E <strong>design</strong> variable 1 - 4<br />
3 Absorber tube<br />
selection<br />
A <strong>design</strong> variable 1 - 4<br />
4 Number of tubes of<br />
the evaporator<br />
Ne <strong>design</strong> variable 440 - 900<br />
5 Number of tubes of<br />
the absorber<br />
Na <strong>design</strong> variable 440 - 900<br />
6 Material cost MC goal dollars<br />
Table 4.5 Material Cost Goal for each Refrigeration Capacity<br />
Refrigeration Capacity Material Cost Goal<br />
T<br />
MCT,goal (dollars)<br />
600 tons 99000<br />
700 tons 112500<br />
800 tons 126000<br />
900 tons 139500<br />
1000 tons 153000<br />
1100 tons 166500<br />
1200 tons 180000<br />
1300 tons 193500<br />
133
Table 4.6 Constraint Values for Product Design<br />
Capacity min UA Evaporator<br />
(BTU/ hr o F)<br />
min UA Absorber<br />
(BTU/ hr o F)<br />
Max Pressure Drop ? P<br />
(ft of water)<br />
600 tons 771100 575035 23<br />
700 tons 905300 677235 23<br />
800 tons 1039500 779435 23<br />
900 tons 1173700 881635 23<br />
1000 tons 1307900 983835 23<br />
1200 tons 1576300 1188235 23<br />
1400 tons 1844708 1392635 23<br />
1600 tons 2113108 1597035 23<br />
Table 4.7 Cost of Raw Material at Each Subassembly Station (in Dollars)<br />
Capacity<br />
T<br />
MCSA1 MCSA2 MCSA3 600 20000 15000 4000<br />
700 22000 16000 4500<br />
800 24000 17000 5000<br />
900 26000 18000 5500<br />
1000 28000 19000 6000<br />
1100 30000 20000 6500<br />
1200 32000 21000 7000<br />
1300 34000 22000 7500<br />
Given length, tubes selection and number of tubes for each capacity, the cost of material<br />
(MCT) is calculated as:<br />
134<br />
MC T ? L T( N eTC e ? N aTC a) ? MC T, SA1 ? MC T,SA2 ? MC T, SA3 (4.1)<br />
where Ce and Ca are the cost of the evaporator and absorber tubes (Tables 4.2 and 4.3). For<br />
each capacity, the goal for product <strong>design</strong> is formulated as follows:<br />
MC T<br />
MC T ,goal<br />
? d ? ? d ? ? 1 (4.2)
135<br />
? ?<br />
where MCT,goal are the target values for cost of material given in Table 4.5, and d and d are<br />
deviation variables used in a compromise DSP.<br />
This information (<strong>design</strong> variables, bounds, goals and constraints) embody the baseline<br />
decision for each product <strong>design</strong>er. A compromise DSP corresponding to these baseline<br />
decisions is shown in Figure 4.9. This compromise DSP has to be solved by each product<br />
<strong>design</strong>er for a specific refrigeration capacity.
Given<br />
?? A refrigeration capacity<br />
?? Heat-transfer models of the absorber-evaporator<br />
?? Commercial tube alternatives (Table 4.2).<br />
Find<br />
?? Length LT of tubes<br />
?? Selection of tube for the evaporator ET<br />
?? Selection of tube for the absorber AT<br />
?? Number of tubes of the evaporator, NeT<br />
?? Number of tubes of the absorber NaT<br />
Satisfy<br />
System Constraints<br />
?? UAEVAP ? UAEVAP min (Table 4.5)<br />
?? UAABS ? UAABS min (Table 4.5)<br />
?? ? P ? ? Pmax (Table 4.5)<br />
System Goals for each capacity (minimize material cost):<br />
??<br />
MC T<br />
MC T ,goal<br />
? d ? ? d ? ? 1 (Equation 4.2)<br />
Bounds for each capacity<br />
?? Na min ? Na ? Na max (Table 4.3)<br />
?? Ne min ? Ne ? Ne max<br />
?? Lmin ? L ? Lmax<br />
?? d ? , d ? ? 0<br />
d ? ?d ? ? 0<br />
Minimize<br />
The deviation function<br />
?? Z ? d ?<br />
Figure 4.9 Baseline Compromise DSP for Product Designers<br />
136
4.3.2 Formulation of Decision for Production System Design<br />
The baseline decision of the production system <strong>design</strong>er is that of specifying the required<br />
number of servers at station S1, S2 and S3 of the production line, and to set a safety stock level<br />
of absorber tubes and evaporator tubes. The performance measurement is <strong>based</strong> on the<br />
expected cycle time and production cost.<br />
The <strong>approach</strong> described in Section 4.2.2 for modeling production systems is now applied<br />
to the production line of the absorber/evaporator module. The steps are as follows:<br />
1. Identify relevant factors that affect processing times and SCVs of the various<br />
stations. In this case the product <strong>design</strong> variables that mainly affect the processing times<br />
are the length and number of tubes of both the absorber and the evaporator. The SCVs are<br />
all considered exponentially distributed, which gives a SCV of one for all the stations<br />
regardless of the value of the <strong>design</strong> variables.<br />
2. Design experiment. A full factorial CCD <strong>design</strong> is selected for fitting second order<br />
response surfaces to the processing times at stations as functions of the product <strong>design</strong><br />
variables LT, Ne,T and Na,T.<br />
3. Fit response surfaces of the mean processing time (???of the various stations. The<br />
statistical analysis software JMP® (SAS, 1995) is used to this aim. The resulting response<br />
surfaces for the mean processing times ? of each station are:<br />
137<br />
? SA1 ? 24.159 ? 0.063 LT ? 0.563 NET ? 0.688 NAT ? 0.045 LT 2 ? 0.125 NET ?LT ? 0.080 NET 2<br />
? 0.125 NAT ?LT ? 0.125 NAT ?NET ? 0.205 NAT 2 (4.3)
138<br />
? SA2 ? 29 .295 ? 0.312 LT ? 2.062 NET ? 1.397 NAT ? 0.023 LT 2 ? 0.125 NET ?LT ? 0.102 NET 2<br />
? 0.125 NAT ?LT ? 0.125 NAT ?NET ? 0.227 NAT 2 (4.4)<br />
? SA3 ? 8.636 ? 1.25 NET ? 1.25 NAT ? 0.056 NET 2 ? 0.25 NAT ?LT ? 0.25NAT ?NET ? 0.068NAT 2 (4.5)<br />
? S1 ? 66.04 ? 1.312 LT ? 3.801 NET ? 4.062 NAT ? 0.147 LT 2 ? 0.125 NET ?LT ? 0.147 NET 2<br />
? 0.125 NAT ?LT ? 0.875 NAT ?NET ? 0.397 NAT 2 (4.6)<br />
? S2 ? 63.272 ? 1.05LT ? 4.375NET ? 4.875NAT ? 0.363 LT 2 ? 0.011NET 2 ? 0.378 NAT 2 ??????<br />
? S3 ? 81.954 ? 2.397 LT ? 6.187 NET ? 5.937 NAT ? 0.102 LT 2 ? 0.125 NET ?LT ? 0.227 NET 2<br />
? 0.125 NAT ?LT ? 0.125 NAT ?NET ? 0.522 NAT 2 (4.8)<br />
where LT, NET, and NAT are normalized values of the length (LT), number of tubes in the<br />
evaporator (Ne,T) and number of tubes of the absorber (Na,T), such that their values are<br />
<strong>with</strong>in the range [-2, 2]. They are calculated using the upper and lower bounds of the<br />
variables (Table 2) as follows:<br />
LT ?<br />
L ? 17<br />
1.75<br />
NET ? N e ? 440<br />
115<br />
NAT ? N a ? 440<br />
115<br />
? 2 (4.9)<br />
? 2 (4.10)<br />
? 2 (4.11)<br />
4. Link the response surfaces (Equations 4.3 to 4.8) in a network using the parametric<br />
decomposition <strong>approach</strong> (Appendix A), considering a SCV of one at all the stations.<br />
The response surface network model allows estimating the processing times (?), the<br />
utilization (u) at each station, and the expected production time (CT), as functions of the<br />
product <strong>design</strong> variables.
Once the mean processing times, ?, are obtained for a particular combination of <strong>design</strong><br />
variables using the response surface network model, the labor cost (LC) can be estimated as<br />
follows:<br />
LC ? ? SA1 Ops SA1 ? ? SA2 Ops SA2 ? ? SA3 Ops SA3 ? (MS1)? S1 Ops S1 ?<br />
(MS2)? S2 Ops S2 ? (MS3)? S3 Ops S3<br />
139<br />
(4.12)<br />
where Opsi is the number of operators at the station i, and MS1, MS2 and MS3 are the number<br />
of servers (the capacity) at Stations S1, S2 and S3 respectively. In this case there are 2<br />
operators at the subassemblies stations SA1, SA2 and SA3, and there is one operator for each<br />
server at stations S1, S2 and S3. These values are fixed.<br />
For simplicity, only inventory, labor and burden (overhead) costs are considered in<br />
estimating the production cost (MfgC):<br />
where BC is the burden cost and IC is the inventory cost.<br />
follows:<br />
MfgC = LC + BC + IC (4.13)<br />
For this example, the burden cost is estimated as being proportional to the labor cost as<br />
BC = k LC (4.14)<br />
where k is a constant coefficient. With the raw material cost, MC (Table 4.7), and the number<br />
of operators at each station the cost of the product at each station i (PCi) can now be
approximated. For example, for the station SA1, the product cost is simple the material cost<br />
MCSA1, whereas for station S1 it is calculated as:<br />
where ICi is estimated as follows:<br />
?<br />
PC S1 ? MC S1 ? MC i ? LC i ? BC i ? IC i<br />
SA1<br />
SA2<br />
SA3<br />
IC i ? ir(PC i )<br />
TH year<br />
140<br />
(4.15)<br />
(4.16)<br />
where ir is the annual interest rate and THyear is the annual throughput of the product. Here is<br />
considered an ir of 0.1. The cost at other stations is calculated similarly. At station S3 the cost<br />
of the tubes is added to the raw material cost given in Table 4.7.<br />
To estimate IC it is necessary to know the safety stock of evaporator and absorber<br />
tubes. In setting a safety stock it is desired a confidence of at least 99% (a fill rate S ? 0.99)<br />
that there will be no delays at S3 due to a stockout of tubes. Given this constraint, and<br />
according to Hopp and Spearman (1996), the problem for specifying a safety stock is<br />
formulated for each tonnage as follows (shown for the evaporator tubes):<br />
Given Cost of tubes Ce (Table 4.2)<br />
Expected demand of tubes per year Dtubes = D(Ne)<br />
Find Safety stock SE<br />
Satisfy Average replenishment frequency ? F<br />
Average fill rate S ? 0.99<br />
Minimize Inventory investment, ICe
The “Satisfy” and “Minimize” statements can be written mathematically as:<br />
Satisfy D tubes<br />
Q<br />
Minimize C e<br />
1 ? n(r)<br />
Q<br />
141<br />
? F (4.18)<br />
? 0.99 (4.19)<br />
?? Q ??<br />
?? ? r ? ? ?? (4.17)<br />
?? 2 ??<br />
where Q is the replenishment quantity (in units), r is the reorder point (in units), ? is the expected<br />
demand during the replenishment lead time (in units), and n(r) is the expected number of<br />
backorders during a cycle. It follows that Q is chosen as:<br />
Q ? D tubes<br />
F (4.19)<br />
Given Q, r is chosen as the smallest value that satisfies the Equation (4.19), that is:<br />
n(r) ? (1? 0.99) D tubes<br />
F (4.20)<br />
Thus, the problem is to find a value of F that minimizes inventory investment (Equation<br />
4.17). In this case example, the replenishment lead-time is 7 weeks, and the possible F values<br />
are 54, 27, 18, 12 and 10.8 replenishments/year.<br />
Once F is chosen, and r is estimated using Equation (4.20), the safety stock of<br />
evaporator tubes SE is estimated as:<br />
SE = r - ? (4.21)<br />
A similar <strong>approach</strong> is used to obtain the safety stock for the absorber tubes.
142<br />
Based on these considerations, the classification of parameters that define the<br />
production system synthesis is shown in Table 4.8. The goals include production cost and cycle<br />
time. Two constraints are considered: a maximum utilization of 0.9 at any station and an<br />
average fill rate at station S3 of 99% in setting the safety stock of tubes. In Table 4.9 the values<br />
of the production targets for each refrigeration capacity are shown.<br />
Table 4.8 Parameters for Production System Design<br />
Parameter Name Symbol Category Range Units<br />
1 Capacity at station SA1 MSA1 constant 1<br />
2 Capacity at station SA2 MSA2 constant 1<br />
3 Capacity at station SA3 MSA3 constant 1<br />
4 Capacity at station S1 MS1 <strong>design</strong> variable 1 ,2,...,5<br />
5 Capacity at station S2 MS2 <strong>design</strong> variable 1 ,2,...,5<br />
6 Capacity at station S3 MS3 <strong>design</strong> variable 1,2,...,5<br />
7 Replenishment<br />
frequency<br />
F <strong>design</strong> variable 10.8,12,<br />
18,27,54<br />
8 Production cost MfgC goal dollars<br />
9 Expected cycle time CT goal hours
Table 4.9 Goals and Constraint Values for Production System Design<br />
Production Cost/Unit Expected Cycle-Time Fill Rate<br />
MfgC (dollars)<br />
CT (hours)<br />
S<br />
Capacity GOAL GOAL CONSTRAINT<br />
600 tons 6000 280 0.99<br />
700 tons 7000 290 0.99<br />
800 tons 8000 300 0.99<br />
900 tons 9000 310 0.99<br />
1000 tons 10000 320 0.99<br />
1100 tons 11000 330 0.99<br />
1200 tons 12000 340 0.99<br />
1300 tons 13000 350 0.99<br />
For the synthesis of the system, the value of the individual goals shown in Table 4.9 are<br />
“aggregated” into two systems goals as follows:<br />
143<br />
GoalAP ? ? Goal T DT (4.22)<br />
T<br />
where T indicates individual products (T=1,2,...,8), and DT is the estimated fraction of the total<br />
demand of each product (Table 4.1). Thus, for the aggregated product, the MfgC target value<br />
is:<br />
MfgCgoal=0.05(6000)+0.15(7000)+0.16(8000)+0.05(9000)+0.15(10000)+0.12(11000)+<br />
0.25(12000)+0.07(13000) = 9810 (dollars) (4.23)<br />
Similarly, for CT the target value is:
144<br />
CTGOAL=0.05(280)+0.15(290)+0.16(300)+0.05(310)+0.15(320)+0.12(330)+<br />
0.25(340)+0.07(350)=318 (hours) (4.24)<br />
Using the target given by Equation (4.23), the production cost goal for a compromise<br />
DSP is formulated as:<br />
MfgC<br />
MfgC goal<br />
Similarly, the CT goal is formulated as follows:<br />
CT<br />
CT goal<br />
? d 1 ? ? d1 ? ? 1 (4.25)<br />
? d 2 ? ? d2 ? ? 1 (4.26)<br />
An objective function Z ? W1(d1 ? ? d1 ? ) ? W2 (d2 ? ? d2 ? ) can now be formulated where Wi is the<br />
weight given to the achievement of the ith goal in an Archimedean formulation of the deviation<br />
function (Section 3.4).<br />
As this is a two-objective problem, the modified <strong>approach</strong> of Rao and Freihet (1992)<br />
described in Section 3.4.2, can be applied. First, two compromise DSPs are solved to<br />
construct a payoff matrix P:<br />
P ? Z ??<br />
11 Z ??<br />
12<br />
?? ???<br />
?? Z21 Z 22??<br />
d ??<br />
1 ??<br />
?? d2 ? *<br />
( X1 )<br />
? *<br />
d1 (X2 )<br />
? *<br />
( X1 )<br />
? *<br />
d2 (X2 )<br />
??<br />
??<br />
?? (4.27)<br />
where X1 * is the <strong>design</strong> vector X1 * =[MS1, MS2, MS3, F]1 obtained when minimizing only the<br />
production cost (i.e., W2=0), and X2 * is the <strong>design</strong> vector obtained when minimizing only the<br />
cycle time (W1=0). di ? ( X) is the value of the deviation variable associated <strong>with</strong> the achievement
of the ith goal when the <strong>design</strong> vector is X. The compromise DSP for obtaining the payoff<br />
matrix is shown in Figure 4.10.<br />
With the values of the payoff matrix (Equation 4.27), a modified deviation function is<br />
formulated as:<br />
145<br />
Z mp = Z - SC+1 (4.28)<br />
where 1 is added to make Z mp always positive, and the other terms are given by:<br />
Z ? W1 ˆ<br />
Z 1 ? W2 ˆ<br />
Z 2 (4.29)<br />
SC ? (1 ? ˆ<br />
Z 1)(1 ? ˆ<br />
Z 2) (4.30)<br />
ˆ<br />
Z 1 ?<br />
ˆ<br />
Z 2 ?<br />
Z1 ? Z1( X1 * )<br />
* * (4.31)<br />
Z1 ( X2 ) ? Z1 (X1 )<br />
Z2 ? Z2 (X2 * )<br />
* * (4.32)<br />
Z 2 ( X1 ) ? Z2 ( X2 )<br />
In this case study, and to facilitate the comparison of solutions obtained <strong>with</strong> different<br />
<strong>approach</strong>es, fixed values of W1 = W2 = 0.5 are used when solving compromise DSPs <strong>with</strong><br />
deviation function Z mp . The resulting baseline compromise DSP for the synthesis of the<br />
production system is shown in Figure 4.11.
Given<br />
?? Production model<br />
?? Inventory model<br />
?? Product <strong>design</strong> variables (LT, AT, ET, Ne,T, Na,T) T=1,2,...,8<br />
Find<br />
?? Capacity (number of servers) at station S1, MS1<br />
?? Capacity at station S2, MS2<br />
?? Capacity at station S3, MS3<br />
?? Replenishment frequency, F<br />
?? Deviation variables di ? ,di ? i=1,2<br />
Satisfy<br />
System Constraints<br />
?? Fill Rate at S3 (for tubes safety stock)<br />
S ? 0.99<br />
?? Maximum utilization<br />
u ? ???<br />
(Equation 4.20)<br />
System Goals for each capacity (minimize average production cost):<br />
?? MfgC ? ?<br />
? d1 ? d1 ? 1 (Equation 4.25)<br />
??<br />
MfgC goal<br />
CT<br />
CT goal<br />
? d 2 ? ? d2 ? ? 1 (Equation 4.26)<br />
Bounds<br />
?? 1? MS1 ? 5<br />
?? 1? MS2 ? 5<br />
?? 1? MS3 ? 5<br />
?? di ? , di ? ? 0<br />
di ? ?di ? ? 0<br />
Minimize<br />
The deviation function<br />
?? Z = Z ? W1d1 ? ? W2d2 ?<br />
Figure 4.10 Compromise DSP for Building a Payoff Matrix for Production System<br />
Design<br />
146
Given<br />
?? Production model<br />
?? Inventory model<br />
?? Product <strong>design</strong> variables (LT, AT, ET, Ne,T, Na,T) T=1,2,...,8<br />
?? Payoff matrix P (Equation 4.27)<br />
Find<br />
?? Capacity (number of servers) at station S1, MS1<br />
?? Capacity at station S2, MS2<br />
?? Capacity at station S3, MS3<br />
?? Replenishment frequency, F<br />
?? Deviation variables di ? ,di ? i=1,2<br />
Satisfy<br />
System Constraints<br />
?? Fill Rate at S3 (for tubes safety stock)<br />
S ? 99%<br />
?? Maximum station utilization<br />
u ? ???<br />
System Goals for each capacity<br />
(Equation 4.20)<br />
?? MfgC ? ?<br />
? d1 ? d1 ? 1 (Equation 4.25)<br />
??<br />
MfgC goal<br />
CT<br />
CT goal<br />
? d 2 ? ? d2 ? ? 1 (Equation 4.26)<br />
Bounds<br />
?? 1? MS1 ? 5<br />
?? 1? MS2 ? 5<br />
?? 1? MS3 ? 5<br />
?? di ? , di ? ? 0<br />
di ? ?di ? ? 0<br />
Minimize<br />
The deviation function<br />
Z mp = Z-SC+1 (Equation 4.28)<br />
Figure 4.11 Baseline Compromise DSP for Production System Design<br />
147
4.3.3 Results of Baseline Approach<br />
In the present case example the deviation function is highly non-linear and exhibits various<br />
minima. In general, for this kind of function it is impossible to know if an optimum is a global<br />
optimum or not. In order to verify at some extent the quality of the optimization algorithm that is<br />
used in this work, three different algorithms are compared, namely, Simulated Annealing (SAn),<br />
a clustering-<strong>based</strong> algorithm (GLOBAL), and the Foraging-Adaptive Linear Programming<br />
(FALP) algorithm.<br />
The first algorithm, SAn, is solved using the commercial optimization software OptdesX ®<br />
(Balling, et al., 1998). The parameters used in the optimization process are: 100 optimization<br />
cycles, a probability of accepting a worse <strong>design</strong> at the beginning of the optimization of 0.5, and<br />
a probability of accepting a worse <strong>design</strong> at the end of the optimization of 1.e-6. For all the<br />
capacities the initial starting point is LT =20.5, Ne,T=670, Na,T= 670, ET=3, and AT=3. The<br />
final values of the <strong>design</strong> variables and material cost for each refrigeration capacity are shown in<br />
Table 4.10.<br />
Table 4.10 Results for Each Capacity using Simulated Annealing<br />
Tons LT (ft) AT ET Na,T Ne,T MC T (dollars)<br />
600 17.01 2 2 443 453 115415<br />
700 18.46 2 2 441 478 128348<br />
800 19.45 2 2 464 457 135363<br />
900 17.06 2 2 498 697 156150<br />
1000 22.51 2 3 448 544 167891<br />
1100 19.47 3 2 637 538 176333<br />
1200 23.82 2 3 489 615 195988<br />
148
1300 23.95 2 3 527 687 214625<br />
Clustering-<strong>based</strong> algorithms are a modified form of the standard multi-start procedure.<br />
Simply explained, in a clustering-<strong>based</strong> algorithm a local search starting from several points<br />
distributed over the entire search domain is performed. The three main steps of the algorithm<br />
are: (1) sample points in the search domain, (2) transform the sampled point to group them<br />
around local minima, and (3) apply a clustering technique to identify groups that represent<br />
neighborhoods of local minima. Here is used the GLOBAL routine of Csendes (Csendes,<br />
1988) to solve the compromise DSP of Figure 4.9 <strong>with</strong> this technique. The local search<br />
procedure used in GLOBAL is a Quasi-Newton type, which uses the so-called DFP (Davidon-<br />
Fletcher-Powell) update formula. Note that the solutions obtained <strong>with</strong> this algorithm are<br />
independent of initial conditions. Because the GLOBAL routine is for continuous variables only,<br />
at each iteration of the continuous variables an exhaustive search on possible combinations of<br />
tubes selection is performed. A penalty function method is used for constraining the solution.<br />
The results obtained <strong>with</strong> this method are shown in Table 4.11.<br />
Table 4.11 Results for Each Capacity Using a Clustering-<strong>based</strong> Algorithm<br />
Tons LT (ft) AT ET Na,T Ne,T MC T (dollars)<br />
600 17.04 2 2 476 440 115942<br />
700 18.11 2 2 523 440 127445<br />
800 18.26 4 2 504 494 141445<br />
900 19.13 2 2 617 522 155719<br />
1000 23.05 3 3 535 440 170238<br />
149
1100 22.285 4 3 528 501 181691<br />
1200 23.70 4 3 530 493 192033<br />
1300 23.90 2 3 747 528 214321<br />
The third algorithm, FALP (Lewis and Mistree, 1996), is a solution scheme for mixed<br />
discrete/continuous <strong>design</strong> problems, which uses a successive approximation programming<br />
method to constrain the solution. The discrete portion of the scheme is <strong>based</strong> on the notion of<br />
foraging of animals and includes dynamic memory structure and schema identification. The<br />
continuous portion of the scheme is solved using the Adaptive Linear Programming (ALP)<br />
algorithm. Both the FALP and the ALP algorithms are part of the DSIDES software (Mistree,<br />
et al., 1993). The initial <strong>design</strong> point for each capacity is the same as <strong>with</strong> the SAn algorithm<br />
and the optimization is performed <strong>with</strong> a maximum of 100 cycles. The results are shown in<br />
Table 4.12, where the solutions marked <strong>with</strong> the symbol “*” are infeasible solutions considering<br />
an allowable relative constraint violation of 0.001. With SAn and the clustering-<strong>based</strong> algorithm<br />
all the solutions are feasible under this criterion.<br />
Table 4.12 - Results for Each Capacity using FALP<br />
Tons LT AT ET Na,T Ne,T MC (dollars)<br />
600 17.30 2 2 440 440 115120<br />
700 18.95 2 2 463 474 131593<br />
800 22.16 2 2 445 488 150806<br />
900 20.51 4 2 463 467 149743*<br />
1000 23.25 3 3 440 440 162973*<br />
150
1100 24.00 3 3 440 440 170020 *<br />
1200 23.05 4 3 486 486 182958 *<br />
1300 22.61 3 3 578 578 203987 *<br />
Observing Table 4.11 to 4.12, it is clear that the value of the <strong>design</strong> variables do not<br />
follow monotonic behavior <strong>with</strong> increasing capacity requirements. It is also observed that, in<br />
general, each algorithm reaches a different minimum. In Figure 4.12 it is shown the value of the<br />
deviation function Z for each capacity.<br />
Z<br />
0.25<br />
0.2<br />
0.15<br />
0.1<br />
0.05<br />
0<br />
* Infeasible<br />
*<br />
600 700 800 900 1000 1100 1200 1300<br />
Capacity<br />
Sim.Ann.<br />
GLOBAL<br />
Figure 4.12 Comparison of the Value of the Deviation Function Z for each Capacity<br />
Using SAn, GLOBAL and FALP Algorithms<br />
Clearly, the solutions obtained for this particular case <strong>with</strong> SAn algorithm are generally<br />
better than those obtained <strong>with</strong> the clustering-<strong>based</strong> method and FALP. Particularly, <strong>with</strong> the<br />
FALP algorithm the solution tends to be infeasible. The reason is that the continuous part of<br />
*<br />
FALP<br />
*<br />
*<br />
*<br />
151
FALP, the ALP algorithm is more suitable for highly constrained problems, which is not the<br />
case of this problem. Based on this observation, the SAn algorithm is used to solve subsequent<br />
compromise DSPs.<br />
The resultants MfgC and CT obtained by solving the compromise DSP shown in Figure<br />
4.11 <strong>with</strong> SAn are presented in Table 4.13.<br />
Table 4.13 - Results for Minimum Production Cost and Cycle Time<br />
MS1 MS2 MS3 F<br />
MfgC<br />
(dollars)<br />
CT<br />
(hours)<br />
Min COST 2 1 2 54 46,762 703<br />
Min CT 5 5 5 54 65,680 448<br />
With these values of MfgC and CT, and the targets of MfgCgoal=9810 and CTgoal=318.1,<br />
obtained <strong>with</strong> Equations (4.23) and (4.24), a payoff matrix (Equation 4.27) is constructed as:<br />
and the modified deviation function is<br />
where<br />
152<br />
Po ? Z11 Z ?? ?? ?? 3.76 5.69??<br />
12<br />
?? ??? ?? ?? (4.35)<br />
?? Z21 Z22 ?? ?? 1.21 0.41??<br />
Z mp = Z - SC+1 (4.36)<br />
Z ? 1<br />
( Z<br />
ˆ<br />
1 ?<br />
2 ˆ<br />
Z 2) (4.37)<br />
SC ? (1 ? ˆ<br />
Z 1)(1 ? ˆ<br />
Z 2) (4.38)<br />
ˆ<br />
Z 1 ? Z 1 ? 3.76<br />
5.69 ? 3.76 (4.39)
153<br />
Z<br />
ˆ<br />
2 ? Z 2 ? 0.41<br />
1.21? 0.41 (4.40)<br />
With the modified deviation function given in Equation 4.36, the compromise DSP shown in<br />
Figure 4.11 is solved, and the results are shown in Table 4.14.<br />
Table 4.14 Baseline Solution for Production System Design<br />
Design Variables Performance Variables<br />
MS1 MS2 MS3 F<br />
MfgC<br />
(dollars)<br />
CT<br />
(hours)<br />
Z mp<br />
3 2 3 54 52430 476 0.652<br />
The results of CT that are shown in Tables 4.13 and 4.14 are those predicted <strong>with</strong> the<br />
response surface network. As a means to validate this model of the production system, in<br />
Tables 4.15 it is shown a comparison of the values predicted <strong>with</strong> this model and the results of a<br />
discrete-event simulation, <strong>based</strong> on the values of length and number of tubes of Table 4.5. The<br />
simulation was carried out using the commercial simulation software ARENA® (Kelton, et al.,<br />
1998), <strong>with</strong> 20 replicates and a simulation time equivalent to 500,000 hrs (57 years) of<br />
operation. The results of CT for the 20 replicates are shown in Table 4.16.<br />
Table 4.15 Cycle Times in Hours Estimated <strong>with</strong> Response Surface Network<br />
Model (RSNM) and Discrete Event Simulation (DES)<br />
600 700 800 900 1000 1100 1200 1300 MEAN
RSNM 395 409 415 469 448 466 478 501 448<br />
DES 404 413 424 459 448 467 471 490 448<br />
Difference<br />
%<br />
2.2 1.0 2.1 2.2 0 0.2 1.5 2.2 0<br />
Table 4.16 Cycle Times in Hours for 20 Replicates of Simulation<br />
Rep CT Rep CT Rep CT Rep CT<br />
1 439 6 440 11 451 16 450<br />
2 448 7 446 12 448 17 442<br />
3 445 8 443 13 457 18 449<br />
4 453 9 453 14 450 19 453<br />
5 436 10 453 15 453 20 456<br />
154<br />
In Table 4.15 it is observed that there is no significant difference between the estimations<br />
of the Response Surfaces Network Model and Discrete Event Simulation.<br />
Once the accuracy of the production system model and the quality of the algorithm for<br />
solving the compromise DSPs have been verified <strong>with</strong> the data presented in this section, the<br />
cooperative and non-cooperative models are applied to the <strong>design</strong> of the absorber/evaporator<br />
module in Sections 4.4 and 4.5.<br />
4.4 COOPERATIVE FORMULATION AND SOLUTION
In this section a full-cooperative model <strong>with</strong> a conventional (non-<strong>probabilistic</strong>) solution scheme<br />
is applied in the <strong>design</strong> of the absorber/evaporator module. The objective in doing so is<br />
threefold:<br />
1. To verify the applicability of <strong>game</strong> <strong>theoretical</strong> representations for modeling enterprise<br />
<strong>design</strong> processes that involve product <strong>design</strong> and production system <strong>design</strong><br />
(Hypothesis 1).<br />
2. To examine how the Compromise DSP is an adequate decision model to embody a<br />
cooperative <strong>design</strong> formulation (Hypothesis 2.2).<br />
3. To develop information to be used in the validation of a non-cooperative solution.<br />
In doing so, the enterprise <strong>design</strong> method described in Section 1.4.4 is applied step by step<br />
<strong>with</strong>in the framework constructed in this work, which is shown in Figure 3.3. The<br />
complementary tasks shown in the right side of Table 4.17 are required to model the integrated<br />
product and process <strong>design</strong> as a cooperative <strong>game</strong> (see Section 2.3).<br />
Table 4.17 Implementation of Cooperative Formulation and Solution<br />
Tasks Enterprise Design Method Complementary Tasks<br />
» Baseline decision<br />
1 Expand scope of decision ?? Identify players of the <strong>game</strong>,<br />
2 Determine Input needed for Analysis ?? Define number of stages, K, and order<br />
and Identify Modeling Techniques of moves<br />
?? Define players’ set of choices, A<br />
?? Identify public and private information<br />
?? Identify probabilities on exogenous<br />
events<br />
3 Define boundaries of decision ?? Define players’ payoffs, fi(X)<br />
155
?? Define strategies<br />
4 Transform and integrate models<br />
5 Generate potential solutions ?? Embody strategy <strong>with</strong> a compromise<br />
DSP<br />
» Recommendation<br />
Following Table 4.17, the input required is the baseline decision for each of the<br />
individuals involved in the <strong>design</strong> process. These baseline decisions are described in Section<br />
4.3, and they are not repeated here. The remaining tasks are described in sections 4.4.1 to<br />
4.4.5.<br />
4.4.1 Task 1: Expand Scope of Decision<br />
The baseline decisions described in Section 4.3 are expanded by three means:<br />
1. By abstracting the <strong>design</strong> process as a <strong>game</strong> <strong>with</strong> multiple players, i.e., identifying other<br />
agents of the enterprise whose interests, needs or constraints should be considered in<br />
making a local decision. In this case, as explained in Section 4.1, there are 9 players<br />
involved in the <strong>game</strong> (N=9): 8 product <strong>design</strong>ers who have to <strong>design</strong> an<br />
absorber/evaporator module for each capacity and a production system <strong>design</strong>er.<br />
2. Adopting a cooperative protocol among all the players that they agree to respect in order to<br />
act in a coordinated manner (this is akin to an implicit coordination mechanism, as discussed<br />
in Section 3.1). In this case, a “full” cooperative protocol is adopted, in which the 9 players<br />
merge their goals and models such that the problem is formulated and solved as a single<br />
156<br />
one. In order to do this they have to adopt system goals, such that the goals of the players
as a group are now to reach some specific targets of total cost (MC+MfgC) and cycle<br />
time.<br />
3. By adopting standardization of tubes for the entire family of absorber/evaporator modules,<br />
such that there will be only one length, L, and tubes selection, E and A, for the 8 individual<br />
products that comprise the family.<br />
4.4.2 Task 2: Determine Input Needed for Analysis and Identify Modeling<br />
Techniques<br />
In the cooperative model that is studied in this section, the 9 players act as a unit and formulate<br />
a single compromise DSP. Thus, the merging of the information required by all of the individual<br />
players in the baseline decision is the information required by the coalition. This information is<br />
presented in Section 4.3. Following Table 4.17, the complementary tasks required at this step<br />
are as follows:<br />
?? Define the order of moves. A full cooperative <strong>game</strong> is, by definition, a static <strong>game</strong> (K=1).<br />
Therefore, there is only one movement at this stage, which is performed by the whole<br />
coalition of players.<br />
?? Define the players’ choices: A player’s set of choices (Ai) is embodied by the combination<br />
of values that the <strong>design</strong> variables that the player controls can take <strong>with</strong>in their bounds. In<br />
this case example, for the cooperative protocol the only <strong>design</strong> variable is a vector X:<br />
? L,<br />
E,<br />
A,<br />
N , N , MS1,<br />
MS2,<br />
MS3,<br />
F ?<br />
X e,<br />
T a,<br />
T<br />
157<br />
? T=1,2,...,8 (4.41)
The bounds and possible values that the elements of this vector can take are the same as in<br />
the baseline decisions (Tables 4.2, 4.3, 4.4 and 4.8). Note that there is only one length, L,<br />
and tubes selections, E and A, but there are 8 different variables Ne,T and Na,T, one for each<br />
capacity.<br />
?? Define what each player knows when making choices. Because all the players act as a<br />
unit, no player has private information. All the information presented during the formulation<br />
of the baseline decisions in the previous section is available to the group as a whole.<br />
Mathematically, if Ii represents the information set that is private to player i, the information<br />
available to each player in the cooperative <strong>game</strong> is given by:<br />
158<br />
I i ? I j j ? 1,..., N (4.42)<br />
?? Identify probability distributions over exogenous events. The only exogenous event<br />
considered in this case is demand, which is considered as a Poisson process <strong>with</strong> a mean<br />
arrival rate of 80 units/year, and <strong>with</strong> a distribution as shown in Table 4.1. Finally, the<br />
modeling techniques to be used in the full cooperative formulation are the same as in the<br />
baseline decisions.<br />
4.4.3 Task 3: Define Boundaries of Decision<br />
The boundaries of decision are embodied by the variables and parameters controlled by a<br />
decision maker that can be altered <strong>with</strong>in the restrictions posed by the organization, which in the<br />
<strong>game</strong> <strong>theoretical</strong> framework being applied are defined as the set of choices A, identified in Task
2. In the cooperative case, the boundaries of decision of the group as a coalition (referred here<br />
as Acoop) are simply the union of the individual sets Ai:<br />
159<br />
A coop ? A i ? A j ? i, j (4.43)<br />
During Task 3 it is necessary to define the players’ payoffs as a function of the <strong>design</strong><br />
variables, fi( X) . As explained in section 2.3, the payoffs are defined in the framework that is<br />
being applied as the value of the deviation function. In order to <strong>design</strong> the entire family<br />
simultaneously, the goals are redefined as “aggregated” goals, i.e., the individual goals of the<br />
players are merged into system goals as follows:<br />
GoalAP ? ? Goal T DT (4.44)<br />
T<br />
where DT is the fraction of the demand associated <strong>with</strong> each capacity of refrigeration, shown in<br />
Table 4.1 (Section 4.3), and the goals for each capacity, GoalT, are shown in Tables 4.5 and<br />
4.8 (Section 4.3).<br />
In this case, the goals are redefined as to achieve some targets of total cost and cycle<br />
time, given the targets of material cost, production cost and cycle time presented in Section 4.3.<br />
In this case the total cost is simply calculated as the sum of the material cost and production<br />
cost.<br />
COST = MC + MfgC (4.45)<br />
Using Equations (4.44) and (4.45) the goals for the system synthesis are formulated as follows:<br />
COST<br />
COST AP, goal<br />
? d 1 ? ? d1 ? ? 1 (4.46)
where the target for COST is calculated as:<br />
And the goal for CT is formulated as:<br />
160<br />
COSTAP, goal ? ? (MCgoal ? MfgCgoal) T DT ? 160245<br />
(4.47)<br />
T<br />
CT<br />
CT goal<br />
? d 2 ? ? d2 ? ? 1 (4.48)<br />
The target of cycle time, CTgoal, remains the same as in the formulation of the baseline decision<br />
for the production system <strong>design</strong>, i.e., CTgoal = 318.<br />
As this is a two-objective problem, the modified <strong>approach</strong> of Rao and Freihet is applied.<br />
The procedure is the same as in the synthesis of the production system <strong>design</strong> described in<br />
Section 4.3, except that the production cost goal is now substituted <strong>with</strong> total cost. Therefore,<br />
the payoff of each player, fi(X), is the same for all of them, and equal to a modified deviation<br />
function Z mp :<br />
where:<br />
fi(X) = Z mp = Z - SC+1 (4.49)<br />
Z ? W1 ˆ<br />
Z 1 ? W2 ˆ<br />
Z 2 (4.50)<br />
SC ? (1 ? ˆ<br />
Z 1)(1 ? ˆ<br />
Z 2) (4.51)<br />
ˆ<br />
Z 1 ?<br />
ˆ<br />
Z 2 ?<br />
Again, W1 and W2 are chosen to be equal to 0.5.<br />
Z1 ? Z1( X1 * )<br />
* * (4.52)<br />
Z1 ( X2 ) ? Z1 (X1 )<br />
Z2 ? Z2 (X2 * )<br />
* * (4.53)<br />
Z 2 ( X1 ) ? Z2 ( X2 )
With the information described up to this point, it is possible to formulate a <strong>design</strong> strategy<br />
s. According to the definition given in Section 2.2.3 of a cooperative solution, the strategy s can<br />
be simply stated as “find a <strong>design</strong> vector X * such that Z mp has the minimum possible value,”<br />
which can be mathematically written as:<br />
161<br />
s : X * ? Z mp (X * ) ? Z mp ( X) (4.54)<br />
where X is defined in Equation 4.41. The algorithm for implementing this strategy is shown in<br />
Figure 4.13.<br />
3<br />
HEAT TRANSFER<br />
ANALYSIS<br />
MODEL<br />
L, E, A, Tons, NeT, NaT<br />
UAEVAP, UAABS, DP<br />
1<br />
L, E, A<br />
2 DETERMINE<br />
MINIMUM<br />
NUMBER OF<br />
TUBES<br />
Tons = 600,..., 1300<br />
4<br />
OPTIMIZATION DRIVER<br />
L, Ne, Na, E, A<br />
MS1<br />
MS2<br />
MS3<br />
F<br />
PRODUCTION MODEL<br />
INVENTORY MODEL<br />
UAEVAP, UAABS, DP<br />
Zmp,<br />
Value of Constraints<br />
5<br />
Evaluate Zmp<br />
Evaluate<br />
constraints<br />
u<br />
COST<br />
CT<br />
Figure 4.13 Implementation of Design Strategy for Cooperative Model
An optimization driver, OptdesX® (Module 1), passes values of L and a selection of<br />
tubes E and A, along <strong>with</strong> values of the production system variables MS1, MS2, MS3 and F.<br />
Then, the minimum number of tubes that can satisfy the heat transfer requirements for each<br />
capacity are calculated (or the number for which the violation of the constraint is as small as<br />
possible). The search of the number of tubes (module 2 of Figure 4.13) is performed using a<br />
bisection search method. Once the number of tubes for each capacity is obtained, the value of<br />
the dependent variables COST and CT are evaluated in module 4, then the value of the<br />
deviation function and the constraints is calculated in module 5, and finally the results are passed<br />
to the optimization driver.<br />
Given a <strong>design</strong> strategy and an <strong>approach</strong> to implement it, the next tasks in the enterprise<br />
<strong>design</strong> method are to transform and integrate models, embody the <strong>design</strong> strategy in a <strong>design</strong><br />
formulation and to generate potential solutions. In this case no transformation of models is<br />
applied and the complete models of all the players are used in the solution process. The<br />
strategy is embodied in a compromise DSP as shown in Figure 4.14. The solution achieved<br />
<strong>with</strong> the cooperative model is described in the next section.<br />
162
163
Given<br />
?? Heat Transfer Models<br />
?? Production model (Section 4.3.2)<br />
?? Inventory model (Section 4.3.2)<br />
?? Commercial tube alternatives (Table 4.2)<br />
?? A payoff matrix<br />
Find<br />
?? Design vector X=[L, E, A, NeT, NaT, MS1, MS2, MS3, F]; T=1,2,..,8<br />
?? Deviation variables di ? ,di ? i=1,2<br />
Satisfy<br />
System Constraints<br />
?? UAEVAP ? UAEVAP min (Table 4.5)<br />
?? UAABS ? UAABS min (Table 4.5)<br />
?? ? P ? ? Pmax (Table 4.5)<br />
?? Fill Rate at S3 (for tubes safety stock)<br />
S ? 99% (Equation 4.20)<br />
?? Maximum station utilization<br />
u ? ???<br />
System Goals for each capacity<br />
??<br />
COST ? ?<br />
? d1 ? d1 ? 1 (Equation 4.46)<br />
??<br />
COST APgoal<br />
CT<br />
CT goal<br />
Bounds<br />
? d 2 ? ? d2 ? ? 1 (Equation 4.48)<br />
?? NaT min ? Na ? NaT max (Table 4.3)<br />
?? NeT min ? Ne ? NeT max<br />
?? Lmin ? L ? Lmax<br />
?? 1? MS1 ? 5 (Table 4.8)<br />
?? 1? MS2 ? 5<br />
?? 1? MS3 ? 5<br />
?? di ? , di ? ? 0<br />
di ? ?di ? ? 0<br />
Minimize<br />
The modified deviation function Z mp (Equation 4.49)<br />
Figure 4.14 Compromise DSP for Cooperative Model<br />
164
4.4.4 Results of Cooperative Model<br />
The compromise DSP shown in Figure 4.14 is solved using a Simulated Annealing algorithm.<br />
The parameters for the optimization process are: 100 cycles, a probability of accepting a worse<br />
<strong>design</strong> of 0.5 at the beginning of the optimization and a probability of 1.e-6 at the end. The<br />
optimization process is carried out starting from three different initial <strong>design</strong> vectors, as shown in<br />
Table 4.18. The final value of the <strong>design</strong> variables and the aggregated cost and cycle time are<br />
shown in Table 4.19. The final number of tubes and the material cost for each capacity are<br />
shown in Table 4.20.<br />
Table 4.18 Initial Value of Design Variables for Optimization Process<br />
Case L(ft) E A MS1 MS2 MS3 F<br />
1 17 1 1 1 1 1 10.8<br />
2 20.5 2 3 3 3 3 18<br />
3 24 4 4 5 5 5 54<br />
Table 4.19 Results of Cooperative Model<br />
Case L(ft) E A MS1 MS2 MS3 F<br />
COST<br />
(dollars)<br />
CT<br />
(hours)<br />
Z mp<br />
1 19.45 2 2 3 2 2 54 220050 530 0.63<br />
2 19.45 2 2 3 2 2 54 220050 530 0.63<br />
3 22.95 3 4 3 2 2 54 219484 466 0.51<br />
165
Table 4.20 Results of Cooperative Model for each Capacity<br />
Case 1 Case 2 Case 3<br />
Capacity Ne Na MC Ne Na MC Ne Na MC<br />
(dollars)<br />
(dollars)<br />
(dollars)<br />
600 440 440 124580 440 440 124580 440 440 150078<br />
700 440 446 128488 440 446 128488 440 440 153578<br />
800 457 504 138086 457 504 138086 440 440 157078<br />
900 510 575 153120 510 575 153120 440 440 160578<br />
1000 568 647 168854 568 647 168854 440 457 165639<br />
1100 626 647 183839 626 647 183839 475 503 178984<br />
1200 687 708 199952 687 708 199952 522 554 194717<br />
1300 745 780 215687 745 780 215687 568 600 209829<br />
It is observed that the algorithm found two different minima, one for Case 1 and Case 2,<br />
and one for Case 3. Clearly, the solution obtained in Case 3 is the best of the three. With an<br />
increment of only 0.9% in the total cost <strong>with</strong> respect to the other solution cycle time is reduced<br />
69 hours (12%). In Figure 4.14 it is shown a plot of contours of the deviation function, Z mp ,<br />
against the variables length and evaporator tube selection for the absorber tube selection A=2,<br />
and A=4, considering other <strong>design</strong> variables fixed at the values given in Table 4.19.<br />
In Figure 4.15 we can observe a strong similarity between the deviation function <strong>with</strong><br />
absorber tube selection A=2 and A=4, where the two minima found are present. Interestingly,<br />
the minima are in the middle of the <strong>design</strong> space for these particular values of A, MS1, MS2,<br />
166<br />
MS3 and F. Further discussion of this solution is carried out in Section 4.7, where these results
are compared <strong>with</strong> those achieved <strong>with</strong> other <strong>approach</strong>es, and the fulfillment of the objectives<br />
posed at the beginning of this section is examined.<br />
A = 2<br />
MS1 = 3<br />
MS2 = 2<br />
MS3 = 2<br />
F = 54<br />
EVAPORATOR TUBE<br />
Design Solution<br />
A = 2<br />
17.0<br />
CONTOURS Z mp<br />
17.0 18.75 20.5 22.25 24.0<br />
LENGTH<br />
LENGTH<br />
20 .5 24.0<br />
22.25<br />
Zmp<br />
1<br />
2<br />
3<br />
EVAPORATOR TUBE<br />
a) Deviation Function <strong>with</strong> Absorber Tube 2<br />
4<br />
A = 4<br />
MS1 = 3<br />
MS2 = 2<br />
MS3 = 2<br />
F = 54<br />
EVAPORATOR TUBE<br />
A = 4<br />
17.0<br />
CONTOURS Z mp<br />
17.0 18.75 20.5 22.25 24.0<br />
LENGTH<br />
LENGTH<br />
20 .5<br />
22.25<br />
24.0<br />
Z mp<br />
1<br />
2<br />
3<br />
EVAPORATOR TUBE<br />
b) Deviation Function <strong>with</strong> Absorber Tube 4<br />
Figure 4.15 Modified Deviation Function Z mp<br />
4.5 NON-COOPERATIVE FORMULATION AND SOLUTION<br />
In this section, the absorber/evaporator <strong>design</strong> problem is solved using a Stackelberg or<br />
Leader/Follower <strong>game</strong> model between the product <strong>design</strong> group and the production system<br />
4<br />
167
<strong>design</strong>er. As explained in Section 2.3, the Stackelberg <strong>game</strong> is a particular class of non-<br />
cooperative <strong>game</strong>s. The objectives of this section are as follows:<br />
1. To continue the verification initiated in Section 4.4 of the applicability of <strong>game</strong><br />
<strong>theoretical</strong> representations for modeling enterprise <strong>design</strong> processes that involve<br />
product <strong>design</strong> and production system <strong>design</strong> (Hypothesis 1).<br />
2. To verify that product <strong>design</strong> and production operations management, which are two<br />
core activities of an enterprise, can be effectively coordinated using non-cooperative<br />
formulations of individual <strong>design</strong> problems (Hypothesis 2.1).<br />
3. To verify that the Compromise DSP is an adequate decision model to embody a non-<br />
cooperative <strong>design</strong> formulation (Hypothesis 2.2).<br />
The implementation is akin to that carried out in the cooperative formulation described in the<br />
previous section. Again, the baseline decisions are those described in Section 4.3. The<br />
remaining tasks are described in Section 4.5.1 to 4.5.4.<br />
4.5.1 Task 1: Expand Scope of Decision<br />
Similarly as in the cooperative formulation, the baseline decisions are expanded by four means:<br />
1. By abstracting the <strong>design</strong> process as a multi-player <strong>game</strong>.<br />
2. Adopting a <strong>game</strong> <strong>theoretical</strong> protocol to coordinate decisions. In this case, the 8 product<br />
<strong>design</strong>ers adopt a full cooperative protocol among them, and they, as a group, act as<br />
168
leaders of the production system <strong>design</strong>er in a Stackelberg (Leader/Follower) <strong>game</strong>. This<br />
protocol is illustrated in Figure 4.16.<br />
3. By adopting standardization of tubes for the entire family of absorber/evaporator modules,<br />
as in the cooperative case.<br />
4. By adopting system goals in making “local” decisions. In this case, both the product <strong>design</strong><br />
group and the production system <strong>design</strong>er adopt total cost and cycle time as performance<br />
measures.<br />
Goal<br />
Minimize Material Cost<br />
Product Designers<br />
600 700 800 900 1000 1100 1200 1300<br />
Capacities<br />
RRS f<br />
DESIGN VARIABLES<br />
DESIGN VARIABLES<br />
Manufacturing Process<br />
Designer<br />
X lg =[L, E, A, N eT , N aT ] T=1,2,…,8<br />
Goal<br />
Minimize Production Cost<br />
Minimize Production Time<br />
X f =[MS1, MS2, MS3, F]<br />
Figure 4.16 Stackelberg Protocol between Product Design and<br />
Production System Design<br />
169
4.5.2. Task 2: Determine Input Needed for Analysis and Identify Modeling<br />
Techniques<br />
Because the product <strong>design</strong> group and the production system <strong>design</strong>er do not act as a single<br />
player, two sets of information have to be distinguished. The information required by the group<br />
of product <strong>design</strong>ers and the information required by the production system <strong>design</strong>er. For<br />
simplicity, in this section and in Section 4.6, the group of 8 product <strong>design</strong>ers is referred as the<br />
leader group, and the production system <strong>design</strong>er as the follower. Once the players are<br />
identified, in order to determine the input needed for analysis, the <strong>game</strong> is set up by the following<br />
tasks:<br />
?? Define the order of moves. In this case the <strong>game</strong> has at least two stages ( K ? 2). In the<br />
first stage the leader group initiates the process while the follower does “nothing.” After a<br />
<strong>design</strong> is obtained by the leader group in the first stage, in the second move the follower<br />
<strong>design</strong>s a production system to meet the required demand using the information generated by<br />
the leader group in the previous stage. In this example only two stages are considered, but<br />
the leader group could proceed to more detailed stages of the <strong>design</strong> process (if necessary)<br />
or they could wait until the follower provides them <strong>with</strong> “updated” information of the<br />
production system.<br />
170<br />
k<br />
?? Define the players’ choices. The set of choices at stage k for the leader group, Alg, are<br />
defined as follows:<br />
A<br />
k<br />
lg<br />
k<br />
k k<br />
? X lg ? Slg<br />
: X f ( X ) ? RRS ( X ) ?<br />
? (4.55)<br />
lg<br />
f<br />
lg
171<br />
k k<br />
where Xlg is the <strong>design</strong> vector of the leader group, and RRSf (Xlg)<br />
is the rational reaction set<br />
of the follower at stage k. In this case, Xlg is as defined as follows:<br />
lg<br />
? L,<br />
E,<br />
A,<br />
N eT , NaT<br />
?<br />
X ? T = 1,2, ...,8 (4.56)<br />
Note that the leader group coordinates decisions <strong>with</strong> the follower via the RRSf. The<br />
k<br />
point Xlg ? RRSf is by definition a point of Nash equilibrium. Thus, the “leader” restricts the<br />
set of alternatives, Alg, to those that form Nash equilibrium <strong>with</strong> the follower. This is a hybrid<br />
of the explicit and implicit coordination mechanisms discussed in Section 3.1. It is implicit<br />
because the leader group agrees on a set of rules defined by the Stackelberg protocol and<br />
restricts choices accordingly. It is explicit because the decision is <strong>based</strong> on an explicit<br />
evaluation of alternatives that satisfy Equation (4.55).<br />
On the other hand, the follower’s options for this specific case are defined as:<br />
Af k ?<br />
??<br />
??<br />
? k ? 1<br />
?? k k ?1<br />
Af ? Xlg<br />
?? ? k ? 1<br />
(4.57)<br />
Equation (4.57) means that at stage 1 the only alternative of the follower is “do nothing,”<br />
whereas for later stages the set of choices depends only on the absorber/evaporator <strong>design</strong>,<br />
k?1<br />
Xlg , obtained by the leader group at the previous stage.<br />
k<br />
?? Define what each player knows when making choices. Let I lg be the set of information<br />
that is private to the group of product <strong>design</strong>ers. This set is the union of the information sets<br />
of the players that comprise the group. The information available to the leader group at stage<br />
k in the Stackelberg protocol is at least , but not restricted to:
172<br />
k k<br />
Ilg ? RRSf<br />
(4.58)<br />
In this work only two stages are considered, thus only the RRSf for stage k=1 is needed.<br />
From this point forward the convention that RRS refers to<br />
1 RRS f is adopted.<br />
For the production system <strong>design</strong>er, the follower, the information available at the<br />
beginning of stage k (for k>1) is as follows:<br />
k? 1<br />
If ? Xlg (4.59)<br />
where If is the private information of the follower, described in Section 4.3.2.<br />
?? Identify probability distributions over exogenous events. This is the same as in the<br />
cooperative model.<br />
Finally, the modeling techniques to be used in the non-cooperative model for each<br />
group are the same as in the baseline decisions. The only addition is that, according to the<br />
procedure described in Section 3.5.3, the RRS is modeled using response surfaces. The<br />
construction of the RRS is described at the end of the present section.<br />
4.5.3 Task 3: Define Boundaries of Decision<br />
In the non-cooperative case, the boundaries of decision of the leader group (Alg) are simply the<br />
union of the individual sets of the 8 different product <strong>design</strong>ers:<br />
A lg ? A i ? A j , i, j ? f (4.60)<br />
where Alg is given in Equation (4.55). For the follower, the boundaries of decision are simply<br />
the alternative choices Af given by Equation 4.57.
During Task 3 it is necessary to define the payoffs of the players as function of the<br />
<strong>design</strong> variables. In this case, the goals and deviation functions for both the leader group<br />
mp mp<br />
( Zlg ) and the follower (Zf ) are the same as in the cooperative case (Equations 4.46, 4.48,<br />
mp mp mp<br />
and 4.49). Thus, Zlg ? Z f = Z .<br />
Now it is possible to define a <strong>design</strong> strategy for the leader group, slg, and the follower, sf.<br />
According to the definition of a Stackelberg solution, (Section 2.3.2), the strategy for the leader<br />
is mathematically written as:<br />
Similarly, the follower’s strategy is:<br />
work.<br />
173<br />
* mp * * mp<br />
slg : Xlg ? Z (Xlg,RRS<br />
f ( Xlg )) ? Z ( Xlg, RRSf (Xlg)) (4.61)<br />
s<br />
f<br />
: X<br />
* f<br />
mp * mp<br />
? Z ( X , X ) ? Z ( X , X )<br />
(4.62)<br />
* f<br />
In Figure 4.17 it is shown the implementation of these two strategies followed in this<br />
lg<br />
f<br />
* lg
2<br />
3<br />
1<br />
FOLLOWER<br />
LEADER GROUP<br />
FOLLOWER<br />
CONSTRUCT RRS<br />
RRS<br />
DETERMINE Xlg<br />
Xlg<br />
DETERMINE Xf<br />
1.1<br />
1.2<br />
1.3<br />
2.1<br />
2.2<br />
L,E,A<br />
2.3<br />
3.1<br />
3.2<br />
DOE Driver<br />
Xl g<br />
Optimization Software<br />
Evaluate Z<br />
Xf<br />
Production Model<br />
mp<br />
Zmp<br />
1.4<br />
Evaluate constraints<br />
Value of Constraints<br />
u<br />
CT,<br />
MfgC<br />
Inventory Model<br />
Optimization Software<br />
Ne,N a<br />
Determine Minimum<br />
Number of Tubes<br />
Tons = 600,..., 1300<br />
Xlg constraints<br />
Heat Transfer Model<br />
Optimization Software<br />
Xf<br />
Production Model<br />
Inventory Model<br />
MfgC, CT<br />
Xlg<br />
constraints<br />
Z mp<br />
Value of Constraints<br />
2.4<br />
Zmp<br />
Value of Constraints<br />
3.3<br />
Evaluate Z mp<br />
Evaluate constraints<br />
Evaluate Z mp<br />
Evaluate constraints<br />
u<br />
CT,<br />
MfgC<br />
Figure 4.17 Implementation of Design Strategies for Non-Cooperative Model<br />
174
In Module 1, the follower constructs the RRS, following the procedure given in Section<br />
3.5.3, as follows:<br />
?? Design an experiment for constructing response surfaces of MfgC and CT as functions of<br />
Xlg using a DOE driver (Module 1.1).<br />
?? For each level of the experiment find the “optimal values” of MfgC and CT using an<br />
optimization algorithm (Module 1.2), and appropriate analyses routines (Modules 1.3 and<br />
1.4).<br />
?? Fit response surfaces to the values of MfgC and CT obtained in step 3 using the DOE<br />
driver. The RRS is the vector [MfgC(Xlg), CT(Xlg)], where MfgC(Xlg) and CT(Xlg) are<br />
the response surfaces.<br />
175<br />
Note than in this example, the RRS of the follower is not the <strong>design</strong> vector Xf(Xlg), i.e.,<br />
the values of MS1, MS2, MS3 and F as functions of Xlg, but the resulting values of the<br />
responses MfgC and CT of the follower. The reason to do so is that the leader group does not<br />
actually need for its calculations to know the value of the follower’s <strong>design</strong> variables. They only<br />
need from the follower the value of CT and MfgC in order to evaluate a payoff function. Using<br />
MfgC and CT as the elements of the RRS vector greatly expedites computation during the<br />
product synthesis.<br />
The compromise DSP used in step one is the same as in the baseline decision for the<br />
production system <strong>design</strong>, illustrated in Figure 4.10 (Section 4.3). The resultant response<br />
surfaces are:
MfgC? 49263 ? 1215L C ? 442E C ? 412 A C ? 2451NE C ? 2208NA C<br />
CT ? 505.<br />
5 ? 11.<br />
1L<br />
- 365 LC EC ? 684 LC AC ? 479LC NAC ? 357EC AC<br />
? 376E C NE C ? 798E C NA C ? 686A C NE C ? 824A C NA C<br />
- 504NEC 2 - 392NEC NAC -345NAC 2<br />
? 3.<br />
63E<br />
C<br />
2<br />
C<br />
? 0.<br />
1E<br />
? 3.<br />
8E<br />
C<br />
C<br />
NA<br />
? 26.<br />
7 A<br />
C<br />
C<br />
? 9.<br />
4NE<br />
? 23.<br />
41NE<br />
2<br />
C<br />
C<br />
? 15.<br />
4NE<br />
? 3.<br />
0L<br />
where LC, EC, AC, NEC and NAC are normalized values obtained as follows:<br />
NEC ?<br />
NAC ?<br />
LC ?<br />
L ? 17<br />
1.75<br />
?? ?2, if E ? 1<br />
??<br />
?? ?1, if E ? 2<br />
EC ? ??<br />
?? 1, if E ? 3<br />
??<br />
?? 2, if E ? 4<br />
?? ?2, if E ? 1<br />
??<br />
?? ?1, if E ? 2<br />
EC ? ??<br />
?? 1, if E ? 3<br />
??<br />
?? 2, if E ? 4<br />
?<br />
T<br />
?<br />
T<br />
C<br />
NA<br />
C<br />
C<br />
E<br />
C<br />
? 4.<br />
5NA<br />
2<br />
C<br />
176<br />
(4.60)<br />
(4.61)<br />
- 2 (4.62)<br />
(4.63)<br />
(4.64)<br />
NeT DT ? 440<br />
115<br />
? 2 (4.65)<br />
N aTD T<br />
115<br />
? 440<br />
? 2 (4.66)<br />
By using Equations (4.62) to (4.66) the value of all the variables for fitting a response<br />
surface are in the same range of [-2,2]. With Equations (4.65) and (4.66), the number of<br />
variables for fitting the response surfaces are reduced from 19 (considering the number of tubes<br />
for both the absorber and evaporator for the 8 different capacities, i.e., 16 variables) to 5, <strong>with</strong>
the consequent reduction in the size of the experiment, and therefore, in the time and<br />
computational effort required to build the response surfaces. In Figures 4.18 and 4.19, and in<br />
Tables 4.21 to 4.24, it is verified that these response surfaces have an adequate accuracy for<br />
practical purposes.<br />
In Table 4.21 it is shown the ANOVA results for the MfgC fitting <strong>with</strong> a confidence level<br />
of 99%. The hypothesis that the model adequately fits the data is accepted under this level of<br />
confidence. The adequacy of the fit can also be observed in Table 4.22 where it is shown the<br />
R 2 value, which can be loosely interpreted as the proportion of the variability in data that can be<br />
“explained” by means of the response surface. Clearly, the R 2 value of 95% is appropriate for<br />
product <strong>design</strong> purposes.<br />
Table 4.21 ANOVA Table for MfgC Regression<br />
Source DF Sum of Squares Mean Square F Ratio Prob>F<br />
Model 17 384.5e+06 22.6e+06 11.6 0.0004<br />
Error 9 17.6e+06 1.9e+06<br />
Total 26 402.0e+06<br />
Table 4.22 Summary of Fit for MfgC<br />
R 2 0.956<br />
Root Mean Square Error 139<br />
Mean of Response 48426<br />
Observations 27<br />
177
MfgC<br />
55000<br />
52500<br />
50000<br />
47500<br />
45000<br />
42500<br />
40000<br />
40000 42500 45000 47500 50000<br />
MfgC Predicted<br />
52500 55000<br />
Figure 4.18 Fit of MfgC Regression<br />
The response surface for CT is also adequate for approximating the CT, as shown in<br />
Figure 4.19, and Tables 4.23 and 4.24. Again, the R 2 value is around 0.95.<br />
Table 4.23 ANOVA Table for CT Regression<br />
Source DF Sum of Squares Mean Square F Ratio Prob>F<br />
Model 10 40967 4097 42
Xlg:<br />
CT<br />
600<br />
550<br />
500<br />
450<br />
450 500 550 600<br />
CT Predicted<br />
Figure 4.19 Fit of CT Regression<br />
179<br />
Once the RRS is built and validated, the leader group uses the RRS to find a <strong>design</strong> vector<br />
Xlg=[L,E,A,NeT,NaT], T=1,2,...,8 (4.67)<br />
An optimization driver (Module 2.1) passes values of L, E, and A, to a subroutine (Module 2.2)<br />
that searches the minimum number of tubes for each capacity that satisfies the constraints or that<br />
violates them the least. Then, the values of Xlg are passed to Module 2.3 where the deviation<br />
function, Zmp, is evaluated.<br />
Once the <strong>design</strong> vector Xlg is obtained by the leader group using the previous<br />
implementation, the follower follows a similar <strong>approach</strong> to find the values of Xf.<br />
Given these <strong>design</strong> strategies, both the leader group and the follower embody their<br />
strategies in a compromise DSP. The compromise DSP for the leader group is shown in
Figures 4.20. The follower’s compromise DSP is the same one shown in Figure 4.10 (Section<br />
4.3).<br />
Given<br />
?? Heat-transfer models of the absorber-evaporator<br />
?? Commercial tube alternatives (Table 4.2)<br />
?? Rational reaction set of follower RRS (Equations 4.60 & 4.61)<br />
?? A payoff matrix P<br />
Find<br />
Design Vector Xlg=[L,E,A,NeT,NaT]<br />
?? Length of tubes, L<br />
?? Selection of tube for the evaporator heat-exchanger, E<br />
?? Selection of tube for the absorber heat-exchanger , A<br />
?? Number of tubes of the evaporator for each capacity , NeT; T=1,2,...,8<br />
?? Number of tubes of the absorber for each capacity , NaT; T=1,2,...,8<br />
?? Deviation variables di ? ,di ? ; i=1,2<br />
Satisfy<br />
System Constraints<br />
?? UAEVAP ? UAEVAP min (BTU/ hr o F) (Table 4.5)<br />
?? UAABS ? UAABS min (BTU/ hr o F) (Table 4.5)<br />
?? ? P ? ? Pmax (Table 4.5)<br />
?? X f (X lg) ? RRS( X lg)<br />
System Goals<br />
COST<br />
??<br />
??<br />
COST APgoal<br />
CT<br />
CT goal<br />
Bounds<br />
? d 1 ? ? d1 ? ? 1 (Equation 4.46)<br />
? d 2 ? ? d2 ? ? 1 (Equation 4.48)<br />
?? NaT min ? Na ? NaT max (Table 4.3)<br />
?? NeT min ? Ne ? NeT max<br />
?? Lmin ? L ? Lmax<br />
?? di ? , di ? ? 0<br />
?? di ? ?di ? ? 0<br />
Minimize<br />
?? The modified deviation function<br />
Z mp (Equation 4.49)<br />
180
Figure 4.20 Stackelberg Compromise DSP for Product Design<br />
181
4.5.4 Results of Non-Cooperative Model<br />
The compromise DSPs for the non-cooperative model are solved using a Simulated Annealing<br />
algorithm, as in the baseline and cooperative solutions. The parameters for the optimization<br />
process are the same as in the cooperative model to facilitate comparison of results in Section<br />
4.7. The optimization process is initiated from three different points, shown in Table 4.25. In<br />
Table 4.26 it is presented the material cost (MC), the total cost (COST) and cycle time (CT)<br />
“predicted” using the RRS, and the value of the deviation function, Z mp , for each solution. The<br />
number of tubes and the material cost for each capacity are shown in Table 4.27 (using the L<br />
value obtained in Case 1).<br />
Table 4.25 Initial Value of Design Variables for Optimization Process<br />
Case L(ft) E A<br />
1 17 1 1<br />
2 20.5 2 3<br />
3 24 4 4<br />
Table 4.26 Results of Product Design: MC and “Predicted” Values of MfgC, COST<br />
and CT<br />
Case L(ft) E A<br />
MC<br />
(dollars)<br />
MfgC<br />
(dollars)<br />
COST<br />
(dollars)<br />
CT<br />
(hours)<br />
Z mp<br />
1 19.24 2 2 167241 49430 216413 461 0.380<br />
2 19.25 2 2 167901 49512 216613 462 0.386<br />
3 19.27 2 2 168201 49553 217754 462 0.396<br />
182
Table 4.27 Results of Non-Cooperative Model for Each Capacity<br />
Case 1<br />
Capacity Ne Na MC<br />
600 440 440 123216<br />
700 440 446 127118<br />
800 457 518 137556<br />
900 518 590 153468<br />
1000 579 647 168376<br />
1100 636 719 183791<br />
1200 697 791 199703<br />
1300 758 863 215615<br />
In Table 4.26 it is observed that the solution practically converges to the same point,<br />
starting from completely different points of the <strong>design</strong> space. This is probably due to the use of<br />
second order response surfaces to model the RRS of the production system <strong>design</strong>er, which<br />
tends to “smooth” the deviation function.<br />
In Table 4.28, it is shown the values of MfgC, COST and CT obtained by the production<br />
system <strong>design</strong>er using the values of the product <strong>design</strong> variables of Case 1.<br />
Table 4.28 Production System Design Results <strong>with</strong> Non-Cooperative Model<br />
MS1 MS2 MS3 F MfgC CT<br />
3 3 3 54 51238 468<br />
183
It is observed that the difference between the values of production cost and cycle time<br />
predicted by the leader group (Table 4.26) and the final values of MfgC and CT estimated by<br />
the follower (Table 4.28) are very similar, <strong>with</strong> an error of 3.4% for MfgC. Given this error,<br />
the <strong>design</strong> process can be carried out again, <strong>with</strong> an “updated” RRS approximated around the<br />
product <strong>design</strong> vector, Xlg, obtained in the first stage, or, if the cost and/or time that this<br />
re<strong>design</strong> process requires are not justified considering this magnitude of error, the <strong>design</strong><br />
process continues to more detailed stages. Generally, differences in the estimate of different<br />
agents are unavoidable if the exact RRS are not known and have to be approximated by<br />
response surfaces or any other means. Thus, iteration and negotiation between agents is<br />
expected and desirable, the difference being now that the negotiation is <strong>based</strong> on quantitative<br />
estimates obtained in a consistently manner. In Section 4.6 the non-cooperative model is<br />
modified <strong>with</strong> the <strong>probabilistic</strong> <strong>design</strong> model discussed in Section 3.6, such that differences on<br />
estimations and other sources of uncertainty and imprecision are formally considered during the<br />
product and manufacturing process <strong>design</strong>. A comparison of results <strong>with</strong> other <strong>approach</strong>es and<br />
further discussion is carried out in Section 4.7.<br />
4.6 NON-COOPERATIVE FORMULATION AND SOLUTION WITH<br />
PROBABILISTIC DESIGN MODEL<br />
The <strong>probabilistic</strong> <strong>design</strong> model presented in Section 3.5 is now applied to the development of a<br />
flexible <strong>design</strong> solution of the absorber/evaporator module. The motivation is threefold:<br />
184
1. To “capture” and deal <strong>with</strong> uncertainty and imprecision that results from the use of<br />
approximations<br />
2. To achieve an absorber/evaporator module <strong>design</strong> that is robust to external<br />
variability, such as changes in the cost of raw material, and<br />
3. To achieve a flexible <strong>design</strong> solution of the absorber/evaporator module to facilitate<br />
coordination of product <strong>design</strong> and the production system <strong>design</strong>, thus verifying<br />
Hypothesis 3.<br />
The only difference <strong>with</strong> the process studied in Section 4.5 is in the formulation of the<br />
compromise DSPs of the leader group (Figure 4.19) and the follower (Figure 4.10). The<br />
necessary modifications are carried out according to the procedure described in Section 3.6.6,<br />
and the <strong>probabilistic</strong> compromise DSPs are shown in Figures 4.22 and 4.23.<br />
For the product <strong>design</strong>:<br />
(1) In Section “Given” of the compromise DSP include:<br />
?? Preference functions for the ranged performance variables.<br />
The preference functions for this case study are defined as shown in Figure 4.21.<br />
185
P(COST)<br />
1.0<br />
0.0<br />
160245<br />
COST (dollars)<br />
320000<br />
P(CT)<br />
1.0<br />
0.0<br />
318<br />
CT (hours)<br />
Figure 4.21 Preference Functions for COST and CT<br />
For the variable COST, a linear preference function is defined as follows:<br />
618<br />
186<br />
? 1<br />
if 160245 ? COST<br />
? 320000 ? COST<br />
P ( COST ) ? ?<br />
? 320000 ? 160245<br />
? 0<br />
if 160245 ? COST ? 320000 (4.68)<br />
otherwise<br />
where “160245” is the value of the COST target and “320000” is some higher value,<br />
such that a <strong>design</strong> that results in COST <strong>with</strong>in this range is acceptable. Similarly, for CT<br />
the preference function is defined as:<br />
? 1<br />
if 318<br />
? COST<br />
? 600 ? CT<br />
P ( CT ) ? ?<br />
? 600 ? 318<br />
? 0<br />
if 318<br />
? COST ? 600<br />
(4.69)<br />
otherwise<br />
?? Probability density functions for COST and CT. To simplify calculations, it is assumed a<br />
normal distribution of the probability density functions of these two variables.<br />
Accordingly, they are defined as follows:<br />
2<br />
1<br />
?<br />
?<br />
? 1 ? COST ? COST ?<br />
f ( COST ) ? exp ? ?<br />
? ?<br />
(4.70)<br />
? 2?<br />
? 2 ?<br />
? ?<br />
COST ? ? ? COST ? ?
(2) In Section “Find”:<br />
187<br />
2<br />
1<br />
?<br />
?<br />
? 1 ? CT ? CT ?<br />
f ( CT ) ? exp ? ? ?<br />
?<br />
?<br />
? ? ?<br />
(4.71)<br />
? 2?<br />
2<br />
? ? ?<br />
CT<br />
CT ? ?<br />
?? Substitute the <strong>design</strong> variables L, Ne,T and Na,T, <strong>with</strong> the average values,<br />
L , N<br />
e,<br />
T<br />
and N<br />
a T,<br />
(3) In Section “Satisfy”:<br />
(3.1) For system constraints:<br />
, and their half-width variation, ? L, ? Ne,T and ? Na,T.<br />
?? Include lower limiting values for the deviation of L:<br />
? L ? 0.25 (ft) (4.72)<br />
?? Add a limiting value for the deviation of performance of COST and CT. It is chosen to<br />
limit the change of COST and CT as 10% of the average value <strong>with</strong>in the <strong>design</strong> range<br />
as follows:<br />
? COST ? 0.10 COST (4.73)<br />
? CT ? 0.10 CT (4.74)<br />
?? Define the constraints using a statistical feasibility formulation as:<br />
UA ? 3 ? ? UA<br />
(4.75)<br />
EVAP<br />
ABS<br />
UA,<br />
EVAP<br />
UA,<br />
ABS<br />
EVAP,<br />
MIN<br />
UA ? 3 ? ? UA<br />
(4.76)<br />
? ? ? ? P<br />
ABS,<br />
MIN<br />
P 3 ? ? P<br />
(4.77)<br />
where the constraint values UAEVAP,MIN, UAABS,MIN, ? PMAX have the same values as in<br />
the baseline decision formulation.<br />
MAX
(3.2) For system goals:<br />
?? Redefine the system goals as:<br />
188<br />
? ?<br />
DPICOST ? d ? d ? 1<br />
(4.78)<br />
1<br />
1<br />
? ?<br />
DPICT ? d ? d ? 1<br />
(4.79)<br />
where the DPI of the variables COST and CT is calculated as:<br />
2<br />
COST?<br />
? COST<br />
2<br />
DPI COST ? ? P(<br />
COST ) f ( COST ) d(<br />
COST )<br />
(4.80)<br />
COST?<br />
? COST<br />
CT ? ? CT<br />
DPI CT ? ? P(<br />
CT ) f ( CT ) d(<br />
CT )<br />
(4.81)<br />
CT ? ? CT<br />
The integral of Equations (4.80) and (4.81) is evaluated numerically during the<br />
optimization process.<br />
(3.3) In Section “Bounds:”<br />
?? Define the bounds for L, Ne,T and Na,T as follows:<br />
e,<br />
T MIN<br />
LMIN MAX<br />
? ? L ? L ? L ? ? L<br />
(4.82)<br />
N ? ? N ? N ? N ? ? N<br />
(4.83)<br />
a,<br />
T MIN<br />
e T,<br />
a,<br />
T<br />
e,<br />
T<br />
a,<br />
T<br />
e,<br />
T MAX<br />
N ? ? N ? N ? N ? ? N<br />
(4.84)<br />
a , T MAX<br />
The aforementioned changes are required for the transformation of the compromise DSP of<br />
Figure 4.19 into a <strong>probabilistic</strong> formulation. Besides these changes, in the formulation of the<br />
<strong>probabilistic</strong> <strong>design</strong> problem it is considered that there are two sources of uncertainty, which are<br />
not considered in the conventional compromise DSP: (1) uncertainty regarding the cost of the<br />
e,<br />
T<br />
a,<br />
T
tubes, and (2) uncertainty that results from approximating the RRS <strong>with</strong> response surfaces. The<br />
first one is included as a means to illustrate the effect of uncertainty caused by factors that are<br />
external to the enterprise, whereas the second is included as a means to <strong>design</strong> considering that<br />
there are errors in the values of MfgC and CT predicted during product <strong>design</strong>.<br />
To include these considerations the following changes are done:<br />
?? In section “Given” include a range for deviation of variables or factors due to<br />
189<br />
uncertainty. In this case example it is considered that the cost of evaporator tubes might<br />
vary <strong>with</strong>in a range of ? 0. 5dollars<br />
around the nominal value, and the absorber tubes<br />
<strong>with</strong>in a range of ? 0. 25 dollars. Also, it is assumed that the relative error of MfgC<br />
calculated <strong>with</strong> Equation (4.60) is around 10%, and the error for CT calculated <strong>with</strong><br />
Equation (4.61) is around 5%.<br />
During the optimization process, for each combination of the <strong>design</strong> variables that is evaluated, a<br />
CCD experiment is performed considering all the possible deviations of factors and variables,<br />
and the standard deviation of any variable is approximated by one sixth of the maximum and<br />
minimum values of the variable obtained in the experiment:<br />
?<br />
VARIABLE<br />
?<br />
1<br />
6<br />
? max ? VARIABLE CCD ? ? min?<br />
VARIABLE CCD ?<br />
And the deviation of constraints and performance variables is then approximated by:<br />
(4.85)<br />
? UAEVAP ?3? UA,EVAP (4.86)<br />
? UAABS ?3? UA,ABS (4.87)<br />
? P ?3? ? P (4.88)
190<br />
? COST ?3? COST (4.89)<br />
? CT ?3? CT (4.90)<br />
These considerations apply also in the formulation and solution of a compromise DSP for<br />
the production system <strong>design</strong>, shown in Figure 4.23. Besides the reformulation of the<br />
compromise DSP the strategy and implementation is the same as that carried out in Section 4.5.
Given<br />
?? Heat-transfer models of the absorber-evaporator<br />
?? Commercial tube alternatives (Table 4.2)<br />
?? Rational reaction set of follower RRS (Equations 4.60 & 4.61)<br />
?? Preference function for COST (Equation 4.68)<br />
?? Preference function for CT (Equation 4.69)<br />
?? Probability density function for COST (Equation 4.70)<br />
?? Probability density function for CT (Equation 4.71)<br />
?? Half-width ranges for cost of tubes<br />
? C e=0.5 (dollars)<br />
? C a=0.25 (dollars)<br />
?? Assumed half-width range of error of the RRS:<br />
? MfgC=0.1MfgC<br />
? CT=0.05CT<br />
?? A payoff matrix P<br />
Find<br />
?? Design vector X lg=[L,E,A,N eT,N aT,? L,? N e,T,? N a,T]; T=1,2,...,8<br />
?? Deviation variables di ? ,di ? ; i=1,2<br />
Satisfy<br />
System Constraints<br />
?? ? L ? 0.25 (ft) (Equation 4.72)<br />
?? ? COST ? 0.10 COST (Equation 4.73)<br />
?? ? CT ? 0.10 CT (Equation 4.74)<br />
?? UA EVAP ? 3 ? UA,<br />
EVAP ? UAEVAP,<br />
MIN<br />
(Equation 4.75)<br />
UA ? 3 ? ? UA<br />
(Equation 4.76)<br />
?? ABS UA,<br />
ABS ABS,<br />
MIN<br />
?? ? P ? ? ? P ? ? PMAX<br />
?? Xf (X lg) ? RRS( Xlg) 3 (Equation 4.77)<br />
System Goals<br />
??<br />
? ?<br />
DPI COST ? d1<br />
? d ? 1 1<br />
(Equation 4.78)<br />
??<br />
? ?<br />
DPI CT ? d ? d ? 1<br />
(Equation 4.79)<br />
Bounds<br />
2<br />
?? ? L ? L ? L ? ? L<br />
2<br />
LMIN ? MAX<br />
(Equation 4.82)<br />
N eT,<br />
MIN ? ? N eT,<br />
? N eT,<br />
? N e,<br />
T MAX ? ? N e<br />
(Equation 4.83)<br />
N ? ? N ? N ? N ? ? N<br />
(Equation 4.84)<br />
?? T,<br />
?? a T, MIN a T, a T, a,<br />
T MAX a T,<br />
?? di ? , di ? ? 0<br />
di ? ?di ? ? 0<br />
Minimize<br />
?? The modified deviation function Z mp (Equation 4.49)<br />
Figure 4.22 Stackelberg Probabilistic Formulation for Product Design<br />
191
Given<br />
?? Production System Model<br />
?? Inventory Model<br />
?? Replenishment frequency alternatives<br />
?? Product Design Vector X lg=[L,E,A,N eT,N aT,? L,? N e,T,? N a,T]; T=1,2,...,8<br />
?? Preference function for COST (Equation 4.68)<br />
?? Preference function for CT (Equation 4.69)<br />
?? Probability density function for COST (Equation 4.70)<br />
?? Probability density function for CT (Equation 4.71)<br />
?? Half-width ranges for cost of tubes<br />
? C e=0.5<br />
? C a=0.25<br />
?? A payoff matrix P<br />
Find<br />
?? Production system <strong>design</strong> vector Xf=[MS1,MS2,MS3,F]<br />
?? Deviation variables di ? ,di ? ; i=1,2<br />
Satisfy<br />
System Constraints<br />
?? ? COST ? 0.10 COST (Equation 4.73)<br />
?? ? CT ? 0.10 CT (Equation 4.74)<br />
?? ? u ? 3 ? u ? 0.<br />
9<br />
System Goals<br />
??<br />
? ?<br />
DPI COST ? d1<br />
? d ? 1 1<br />
(Equation 4.78)<br />
??<br />
? ?<br />
DPI CT ? d ? d ? 1<br />
(Equation 4.79)<br />
Bounds<br />
?? 1? MS1 ? 5<br />
?? 1? MS2 ? 5<br />
?? 1? MS3 ? 5<br />
?? di ? , di ? ? 0<br />
di ? ?di ? ? 0<br />
Minimize<br />
2<br />
2<br />
?? The modified deviation function Z mp (Equation 4.49)<br />
Figure 4.23 Stackelberg Probabilistic Formulation for Production System Design<br />
192
The solution of the compromise DSP of product <strong>design</strong> (Figure 4.22) is shown in Table<br />
4.29, along <strong>with</strong> the average values of MC, and predicted MfgC , COST and CT . The range<br />
of the number of tubes and material cost for each capacity are shown in Table 4.30. The<br />
production system variables and the average values of MfgC, COST and CT obtained by the<br />
production system <strong>design</strong>er are shown in Table 4.31.<br />
Table 4.29 Results of Product Design: MC and<br />
Predicted Values of MfgC , COST and CT<br />
L (ft) E A MC<br />
dollars<br />
MfgC<br />
dollars<br />
COST<br />
dollars<br />
CT<br />
hours<br />
? COST<br />
dollars<br />
? CT<br />
hours<br />
193<br />
mp Z<br />
18.6-19.39 2 2 175043 46873 221916 460 21370 11.5 0.39<br />
Table 4.30 Results of Non-Cooperative Probabilistic Model for Each Capacity<br />
Capacity Ne Na MC(dollars)<br />
600 440-440 440-440 120928-124272<br />
700 440-440 446-468 126253-129179<br />
800 453-468 518-540 138200-139528<br />
900 511-529 575-611 152873-155522<br />
1000 572-590 647-683 168941-171460<br />
1100 629-651 719-755 184505-187520<br />
1200 690-715 780-827 199827-203095<br />
1300 752-777 852-899 216021-219154<br />
Table 4.31 Production System Design <strong>with</strong> Probabilistic Non-Cooperative Model
MS1 MS2 MS3 F MfgC<br />
dollars<br />
COST<br />
dollars<br />
CT<br />
hour<br />
s<br />
? COST<br />
dollars<br />
? CT<br />
hours<br />
mp Z<br />
194<br />
3 3 3 54 51367 220115 471 2335 2 0.91<br />
In Tables 4.29 and 4.31 is observed that the difference between the cost and the cycle<br />
times estimated during product <strong>design</strong> <strong>with</strong> the RRS is less than 1% for COST and around 2.5%<br />
for CT. Note that the largest deviation of the COST from the average <strong>with</strong>in the interval would<br />
be of 21370 dollars or 10%. Similarly, the largest change for CT would be of 2.5%. That is, if<br />
the variable length were not chosen to be 19 ft (the average) but some other value <strong>with</strong>in the<br />
interval [18.6-19.39] then the COST would not vary more than 10% and the CT more than<br />
2.5%. This is further discussed in the next section where these results are compared <strong>with</strong> those<br />
achieved in the previous sections.<br />
4.7 COMPARISON OF SOLUTIONS<br />
In Section 4.3 to 4.6 four different solutions are obtained for the <strong>design</strong> of the<br />
absorber/evaporator module. In Figure 4.23 it is shown the deviation of the material cost from<br />
the original targets for the various refrigeration capacities obtained <strong>with</strong> each solution<br />
(Z=MC/MCgoal-1). The value shown for the <strong>probabilistic</strong> solution is that estimated <strong>with</strong> the<br />
average length of 19 ft. Interestingly, in the cooperative solution there is a large increment of the<br />
material cost for the smallest capacities. This effect is even more evident in the average or<br />
“aggregated” cost, which is shown in Figure 4.24. As expected, the baseline solution in which<br />
each product <strong>design</strong>er strikes at minimizing the cost of material for each capacity is the one <strong>with</strong>
the lowest material cost, but the solutions achieved <strong>with</strong> the standardization of tubes, either<br />
cooperative or non-cooperative, are almost as good as the baseline one, the largest difference<br />
being approximately of 8000 dollars (4%).<br />
Z MC<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0<br />
600 700 800 900 1000 1100 1200 1300<br />
Capacity<br />
Baseline<br />
Coop<br />
Conv-NC<br />
Prob-NC<br />
Figure 4.23 Deviation from Material Cost Target for Each Capacity<br />
190000<br />
187500<br />
185000<br />
182500<br />
180000<br />
177500<br />
175000<br />
Baseline Coop Conv-NC Prob-NC<br />
Figure 4.24 Average Material Cost (in Dollars) of the Various Solutions<br />
195
In Figure 4.25 it is illustrated the average production cost. In this case, it is the<br />
cooperative solution (Coop) the one <strong>with</strong> the lowest cost, and the baseline is the one <strong>with</strong> the<br />
highest cost, the difference being approximately 6,000 dollars/unit. However, when considering<br />
total cost the baseline solution is slightly better. This is shown in Figure 4.26, where it is<br />
observed how both the baseline and the conventional non-cooperative (Conv-NC) solutions<br />
achieve a lower average total cost than the cooperative one. The <strong>probabilistic</strong> non-cooperative<br />
(Prob-NC) solution, for the mean length of 19 ft, has a cost as high as the one of the<br />
cooperative solution. However, it has to be remembered that the mean value is not necessarily<br />
the best value of the interval of solution. Actually, the solution obtained <strong>with</strong> the conventional<br />
non-cooperative model (L=19.14) is <strong>with</strong>in the interval of solution obtained <strong>with</strong> the<br />
<strong>probabilistic</strong> model (18.6? L ? 19.39).<br />
53000<br />
52000<br />
51000<br />
50000<br />
49000<br />
48000<br />
47000<br />
46000<br />
45000<br />
44000<br />
43000<br />
42000<br />
Baseline Coop Conv-NC Prob-NC<br />
Figure 4.25 Average Production Cost (in Dollars) of the Various Solutions<br />
196
221000<br />
220000<br />
219000<br />
218000<br />
217000<br />
216000<br />
215000<br />
Baseline Coop Conv-NC Prob-NC<br />
Figure 4.26 Average Total Cost (in Dollars) of the Various Solutions<br />
The values of average production time estimated for the various solutions are shown in<br />
Figure 4.27. The cooperative solution is in this case much better than the others. Particularly,<br />
the time estimated in the non-cooperative solutions and the baseline one are 40 hours above.<br />
480<br />
470<br />
460<br />
450<br />
440<br />
430<br />
420<br />
410<br />
400<br />
Baseline Coop Conv-NC Prob-NC<br />
Figure 4.27 Average Production Time (in Hours) of the Various Solutions<br />
197
In Figure 4.28 is shown the deviation of total cost and cycle time from the original<br />
targets (Z=Variable/Target –1 .0). It is observed that the difference between the cost<br />
achieved between the various solutions is minimal but it is in the reduction in cycle time that the<br />
cooperative solution overrides the others. The performance of the non-cooperative solutions is<br />
very similar to the baseline one. However, the real benefit of the <strong>approach</strong> is in the reduction of<br />
components’ variety. With the baseline solution there is a large variety of tubes. The<br />
manufacturer would have to ask the supplier to provide 8 different tubes for the absorber and 8<br />
different tubes for the evaporator, thus increasing the complexity of the system. On the other<br />
hand, <strong>with</strong> the cooperative and non-cooperative solutions there is a single combination of length<br />
and tubes selection for the evaporator and one for the absorber. This advantage is evident in<br />
Table 4.32, where the required safety stock for each solution is presented.<br />
Z<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0<br />
Baseline Coop Conv-NC Prob-NC<br />
COST<br />
Figure 4.28 Deviation from Targets of Total Cost and Production Time of the<br />
Various Solutions<br />
CT<br />
198
Table 4.32 Safety Stock of Evaporator and Absorber Tubes in Number of Tubes<br />
Baseline Cooperative Conv-NC Prob-NC<br />
SE SA SE SA SE SA SE SA<br />
600 1577 1542<br />
700 2124 1959<br />
800 1983 2013<br />
900 2426 1733 4561 4715 5561 6187 5673 6318<br />
1000 2417 1990<br />
1100 2558 3029<br />
1200 3325 2643<br />
1300 2249 1725<br />
Total 18659 16634 4561 4715 5561 6187 5673 6318<br />
Investment<br />
(dollars)<br />
2892530 1343690 770800 455474 727781 436173 736957 442260<br />
The number of tubes to keep as safety stock, in order to have the desired fill rate of<br />
0.99 at station S3, is approximately three to four times the inventory required <strong>with</strong> the other<br />
solutions. Clearly, the material handling and storage is facilitated <strong>with</strong> this reduction. The<br />
difference in the investment in safety stock is also important. Due to the variety of tubes, the<br />
safety stock in the baseline solution is as large as 3,300,000 dollars, whereas for the other<br />
solutions is in the order of 1,100,000 dollars.<br />
199<br />
An important issue is the sensitivity of the solutions to future changes of the <strong>design</strong><br />
variables, particularly the length. The sensitivity of the total cost and cycle time for an interval of<br />
? 0. 38 ft around the value of the length of 22.95 ft obtained <strong>with</strong> the cooperative model are<br />
shown in Figure 4.29. Similarly, the sensitivity of the <strong>probabilistic</strong> solution is shown in Figure<br />
4.30. The cooperative solution is, as expected, more sensitive to changes on the length,<br />
whereas the <strong>probabilistic</strong> solution shows a smooth monotonic behavior <strong>with</strong>in the range.
COST (dollars)<br />
220600<br />
220400<br />
220200<br />
220000<br />
219800<br />
219600<br />
219400<br />
219200<br />
219000<br />
218800<br />
22.57 22.722 22.874 23.026 23.178<br />
Length (ft)<br />
Cost<br />
CT (hours)<br />
Figure 4.29 Sensitivity of Cooperative Solution to Changes in the Length<br />
COST (dollars)<br />
221500<br />
221000<br />
220500<br />
220000<br />
219500<br />
219000<br />
218500<br />
218000<br />
217500<br />
Cost<br />
CT<br />
Cooperative<br />
Solution<br />
Range of Probabilistic Solution<br />
18.62 18.772 18.924 19.076 19.228<br />
Length (ft)<br />
CT<br />
Non-Probabilistic<br />
Solution<br />
455<br />
450<br />
445<br />
440<br />
435<br />
430<br />
425<br />
420<br />
CT (hours)<br />
Figure 4.30 Sensitivity of Probabilistic Non-Cooperative Solution<br />
to Changes in the Length<br />
480<br />
475<br />
470<br />
465<br />
460<br />
455<br />
200
study:<br />
In summary, the following observations are made regarding the results obtained in the case<br />
?? On the <strong>game</strong> <strong>theoretical</strong> representation of the integrated product and process <strong>design</strong><br />
(Hypothesis 1).<br />
The integrated <strong>design</strong> process of the absorber/evaporator module was successfully modeled<br />
using either a cooperative or a non-cooperative model depending on the desired kind of<br />
coordination. The mathematics of <strong>game</strong> theory provided an appropriate framework for<br />
expanding decisions and identifying <strong>design</strong> strategies <strong>with</strong>in the integrated <strong>design</strong> process.<br />
?? On the use of a non-cooperative formulation to coordinate decisions between product<br />
<strong>design</strong> and manufacturing process <strong>design</strong> (Hypothesis 2.1).<br />
The solution achieved <strong>with</strong> a cooperative formulation is clearly better than the one achieved<br />
by coordinating the integrated product and process <strong>design</strong> using a non-cooperative<br />
formulation. Even when it is expected that the solution achieved by integrating all the<br />
models into a joint solution be better, at least for small problems, the use of a second order<br />
response surface to approximate the rational reaction set of the production system <strong>design</strong>er<br />
contributes to this deviation of results because local minima are “lost” <strong>with</strong> the<br />
approximation. However, it is also clear that <strong>with</strong> the non-cooperative formulation a<br />
solution was achieved that was as good as the baseline and the cooperative solutions in<br />
terms of total cost, and better than the baseline one in terms of reduction of cycle time and<br />
components’ variety.<br />
201
It has to be remarked that the cooperative solution is very good for the whole system<br />
performance, but it is so by requiring a large “loss” for some of the individual products (see<br />
Figure 4.23), whereas the non-cooperative solution obtained is more equitable for the<br />
various product <strong>design</strong>ers, thus achieving a better balance between their interests.<br />
?? On the use of the Compromise DSP to embody the cooperative and non-cooperative<br />
formulations (Hypothesis 2.2).<br />
The Compromise DSP proved to be a very flexible decision model, which can be easily<br />
used to formulate cooperative or non-cooperative <strong>design</strong> strategies. The goal-oriented<br />
structure of this formulation provided an adequate model to embody the integrated product<br />
and process <strong>design</strong> of the absorber/evaporator module.<br />
?? On the use of a <strong>probabilistic</strong> <strong>design</strong> model to enhance the coordination of product<br />
<strong>design</strong> and manufacturing process <strong>design</strong> of the absorber/evaporator module<br />
(Hypothesis 3).<br />
202<br />
The interval representation of the <strong>design</strong> solution allows the group of product <strong>design</strong>ers<br />
to maintain their <strong>design</strong> freedom while enabling the production system <strong>design</strong>er to have a<br />
preliminary <strong>design</strong> of the production system that is robust <strong>with</strong>in the interval of solution.<br />
There are various advantages of the <strong>probabilistic</strong> <strong>approach</strong>:<br />
?? It was possible to capture uncertainty associated <strong>with</strong> factors that are external to the<br />
enterprise.
203<br />
?? It was possible to capture uncertainty associated <strong>with</strong> inaccuracy of the RRS<br />
approximations.<br />
?? It is possible to obtain robust solutions that are insensitive to the imprecision<br />
associated <strong>with</strong> this early stage of the absorber/evaporator <strong>design</strong> process, thus<br />
facilitating the integrated product and production system <strong>design</strong>.<br />
The main disadvantage of the <strong>approach</strong> is in the need to evaluate the distribution of<br />
performance at every iteration of the optimization process, which can become time<br />
consuming.<br />
4.8 CLOSURE OF CASE STUDY AND OPEN ISSUES<br />
4.8.1 Review of Case Study<br />
Make-to-order systems can be effectively improved by adopting a systems perspective in the<br />
<strong>design</strong> of their products. This perspective is implemented in this chapter by two means. First,<br />
by applying the enterprise <strong>design</strong> framework constructed in this thesis and second, by applying<br />
the <strong>design</strong> principles of standardization and robustness.<br />
The enterprise <strong>design</strong> framework is applied by:<br />
?? Modeling the <strong>design</strong> process as a <strong>game</strong> between product <strong>design</strong>ers and the production<br />
system <strong>design</strong>er.<br />
?? Expanding the scope of individual decisions by adopting system’s goals.
?? Coordinating the decisions of the players using either a cooperative or a non-cooperative<br />
formulation of strategies.<br />
?? Embodying the strategy of the players <strong>with</strong> a common decision model, the Compromise<br />
DSP.<br />
?? Providing ranged solutions using a <strong>probabilistic</strong> <strong>design</strong> model as a means to enhance the<br />
coordination of the individual’s decisions.<br />
Thus, the various hypotheses posed in Chapter 1 are being tested in this example. In applying<br />
this framework a family of products is <strong>design</strong>ed as efficiently as a single product while<br />
incorporating production issues since a very early stage of the <strong>design</strong> process.<br />
4.8.2 Open Issues<br />
An important consideration that has to be done regarding the use of the non-cooperative<br />
<strong>game</strong> theory in coordinating enterprise <strong>design</strong> decisions is the issue of building rational reaction<br />
sets. In this work response surfaces are used to construct them, but generally speaking, when<br />
the number of players and/or <strong>design</strong> variables become large, the construction of the RRS using<br />
<strong>design</strong> of experiments might be extremely cumbersome. It is necessary to look for alternative<br />
methods for building the RRS.<br />
In the case on an integrated product-process development it is necessary to somehow<br />
estimate the effect of changes in the <strong>design</strong> variables in the production system. In this thesis it is<br />
proposed to do so by modeling the floor shop as a response surface network. This <strong>approach</strong><br />
204
proved to be effective for this case example in which the number of variables that largely affect<br />
the processing times were small. However, when the functions are highly nonlinear, the space is<br />
relatively unconstrained and there is large freedom in choosing <strong>design</strong> variables, response<br />
surfaces might not provide appropriate fit.<br />
Regarding the <strong>design</strong> of product families, in this case study it is proposed to do so using<br />
the notion of an “aggregated” product that is used as a product platform. The <strong>design</strong> of the<br />
family is carried out by applying adaptive <strong>design</strong> on this aggregated product to satisfy the<br />
requirements of the individual products of the family. This greatly facilitates the family <strong>design</strong><br />
process. Here it is suggested to do the “aggregation” <strong>based</strong> on demand, because it facilitates the<br />
analysis of production performance. It is difficult to assess at the present the effectiveness of<br />
this <strong>approach</strong> but certainly it opens an interesting avenue for future research.<br />
205
CHAPTER 5<br />
CLOSURE<br />
201<br />
This thesis is motivated by the need of developing a mathematical framework for<br />
enterprise integration. In this chapter the thesis argument developed throughout this work to<br />
answer the posed research questions, and applied to a case study in Chapter 4, is critically<br />
reviewed, namely, the merging of <strong>game</strong> theory, an enterprise <strong>design</strong> method <strong>based</strong> on a<br />
common model for decision support, the Compromise DSP and a <strong>probabilistic</strong> <strong>design</strong> model for<br />
the appropriate coordination of <strong>design</strong> decisions. Emphasis is given to potential shortcomings of<br />
using <strong>game</strong> theory as a mathematical foundation for enterprise <strong>design</strong> and integration.<br />
The contributions and relevance of the work are listed and discussed in this chapter, and<br />
areas of future work are identified.
5.1 CRITICAL REVIEW OF THE THESIS: DISCUSSION OF RESEARCH<br />
QUESTIONS AND SHORTCOMINGS.<br />
Presently, it is recognized that an efficacious integration of the various enterprise activities<br />
improves the competitiveness of manufacturing organizations. The fundamental thesis developed<br />
in this work, which is <strong>based</strong> on a decision-<strong>based</strong> perspective, is that the integration of the<br />
enterprise <strong>design</strong> processes, namely, process <strong>design</strong>, manufacturing process <strong>design</strong> and<br />
organization <strong>design</strong>, should be <strong>based</strong> on an appropriate formulation and solution of individual<br />
<strong>design</strong> decisions. Accordingly, two key objectives are posed in this work:<br />
1. To establish a mathematically supported cooperative framework that enhances a practical,<br />
effective and efficient integration of the enterprise <strong>design</strong> processes, and<br />
2. To provide an appropriate model to support formulation and solution of <strong>design</strong> problems,<br />
consistent <strong>with</strong> the aforementioned cooperative framework.<br />
These fundamental objectives are embodied by three central research questions:<br />
1. How can an enterprise <strong>design</strong> process be modeled mathematically in the<br />
context of decision-<strong>based</strong> <strong>design</strong>?<br />
2. How can enterprise decisions be coordinated through a <strong>design</strong> formulation<br />
<strong>based</strong> on a <strong>game</strong> <strong>theoretical</strong> representation of the enterprise <strong>design</strong> process?<br />
3. How can the <strong>design</strong> problems be solved to enhance the coordination of<br />
enterprise <strong>design</strong> decisions?<br />
202
203<br />
In answering Question 1, the enterprise activity is perceived as a decision making process,<br />
and therefore it is necessary to use a model that captures the dynamics of decisions.<br />
Mathematically modeling of a system behavior, where one decision-maker’s action depends on<br />
decisions by others, as occur in any manufacturing organization, is well established in <strong>game</strong><br />
theory. Thus, it is proposed to model the enterprise <strong>design</strong> process as a <strong>game</strong>, particularly as<br />
an extensive <strong>game</strong>. This assertion has been extensively discussed in Chapter 2 and applied in<br />
the integrated product and process <strong>design</strong> of the absorber/evaporator module presented in<br />
Chapter 4. Organization <strong>design</strong> problems are not explicitly integrated in this case study.<br />
However, it is apparent that the proposed framework for coordinating decisions has important<br />
implications for organization <strong>design</strong>, particularly in the setting of reward and information systems,<br />
as discussed in Section 3.2. Such a <strong>design</strong> should promote the achievement of a particular kind<br />
of equilibrium depending on the protocol that is used. In other words, it is necessary to prepare<br />
the appropriate field for the specific <strong>game</strong> that is played by the various agents that comprise the<br />
enterprise.<br />
It is noted that <strong>game</strong> theory has some shortcomings as a foundation for enterprise <strong>design</strong>.<br />
Particularly, a <strong>game</strong> <strong>theoretical</strong> model is only a descriptive model and not a normative one. That<br />
is, <strong>game</strong> theory identifies optimal strategies for the players in a <strong>game</strong> but does not address the<br />
issue of how to structure the <strong>game</strong>. This is akin to asking (quoted from Kreps, 1990): “if the<br />
players in the <strong>game</strong> are so smart, why don’t they change the rules and play a <strong>game</strong> where they
can do better?” In this sense, it is descriptive and does not provide a normative theory for<br />
enterprise management and organization <strong>design</strong>.<br />
A second problem that poses the use of <strong>game</strong> theory is that it is <strong>based</strong> on the assumption<br />
of rationality of individuals that have transitive preferences. Kenneth Arrow has shown that this<br />
condition is not always sufficient for group decision making (see Hazelrigg, 1996). Although the<br />
group consists of rational individuals, the group itself may be irrational. In general, the more<br />
members that comprise the group and the more alternatives among which the group must<br />
choose, the more likely it is that the group will be irrational. Most manufacturing organizations<br />
consist of several individuals pursuing the selection of a preferred course of action from among a<br />
set of alternatives. In such a case and in the absence of an externally imposed preference<br />
ordering, it is a virtual certainty that the <strong>design</strong> team will exhibit intransitive preferences. This<br />
condition will lead to a course of action that is guided by the intransitive preferences of the<br />
group. In short, the decision might not be rational. Thus, Arrow’s Impossibility Theorem (AIT)<br />
shows the suboptimality of many of the answers the proposed <strong>approach</strong> gets. Hence, why is<br />
the proposed <strong>approach</strong> useful?<br />
204<br />
In answering this question it is necessary to examine the requirements that the AIT<br />
imposes to an organization. First, it is necessary for the organization to state explicitly the utility<br />
against which the <strong>design</strong> should be judged. Second, the organization must create a set of<br />
incentives and rewards that is in the best interest of each individual to make use of the stated<br />
utility measure. However, uncertainty and bounded rationality prevents the organization to have
a single utility function to be used by any individual. That is, the fundamental goal of any<br />
organization is to profit; thus, it is the only utility measure that could be used to this aim.<br />
However, because each individual has different information, what one would be perceive as an<br />
optimal decision to maximize profit, might not be perceived as such by others. Also, the profit<br />
that a decision causes depends on future events that cannot be predicted <strong>with</strong> certainty.<br />
Consider the case example of Chapter 4. Two system goals are posed: to minimize cost and to<br />
minimize cycle time of production. The reason to do so is that it is difficult to predict accurately<br />
how does a reduction of the cycle time benefit sales and therefore it is difficult to forecast the<br />
real value of cycle time reduction for the organization in units of dollars.<br />
Therefore, because of uncertainty and bounded rationality, the problem changes from<br />
choosing the right course of action which maximizes profit to finding a way of calculating, very<br />
approximately, where a good course of action lies. The logic on using <strong>game</strong> theory for<br />
coordinating decisions is that such a course of action usually lies on a point of equilibrium of the<br />
different individuals’ preferences. Game theory provides the mathematical tools to estimate<br />
such equilibrium. However, particularly at early stages, such equilibrium may not exist because<br />
individuals begin <strong>with</strong> different expectations and are trying to learn what to expect of others.<br />
Hence, it would seem clear that it is necessary to complement this <strong>approach</strong> <strong>with</strong> other means.<br />
Thus, the proposed framework does not exclude other ways of enterprise integration, such as<br />
team meetings or CIM, that help us to better estimate the preferences and motivations of other<br />
members of the enterprise.<br />
205
A third problem on the use of <strong>game</strong> theory for coordination of decisions is the need for<br />
precise protocols and mathematical models. Because human behavior cannot be reduced to a<br />
set of equations, manufacturing organizations are only partially amenable to mathematical<br />
analysis and modeling. Moreover, intuition and experience drive the individual’s sense of what<br />
is satisfactory. Thus, the equilibrium that arises <strong>with</strong> a particular group of individuals might be<br />
different to the one achieved <strong>with</strong> a different group even if all other conditions remain the same.<br />
However, by adopting a common model for decision support and formulating the resulting<br />
<strong>design</strong> problem using a common decision model it is expected this problem to be ameliorated.<br />
It has to be noted that the use of a rigorous decision model should not aim at suppressing<br />
intuition and experience but to provide a foundation to support them and therefore to improve<br />
the decision making process. Thus, the answer to the research question 2 that is explored in this<br />
thesis is that enterprise integration can be effectively pursued by formulating the <strong>design</strong> problems<br />
using a non-cooperative-<strong>game</strong> formulation of the <strong>design</strong> problems. It is recommended the use<br />
of the Compromise Decision Support Problem (DSP) to embody the resulting <strong>design</strong><br />
formulations. The reason to do so is that, as previously argued, it is difficult to measure the<br />
profitability of a decision using a single scalar measure. Thus, a multiobjective decision model is<br />
needed. The Compromise DSP is a rigorous tool that fulfills this need.<br />
206<br />
The concern that motivates the third research question is that for most real-world<br />
problems it is not possible to find the optimum solution. Even the prediction of points of non-<br />
cooperative equilibrium is difficult in actuality. Therefore, it is characteristic of the enterprise
<strong>design</strong> activity that the final outcome is built from a search of a satisficing solution that is<br />
gradually refined. Thus, the idea of a fixed goal is inconsistent <strong>with</strong> our limited ability to<br />
determine the best point of equilibrium. The real results of our actions are to establish initial<br />
conditions for the next succeeding stage of action. What are the goals for a given stage are in<br />
fact criteria for choosing the initial conditions that we will leave to our successors. An attractive<br />
alternative in conducting this search of a solution is to leave as many alternatives as possible for<br />
future decision making, avoiding irreversible commitments that cannot be undone, through<br />
flexible solutions. A <strong>probabilistic</strong> <strong>design</strong> model is proposed as a formal means to develop<br />
flexible solutions to meet a ranged set of <strong>design</strong> requirements, ideally facilitating the convergence<br />
of the group decision making process to an equilibrium that is satisfactory for the individuals that<br />
comprise the group.<br />
In summary, the proposed framework developed in this thesis for the integration of the<br />
enterprise activities <strong>based</strong> on the formulation of individuals <strong>design</strong> problems is <strong>based</strong> on the<br />
following elements:<br />
?? A <strong>game</strong> <strong>theoretical</strong> representation of the enterprise activity<br />
?? A common paradigm for decision support which builds upon the strengths of operations<br />
research, engineering <strong>design</strong> and systems theory.<br />
?? An enterprise <strong>design</strong> method <strong>based</strong> on a philosophy of empowerment through decision<br />
making.<br />
?? A common decision model, the Compromise DSP.<br />
207
?? A <strong>probabilistic</strong> <strong>design</strong> model to attain flexible and robust decisions, and to handle<br />
uncertainty and imprecision.<br />
These elements are merged in this work into a consistent framework for enterprise <strong>design</strong> and<br />
integration, as is explained in Section 3.6. This framework has been exercised in a case study in<br />
Chapter 4: the integrated product and process <strong>design</strong> of an absorber/evaporator module for a<br />
family of absorption chillers. Even when the results of the case study are encouraging, the<br />
effectiveness of the <strong>approach</strong> for larger and more complex problems has yet to be proved.<br />
5.2 CONTRIBUTIONS AND RELEVANCE OF THE THESIS<br />
A fundamental foundation of this work is the tenet of enterprise <strong>design</strong> as a scientific discipline.<br />
In Chapter 1 it is argued that a scientific <strong>design</strong> practice should incorporate:<br />
1. Representation of <strong>design</strong> problems.<br />
2. A theory of structure and <strong>design</strong> organization.<br />
3. A utility theory as a logical framework for rational choice among alternatives, and tools for<br />
actually making the computations.<br />
4. The body of techniques for actually deducing which of the available alternatives is the<br />
optimum.<br />
5. Algorithms to search for solutions.<br />
This work provides a contribution on all these elements. First, it provides a <strong>game</strong> <strong>theoretical</strong><br />
representation of the enterprise <strong>design</strong> process <strong>based</strong> on viewing the enterprise activity as a<br />
208
group decision making process. This representation focuses attention on issues of cooperative<br />
dynamics <strong>with</strong>in an organization, so that the “physics” of interaction between the agents that<br />
comprise it come to the fore. As commented in Chapter 2, if the mechanics of interaction<br />
between individuals matter for achieving enterprise integration, this representation is by itself a<br />
relevant contribution.<br />
This work contributes to a theory of structure and enterprise <strong>design</strong> organization as it<br />
poses a mathematically <strong>based</strong> theory for collaboration of individuals <strong>based</strong> on the appropriate<br />
formulation of their <strong>design</strong> problems. This concept is built upon the constructs of <strong>game</strong> theory,<br />
operations research and systems theory. By using a hybrid paradigm for decision support that is<br />
domain independent, this theory capitalizes and integrates the structure of engineering <strong>design</strong><br />
<strong>based</strong> on modeling, analysis and synthesis.<br />
On the need of a utility theory and the body of techniques for the synthesis of <strong>design</strong>s, the<br />
application of <strong>game</strong> theory provides clear objectives <strong>based</strong> on the notion of Nash and Pareto<br />
equilibrium. Depending on the kind of solution that is desired, <strong>game</strong> theory provides the basic<br />
theory that is required for identifying such equilibrium.<br />
The proposed framework provides also a solution technique that is <strong>based</strong> on theory of<br />
optimization that translates into specific algorithms. Thus, it provides a foundation for rigorously<br />
implementing enterprise <strong>design</strong> and integration.<br />
209<br />
A secondary, but also important contribution of this work, is the development of a method<br />
for modeling production systems for product <strong>design</strong> purposes. Simulation <strong>based</strong> models are too
cumbersome to be used during product optimization at early <strong>design</strong> stages, thus they do not<br />
facilitate an integrated product/process development. Instead, it is proposed to model the<br />
system as a response surface network, in which the response surfaces of individual processing<br />
stations relate product characteristics <strong>with</strong> production time and processing variability. These<br />
response surfaces are the modules for the construction of a queuing model. This method allows<br />
a quickly evaluation of the impact of changes of product variables in the production system, and<br />
enhances the practical implementation of the integration of product <strong>design</strong> and manufacturing<br />
process <strong>design</strong> using a non-cooperative protocol because it greatly facilitates building the<br />
required rational reaction sets.<br />
Finally, the case study provides specific contributions as well. It deals <strong>with</strong> the important<br />
issue of developing <strong>design</strong> methods for products that are produced in make-to-order systems.<br />
Low volume, uncertain demand and high product variety characterize these systems. These<br />
characteristics result in long lead times, large work in process and high manufacturing costs.<br />
The <strong>approach</strong> developed on this case study, which focus on standardization, modularity and<br />
robustness, is a steppingstone for further progress in this area.<br />
In summary, this thesis provides a formal means to coordinate decisions of the individuals<br />
that comprise a manufacturing system <strong>based</strong> on an <strong>approach</strong> that aims at framing the tradeoffs<br />
needed to coordinate decisions through controls that are consistent, not contrary to employee<br />
empowerment. Because the major opportunities for improvement of manufacturing enterprises<br />
210
lie at the interfaces (e.g., between product <strong>design</strong> and manufacturing), manufacturing<br />
organizations can greatly benefit from this <strong>approach</strong>.<br />
5.3 RECOMMENDATIONS FOR FUTURE WORK<br />
Even when the proposed framework seems to be consistent there are several issues that have<br />
arisen through the development of this work and remain open. Maybe the most challenging<br />
issue is that of organization <strong>design</strong>. Unfortunately, organization <strong>design</strong> is, in general, not<br />
amenable to mathematical support and analysis. The proposed <strong>approach</strong> for enterprise<br />
integration requires the mathematical formulation and solution of problems. Therefore, it is<br />
extremely important to develop mathematical-<strong>based</strong> methods for the modeling, analysis and<br />
synthesis of organization <strong>design</strong> problems.<br />
Another problem is the difficulty to deal <strong>with</strong> large and complex problems. Cooperative-<br />
<strong>based</strong> integration <strong>based</strong> on large models is not practically due to computational and cognitive<br />
limitations. Non-cooperative-<strong>based</strong> coordination seems to be more practical, but building<br />
rational reaction sets when several agents are involved in making a decision is also very<br />
cumbersome. This is a critical problem that has to be studied if non-cooperative <strong>game</strong> theory is<br />
intended to be the foundation of enterprise integration.<br />
In the specific issue of manufacturing systems modeling, the proposed response surface<br />
network modeling method has important shortcomings, particularly at the use of response<br />
surfaces to approximate behavior of process stations when mixed continuous/discrete product<br />
211
variables are present. Alternative approximation techniques are required to tackle this problem.<br />
Also, for large complex production systems, parametric decomposition-<strong>based</strong> models are<br />
usually not accurate.<br />
212
REFERENCES<br />
Agrell, P. J., 1994, "A Multicriteria Approach to Concurrent Engineering," Int. J. Production<br />
Economics, Vol. 34, pp. 99-113.<br />
Antonsson, E. K. and Otto, K. N., 1995, “Imprecision in Engineering Design”, ASME Journal<br />
of Mechanical Engineering, Vol. 117, pp. 25-32.<br />
Baird, S. P. and Leavy J. J., 1994, “Simulation Modeling Using PROMODEL for Windows,”<br />
Proceedings of the 1994 Winter Simulation Conference, IEEE, Piscataway, NJ, pp.<br />
527-532.<br />
260<br />
Balling, R., Borup, L., Busaker B., Chambers T., Davidson, D., Gritton, G., Free, J.,<br />
Parkinson, A., Talbert, J., Warren, D., 1998, OptdesX: A Software System for<br />
Optimal Engineering Design, Users Manual, Release 2.0.4., Design Synthesis Inc.,<br />
Provo, UT.<br />
Bascaran, E., Mistree, F. and Bannerot, R. B., 1987, “Compromise: An Effective Approach for<br />
Solving Multi-Objective Thermal Design Problems,” Engineering Optimization, Vol.<br />
12, No. 3, pp. 175-189.<br />
Bitran, G. R. and Tirupati, D., 1988, "Multiproduct Queueing Networks <strong>with</strong> Deterministic<br />
Routing: Decomposition Approach and the Notion of Interference," Management<br />
Science, Vol. 34, No. 1, pp. 75-100.<br />
Boothroyd, G., 1994, "Product <strong>design</strong> for manufacture and assembly," Computer Aided<br />
Design, Vol. 26, No. 7, pp. 505-520.<br />
Box, G. E. P. and Draper, N. R., 1987, Empirical Model-Building and Response Surfaces,<br />
John Wiley & Sons, New York.<br />
Browning, T. R., 1997, “Exploring Integrative Mechanisms <strong>with</strong> a View towards Design for<br />
Integration.” Advances in Concurrent Engineering (Ganesan, S., ed.), Technomic<br />
Publishing Co., pp. 83-89.
261
Buzacott, J. A., and Shanthikumar, J. G., 1993, Stochastic Models of Manufacturing<br />
Systems, Prentice Hall, Englewood Cliffs, NJ.<br />
Chang, K. J., Haftka, R. T., Giles, G. L. and Kao, P. J., 1993, “Sensitivity-<strong>based</strong> Scaling for<br />
Approximating Structural Response,” Journal of Aircraft, Vol. 30, pp. 283-287.<br />
Chen, W. and Yuan C., 1997, “A Probabilistic-Based Design Model for Achieving Flexibility in<br />
Design”, Design Theory and Methodology, Sacramento, CA, ASME Paper No. DTM-<br />
3882.<br />
Chen, W., Allen, J. K., Mistree, F. and Tsui, K. L., 1995, “Integration of Response Surface<br />
Method <strong>with</strong> the Compromise DSP in Developing a General Robust Design Procedure,”<br />
Advances in Design Automation, ASME DE, Vol. 82, No. 2, pp.485-492.<br />
Colquhoun, G. J., Baines R. W., and Crossley R., 1993, “A State of the Art of IDEF0,”<br />
International Journal of Computer Integrated Manufacturing,” Vol. 6, No. 4, pp.<br />
252-264.<br />
262<br />
Dixon, J. R., Duffey, M. R., Irani, R. K., Meunier, K. L. and Orelup, M. F., 1988, "A<br />
Proposed Taxonomy of Mechanical Design Problems," Proceedings ASME Computers<br />
in Engineering Conference, San Francisco, California, pp. 41-46.<br />
Doumeingts, G., Vallespir B., and Chen D., 1995, “Methodologies for Designing CIM systems:<br />
A Survey.” Computers in Industry, No. 25, pp. 263-280.<br />
Drevna, J. J. and Kasales, C. J., 1994, “Introduction to ARENA,” Proceedings of the 1994<br />
Winter Simulation Conference, IEEE, Piscataway, NJ, pp. 77-80.<br />
Fudenburg D. and Tirole, J., 1993, Game Theory, The MIT Press, London, U.K.<br />
Gane, C. and Sarson, T., 1979, Structured Systems Analysis-Tools and techniques.<br />
Prentice Hall, Englewood Cliffs, NJ.<br />
Goicoechea, A., Hansen, D. R. and Duckstein, L., 1982, Multiobjective Decision Analysis<br />
<strong>with</strong> Engineering and Business Applications, Wiley, NY.<br />
Hazelrigg, G. A., 1996, “Systems Engineering: A New Framework for Engineering Design”,<br />
Engineering Systems, ASME DSC, Vol. 60, pp. 39-46.<br />
Hopp, W. J. and Spearman, M. L., 1996, Factory Physics, IRWIN, Boston.
Hwang, C. L., Paidy, S. R., Yoon, K. and Masud, A. S. M., 1980, "Mathematical<br />
Programming <strong>with</strong> Multiple Objectives: a Tutorial," Comp. Oper. Res., No. 7, pp. 5-31.<br />
Ignizio, J. P., 1985, Introduction to Linear Goal Programming, Sage University Papers,<br />
Beverly Hills, California.<br />
263<br />
Joryz, R. H. and Vernadat, F. B., 1990, “CIM-OSA Part I: Total Enterprise Modeling and<br />
Function View,” International Journal of Computer Integrated Manufacturing, Vol.<br />
3, No. 3, pp. 144-156.<br />
Karandikar, H. M. and Mistree, F., 1991, "The Integration of Information from Design and<br />
Manufacture Through Decision Support Problems," Vol. 44, No. 11, Part 2, pp. 150-<br />
159.<br />
Kelton D. W., Sadowski R. P., and Sadowski D. A., 1998, Simulation <strong>with</strong> Arena,<br />
McGraw-Hill, Boston, MA.<br />
Lewis, K., 1997, An Algorithm for Integrated Subsystem Embodiment and System<br />
Synthesis, Ph. D. Dissertation, The George W. Woodruff School of Mechanical<br />
Engineering, Georgia Institute of Technology, Atlanta, GA.<br />
Lewis, K. and Mistree, F., 1997, "Collaborative, Sequential, and Isolated Decisions in Design,"<br />
ASME Design Theory and Methodology, (Shah, J. Ed.), New York, ASME97-<br />
DETC97/DTM-3883<br />
Lewis, R., Parkinson A., 1998, “Robust Optimal Design Using a Second Order Tolerance<br />
Model,” Research in Engineering Design, in press, copies available from Design<br />
Synthesis Inc. (Provo, UT).<br />
Little, D.D., 1992, “Are Threre ‘Laws’ of Manufacturing?” Manufacturing Systems:<br />
Foundations of World-Class Practice, J. A., Heim and W. D. Compton, Washington,<br />
D.C., National Academy Press, pp. 180-188.<br />
Lizotte S., and Chaib-draa B., 1997, “Coordination in CE Systems: An Approach Based on<br />
the Management of Dependences between Activities,” Concurrent Engineering:<br />
Research and Applications, Vol. 5, No. 4, pp. 367-376.
Lu, S. C. Y. and Subramanyam, A., 1988, “Computer-<strong>based</strong> Environment for Simultaneous<br />
Product and Process Design,” ASME Production Eng. Div. Publ. PED, No. 31, pp.<br />
35-46.<br />
Kateel, G., Kamath, M. and Pratt, D., 1996, "An Overview of CIM Enterprise Modeling<br />
Methodologies," IEEE Winter Simulation Conference Proceedings, Coronado, CA,<br />
pp. 1000-1007.<br />
264<br />
Kim C., Kim K., and Choi, I., 1993, “An Object Oriented Information Modeling Methodology<br />
for Manufacturing Information Systems,” Computers and Industrial Engineering, Vol.<br />
24, No. 3, pp. 337-353.<br />
Kirsch, U., 991, “Reduced Basis Approximations of Structural Displacements for Optimal<br />
Design,” AAIA Journal, Vol. 29, pp. 1751-1758.<br />
Kodiyalam, S. and Varderplaats, G. N., 1989, “Shape Optimization of 3D Continuum<br />
Structures Via Force Approximation Technique,” AIAA Journal, Vol. 27, pp. 1256-<br />
1263.<br />
Mesterton-Gibbons, M., 1992, An Introduction to Game-Theoretic Modeling, Addison-<br />
Wesley Publishing Company, Rewood City, CA.<br />
Metwalli, S.M., Shawki, G. S. A., Mokhtar M. O., Sheif, M. A. A., 1984, “Multiple Design<br />
Objectives in Hydrodynamic Journal Bearing Optimization,” ASME Journal of<br />
Mechanisms, Transmisions and Automation in Design, Vol. 106, No. 1, pp. 54-60.<br />
Mistree, F., Smith, W. F., Bras, B., Allen, J. K. and Muster, D., 1990, "Decision-Based<br />
Design: A Contemporary Paradigm for Ship Design," Transactions, Society of Naval<br />
Architects and Marine Engineers, New Jersey, pp. 565-597.<br />
Mistree, F., Hughes, O. F. and Bras, B. A., 1993, "The Compromise Decision Support<br />
Problem and the Adaptive Linear Programming Algorithm," Structural Optimization:<br />
Status and Promise, AIAA, Washington, D.C., pp. 247-286.<br />
Muster, D. and Mistree, F., 1988, “The Decision Support Problem Technique in Engineering<br />
Design,” International Journal of Applied Engineering Education, Vol. 4, No. 1, pp.<br />
23-33.<br />
Myerson, R.B., 1991, Game Theory: Analysis of Conflict, Harvard University Press,<br />
Cambridge, MA.
O’Sullivan, D., 1994, Manufacturing Systems Re<strong>design</strong>: Creating the Integrated<br />
Manufacturing Environment, Prentice Hall, Englewood Cliffs, NJ.<br />
Osyczka, A., 1984, Multicriterion Optimization in Engineering <strong>with</strong> FORTRAN Programs,<br />
Halsted Press, NY.<br />
265<br />
Pandya, K. V., 1995, “Review of Modeling Techniques and tools for Decision Making in<br />
Manufacturing Management,” IEE Proceedings on Science, Measurement, and<br />
Technology, Vol. 142, No. 5, pp. 371-377.<br />
Parkinson, A., 1993, “A General Approach for Robust Optimal Design,” J. of Mechanical<br />
Design, Vol. 115, No. 1, pp. 74-80.<br />
Peplinski, J. D., 1997, Enterprise Design: Extending Product Design to Include<br />
Manufacturing Process Design and Organization Design.” Ph.D. Dissertation, The<br />
George W. Woodruff School of Mechanical Engineering, Georgia Institute of<br />
Technology, Atlanta, Georgia.<br />
Peplinski, J. D., and Mistree, F., 1997, "A Decision-Based Approach to Enterprise Design and<br />
Integration," Advances in Concurrent Engineering, Rochester, Michigan, pp. 65-74.<br />
Peplinski, J. D., Allen, J. K. and Mistree, F., 1996, "Integrating Product Design <strong>with</strong><br />
Manufacturing Process Design using The Robust Concept Exploration Method," Design<br />
Theory and Methodology, Irvine, CA, ASME, Paper No. 96-DETC/DTM-1502.<br />
Kreps D. M., 1990, A Course in Microeconomic Theory, Princeton University Press, NJ.<br />
Rao, J.R., Badhrinath K., Pakala, R., Mistree, F., 1997, “A Study of Optimal Design Under<br />
Conflict Using Models of Multi-Player Games,” Engineering Optimization, Vol. 28,<br />
No. 1-2, pp. 63-94.<br />
Rao, S. S. and Freiheit, T. I., 1991, "A Modified Game Theory Approach to Multiobjective<br />
Optimization," Journal of Mechanical Design, Vol. 113, September, pp. 286-291.<br />
Rao, S.S. and Hati, S. K., 1984, “Optimum Design of Shock and Vibration Isolation Systems<br />
Using Game Theory,’ Engineering Optimization, Vol. 4, No. 4, pp. 215-226.<br />
Reiser, M., and Kobayashi, H., 1974, “Accuracy of Diffusion Approximations for Some<br />
Queueing Systems,” IBM J. Res. Develop., Vol. 18, pp. 110-124.
Rosenman, M. A. and Gero, J. S., 1983, “Pareto Optimal Serial Dynamic Programming,”<br />
Engineering Optimization, Vol. 6, pp. 161-175.<br />
Shanthikumar, J. G., and Buzacott, J. A., 1981, “Open Queueing Network Models of Dynamic<br />
Jobshops,” Internat. J. Production Res., Vol. 19, No. 3, pp. 255-266.<br />
266<br />
Shoham, Y. and Tennenholtz, M., 1992, “On the Synthesis of Useful Social Laws for Artificial<br />
Agent Societies,” Proceedings of the 10th. National Conference on Artificial<br />
Intelligence, Menlo Park, CA, pp. 276-281.<br />
Shupe, J. A. and Mistree, F., 1987, ‘Compromise: An Effective Approach for the Design of<br />
Damage Tolerant Structural Systems,” Computers and Structures, Vol. 27, No. 3, pp.<br />
407-415.<br />
Simon, H.A., 1996, The Sciences of the Artificial, The MIT Press, Cambridge, Mass.<br />
Sobieszczanski-Sobieski J. and Haftka, R. T., 1997, “Multidisciplinary Aerospace Design<br />
Optimization: Survey of Recent Developments,” Structural Optimization, Vol. 14,<br />
p.1-23.<br />
Thurston, D. L. and Liu, T., 1991, “Design Evaluation of Multiple Attributes Under<br />
Uncertainty,” International Journal of Systems Automation: research and Applications,<br />
Vol. 1, No. 2, pp. 143-159.<br />
Toropov, V. and Markine V. L., 1996, “The Use of Simplified Numerical Models as Mid-<br />
Range Approximations,” Proceedings of the 6th. AIAA/NASA/ISSMO Symposium on<br />
Multidisciplinary Analysis and Optimization, Bellevue, WA, pp. 941-951.<br />
Unger, E. R., Hutchison, M. G., Rais-Rohani, M., Haftka, R. T. and Grossman, B., 1992,<br />
“Variable-complexity Multidisciplinary Design of a Transport Wing,” International<br />
Journal of System Automation: Research and Applications, Vol. 2, pp. 87-113.<br />
Vadde, S., Krishnamachari, R. S., Allen, J. K. and Mistree, F., 1994, "The Bayesian<br />
Compromise Decision Support Problem for Hierarchical Design Involving Uncertainty,"<br />
ASME Journal of Mechanical Design, Vol. 116, pp. 388-395.<br />
Vanderpooten, D. and Vincke, P., 1989, “Description and Analysis of Some Representative<br />
Interactive Procedures,” Mathematical Computer Modelling, Vol. 12, No. 10-11, pp.<br />
1221-1238.
Vernadat, F. B., 1996, “Enterprise Integration: On Business Process and Enterprise Activity<br />
Modelin,” Concurrent Engineering: Research and Applications, Vol. 4, No. 3, pp.<br />
219-228.<br />
Vincent, T. L., 1983, "Game Theory as a Design Tool," ASME Journal of Mechanisms,<br />
Transmissions, and Automation in Design, Vol. 105, June, pp. 165-170.<br />
Vliegen, J. J. W. and Van Mal, H. H., 1990, "Rational Decision Making: Structuring of Design<br />
Meetings," IEEE Trans. Eng. Managem., Vol. 37, No. 3, pp. 185-190.<br />
Wallace, D. R. and Jakiela, M., 1995, “Design Search under Probabilistic Specifications using<br />
Genetic Algorithms” Computer-Aided Design (in press).<br />
Webster’s Collegiate Dictionary, Riverside Publishing Co., Boston, MA, 1984.<br />
Whitt, W., 1983a, “The Queueing Network Analyzer,” Bell System Tech. J., Vol. 62,<br />
No. 9, pp. 2779-2815.<br />
Whitt, W., 1983b, “Performance of the Queueing Network Analyzer,” Bell System Tech. J.,<br />
Vol. 62, No. 9, pp. 2817-2843.<br />
Whitt, W., 1995, “Variability Functions for Parametric-decomposition Approximations of<br />
Queueing Networks,” Management Science, Vol. 41, No. 10, pp. 1704-1715.<br />
Wright, R., 1989, Systems Thinking: A Guide to Managing in a Changing Environment,<br />
SME, Dearborn, Mich.<br />
Zhou, Q.-J., Allen, J. K. and Mistree, F., 1992, "Decisions under Uncertainty: The Fuzzy<br />
Compromise Decision Support Problem," Engineering Optimization, Vol. 20, pp. 21-<br />
43.<br />
267
APPENDIX A<br />
THE PARAMETRIC DECOMPOSITION APPROACH<br />
Queuing networks have been used to model the performance of a variety of production<br />
systems. Since exact results exist for only a limited class of networks, the decomposition<br />
methodology has been used extensively to obtain approximate results. In this Appendix it is<br />
given a brief overview of the parametric decomposition <strong>approach</strong>, which is a means to<br />
approximating the steady-state performance measures of open queueing networks <strong>with</strong> general<br />
arrival processes and general service-time distributions. This modeling <strong>approach</strong> is combined<br />
<strong>with</strong> a response surface technique in order to develop efficacious models of production systems<br />
in Chapter 4.<br />
212
Queueing networks have been used to model the performance of a variety of complex systems<br />
such as production job shops, flexible manufacturing systems, service sytems, etc. However,<br />
exacts results exist only for the limited class of product-form type (Jackson type) networks.<br />
The exact results encountered in the literature basically rely on the following assumptions (Whitt,<br />
1983a):<br />
?? Exponential distribution for service times.<br />
?? Homogeneous service requirement at all stations (the distribution of service time is<br />
independent of the product type).<br />
?? Priority discipline independent of customer class.<br />
?? Poisson arrivals.<br />
These are overly restrictive in many practical situations. Since the exact results do not extend to<br />
more general networks, it has led to the development of approximation schemes. These may be<br />
broadly described under the following four categories:<br />
?? Diffusion approximation.<br />
?? Mean value analysis<br />
?? Operational analysis<br />
?? Decomposition methods<br />
The first three <strong>approach</strong>es are not directly related to this work, and are mentioned here only for<br />
completeness. Bitran and Tirupati (1993) provide a detailed list of references on these<br />
schemes.<br />
213
Decomposition Methods are essentially attempts to generalize the notion of independence<br />
and product form results for Jackson networks to more general systems. This <strong>approach</strong> relies<br />
on two notions:<br />
1. The nodes can be treated as being stochastically independent, and<br />
2. two parameters (typically mean and variance) approximations provide reasonable accurate<br />
results.<br />
This <strong>approach</strong> was first proposed by Reiser and Kobayashi (1974) and has been used by<br />
several researchers in developing approximations. The work by Buzacott and Shanthikumar<br />
(1993) and Whitt (1983a, 1995) are representative of this <strong>approach</strong>. Essentially the method<br />
involves three steps:<br />
Step 1. Analysis of interaction between stations.<br />
Step 2. Decomposition of the network into subsystems of individual stations and their<br />
analysis.<br />
Step 3. Recomposition of the results to obtain the network performance.<br />
The first step is critical to the decomposition <strong>approach</strong>. The interaction between stations is<br />
analyzed by looking at the network as a composite of three basic processes:<br />
(a) superposition or merging,<br />
(b) flow through a queue or a station, and<br />
(c) splitting or decomposition.<br />
214
The superposition process (a) typically represents the arrivals at a station while (b) describes the<br />
departures from a station. The splitting and the merging processes model the product routing in<br />
the network.<br />
(a) The merging process considers superposition of arrivals at a station (say station j). Let ? ij<br />
and caij be the arrival rate and the square coefficient of variation 1 (scv) of interarrival time of the<br />
flow from station i to j (i = 0 for external arrivals). The resulting mixture is approximately<br />
described <strong>with</strong> an arrival rate ? j and scv caj as follows<br />
?<br />
j<br />
?<br />
N<br />
N ij<br />
? ? ij , ca j ? ?<br />
i?<br />
0<br />
i? 0?<br />
j<br />
?<br />
ca<br />
ij<br />
215<br />
(A.1)<br />
(b) The flow process describes approximately the scv of the departure stream from a station. If<br />
caj and csj are the scv of the interarrival and service time at station j <strong>with</strong> a utilization ? j the scv<br />
of the departures is approximately given by<br />
cd<br />
j<br />
j<br />
ij<br />
2 ? ? j ? ca j<br />
2 ? ? cs ? 1 ?<br />
(A.2)<br />
When there is more than one machine m at a station, the following is a reasonable way to<br />
estimate cdj [2]:<br />
2<br />
2 ? j<br />
? 1?<br />
( 1?<br />
? j )( ca j ? 1)<br />
? ( cs ? 1)<br />
(A.3)<br />
m<br />
cd j<br />
j<br />
where m is the number of machines at the station. Notice that this reduces to (A.2) when m=1.<br />
(c) In the decomposition process a product stream <strong>with</strong> arrival rate ? j and scv of cdj is split<br />
into a number of substreams, each stream representing the flow from station j to other stations in
the network. The routing is assumed to be Markovian. If pi is the probability that a job would<br />
follow pat l , then the substream i is characterized by the following parameters:<br />
216<br />
? ij ? p i ? j , cd ji ? p i cd j ? 1? p i (A.4)<br />
A rationale for these approximations can be found in Whitt [19]. He examines two methods -<br />
the stationary interval method and the asymptotic method in approximating a point process by a<br />
renewal process. These methods are used to determine the moments of the renewal process<br />
that approximates the point process. For example, the asymptotic method to approximate the<br />
superposition process leads to (A.1). Also shows that (A.2) can be interpreted as an<br />
approximation resulting from the stationary interval method. This explanation involves the use of<br />
an approximation for the mean waiting time in a GI/G/1 2 queue. Expression (A.4) implicitly<br />
assumes that interdeparture intervals of the product stream are iid <strong>with</strong> mean 1/? j and scv of<br />
cdj. (Note that caji = cdji) Whitt (1983a) explain that (A.3) is in fact exact under the<br />
assumption of Markovian routing.<br />
The interactions among stations are reflected by the scv of the arrivals at each station.<br />
Combining the approximations (considering m=1) for the three basic processes described<br />
above leads to the following linear systems:<br />
N<br />
? i ? ? r0i ? ? ? jr ji , i = 1, 2,..., N<br />
(A.5)<br />
j?1<br />
1<br />
The square coefficient of variation, scv, of a random variable is defined as the ratio of its variance to the<br />
square of the mean<br />
2<br />
Kendall’s notation which characterizes a station. In this case a one-machine station, <strong>with</strong> general<br />
exponentially distributed interarrival times, and generally dis tributed process times.
?<br />
where<br />
i<br />
ca<br />
i<br />
N<br />
N<br />
2<br />
2<br />
? j ? ? j ? r jica<br />
j ? ? ? r0i<br />
? ? ? jr<br />
ji ? j rjics<br />
j ? 1?<br />
r ji ? ,<br />
217<br />
2 ? ? 1 ?<br />
i = 1,2,..., N (A.6)<br />
j?<br />
1<br />
? = arrival rate into the system,<br />
? j = net arrival rate at station j,<br />
? j = utilization at station j<br />
caj = scv of arrivals<br />
csj = scv of service at station j,<br />
j?<br />
1<br />
Notice that (A.5) consists of standard flow balance equations to determine the net arrivals at<br />
each station. The internal flows ? ij are given by ? i?rij. (A.6) is the approximation to compute<br />
the scv of arrivals at each station.<br />
In Step 2, each station is analyzed <strong>based</strong> on the partial information obtained in Step 1.<br />
Shanthikumar and Buzacott examine M/G/1 and GI/G/1 approximation to evaluate the station at<br />
the performance. Whitt (1983a, 1983b) also examines the GI/G/m queue. In particular the<br />
GI/G/1 and the G/G/m approximations require only the mean and scv of the interarrival and<br />
service times.<br />
Finally in Step 3 these results are synthesized and performance measures for the network<br />
are estimated. This procedure is straightforward for mean queue lengths, and lead times.<br />
In developing the queueing network analyzer (QNA) Whitt (1983a) essentially used an<br />
<strong>approach</strong> similar to the one described above in analyzing the interactions between the nodes.<br />
The superposition or the merging process has been modified as follows:
chj = wcaj + 1 -w<br />
where chj is the modified scv of the merged stream, caj is determined by (A.1) and w is a<br />
weighting function that depends on the station utilization ? j and ? ij. In the QNA the weight w is<br />
determined as follows:<br />
2<br />
? 1 ? 4(<br />
1 ? ? ) ( ? ) ?<br />
m ?<br />
? 1<br />
where ? ? ?<br />
? 1?<br />
218<br />
? 1 w ?<br />
j ? 1<br />
(A.7)<br />
2<br />
kj<br />
2<br />
k j<br />
(A.8)<br />
This modification is motivated by the observation that neither the asymptotic method (which<br />
leads to Equation A.1) nor the stationary interval <strong>approach</strong> by itself perform well over a wide<br />
range of ca for approximating the superposition process. (A.7) above is a hybrid approximation<br />
of both <strong>approach</strong>es. In the QNA a modified form of (A.2), similar to (A.3), is used to account<br />
for multiple servers at each station. The splitting process is <strong>based</strong> on Markovian routing. The<br />
estimation of station performance measures in Step 2 is also modified to provide for multiple<br />
servers at each station. However, the approximations continue to use the first two moments of<br />
the arrival and service distributions.<br />
In Witt (1995) it is proposed to replace the variability parameters of arrival processes<br />
by variability functions that depend on the utilization ?. These variability functions provide an<br />
improved framework for parametric-decomposition approximations.
Bitran and Tirupati (1988) propose a modification to the parametric decomposition<br />
<strong>approach</strong> for the analysis of multiproduct queueing networks <strong>with</strong> deterministic routings, which<br />
can be viewed as a generalization of the one used by Shanthikumar and Buzacott and Whitt. It<br />
is <strong>based</strong> on the notion that the scv of product departures can be approximated as the sum of<br />
two terms, one that represents the influence of station congestion and service time and one that<br />
represents the interference due to the presence of multiple products. Since the estimation of the<br />
interference effect is difficult to evaluate exactly, they propose three approximations that are<br />
asymptotically exact. The approximation used in this work is <strong>based</strong> on the notion that the<br />
aggregation of the arrival streams of a large number of products may be approximated by a<br />
Poisson process, resulting in the following scv of departure for product i at any station:<br />
where<br />
i<br />
i<br />
i<br />
? p ? ( p ) ca ?<br />
i<br />
i<br />
i<br />
219<br />
cd ? p cd ? ( 1 ? p ) 1?<br />
(A.9)<br />
pi = ? i?? ?<br />
? i = arrival rate of product I<br />
????total arrival rate at the station = ? ? i<br />
cdi = scv of departure for product i at station<br />
cd = scv of departure of station<br />
cai = scv of arrival of product i at station<br />
Details on this approximation can be studied in Bitran and Tirupati (1988).<br />
Buzacott and Shanthikumar (1991), Whitt (1995), and Bitran and Tirupati (1988)<br />
report encouraging results <strong>based</strong> on experiments comparing their respective approximations<br />
<strong>with</strong> simulation.
APPENDIX B<br />
FORTRAN FILES<br />
220
221
B.1 SAMPLE OF FORTRAN FILES FOR OptdesX<br />
222<br />
C FORTRAN FILE FOR OptdesX FOR PRODUCT DESIGN - BASELINE<br />
C FORMULATION<br />
c ANALYSIS SUBROUTINES FOR ABSORPTION CHILLER OPTIMIZATION WITH c<br />
OptdesX<br />
c Gabriel Hernandez<br />
C June 05, 1998<br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anapre(modelN)<br />
character*17 modelN<br />
modelN = "Absorber/Evaporator Design"<br />
return<br />
end<br />
c-----------------------------------------------------------------------<br />
--<br />
C LIST OF VARIABLES<br />
c A = Scaled value for number of absorber tubes [-2,2]<br />
c CAPACITY = refrigeration capacity (tons)<br />
c COST = Material Cost (dollars) of the Tubes of the<br />
Absorber/Evaporator<br />
c COSTA = Cost/ft of different Absorber Tubes (Vector)<br />
c COSTE = Cost/ft of different Evaporator Tubes (Vector)<br />
c DIE = Internal diameter (in) of evaporator tubes (vector)<br />
c DOA = External diameter (in) of absorber tubes (vector)<br />
c DOE = External diameter (in) of evaporator tubes (vector)<br />
c DPE = pressure drop of chilled water at evaporator<br />
c DPE_MIN = minimum DPE required<br />
c DR = Root diameter (in) of evaporator tubes (vector)<br />
c E = Scales value for number of tubes of the evaporator [-2,2]<br />
c FH = Height of fin (in) of evaporator tubes (vector)<br />
c FPI = Number of fins/in of evaporator tubes (vector)<br />
c FW = Width of fin (in) of evaporator tubes (vector)<br />
c I = Integer variable of evaporator tubes selection<br />
c J = Integer variable of absorber tubes selection<br />
c L = Length of tube in ft<br />
c LT = Scaled value for length of tubes [2,2]<br />
c NABSTA= Number of absorber tubes options<br />
c NAT = Number of tubes of the Absorber<br />
c NEVPTA= Number of evaporator tubes options<br />
c NET = Number of tubes of the Evaporator<br />
c NRA = Number of rows of the absorber<br />
c NATR= Number of tubes/row of the absorber<br />
c THKA= Thickness (in) of tubes of the absorber<br />
c THKE= Thickness (in) of tubes of the evaporator<br />
c TONS= Required capacity<br />
c UAA = Heat Transfer Exchange Coefficient multiplied by<br />
c area of heat exchange for the absorber
223<br />
c UAANOM= UAA required<br />
c UAE = Heat Transfer Exchange Coefficient multiplied by<br />
c area of heat exchange for the evaporator<br />
c UAENOM= UAE required<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAFUN<br />
c sample analysis routine for chiller <strong>design</strong><br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anafun<br />
c ... variables passed between OptdesX and these subroutines<br />
DOUBLE PRECISION DLT,DE,DA,DCAP,DCOST,DUAE,DUAA,DDP,<br />
+ DSELE,DSELA,DL,DE,DA,<br />
+ DTONS,DUAENOM,DUAANOM<br />
c ----------------------------------------------------------------------<br />
-----<br />
INTEGER I,J,NEVPTA,NABSTA,<br />
+ NRA,NATR<br />
PARAMETER (NEVPTA=4,NABSTA=4)<br />
REAL A,<br />
+ COST,COSTE(NEVPTA),COSTA(NABSTA),<br />
+ DIE(NEVPTA),DOA(NABSTA),DOE(NEVPTA),DR(NEVPTA),<br />
+ E,<br />
+ FH(NEVPTA),FPI(NEVPTA),FW(NEVPTA),<br />
+ L,LT,<br />
+ NAT,NET,<br />
+ THKA(NABSTA),THKE(NEVPTA),TONS,<br />
+ UAA,UAANOM,UAE,UAENOM<br />
C -----------------------------------------------------------<br />
C TUBES ALTERNATIVES<br />
DATA (DOE(I), I=1,NEVPTA) /0.625,0.75,0.875,1.0/,<br />
& (THKE(I), I=1,NEVPTA) /0.025,0.025,0.02,0.025/,<br />
& (FPI(I), I=1,NEVPTA) /30,30,30,30/,<br />
& (DIE(I), I=1,NEVPTA) /0.444,0.569,0.694,0.819/,<br />
& (DR(I), I=1,NEVPTA) /0.5,0.625,0.75,0.875/,<br />
& (FW(I), I=1,NEVPTA) /0.02,0.02,0.02,0.02/,<br />
& (FH(I), I=1,NEVPTA) /0.05,0.05,0.05,0.05/,<br />
& (COSTE(I),I=1,NEVPTA) /6.0,6.5,7.0,7.5/,<br />
& (DOA(I), I=1,NABSTA) /0.625,0.75,0.875,1.0/,<br />
& (THKA(I), I=1,NABSTA) /0.03,0.03,0.03,0.03/,<br />
& (COSTA(I),I=1,NABSTA) /3.25,3.50,3.75,4.0/<br />
c...input scalar AV values<br />
call avdsca(DLT,'length')<br />
call avdsca(DE,'NET')<br />
call avdsca(DA,'NAT')<br />
call avdsca(DSELE,'EvapTube')<br />
call avdsca(DSELA,'AbsTube')<br />
call avdsca(DTONS,'Tons')<br />
c ... Scaled Length
LT=REAL(DLT)<br />
c ... Scaled number of tubes<br />
E=REAL(DE)<br />
A=REAL(DA)<br />
c<br />
c...Selection of Tubes<br />
c ... For the evaporator<br />
I=INT(DSELE)<br />
IF(I.LT.1) I=1<br />
IF(I.GT.NEVPTA) I=NEVPTA<br />
c ... For the absorber<br />
J=INT(DSELA)<br />
IF(J.LT.1) J=1<br />
IF(J.GT.NABSTA) J=NABSTA<br />
c<br />
c ---------------------------------------------------------------<br />
c... value of constraints<br />
TONS=DTONS<br />
c... linear relationship between required capacity and UA required<br />
UAENOM=1372.183*TONS-34091.8<br />
UAANOM=1021.806*TONS-38163.0<br />
DPE_MIN=23.0<br />
C ****************************************************************<br />
c... Transform scaled variables to actual values<br />
L=1.75*(LT+2)+17.0<br />
NET=115*(E+2)+440<br />
NAT=115*(A+2)+440<br />
DL=L<br />
DE=NET<br />
DA=NAT<br />
c -----------------------------------------------------------------<br />
C...compute heat transfer functions<br />
C...EVAPORATOR<br />
c ... pass TONS,DOE,THKE,FPI,DIE,DR,FW,FH,L,NET<br />
C ... get UAE and DPE<br />
CALL EVAP(TONS,DOE(I),THKE(I),FPI(I),DIE(I),DR(I),<br />
& FW(I),FH(I),L,NET,UAE,DPE)<br />
C...ABSORBER<br />
C NOTE: THE NUMBER OF ROWS IS FIXED AS 12<br />
NRA=12<br />
NATR = INT(NAT/NRA)<br />
c ... pass TONS,DOA,THKA,L,NRA,NATR<br />
c ... get UAA and CAPACITY<br />
CALL ABSE(TONS,DOA(J),THKA(J),L,NRA,NATR,UAA,CAPACITY)<br />
C ------------------------------------------------------------<br />
c<br />
c ... Compute cost of tubes<br />
COST =COSTE(I)*(L*NET)+COSTA(J)*(L*NAT)<br />
c<br />
c ************************************************************<br />
C Value of dependent variables for optimization driver<br />
224
DUAA=UAA<br />
DUAE=UAE<br />
DDP=DPE<br />
DCAP=CAPACITY<br />
DCOST=COST<br />
DUAENOM=UAENOM<br />
DUAANOM=UAANOM<br />
C ***************************************************************<br />
c...output scalar AF values<br />
call afdsca(DCAP,'capacity')<br />
call afdsca(DUAE,'UAevp')<br />
call afdsca(DUAA,'UAabs')<br />
call afdsca(DDP,'DP')<br />
call afdsca(DCOST,'Cost')<br />
call afdsca(DL,'Length')<br />
call afdsca(DE,'Evap')<br />
call afdsca(DA,'Abs')<br />
call afdsca(DUAENOM,'UAENOM')<br />
call afdsca(DUAANOM,'UAANOM')<br />
return<br />
end<br />
225<br />
C FORTRAN FILE FOR PRODUCTION SYSTEM DESIGN WITH<br />
C MODIFIED DEVIATION FUNCTION - BASELINE FORMULATION<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAPRE<br />
c preprocessing routine<br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anapre(modelN)<br />
character*17 modelN<br />
modelN = "Chiller Design"<br />
return<br />
end<br />
c-----------------------------------------------------------------------<br />
-------<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAFUN<br />
c sample analysis routine for chiller <strong>design</strong><br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anafun<br />
c product variables<br />
DOUBLE PRECISION DCOST,<br />
+ DMatCOST,DU,DZ,DFN1,DFN2<br />
DOUBLE PRECISION DCT,DF,DMfgCOST,DMS1,DMS2,DMS3<br />
c production variables
INTEGER I,J,K,NABSTA,NEVPTA,A(8),E(8)<br />
PARAMETER (NEVPTA=4,NABSTA=4)<br />
REAL COSTE(NEVPTA),COSTA(NABSTA),<br />
+ FN1,FN2,<br />
+ L(8),LT(8),<br />
+ NAT(8),NET(8),NE(8),NA(8),<br />
+ PROB(8),<br />
+ TONS(8),TCOST,<br />
+ Z<br />
REAL CT,<br />
+ F,<br />
+ MS1,MS2,MS3,MfgCOST,<br />
+ TCT,<br />
+ U<br />
c -----------------------------------------------------------<br />
c ... PRODUCT DATA<br />
DATA (TONS(K), K=1,8) /600,700,800,900,1000,1100,1200,1300/,<br />
& (PROB(K), K=1,8) /0.05,0.15,0.16,0.05,<br />
+ 0.15,0.12,0.25,0.07/<br />
DATA (COSTE(I),I=1,NEVPTA) /6.0,6.5,7.0,7.5/,<br />
& (COSTA(I),I=1,NABSTA) /3.25,3.50,3.75,4.0/<br />
c ... baseline solution<br />
c & (L(I), I=1,8) /17.01,18.46,19.45,17.06,22.51,19.25,<br />
c + 23.82,23.95/,<br />
c & (E(I) ,I=1,8) /2,2,2,2,3,2,3,3/,<br />
c & (A(I) ,I=1,8) /2,2,2,2,2,2,2,2/<br />
c & (NE(I),I=1,8) /443,441,464,498,445,645,489,527/,<br />
c & (NA(I),I=1,8) /453,478,457,697,549,727,615,687/<br />
c Z11 = 46762.<br />
c Z12 = 65680.<br />
c Z21 = 703.6<br />
c Z22 = 447.7<br />
c<br />
c ... regular cooperative solution<br />
c DO 5 I=1,8<br />
c L(I) = 22.06<br />
c E(I) = 3<br />
c A(I) = 2<br />
c5 CONTINUE<br />
c DATA (NE(I),I=1,8) /440,440,440,440,460,511,557,608/,<br />
c & (NA(I),I=1,8) /440,440,446,504,554,612,662,719/<br />
c Z11 = 219817<br />
c Z12 = 239155.<br />
c Z21 = 709.3<br />
c Z22 = 458.3<br />
c<br />
C ... modifed cooperative solution<br />
c DO 5 I=1,8<br />
c L(I) = 19.25<br />
c E(I) = 2<br />
c A(I) = 2<br />
226
c5 CONTINUE<br />
c DATA (NE(I),I=1,8) /440,440,440,440,446,493,539,586/,<br />
c & (NA(I),I=1,8) /440,440,440,493,540,601,647,708/<br />
c Z11 = 219965.<br />
c Z12 = 239301.<br />
c Z21 = 697.<br />
c Z22 = 455.<br />
c<br />
c ... non cooperative solution<br />
DO 5 I=1,8<br />
L(I) = 19.25<br />
E(I) = 2<br />
A(I) = 2<br />
5 CONTINUE<br />
DATA (NE(I),I=1,8) /440,440,457,518,579,636,697,758/,<br />
& (NA(I),I=1,8) /440,446,518,590,647,719,791,863/<br />
Z11 = 216630.<br />
Z12 = 235456.<br />
Z21 = 614.<br />
Z22 = 494.<br />
c<br />
c...input scalar AV values<br />
c<br />
call avdsca(DMS1,'MS1')<br />
call avdsca(DMS2,'MS2')<br />
call avdsca(DMS3,'MS3')<br />
call avdsca(DF,'F')<br />
c<br />
c product analysis<br />
c<br />
c...compute scalar AF values<br />
c<br />
c...intermediate constants<br />
C **********************************************************<br />
C...compute functions<br />
MatCOST=0.0<br />
DO 100 K=1,8<br />
I=E(K)<br />
J=A(K)<br />
LT(K)=1.75*(L(K)-17.0)-2.0<br />
NET(K)=(NE(K)-440)/115.0-2<br />
NAT(K)=(NA(K)-440)/115.0-2<br />
MatCOST=PROB(K)*(COSTE(I)*(L(k)*NE(K))+<br />
+ COSTA(J)*(L(k)*NA(K)))+MatCOST<br />
100 CONTINUE<br />
MatCOST=MatCOST+52335.0<br />
c<br />
c<br />
c production analysis<br />
MS1=REAL(DMS1)<br />
MS2=REAL(DMS2)<br />
227
c<br />
MS3=REAL(DMS3)<br />
F =REAL(DF)<br />
CALL QNA(MS1,MS2,MS3,F,LT,NET,NAT,I,J,U,CT,MfgCOST)<br />
228<br />
COST=MatCOST+MfgCOST<br />
c ************************************************************<br />
C modified Pareto objective<br />
TCOST = 169810.0<br />
TCT = 318.1<br />
Z11 = Z11/TCOST-1.0<br />
Z12 = Z12/TCOST-1.0<br />
Z21 = Z21/TCT-1.0<br />
Z22 = Z22/TCT-1.0<br />
W1 = 0.5<br />
W2 = 0.5<br />
c<br />
c ... COST GOAL<br />
F1 = COST/TCOST-1.0<br />
FN1 = (F1-Z11)/(Z12-Z11)<br />
c ... CT Goal<br />
F2 = CT/TCT-1.0<br />
FN2 = (F2-Z22)/(Z21-Z22)<br />
c<br />
c ... Formulate Supercriterion<br />
S = (1-FN1)*(1-FN2)<br />
c ... Formulate a Pareto Optimal objective FC<br />
FC = W1*FN1+W2*FN2<br />
c ... Modified Objective Function<br />
Z = FC-S+1<br />
C EVALUATE DEVIATION FUNCTION USING AN ARCHIMEDEAN FORMULATION<br />
C WEIGHT OF EACH GOAL AT EACH PRIORITY LEVEL<br />
DCOST = COST<br />
DMatCOST = MatCOST<br />
DU = U<br />
DMfgCOST = MfgCOST<br />
DCT = CT<br />
DFN1 = FN1<br />
DFN2 = FN2<br />
DZ = Z<br />
C<br />
C<br />
************************************************************************<br />
****<br />
c...output scalar AF values<br />
call afdsca(DU,'Utilization')<br />
call afdsca(DCOST,'COST')<br />
call afdsca(DMatCOST,'MatCOST')<br />
call afdsca(DMfgCOST,'MfgCOST')<br />
call afdsca(DCT,'CT')<br />
call afdsca(DFN1,'FN1(COST)')
call afdsca(DFN2,'FN2(CT)')<br />
call afdsca(DZ,'Z')<br />
return<br />
end<br />
229<br />
C FORTRAN FILE FOR OptdesX for COOPERATIVE SOLUTION WITH MODIFIED<br />
C DEVIATION FUNCTION<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAPRE<br />
c preprocessing routine<br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anapre(modelN)<br />
character*17 modelN<br />
modelN = "Chiller Design"<br />
return<br />
end<br />
c-----------------------------------------------------------------------<br />
-------<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAFUN<br />
c sample analysis routine for chiller <strong>design</strong><br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anafun<br />
c product variables<br />
DOUBLE PRECISION DLT,DE,DA,DCOST,DFEAS,<br />
+ DSELE,DSELA,DL,DE,DA,<br />
+ DMatCOST,DU<br />
DOUBLE PRECISION DCT,DF,DMfgCOST,DMS1,DMS2,DMS3<br />
DOUBLE PRECISION DFN1,DFN2,DZ<br />
c production variables<br />
INTEGER I,J,K,NABSTA,NEVPTA<br />
PARAMETER (NEVPTA=4,NABSTA=4)<br />
REAL A,<br />
+ COSTE(NEVPTA),COSTA(NABSTA),<br />
+ E,FEASE,FEASA,FEAS,<br />
+ L,LT,<br />
+ NAT(8),NET(8),NE,NA,<br />
+ PROB(8),<br />
+ TONS(8),TCOST<br />
REAL CT,<br />
+ F,<br />
+ MS1,MS2,MS3,MfgCOST,<br />
+ TCT,<br />
+ U<br />
REAL F1,F2,FN1,FN2,FC,
+ S,<br />
+ W1,W2,<br />
+ Z11,Z12,Z21,Z22,Z<br />
c -----------------------------------------------------------------<br />
DATA (TONS(K), K=1,8) /600,700,800,900,1000,1100,1200,1300/,<br />
& (PROB(K), K=1,8) /0.05,0.15,0.16,0.05,<br />
+ 0.15,0.12,0.25,0.07/<br />
DATA (COSTE(I),I=1,NEVPTA) /6.0,6.5,7.0,7.5/,<br />
& (COSTA(I),I=1,NABSTA) /3.25,3.50,3.75,4.0/<br />
c -----------------------------------------------------------------<br />
c...input scalar AV values<br />
call avdsca(DLT,'length')<br />
call avdsca(DSELE,'EvapTube')<br />
call avdsca(DSELA,'AbsTube')<br />
c<br />
call avdsca(DMS1,'MS1')<br />
call avdsca(DMS2,'MS2')<br />
call avdsca(DMS3,'MS3')<br />
call avdsca(DF,'F')<br />
c<br />
c...compute scalar AF values<br />
c...intermediate constants<br />
C **********************************************************<br />
c ... product analysis<br />
LT=REAL(DLT)<br />
E=REAL(DE)<br />
A=REAL(DA)<br />
I =INT(DSELE)<br />
IF(I.LT.1) I=1<br />
IF(I.GT.NEVPTA) I=NEVPTA<br />
J =INT(DSELA)<br />
IF(J.LT.1) J=1<br />
IF(J.GT.NABSTA) J=NABSTA<br />
L = 1.75*(LT+2)+17.0<br />
DL= L<br />
C...compute functions<br />
MatCOST= 0.0<br />
FEAS = 16.0<br />
DO 100 K=1,8<br />
c ............. Find minimum number of tubes for EVAP<br />
CALL MINEVP(L,I,TONS(K),NE,FEASE)<br />
NET(K)=(NE-440)/115.0-2<br />
c ............. Find minimum number of tubes for ABS<br />
CALL MINABS(L,J,TONS(K),NA,FEASA)<br />
NAT(K)=(NA-440)/115.0-2<br />
FEAS = FEAS-FEASE-FEASA<br />
MatCOST=PROB(K)*(COSTE(I)*(L*NE)+COSTA(J)*(L*NA))+MatCOST<br />
100 CONTINUE<br />
MatCOST=MatCOST+52335.0<br />
c<br />
230
231<br />
c ----------------------------------------------------------------------<br />
-------<br />
c ... production analysis<br />
MS1=REAL(DMS1)<br />
MS2=REAL(DMS2)<br />
MS3=REAL(DMS3)<br />
F =REAL(DF)<br />
CALL QNA(MS1,MS2,MS3,F,LT,NET,NAT,I,J,U,CT,MfgCOST)<br />
c<br />
COST=MatCOST+MfgCOST<br />
c<br />
************************************************************************<br />
*****<br />
C modified Pareto objective<br />
TCOST = 160000.0<br />
TCT = 318.1<br />
Z11 = 214698/TCOST-1.0<br />
Z12 = 233959/TCOST-1.0<br />
Z21 = 769/TCT-1.0<br />
Z22 = 394/TCT-1.0<br />
W1 = 0.5<br />
W2 = 0.5<br />
c<br />
c ... COST GOAL<br />
F1 = COST/TCOST-1.0<br />
FN1 = (F1-Z11)/(Z12-Z11)<br />
c ... CT Goal<br />
F2 = CT/TCT-1.0<br />
FN2 = (F2-Z22)/(Z21-Z22)<br />
c<br />
c ... Formulate Supercriterion<br />
S = (1-FN1)*(1-FN2)<br />
c ... Formulate a Pareto Optimal objective FC<br />
FC = W1*FN1+W2*FN2<br />
c ... Modified Objective Function<br />
Z = FC-S+1<br />
c<br />
c<br />
************************************************************************<br />
*****<br />
DCOST = COST<br />
DMatCOST = MatCOST<br />
DFEAS = FEAS<br />
DU = U<br />
DMfgCOST = MfgCOST<br />
DCT = CT<br />
DFN1 = FN1<br />
DFN2 = FN2<br />
DZ = Z<br />
c<br />
C
232<br />
C<br />
************************************************************************<br />
****<br />
c...output scalar AF values<br />
call afdsca(DL,'Length')<br />
call afdsca(DFEAS,'PRODUCT_FEAS')<br />
call afdsca(DU,'Utilization')<br />
call afdsca(DCOST,'COST')<br />
call afdsca(DMatCOST,'MatCOST')<br />
call afdsca(DMfgCOST,'MfgCOST')<br />
call afdsca(DCT,'CT')<br />
call afdsca(DFN1,'FN1(COST)')<br />
call afdsca(DFN2,'FN2(CT)')<br />
call afdsca(DZ,'Z')<br />
return<br />
end<br />
c FORTRAN FILE FOR OptdesX for BUILDING RATIONAL REACTION SETS<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAPRE<br />
c preprocessing routine<br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anapre(modelN)<br />
character*17 modelN<br />
modelN = "Non-cooperative Chiller Design"<br />
return<br />
end<br />
c-----------------------------------------------------------------------<br />
-------<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAFUN<br />
c sample analysis routine for chiller <strong>design</strong><br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anafun<br />
c product variables<br />
DOUBLE PRECISION DLT,DNET,DNAT,<br />
+ DSELE,DSELA<br />
DOUBLE PRECISION DCT,DF,DMfgCOST,DMS1,DMS2,DMS3,DU<br />
c production variables<br />
INTEGER I,J,NABSTA,NEVPTA<br />
PARAMETER (NEVPTA=4,NABSTA=4)<br />
REAL LT,<br />
+ NAT,NET
REAL CT,<br />
+ F,<br />
+ MS1,MS2,MS3,MfgCOST,<br />
+ U<br />
c -----------------------------------------------------------<br />
c...input scalar AV values<br />
c.....values of experiment<br />
call avdsca(DLT,'length')<br />
call avdsca(DSELE,'EvapTube')<br />
call avdsca(DSELA,'AbsTube')<br />
call avdsca(DNET,'NET')<br />
call avdsca(DNAT,'NAT')<br />
c<br />
c .... values of rational reaction set<br />
call avdsca(DMS1,'MS1')<br />
call avdsca(DMS2,'MS2')<br />
call avdsca(DMS3,'MS3')<br />
call avdsca(DF,'F')<br />
c<br />
c product analysis<br />
LT=REAL(DLT)<br />
NET=REAL(DNET)<br />
NAT=REAL(DNAT)<br />
I=INT(DSELE)<br />
IF(I.LT.1) I=1<br />
IF(I.GT.NEVPTA) I=NEVPTA<br />
J=INT(DSELA)<br />
IF(J.LT.1) J=1<br />
IF(J.GT.NABSTA) J=NABSTA<br />
c...compute scalar AF values<br />
c<br />
c...intermediate constants<br />
c **********************************************************<br />
C...compute functions<br />
c.... production analysis<br />
MS1=REAL(DMS1)<br />
MS2=REAL(DMS2)<br />
MS3=REAL(DMS3)<br />
F =REAL(DF)<br />
CALL QNA(MS1,MS2,MS3,F,LT,NET,NAT,I,J,U,CT,MfgCOST)<br />
c<br />
c ************************************************************<br />
C EVALUATE DEVIATION FUNCTION USING AN ARCHIMEDEAN FORMULATION<br />
C WEIGHT OF EACH GOAL AT EACH PRIORITY LEVEL<br />
DU = U<br />
DMfgCOST= MfgCOST<br />
DCT = CT<br />
c<br />
C<br />
233
234<br />
C<br />
************************************************************************<br />
****<br />
c...output scalar AF values<br />
call afdsca(DU,'Utilization')<br />
call afdsca(DMfgCOST,'MfgCOST')<br />
call afdsca(DCT,'CT')<br />
return<br />
end<br />
C FORTRAN FILE FOR OptdesX FOR FINDING PAYOFF MATRIX FOR<br />
C PRODUCT DESIGN - NON COOPERATIVE<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAPRE<br />
c preprocessing routine<br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anapre(modelN)<br />
character*17 modelN<br />
modelN = "Chiller Design"<br />
return<br />
end<br />
c-----------------------------------------------------------------------<br />
-------<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAFUN<br />
c sample analysis routine for chiller <strong>design</strong><br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anafun<br />
c product variables<br />
DOUBLE PRECISION DLT,DCOST,DFEAS,<br />
+ DSELE,DSELA,DL,DNAT,DNET,DNE,DNA,<br />
+ DMatCOST,DZ1<br />
DOUBLE PRECISION DCT,DMfgCOST,DZ2<br />
c production variables<br />
INTEGER I,J,K,NABSTA,NEVPTA<br />
PARAMETER (NEVPTA=4,NABSTA=4)<br />
REAL A,<br />
+ COSTE(NEVPTA),COSTA(NABSTA),<br />
+ E,FEASE,FEASA,FEAS,<br />
+ L,LT,<br />
+ NAT,NET,NE,NA,<br />
+ PROB(8),<br />
+ TONS(8),TCOST<br />
REAL CT,
+ MfgCOST,<br />
+ TCT,<br />
+ XE(4),XA(4)<br />
c<br />
c -----------------------------------------------------------<br />
DATA (TONS(K), K=1,8) /600,700,800,900,1000,1100,1200,1300/,<br />
& (PROB(K), K=1,8) /0.05,0.15,0.16,0.05,<br />
+ 0.15,0.12,0.25,0.07/<br />
DATA (COSTE(I),I=1,NEVPTA) /6.0,6.5,7.0,7.5/,<br />
& (COSTA(I),I=1,NABSTA) /3.25,3.50,3.75,4.0/<br />
DATA (XA(I), I=1,4) /-2,-1,1,2/,<br />
& (XE(I), I=1,4) /-2,-1,1,2/<br />
c...input scalar AV values<br />
call avdsca(DLT,'length')<br />
call avdsca(DSELE,'EvapTube')<br />
call avdsca(DSELA,'AbsTube')<br />
c<br />
c product analysis<br />
TCOST = 150435.0<br />
TCT = 318.1<br />
LT=REAL(DLT)<br />
I=INT(DSELE)<br />
IF(I.LT.1) I=1<br />
IF(I.GT.NEVPTA) I=NEVPTA<br />
J=INT(DSELA)<br />
IF(J.LT.1) J=1<br />
IF(J.GT.NABSTA) J=NABSTA<br />
c...compute scalar AF values<br />
c...intermediate constants<br />
C **********************************************************<br />
L=1.75*(LT+2)+17.0<br />
DL=L<br />
C...compute functions<br />
MatCOST= 0.0<br />
FEAS =16.0<br />
NET = 0.0<br />
NAT = 0.0<br />
DO 100 K=1,8<br />
c ............. Find minimum number of tubes for EVAP<br />
CALL MINEVP(L,I,TONS(K),NE,FEASE)<br />
NET=NET+PROB(K)*NE<br />
c ............. Find minimum number of tubes for ABS<br />
CALL MINABS(L,J,TONS(K),NA,FEASA)<br />
NAT=NAT+PROB(K)*NA<br />
FEAS = FEAS-FEASE-FEASA<br />
MatCOST=PROB(K)*(COSTE(I)*(L*NE)+COSTA(J)*(L*NA))+MatCOST<br />
100 CONTINUE<br />
c ... code vars<br />
NET=(NET-440)/115.0-2<br />
NAT=(NAT-440)/115.0-2<br />
235
236<br />
E=XE(I)<br />
A=XA(J)<br />
DNET = NET<br />
DNAT = NAT<br />
DNE = NE<br />
DNA = NA<br />
c<br />
MatCOST=MatCOST+52335.0<br />
c<br />
c ... Rational Reaction Sets of Production Designer<br />
MfgCOST = 49263.35+1215.23*LT-442.57*E-<br />
412.04*A+2451.33*NET+2208.98*NAT<br />
+ -365.50*LT*E-684.65*LT*A-479.31*LT*NET-<br />
768.31*LT*NAT+357.44*E*A<br />
+ +376.55*E*NET+798.4*E*NAT+686.45*A*NET+823.85*A*NAT<br />
+ -504.06*NET**2-392.44*NET*NAT-344.81*NAT**2<br />
CT = 505.48+11.08*LT+0.09*E+26.67*A+23.41*NET-<br />
3.0*LT*NET+3.63*E**2<br />
+ -3.78*E*NAT+9.38*NET**2+15.38*NET*NAT+4.50*NAT**2<br />
c<br />
COST=MatCOST+MfgCOST<br />
c ************************************************************<br />
C EVALUATE DEVIATION FUNCTION USING AN ARCHIMEDEAN FORMULATION<br />
C WEIGHT OF EACH GOAL AT EACH PRIORITY LEVEL<br />
DCOST = COST<br />
DMatCOST = MatCOST<br />
DFEAS = FEAS<br />
DMfgCOST = MfgCOST<br />
DCT = CT<br />
DZ1 = COST/TCOST-1.0<br />
DZ2 = CT/TCT-1.0<br />
c<br />
C<br />
C<br />
************************************************************************<br />
****<br />
c...output scalar AF values<br />
call afdsca(DL,'Length')<br />
call afdsca(DNET,'NET')<br />
call afdsca(DNAT,'NAT')<br />
call afdsca(DNE,'NE')<br />
call afdsca(DNA,'NA')<br />
call afdsca(DFEAS,'PRODUCT_FEAS')<br />
call afdsca(DCOST,'COST')<br />
call afdsca(DMatCOST,'MatCOST')<br />
call afdsca(DMfgCOST,'MfgCOST')<br />
call afdsca(DCT,'CT')<br />
call afdsca(DZ1,'Z1(COST)')<br />
call afdsca(DZ2,'Z2(CT)')<br />
return
end<br />
237<br />
C FORTRAN FILE FOR OptdesX FOR PRODUCT SYSTEM DESIGN WITH<br />
C MODIFIED DEVIATION FUNCTION<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAPRE<br />
c preprocessing routine<br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anapre(modelN)<br />
character*17 modelN<br />
modelN = "Chiller Design"<br />
return<br />
end<br />
c-----------------------------------------------------------------------<br />
-------<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAFUN<br />
c sample analysis routine for chiller <strong>design</strong><br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anafun<br />
c product variables<br />
DOUBLE PRECISION DLT,DCOST,DFEAS,<br />
+ DSELE,DSELA,DL,<br />
+ DMatCOST,DNET,DNAT,DNE,DNA<br />
DOUBLE PRECISION DCT,DMfgCOST<br />
DOUBLE PRECISION DFN1,DFN2,DZ<br />
c production variables<br />
INTEGER I,J,K,NABSTA,NEVPTA<br />
PARAMETER (NEVPTA=4,NABSTA=4)<br />
REAL A,<br />
+ COSTE(NEVPTA),COSTA(NABSTA),<br />
+ E,FEASE,FEASA,FEAS,<br />
+ L,LT,<br />
+ NAT,NET,NE,NA,<br />
+ PROB(8),<br />
+ TONS(8),TCOST<br />
REAL CT,<br />
+ MfgCOST,<br />
+ TCT<br />
REAL F1,F2,FN1,FN2,FC,<br />
+ S,<br />
+ W1,W2,<br />
+ XE(4),XA(4),<br />
+ Z11,Z12,Z21,Z22,Z<br />
c -----------------------------------------------------------------
DATA (TONS(K), K=1,8) /600,700,800,900,1000,1100,1200,1300/,<br />
& (PROB(K), K=1,8) /0.05,0.15,0.16,0.05,<br />
+ 0.15,0.12,0.25,0.07/<br />
DATA (COSTE(I),I=1,NEVPTA) /6.0,6.5,7.0,7.5/,<br />
& (COSTA(I),I=1,NABSTA) /3.25,3.50,3.75,4.0/<br />
DATA (XA(I), I=1,4) /-2,-1,1,2/,<br />
& (XE(I), I=1,4) /-2,-1,1,2/<br />
c -----------------------------------------------------------------<br />
c...input scalar AV values<br />
call avdsca(DLT,'length')<br />
call avdsca(DSELE,'EvapTube')<br />
call avdsca(DSELA,'AbsTube')<br />
c<br />
c...compute scalar AF values<br />
c...intermediate constants<br />
C **********************************************************<br />
c ... product analysis<br />
LT=REAL(DLT)<br />
I =INT(DSELE)<br />
IF(I.LT.1) I=1<br />
IF(I.GT.NEVPTA) I=NEVPTA<br />
J =INT(DSELA)<br />
IF(J.LT.1) J=1<br />
IF(J.GT.NABSTA) J=NABSTA<br />
L = 1.75*(LT+2)+17.0<br />
DL= L<br />
C...compute functions<br />
MatCOST= 0.0<br />
FEAS =16.0<br />
NET = 0.0<br />
NAT = 0.0<br />
DO 100 K=1,8<br />
c ............. Find minimum number of tubes for EVAP<br />
CALL MINEVP(L,I,TONS(K),NE,FEASE)<br />
NET=NET+PROB(K)*NE<br />
c ............. Find minimum number of tubes for ABS<br />
CALL MINABS(L,J,TONS(K),NA,FEASA)<br />
NAT=NAT+PROB(K)*NA<br />
FEAS = FEAS-FEASE-FEASA<br />
MatCOST=PROB(K)*(COSTE(I)*(L*NE)+COSTA(J)*(L*NA))+MatCOST<br />
100 CONTINUE<br />
c ... code vars<br />
NET=(NET-440)/115.0-2<br />
NAT=(NAT-440)/115.0-2<br />
E=XE(I)<br />
A=XA(J)<br />
DNET = NET<br />
DNAT = NAT<br />
DNE = NE<br />
DNA = NA<br />
238
239<br />
c<br />
MatCOST=MatCOST+52335.0<br />
c<br />
c ... Rational Reaction Sets of Production Designer<br />
MfgCOST = 49263.35+1215.23*LT-442.57*E-<br />
412.04*A+2451.33*NET+2208.98*NAT<br />
+ -365.50*LT*E-684.65*LT*A-479.31*LT*NET-<br />
768.31*LT*NAT+357.44*E*A<br />
+ +376.55*E*NET+798.4*E*NAT+686.45*A*NET+823.85*A*NAT<br />
+ -504.06*NET**2-392.44*NET*NAT-344.81*NAT**2<br />
CT = 505.48+11.08*LT+0.09*E+26.67*A+23.41*NET-<br />
3.0*LT*NET+3.63*E**2<br />
+ -3.78*E*NAT+9.38*NET**2+15.38*NET*NAT+4.50*NAT**2<br />
c<br />
COST=MatCOST+MfgCOST<br />
c<br />
************************************************************************<br />
*****<br />
C modified Pareto objective<br />
TCOST = 169810.0<br />
TCT = 318.1<br />
Z11 = 211837/TCOST-1.0<br />
Z12 = 233385/TCOST-1.0<br />
Z21 = 631.6/TCT-1.0<br />
Z22 = 448.0/TCT-1.0<br />
W1 = 0.5<br />
W2 = 0.5<br />
c<br />
c ... COST GOAL<br />
F1 = COST/TCOST-1.0<br />
FN1 = (F1-Z11)/(Z12-Z11)<br />
c ... CT Goal<br />
F2 = CT/TCT-1.0<br />
FN2 = (F2-Z22)/(Z21-Z22)<br />
c<br />
c ... Formulate Supercriterion<br />
S = (1-FN1)*(1-FN2)<br />
c ... Formulate a Pareto Optimal objective FC<br />
FC = W1*FN1+W2*FN2<br />
c ... Modified Objective Function<br />
Z = FC-S+1<br />
c<br />
c<br />
************************************************************************<br />
*****<br />
DCOST = COST<br />
DMatCOST = MatCOST<br />
DFEAS = FEAS<br />
DU = U<br />
DMfgCOST = MfgCOST<br />
DCT = CT
240<br />
DFN1 = FN1<br />
DFN2 = FN2<br />
DZ = Z<br />
c<br />
C<br />
C<br />
************************************************************************<br />
****<br />
c...output scalar AF values<br />
call afdsca(DL,'Length')<br />
call afdsca(DNET,'NET')<br />
call afdsca(DNAT,'NAT')<br />
call afdsca(DNE,'NE')<br />
call afdsca(DNA,'NA')<br />
call afdsca(DFEAS,'PRODUCT_FEAS')<br />
call afdsca(DCOST,'COST')<br />
call afdsca(DMatCOST,'MatCOST')<br />
call afdsca(DMfgCOST,'MfgCOST')<br />
call afdsca(DCT,'CT')<br />
call afdsca(DFN1,'FN1(COST)')<br />
call afdsca(DFN2,'FN2(CT)')<br />
call afdsca(DZ,'Z')<br />
return<br />
end<br />
C FORTRAN FILE FOR OptdesX FOR PROBABILISTIC PRODUCT DESIGN<br />
C WITH MODIFIED DEVIATION FUNCTION<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAPRE<br />
c preprocessing routine<br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anapre(modelN)<br />
character*17 modelN<br />
modelN = "Chiller Design"<br />
return<br />
end<br />
c-----------------------------------------------------------------------<br />
-------<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAFUN<br />
c sample analysis routine for chiller <strong>design</strong><br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anafun<br />
c product variables<br />
DOUBLE PRECISION DCHL,DCHCOST,DCHCT,DCOST,DCT,
241<br />
+ DFEAS,<br />
+ DL,DLT,DPICOST,DPICT,<br />
+ DMatCOST,DMfgCOST,<br />
+ DSELE,DSELA<br />
DOUBLE PRECISION DFN1,DFN2,DZ<br />
c production variables<br />
INTEGER I,J,NABSTA,NEVPTA<br />
PARAMETER (NEVPTA=4,NABSTA=4)<br />
REAL ACHCT,ACHCOST,<br />
+ COST,CT,<br />
+ CHL,CHMfgC,CHCOST,CHCT,CHDCT,<br />
+ FEAS,<br />
+ L,LT,<br />
+ MfgCOST<br />
REAL F1,F2,FN1,FN2,FC,<br />
+ S,<br />
+ W1,W2,<br />
+ Z11,Z12,Z21,Z22,Z<br />
c -----------------------------------------------------------<br />
c...input scalar AV values<br />
call avdsca(DLT,'Length')<br />
call avdsca(DSELE,'EvapTube')<br />
call avdsca(DSELA,'AbsTube')<br />
call avdsca(DCHL,'DLength')<br />
c ************************************************************<br />
CHMfgC = 0.1<br />
CHDCT = 0.05<br />
ACHCOST= 0.1<br />
ACHCT = 0.1<br />
c ************************************************************<br />
I=INT(DSELE)<br />
IF(I.LT.1) I=1<br />
IF(I.GT.NEVPTA) I=NEVPTA<br />
J=INT(DSELA)<br />
IF(J.LT.1) J=1<br />
IF(J.GT.NABSTA) J=NABSTA<br />
LT = REAL(DLT)<br />
CHL = REAL(DCHL)<br />
L = 1.75*(LT+2)+17.0<br />
DL = L<br />
C...compute functions<br />
c ... Evaluate DPIs<br />
CALL EVAL_DPI(CHL,LT,I,J,<br />
+ COST,MfgCOST,CT,<br />
+ CHCOST,CHMfgC,CHDCT,CHCT,<br />
+ FEAS,<br />
+ DPI_COST,DPI_CT)<br />
c ----------------------------------------------------------------------<br />
---<br />
C EVALUATE DEVIATION FUNCTION USING AN ARCHIMEDEAN FORMULATION<br />
C WEIGHT OF EACH GOAL AT EACH PRIORITY LEVEL
242<br />
C modified Pareto objective<br />
Z11 = 0.178<br />
Z12 = 0.431<br />
Z21 = 0.994<br />
Z22 = 0.477<br />
W1 = 0.5<br />
W2 = 0.5<br />
c<br />
c ... COST GOAL<br />
F1 = 1-DPI_COST<br />
c WRITE(*,*) Z11,Z12,F1<br />
FN1 = (F1-Z11)/(Z12-Z11)<br />
c WRITE(*,*) FN1<br />
c ... CT Goal<br />
F2 = 1-DPI_CT<br />
FN2 = (F2-Z22)/(Z21-Z22)<br />
c<br />
c ... Formulate Supercriterion<br />
S = (1-FN1)*(1-FN2)<br />
c ... Formulate a Pareto Optimal objective FC<br />
FC = W1*FN1+W2*FN2<br />
c ... Modified Objective Function<br />
Z = FC-S+1<br />
c<br />
DCHCOST = 1.0-CHCOST/(ACHCOST*COST)<br />
DCHCT = 1.0-CHCT/(ACHCT*CT)<br />
DFN1 = FN1<br />
DFN2 = FN2<br />
DCOST = COST<br />
DCT = CT<br />
DFEAS = FEAS<br />
DMatCOST = COST-MfgCOST<br />
DMfgCOST = MfgCOST<br />
DPICOST = DPI_COST<br />
DPICT = DPI_CT<br />
DZ = Z<br />
C<br />
************************************************************************<br />
****<br />
c...output scalar AF values<br />
call afdsca(DFEAS ,'FEAS')<br />
call afdsca(DCHCOST ,'DCOST')<br />
call afdsca(DCHCT ,'DCT')<br />
call afdsca(DFN1 ,'FN1')<br />
call afdsca(DFN2 ,'FN2')<br />
call afdsca(DPICOST ,'DPICOST')<br />
call afdsca(DPICT ,'DPICT')<br />
call afdsca(DZ ,'Z')<br />
call afdsca(DCOST ,'COST')<br />
call afdsca(DCT ,'CT')<br />
call afdsca(DMatCOST,'MatCOST')
call afdsca(DMfgCOST,'MfgCOST')<br />
return<br />
end<br />
SUBROUTINE EVAL_DPI(DL,LT,I,J,<br />
+ AvgCOST,MfgC,AvgCT,<br />
+ DCOST,DMfgC,DDCT,DCT,<br />
+ FEASTOT,<br />
+ DPI_COST,DPI_CT)<br />
INTEGER I,J,K,M,N,P<br />
PARAMETER(N=27)<br />
INTEGER CCD(N,5)<br />
REAL A,AvgCOST,AvgCT,<br />
+ COSTA(8),COSTE(8),CA,CAT,CE,CET,COST,CT,<br />
+ DCA,DCE,DCOST,DCT,DDC,DDCT,DL,DLT,DLTC,<br />
+ DMC,DMfgC,DPI_COST,DPI_CT,<br />
+ FEAS,FEASE,FEASA,FEASTOT,<br />
+ E,<br />
+ LC,LT,LTC,<br />
+ MatCOST,MAXCOST,MAXCT,MINCOST,MINCT,MfgC,Mfg,<br />
+ NA,NAT,NE,NET,<br />
+ PROB(8),<br />
+ TONS(8),<br />
+ VAL(N),<br />
+ XA(4),XE(4),<br />
+ YCOST(N),YCT(N)<br />
c ... Fractional Factorial Experiment w/resolution 5<br />
DATA ((CCD(K,M),M=1,5),K=1,N)<br />
+ /-1,-1,-1,-1, 1,<br />
+ -1,-1,-1, 1,-1,<br />
+ -1,-1, 1,-1,-1,<br />
+ -1,-1, 1, 1, 1,<br />
+ -1, 1,-1,-1,-1,<br />
+ -1, 1,-1, 1, 1,<br />
+ -1, 1, 1,-1, 1,<br />
+ -1, 1, 1, 1,-1,<br />
+ 1,-1,-1,-1,-1,<br />
+ 1,-1,-1, 1, 1,<br />
+ 1,-1, 1,-1, 1,<br />
+ 1,-1, 1, 1,-1,<br />
+ 1, 1,-1,-1, 1,<br />
+ 1, 1,-1, 1,-1,<br />
+ 1, 1, 1,-1,-1,<br />
+ 1, 1, 1, 1, 1,<br />
+ -2, 0, 0, 0, 0,<br />
+ 2, 0, 0, 0, 0,<br />
+ 0,-2, 0, 0, 0,<br />
+ 0, 2, 0, 0, 0,<br />
+ 0, 0,-2, 0, 0,<br />
+ 0, 0, 2, 0, 0,<br />
+ 0, 0, 0,-2, 0,<br />
+ 0, 0, 0, 2, 0,<br />
243
+ 0, 0, 0, 0,-2,<br />
+ 0, 0, 0, 0, 2,<br />
+ 0, 0, 0, 0, 0/<br />
DATA (TONS(K), K=1,8) /600,700,800,900,1000,1100,1200,1300/,<br />
& (PROB(K), K=1,8) /0.05,0.15,0.16,0.05,<br />
+ 0.15,0.12,0.25,0.07/<br />
DATA (COSTE(K),K=1,4) /6.0,6.5,7.0,7.5/,<br />
& (COSTA(K),K=1,4) /3.25,3.50,3.75,4.0/<br />
DATA (XA(K), K=1,4) /-2,-1,1,2/,<br />
& (XE(K), K=1,4) /-2,-1,1,2/<br />
C -----------------------------------------------------------<br />
c ... high and low values for preference functions<br />
MINCOST =169180.0<br />
MINCT =318.1<br />
MAXCOST =320000.0<br />
MAXCT =600.0<br />
c -----------------------------------------------------------<br />
DLTC=4.0/7.0*DL<br />
CET =COSTE(I)<br />
CAT =COSTA(J)<br />
FEASTOT= 0.0<br />
DO 200 K=1,N<br />
IF (CCD(K,1).EQ.-2) THEN<br />
DLT=-DLTC<br />
ELSE IF(CCD(K,1).EQ.-1) THEN<br />
DLT=-DLTC/2.0<br />
ELSE IF(CCD(K,1).EQ.0) THEN<br />
DLT=0.0<br />
ELSE IF(CCD(K,1).EQ.1) THEN<br />
DLT=DLTC/2.0<br />
ELSE<br />
DLT=DLTC<br />
END IF<br />
IF (CCD(K,2).EQ.-2) THEN<br />
DCE=-0.50<br />
ELSE IF(CCD(K,2).EQ.-1) THEN<br />
DCE=-0.25<br />
ELSE IF(CCD(K,2).EQ.0) THEN<br />
DCE= 0.0<br />
ELSE IF(CCD(K,2).EQ.1) THEN<br />
DCE= 0.25<br />
ELSE<br />
DCE= 0.50<br />
END IF<br />
IF (CCD(K,3).EQ.-2) THEN<br />
DCA=-0.25<br />
ELSE IF(CCD(K,3).EQ.-1) THEN<br />
DCA=-0.125<br />
ELSE IF(CCD(K,3).EQ.0) THEN<br />
DCA= 0.0<br />
ELSE IF(CCD(K,3).EQ.1) THEN<br />
244
245<br />
DCA= 0.125<br />
ELSE<br />
DCA= 0.25<br />
END IF<br />
IF (CCD(K,4).EQ.-2) THEN<br />
DM=-DMfgC<br />
ELSE IF(CCD(K,4).EQ.-1) THEN<br />
DM=-DMfgC/2.0<br />
ELSE IF(CCD(K,4).EQ.0) THEN<br />
DMC= 0.0<br />
ELSE IF(CCD(K,4).EQ.1) THEN<br />
DMC= DMfgC/2.0<br />
ELSE<br />
DMC= DMfgC<br />
END IF<br />
IF (CCD(K,5).EQ.-2) THEN<br />
DDC=-DDCT<br />
ELSE IF(CCD(K,5).EQ.-1) THEN<br />
DDC=-DDCT/2.0<br />
ELSE IF(CCD(K,5).EQ.0) THEN<br />
DDC= 0.0<br />
ELSE IF(CCD(K,5).EQ.1) THEN<br />
DDC= DDCT/2.0<br />
ELSE<br />
DDC= DDCT<br />
END IF<br />
LTC=LT+DLT<br />
LC =1.75*(LTC+2)+17.0<br />
CE=(1+DCE)*CAT<br />
CA=(1+DCA)*CET<br />
C ... Calculate COST & MfgC<br />
MatCOST= 0.0<br />
FEAS =16.0<br />
NET = 0.0<br />
NAT = 0.0<br />
DO 100 P=1,8<br />
c ............. Find minimum number of tubes for EVAP (99.7% CI)<br />
CALL MINEVP(LC,I,TONS(P),NE,FEASE)<br />
NET=NET+PROB(P)*NE<br />
c ............. Find minimum number of tubes for ABS (99.7% CI)<br />
CALL MINABS(LC,J,TONS(P),NA,FEASA)<br />
NAT=NAT+PROB(P)*NA<br />
FEAS = FEAS-FEASE-FEASA<br />
MatCOST=PROB(P)*(CE*(LC*NE)+CA*(LC*NA))+MatCOST<br />
100 CONTINUE<br />
FEASTOT=FEASTOT+FEAS<br />
MatCOST=MatCOST+52335.0<br />
c ----------------------------------------------------------------------<br />
--c<br />
... Rational Reaction Sets<br />
c ... code vars
246<br />
NET=(NET-440)/115.0-2<br />
NAT=(NAT-440)/115.0-2<br />
E=XE(I)<br />
A=XA(J)<br />
MfgC = 49263.35+1215.23*LTC-442.57*E-412.04*A+2451.33*NET<br />
+ +2208.98*NAT<br />
+ -365.50*LTC*E-684.65*LTC*A-479.31*LTC*NET-768.31*LTC*NAT<br />
+ +357.44*E*A<br />
+ +376.55*E*NET+798.4*E*NAT+686.45*A*NET+823.85*A*NAT<br />
+ -504.06*NET**2-392.44*NET*NAT-344.81*NAT**2<br />
CT = 505.48+11.08*LTC+0.09*E+26.67*A+23.41*NET-3.0*LTC*NET<br />
+ +3.63*E**2-<br />
3.78*E*NAT+9.38*NET**2+15.38*NET*NAT+4.50*NAT**2<br />
c ----------------------------------------------------------------------<br />
---<br />
MfgC=(1+DMC)*MfgC<br />
CT =(1+DDC)*CT<br />
COST=MatCOST+MfgC<br />
c<br />
C ESTIMATE DEVIATION VARIABLES<br />
C modified Pareto solution<br />
C<br />
YCOST(K)= COST<br />
YCT(K) = CT<br />
C<br />
C ************************************************************<br />
200 CONTINUE<br />
c ... DPI for COST<br />
CALL SORT(YCOST,VAL,AvgCOST,DCOST,MAXCOST,MINCOST,N)<br />
CALL INTEGRAL(DPI_COST,YCOST,VAL,N)<br />
C ... DPI for CT<br />
CALL SORT(YCT,VAL,AvgCT,DCT,MAXCT,MINCT,N)<br />
CALL INTEGRAL(DPI_CT,YCT,VAL,N)<br />
c ------------------------------------------------------------<br />
RETURN<br />
END<br />
SUBROUTINE SORT(Y,VAL,AVG,DY,TMAX,TMIN,N)<br />
INTEGER N,FIRST, PASS, K<br />
REAL AUX,AVG,<br />
+ DY,<br />
+ F,<br />
+ PY,<br />
+ STDDEV,<br />
+ TEMP,TMAX,TMIN,<br />
+ VAL(N),<br />
+ Y(N)<br />
LOGICAL SORTED<br />
PASS=1<br />
SORTED=.FALSE.<br />
DO WHILE(.NOT.SORTED)<br />
SORTED=.TRUE.
DO 10 FIRST=1,N-PASS<br />
IF(Y(FIRST).GT.Y(FIRST+1)) THEN<br />
TEMP=Y(FIRST)<br />
Y(FIRST)=Y(FIRST+1)<br />
Y(FIRST+1)=TEMP<br />
SORTED=.FALSE.<br />
END IF<br />
10 CONTINUE<br />
PASS=PASS+1<br />
END DO<br />
STDDEV=(Y(N)-Y(1))/6.0<br />
STDDEV=ABS(STDDEV)<br />
DY=3*STDDEV<br />
AVG =(Y(N)+Y(1))/2.0<br />
DO 20 K=1,N<br />
IF(Y(K).LT.TMIN) THEN<br />
PY=0.0<br />
ELSE IF(Y(K).GT.TMAX) THEN<br />
PY=0.0<br />
ELSE<br />
PY=(TMAX-Y(K))/(TMAX-TMIN)<br />
END IF<br />
F=-((Y(K)-AVG)/STDDEV)**2.0/2.0<br />
F=EXP(F)<br />
AUX=SQRT(6.2832)<br />
F=1.0/(AUX*STDDEV)*F<br />
VAL(K)=F*PY<br />
20 CONTINUE<br />
RETURN<br />
END<br />
SUBROUTINE INTEGRAL(DPI,X,Y,N)<br />
INTEGER I,N<br />
REAL X(N),Y(N),DPI<br />
DPI=0.0<br />
DO 10 I=1,N-1<br />
DPI=DPI+((Y(I+1)-Y(I))/2.0+Y(I))*(X(I+1)-X(I))<br />
10 CONTINUE<br />
RETURN<br />
END<br />
247<br />
c FORTRAN FILE FOR PRODUCT DESIGN WITH PROBABILISTIC MODEL<br />
C AND MODIFIED DEVIATION FUNCTION<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAPRE<br />
c preprocessing routine
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anapre(modelN)<br />
character*17 modelN<br />
248<br />
modelN = "Chiller Design"<br />
return<br />
end<br />
c-----------------------------------------------------------------------<br />
-------<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAFUN<br />
c sample analysis routine for chiller <strong>design</strong><br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anafun<br />
c product variables<br />
DOUBLE PRECISION DCHCOST,DCHCT,DCOST,DCT,<br />
+ DF,<br />
+ DMatCOST,DMfgCOST,DMS1,DMS2,DMS3,<br />
+ DU<br />
DOUBLE PRECISION DFN1,DFN2,DZ,DZCOST,DZCT<br />
c production variables<br />
INTEGER I,J<br />
REAL ACHCT,ACHCOST,<br />
+ COST,CT,<br />
+ CHL,CHCOST,CHCT,<br />
+ F,<br />
+ L,<br />
+ MS1,MS2,MS3,MfgCOST,<br />
+ U<br />
REAL F1,F2,FN1,FN2,FC,<br />
+ S,<br />
+ W1,W2,<br />
+ Z11,Z12,Z21,Z22,Z<br />
c<br />
c ... product <strong>design</strong> solution<br />
c ... Note: LT is the normalized value of L[17,24] -> LT[-2,2]<br />
L = 19.09<br />
c ... CHL is the half-width range of L<br />
CHL = 0.26<br />
c ... I is the evaporator tube, J is the absorber tube<br />
I = 2<br />
J = 4<br />
5 CONTINUE<br />
c -----------------------------------------------------------<br />
c...input scalar AV values<br />
call avdsca(DMS1,'MS1')<br />
call avdsca(DMS2,'MS2')<br />
call avdsca(DMS3,'MS3')
249<br />
call avdsca(DF,'F')<br />
c ************************************************************<br />
ACHCOST= 0.1<br />
ACHCT = 0.1<br />
c ************************************************************<br />
C...compute functions<br />
c ... Evaluate DPIs<br />
MS1 = DMS1<br />
MS2 = DMS2<br />
MS3 = DMS3<br />
F = DF<br />
CALL EVAL_DPI(CHL,L,I,J,<br />
+ COST,MfgCOST,CT,<br />
+ CHCOST,CHCT,<br />
+ MS1,MS2,MS3,F,<br />
+ U,<br />
+ DPI_COST,DPI_CT)<br />
c ----------------------------------------------------------------------<br />
---<br />
C EVALUATE DEVIATION FUNCTION USING AN ARCHIMEDEAN FORMULATION<br />
C WEIGHT OF EACH GOAL AT EACH PRIORITY LEVEL<br />
C modified Pareto objective<br />
Z11 = 0.170<br />
Z12 = 0.304<br />
Z21 = 1.000<br />
Z22 = 0.417<br />
W1 = 0.5<br />
W2 = 0.5<br />
c<br />
c ... COST GOAL<br />
F1 = 1-DPI_COST<br />
FN1 = (F1-Z11)/(Z12-Z11)<br />
c ... CT Goal<br />
F2 = 1-DPI_CT<br />
FN2 = (F2-Z22)/(Z21-Z22)<br />
c<br />
c ... Formulate Supercriterion<br />
S = (1-FN1)*(1-FN2)<br />
c ... Formulate a Pareto Optimal objective FC<br />
FC = W1*FN1+W2*FN2<br />
c ... Modified Objective Function<br />
Z = FC-S+1<br />
c<br />
DCHCOST = 1.0-CHCOST/(ACHCOST*COST)<br />
DCHCT = 1.0-CHCT/(ACHCT*CT)<br />
DCOST = COST<br />
DCT = CT<br />
DFN1 = FN1<br />
DFN2 = FN2<br />
DU = U<br />
DMatCOST = COST-MfgCOST
250<br />
DMfgCOST = MfgCOST<br />
DZCOST = 1.0-DPI_COST<br />
DZCT = 1.0-DPI_CT<br />
DZ = Z<br />
C<br />
************************************************************************<br />
****<br />
c...output scalar AF values<br />
call afdsca(DU ,'Utilization')<br />
call afdsca(DCHCOST ,'DCOST')<br />
call afdsca(DCHCT ,'DCT')<br />
call afdsca(DFN1 ,'FN1')<br />
call afdsca(DFN2 ,'FN2')<br />
call afdsca(DZ ,'Z')<br />
call afdsca(DZCOST ,'Z1(COST)')<br />
call afdsca(DZCT ,'Z2(CT)')<br />
call afdsca(DCOST ,'COST')<br />
call afdsca(DCT ,'CT')<br />
call afdsca(DMatCOST,'MatCOST')<br />
call afdsca(DMfgCOST,'MfgCOST')<br />
return<br />
end<br />
c=======================================================================<br />
=======<br />
c...subroutine ANAPOS<br />
c postprocessing routine<br />
c-----------------------------------------------------------------------<br />
-------<br />
subroutine anapos<br />
return<br />
end<br />
SUBROUTINE EVAL_DPI(DL,L,I,J,<br />
+ AvgCOST,MfgC,AvgCT,<br />
+ DCOST,DCT,<br />
+ MS1,MS2,MS3,F,<br />
+ MAXU,<br />
+ DPI_COST,DPI_CT)<br />
INTEGER I,ISEED,J,K,M,N,LTABLE<br />
PARAMETER(N=5)<br />
REAL AvgCOST,AvgCT,<br />
+ COST,CT,<br />
+ DCOST,DCT,DL,<br />
+ DPI_COST,DPI_CT,COSTE(4),COSTA(4),<br />
+ F,<br />
+ L,LC,LT(8),L_RAND,<br />
+ MatCOST,MAXCOST,MAXCT,MINCOST,MINCT,MfgC,MAXU,
+ MS1,MS2,MS3,<br />
+ NA(8),NAT(8),NE(8),NET(8),NETABLE(8,5),NATABLE(8,5),<br />
+ PROB(8),<br />
+ TONS(8),<br />
+ VAL(N),<br />
+ YCOST(N),YCT(N)<br />
DATA (TONS(K), K=1,8) /600,700,800,900,1000,1100,1200,1300/,<br />
& (PROB(K), K=1,8) /0.05,0.15,0.16,0.05,<br />
+ 0.15,0.12,0.25,0.07/<br />
DATA (COSTE(I),I=1,4) /6.0,6.5,7.0,7.5/,<br />
& (COSTA(I),I=1,4) /3.25,3.50,3.75,4.0/<br />
ISEED=20029<br />
DATA ((NETABLE(M,K),K=1,5),M=1,8)<br />
+ /440,440,440,440,440,<br />
+ 440,440,440,440,440,<br />
+ 475,467,464,457,453,<br />
+ 536,532,525,518,514,<br />
+ 600,593,586,579,572,<br />
+ 661,654,647,640,633,<br />
+ 726,716,708,701,694,<br />
+ 787,780,769,762,751/<br />
DATA ((NATABLE(M,K),K=1,5),M=1,8)<br />
+ /440,440,440,440,440,<br />
+ 440,440,440,440,440,<br />
+ 482,482,482,467,467,<br />
+ 554,539,539,528,528,<br />
+ 611,611,601,601,590,<br />
+ 672,672,662,661,647,<br />
+ 744,734,734,719,708,<br />
+ 805,805,791,780,781/<br />
C -----------------------------------------------------------<br />
c ... high and low values for preference functions<br />
MINCOST =169180.0<br />
MINCT =318.1<br />
MAXCOST =320000.0<br />
MAXCT =600.0<br />
c -----------------------------------------------------------<br />
MAXU = 0.0<br />
DO 200 K=1,N<br />
c<br />
c random length [L-DL,L+DL]<br />
L_RAND=(L-DL)+2*DL*RANDOM(ISEED)<br />
LC=(L_RAND-17.0)/1.75-2.0<br />
MatCOST=0.0<br />
DO 90 M=1,8<br />
LT(M)=LC<br />
IF(LC.LT.-1.0) THEN<br />
LTABLE=1<br />
ELSE IF(LC.LT.0) THEN<br />
LTABLE=2<br />
ELSE IF(LC.LT.1) THEN<br />
251
c<br />
c<br />
c<br />
LTABLE=3<br />
ELSE IF(LC.LE.2) THEN<br />
LTABLE=4<br />
END IF<br />
x=L-DL+LTABLE*DL/2.0<br />
NE(M)=(NETABLE(M,LTABLE+1)-NETABLE(M,LTABLE))/DL<br />
NE(M)=NE(M)*(L_RAND-x)+NETABLE(M,LTABLE+1)<br />
NET(M)=(NE(M)-440)/115.0-2.0<br />
NA(M)=(NATABLE(M,LTABLE+1)-NATABLE(M,LTABLE))/DL<br />
NA(M)=NA(M)*(L_RAND-x)+NATABLE(M,LTABLE+1)<br />
NAT(M)=(NA(M)-440)/115.0-2.0<br />
252<br />
MatCOST=PROB(M)*(COSTE(I)*L_RAND*NE(M)+<br />
+ COSTA(J)*L_RAND*NA(M))+MatCOST<br />
90 CONTINUE<br />
MatCOST=MatCOST+52335.0<br />
c<br />
CALL QNA(MS1,MS2,MS3,F,LT,NET,NAT,I,J,U,CT,MfgC)<br />
IF(U.GT.MAXU) MAXU=U<br />
c<br />
COST=MfgC+MatCOST<br />
c ----------------------------------------------------------------------<br />
---<br />
C<br />
YCOST(K)= COST<br />
YCT(K) = CT<br />
C<br />
C ************************************************************<br />
200 CONTINUE<br />
c ... DPI for COST<br />
CALL SORT(YCOST,VAL,AvgCOST,DCOST,MAXCOST,MINCOST,N)<br />
CALL INTEGRAL(DPI_COST,YCOST,VAL,N)<br />
C ... DPI for CT<br />
CALL SORT(YCT,VAL,AvgCT,DCT,MAXCT,MINCT,N)<br />
CALL INTEGRAL(DPI_CT,YCT,VAL,N)<br />
c ------------------------------------------------------------<br />
RETURN<br />
END<br />
B.2 SUBROUTINES TO FIND MINIMUM NUMBER OF TUBES GIVEN LENGTH AND TUBES<br />
SELECTION<br />
C SUBROUTINE TO FIND MINIMUM NUMBER OF TUBES OF EVAPORATOR<br />
SUBROUTINE MINEVP(L,I,TONS,MinNET,FEAS)<br />
c production variables<br />
INTEGER I,NEVPTA<br />
c product variables<br />
PARAMETER (NEVPTA=4)
REAL DPE,<br />
+ DIE(NEVPTA),DOE(NEVPTA),DR(NEVPTA),<br />
+ FH(NEVPTA),FPI(NEVPTA),FW(NEVPTA),FEAS,<br />
+ L,FLOOR,ROOF,<br />
+ NET,MinNET,<br />
+ THKE(NEVPTA),TONS,<br />
+ UAE,UAENOM<br />
C TUBES ALTERNATIVES<br />
c evaporator<br />
DATA (DOE(I), I=1,NEVPTA) /0.625,0.75,0.875,1.0/,<br />
& (THKE(I), I=1,NEVPTA) /0.025,0.025,0.02,0.025/,<br />
& (FPI(I), I=1,NEVPTA) /30,30,30,30/,<br />
& (DIE(I), I=1,NEVPTA) /0.444,0.569,0.694,0.819/,<br />
& (DR(I), I=1,NEVPTA) /0.5,0.625,0.75,0.875/,<br />
& (FW(I), I=1,NEVPTA) /0.02,0.02,0.02,0.02/,<br />
& (FH(I), I=1,NEVPTA) /0.05,0.05,0.05,0.05/<br />
UAENOM=1372.183*TONS-34091.8<br />
FLOOR=440<br />
ROOF =900<br />
MinNET=900<br />
FEAS=0<br />
CALL EVAP(TONS,DOE(I),THKE(I),FPI(I),DIE(I),DR(I),<br />
& FW(I),FH(I),L,ROOF,UAE,DPE)<br />
IF(UAE. GE. UAENOM. AND. DPE.LE.23.0) THEN<br />
FEAS=1<br />
CALL EVAP(TONS,DOE(I),THKE(I),FPI(I),DIE(I),DR(I),<br />
& FW(I),FH(I),L,FLOOR,UAE,DPE)<br />
IF(UAE. GE. UAENOM. AND. DPE.LE.23.0) MinNET=FLOOR<br />
END IF<br />
IF(FEAS.EQ.1 .AND. MinNET.GT.FLOOR) THEN<br />
NET=FLOOR<br />
DO 100 WHILE(ABS(MinNET-NET). GT. 1.0)<br />
NET=(FLOOR+ROOF)/2.0<br />
MinNET=(NET+ROOF)/2.0<br />
CALL EVAP(TONS,DOE(I),THKE(I),FPI(I),DIE(I),DR(I),<br />
& FW(I),FH(I),L,NET,UAE,DPE)<br />
IF(UAE. GT. UAENOM. AND. DPE.LE.23.0) THEN<br />
ROOF=NET<br />
ELSE<br />
FLOOR=NET<br />
END IF<br />
100 CONTINUE<br />
ELSE<br />
END IF<br />
RETURN<br />
END<br />
C SUBROUTINE TO FIND MINIMUM NUMBER OF TUBES OF ABSORBER<br />
SUBROUTINE MINABS(L,J,TONS,MinNAT,FEAS)<br />
c production variables<br />
INTEGER I,J,NABSTA,NRA,NATR<br />
253
254<br />
c product variables<br />
PARAMETER (NABSTA=4)<br />
REAL DOA(NABSTA),THKA(NABSTA),<br />
+ L,FLOOR,ROOF,FEAS,<br />
+ NAT,MinNAT,<br />
+ TONS,<br />
+ UAA, UAANOM<br />
C TUBES ALTERNATIVES<br />
DATA (DOA(I), I=1,NABSTA) /0.625,0.75,0.875,1.0/,<br />
& (THKA(I), I=1,NABSTA) /0.03,0.03,0.03,0.03/<br />
c evaporator<br />
UAANOM=1021.806*TONS-38163.0<br />
FLOOR=440<br />
ROOF =900<br />
MinNAT=900<br />
FEAS=0<br />
NRA=12<br />
C CALCULATE EVAPORATOR UA<br />
NATR = INT(ROOF/NRA)<br />
CALL ABSE(TONS,DOA(J),THKA(J),L,NRA,NATR,UAA,CAPACITY)<br />
IF(UAA. GE. UAANOM. AND. CAPACITY.GE.TONS) THEN<br />
FEAS=1<br />
NATR = INT(FLOOR/NRA)<br />
CALL ABSE(TONS,DOA(J),THKA(J),L,NRA,NATR,UAA,CAPACITY)<br />
IF(UAA. GE. UAANOM. AND. CAPACITY.GE.TONS)<br />
MinNAT=FLOOR<br />
END IF<br />
IF(FEAS.EQ.1 .AND. MinNAT.GT.FLOOR) THEN<br />
NAT=FLOOR<br />
DO 100 WHILE(ABS(MinNAT-NAT). GT. 1.0)<br />
NAT=(FLOOR+ROOF)/2.0<br />
MinNAT=(NAT+ROOF)/2.0<br />
NATR = INT(NAT/NRA)<br />
CALL ABSE(TONS,DOA(J),THKA(J),L,NRA,NATR,UAA,CAPACITY)<br />
IF(UAA. GT. UAANOM. AND. CAPACITY.GE.TONS) THEN<br />
ROOF=NAT<br />
ELSE<br />
FLOOR=NAT<br />
END IF<br />
100 CONTINUE<br />
ELSE<br />
END IF<br />
RETURN<br />
END
B.3 RESPONSE SURFACE NETWORK MODEL<br />
C RESPONSE SURFACE NETWORK MODEL QNA<br />
SUBROUTINE QNA(MS1,MS2,MS3,F,LT,NTE,NTA,ET,AT,MAXU,CTTOT,COST)<br />
C ***************************************************************<br />
INTEGER I,J,K,L,NODES,R,ET,AT<br />
PARAMETER (NODES=5,R=8)<br />
REAL A(R),ALFA(NODES),AR(R),ARN(NODES),ARRIVAL,AVAL,<br />
+ ASSYM,<br />
+ B(NODES,NODES),BS(NODES),<br />
+ CA2(R),CA2N(NODES),CT(R),CTtot,<br />
+ CS2(R,10),CS2N(NODES),C02N(NODES),<br />
+ DN(NODES),DRN(NODES),<br />
+ EN(NODES),EW(NODES),<br />
+ FR(NODES,NODES),FRN(NODES),F,<br />
+ MATRIX(NODES,NODES),MF(NODES),MR(NODES),<br />
+ MAXU,MS1,MS2,MS3,<br />
+ NTE(R),NTA(R),<br />
+ P(NODES,NODES),P0(NODES),PCT(R),<br />
+ Q(NODES,NODES),QAUX(NODES),<br />
+ TSN(NODES),TS(R,10),<br />
+ U(NODES),U0(NODES),<br />
+ V(NODES),VAVG(NODES),<br />
+ W(NODES),WAVG(NODES),X(NODES),<br />
+ D,S1,S2,S3,LT,NET,NAT<br />
REAL CTQ,AUX,S1AUX,USA1,USA2,USA3,CTQSA1,CTQSA2,CTQSA3<br />
REAL PCSA1,PCSA2,PCSA3,PCS1,PCS2,PCS3,PCWTM1,PCWTM2,PCWTM3,<br />
+ MCSA1,MCSA2,MCSA3,<br />
+ ICSA1,ICSA2,ICSA3,ICS1,ICS2,ICS3,<br />
+ LCSA1,LCSA2,LCSA3,LCS1,LCS2,LCS3,<br />
+ WIPSA1,WIPSA2,WIPSA3,WIPWTM1,WIPWTM2,WIPWTM3,<br />
+ TSA1,TSA2,TSA3,<br />
+ COST,PCTUBE<br />
INTEGER M(NODES),NN(R),N(R,10),Z<br />
C --------------------------------------------------------------<br />
C INPUT DATA<br />
DATA (BS(I),I=1,NODES) /1.0,1.0,1.0,1.0,1.0/,<br />
& (M(I),I=1,NODES) /1,1,3,2,3/,<br />
& (MF(I),I=1,NODES) /80.0,0.0,80.0,80.0,80.0/,<br />
& (MR(I),I=1,NODES) /88.0,88.0,88.0,88.0,88.0/,<br />
& (PCT(I),I=1,R) /0.05,0.15,0.16,0.05,0.15,<br />
& 0.12,0.25,0.07/<br />
C<br />
C<br />
C Specify route<br />
M(3) = MS1<br />
M(4) = MS2<br />
M(5) = MS3<br />
TSA1=0.0<br />
TSA2=0.0<br />
TSA3=0.0<br />
255
ARRIVAL=80.0/365./24.<br />
DO 10 Z=1,R<br />
NN(Z)=5<br />
N(Z,1)=1<br />
N(Z,2)=2<br />
N(Z,3)=3<br />
N(Z,4)=4<br />
N(Z,5)=5<br />
AR(Z)=ARRIVAL*PCT(Z)<br />
NET=NTE(Z)<br />
NAT=NTA(Z)<br />
AVAL=0.59<br />
C --------------------------------------------------------<br />
C RESPONSE SURFACES<br />
TS(Z,1)=24.159+0.063*LT+0.563*NET+0.688*NAT-0.045*LT**2<br />
+ +0.125*NET*LT+0.080*NET**2+0.125*NAT*LT<br />
+ -0.125*NAT*NET+0.205*NAT**2<br />
TS(Z,1)=TS(Z,1)/AVAL<br />
C This function calculates the time before the TACK-UP<br />
C It is considered that the fabrication of the subcomponents<br />
C start at the same time and the processing time of each is<br />
C exponentially distributed and independent of the others<br />
C MEAN TIMES OF THE PRODUCT<br />
S1=24.159+0.063*LT+0.563*NET+0.688*NAT-0.045*LT**2<br />
+ +0.125*NET*LT+0.080*NET**2+0.125*NAT*LT<br />
+ -0.125*NAT*NET+0.205*NAT**2<br />
S1=S1/AVAL<br />
S2=29.295-0.312*LT+2.062*NET+1.397*NAT-0.023*LT**2<br />
+ -0.125*NET*LT-0.102*NET**2-0.125*NAT*LT<br />
+ -0.125*NAT*NET-0.227*NAT**2<br />
S2=S2/AVAL<br />
S3=8.636+1.25*NET+1.25*NAT-0.056*LT**2<br />
+ -0.056*NET**2-0.25*NAT*LT<br />
+ +0.25*NAT*NET+0.068*NAT**2<br />
S3=S3/AVAL<br />
TSA1=TSA1+S1*PCT(Z)<br />
TSA2=TSA2+S2*PCT(Z)<br />
TSA3=TSA3+S3*PCT(Z)<br />
S1AUX=S1<br />
S1=1.0/S1-ARRIVAL<br />
S2=1.0/S2-ARRIVAL<br />
S3=1.0/S3-ARRIVAL<br />
ASSYM=1.0/S1+1.0/S2+1.0/S3<br />
+ -1.0/(S1+S2)-1.0/(S1+S3)<br />
+ -1.0/(S2+S3)+1.0/(S1+S2+S3)<br />
ASSYM=ASSYM-S1AUX<br />
TS(Z,2)=ASSYM<br />
TS(Z,3)=66.04+1.312*LT+3.801*NET+4.062*NAT+0.147*LT**2<br />
+ -0.125*NET*LT+0.147*NET**2-0.125*NAT*LT<br />
+ +0.875*NAT*NET+0.397*NAT**2<br />
TS(Z,3)=TS(Z,3)/AVAL<br />
256
TS(Z,4)=63.272+1.05*LT+4.375*NET+4.875*NAT-0.363*LT**2<br />
+ +0.011*NET**2-0.378*NAT**2<br />
TS(Z,4)=TS(Z,4)/AVAL<br />
TS(Z,5)=81.954+2.937*LT+6.187*NET+5.937*NAT+0.102*LT**2<br />
+ -0.125*NET*LT+0.227*NET**2-0.125*NAT*LT<br />
+ +0.125*NAT*NET-0.522*NAT**2<br />
TS(Z,5)=TS(Z,5)/AVAL<br />
CS2(Z,1)=1.0<br />
CS2(Z,2)=0.0<br />
CS2(Z,3)=1.0<br />
CS2(Z,4)=1.0<br />
CS2(Z,5)=1.0<br />
10 CONTINUE<br />
CLOSE(15)<br />
C Obtain external arrival rates to each node I<br />
S1=0.0<br />
DO 110 J=1,NODES<br />
ARN(J)=0.0<br />
C Check if the first node of customer j is i<br />
DO 100 K=1,R<br />
IF(N(K,1).EQ.J) ARN(J)=ARN(J)+AR(K)<br />
100 CONTINUE<br />
S1=S1+ARN(J)<br />
110 CONTINUE<br />
DO 115 I=1,NODES<br />
P0(I)=ARN(I)/S1<br />
115 CONTINUE<br />
C Obtain FR(i,j)<br />
DO 150 I=1,NODES<br />
DO 140 J=1,NODES<br />
FR(I,J)=0.0<br />
DO 130 K=1,R<br />
DO 120 L=1,NN(K)-1<br />
IF(N(K,L).EQ.I. AND .N(K,L+1).EQ.J) THEN<br />
FR(I,J)=FR(I,J)+AR(K)<br />
END IF<br />
120 CONTINUE<br />
130 CONTINUE<br />
140 CONTINUE<br />
150 CONTINUE<br />
C Obtain departure rates from node I out of network<br />
DO 160 J=1,NODES<br />
DRN(J)=0.0<br />
C Check if the first node of customer j is i<br />
DO 155 K=1,R<br />
IF(N(K,NN(K)).EQ.J) DRN(J)=DRN(J)+AR(K)<br />
155 CONTINUE<br />
160 CONTINUE<br />
C<br />
C Obtain the routing matrix Q<br />
DO 190 I=1,NODES<br />
257
DO 180 J=1, NODES<br />
S1=0.0<br />
DO 170 K=1,NODES<br />
S1=S1+FR(I,K)<br />
170 CONTINUE<br />
Q(I,J)=REAL(FR(I,J))/REAL(DRN(I)+S1)<br />
180 CONTINUE<br />
190 CONTINUE<br />
C Obtain service-time for the nodes<br />
USA1=ARRIVAL*TSA1<br />
USA2=ARRIVAL*TSA2<br />
USA3=ARRIVAL*TSA3<br />
WIPSA1=USA1/(1-USA1)<br />
WIPSA2=USA2/(1-USA2)<br />
WIPSA3=USA3/(1-USA3)<br />
CTQSA1=TSA1*WIPSA1<br />
CTQSA2=TSA2*WIPSA2<br />
CTQSA3=TSA3*WIPSA3<br />
DO 220 J=1,NODES<br />
S1=0.0<br />
S2=0.0<br />
S3=0.0<br />
DO 210 K=1,R<br />
DO 200 L=1,NN(K)<br />
IF(N(K,L).EQ.J) THEN<br />
S1=S1+AR(K)*TS(K,L)<br />
S2=S2+AR(K)*(TS(K,L)**2)*(CS2(K,L)+1)<br />
S3=S3+AR(K)<br />
END IF<br />
200 CONTINUE<br />
210 CONTINUE<br />
TSN(J)=S1/S3<br />
CS2N(J)=S2/S3/TSN(J)**2-1.0<br />
220 CONTINUE<br />
C --------------------------------------------------------<br />
C Eliminate immediate feedback<br />
DO 225 I=1,NODES<br />
QAUX(I)=Q(I,I)<br />
TSN(I)=TSN(I)/(1.0-Q(I,I))<br />
CS2N(I)=CS2N(I)*(1.0-Q(I,I))+Q(I,I)<br />
Q(I,I)=0.0<br />
DO 224 J=1,NODES<br />
IF(J.NE.I) Q(I,J)=Q(I,J)/(1-QAUX(I))<br />
224 CONTINUE<br />
225 CONTINUE<br />
C --------------------------------------------------------<br />
C SOLVE SYSTEM OF TRAFFIC-RATE EQUATIONS<br />
DO 240 I=1,NODES<br />
DO 230 J=1,NODES<br />
IF(I.EQ.J) THEN<br />
MATRIX(I,J)=1.0-BS(I)*Q(I,I)<br />
258
ELSE<br />
MATRIX(I,J)=-Q(J,I)<br />
END IF<br />
230 CONTINUE<br />
240 CONTINUE<br />
OPEN(UNIT=15,FILE='matrix')<br />
write(15,*) NODES<br />
do 242,i=1,nodes<br />
do 241,j=1,nodes<br />
write(15,*) MATRIX(I,J)<br />
241 CONTINUE<br />
242 CONTINUE<br />
DO 243 J=1,NODES<br />
WRITE(15,*) ARN(J)<br />
243 CONTINUE<br />
CLOSE(15)<br />
CALL SYSTEM('solveqs.out')<br />
OPEN(UNIT=15,FILE='lambdas')<br />
DO 260 I=1,NODES<br />
C Calculate utilization and server load<br />
READ(15,*) FRN(I)<br />
U0(I)=FRN(I)*TSN(I)/REAL(M(I))<br />
IF(U0(I).GT.0.999) U0(I)=0.999<br />
ALFA(I)=FRN(I)*TSN(I)<br />
260 CONTINUE<br />
CLOSE(15)<br />
C ---------------------------------------------------------<br />
C<br />
C Calculate arrival rate to node j from node i<br />
C proportion of arrivals to j from i<br />
C departure rate out of network from i<br />
C total departure rate<br />
C Calculate modified variability function U(i)<br />
D=0.0<br />
MAXU=0.0<br />
DO 280 I=1,NODES<br />
S1=0.0<br />
IF (I.LT.NODES) THEN<br />
AUX=(1-U0(I+1))**2/(1-U0(I))**2<br />
U(I)=U0(I)*MIN(1.0,AUX)<br />
ELSE<br />
U(I)=U0(I)<br />
END IF<br />
DO 270 J=1, NODES<br />
FR(I,J)=FRN(I)*BS(I)*Q(I,J)<br />
P(I,J)=FR(I,J)/FRN(J)<br />
S1=S1+Q(I,J)<br />
270 END DO<br />
DN(I)=FRN(I)*BS(I)*(1-S1)<br />
D=D+DN(I)<br />
IF(U(I).GT.MAXU) MAXU=U(I)<br />
259
280 CONTINUE<br />
DO 300 J=1,NODES<br />
V(J)=0.0<br />
DO 290 I=1,NODES<br />
V(J)=V(J)+P(I,J)**2<br />
290 CONTINUE<br />
V(J)=1.0/(V(J)+P0(J)**2)<br />
W(J)=1.0/(1+4*((1-U(J))**2)*(V(J)-1))<br />
S1=REAL(M(J))<br />
X(J)=1+1.0*(AMAX1(CS2N(J),0.2)-1.0)/SQRT(S1)<br />
300 CONTINUE<br />
C ----------------------------------------------------------<br />
C Calculate VAVG(J) AND WAVG(J)<br />
DO 330 J=1,NODES<br />
VAVG(J)=0.0<br />
S2=0.0<br />
DO 320 K=1,R<br />
S1=0.0<br />
DO 310 L=1,R<br />
IF(N(L,1).EQ.J) S1=S1+AR(L)<br />
310 CONTINUE<br />
IF(N(K,1).EQ.J) THEN<br />
VAVG(J)=VAVG(J)+(AR(K)/S1)**2<br />
S2=S2+(AR(K)/S1)*CA2(K)<br />
END IF<br />
320 CONTINUE<br />
IF(VAVG(j).EQ.0) THEN<br />
VAVG(J)=0.0<br />
WAVG(J)=0.0<br />
C02N(J)=0.0<br />
ELSE<br />
VAVG(J)=1/VAVG(J)<br />
WAVG(J)=1.0/(1.0+4*(1-U(J))**2*(VAVG(J)-1))<br />
C02N(J)=1-WAVG(J)+WAVG(J)*S2<br />
END IF<br />
330 CONTINUE<br />
C ------------------------------------------------------------<br />
C NOTE: MODIFICATION OF A(J) AND B(I,J) BASED ON<br />
C GENERAL FORMULA OF BITRAN AND TIRUPATI (1988) FOR<br />
C MULTIPRODUCT QUEUING NETWORKS<br />
OPEN(UNIT=15,FILE='matrix')<br />
write(15,*) NODES<br />
DO 350 J=1,NODES<br />
S1=0.0<br />
DO 340 I=1, NODES<br />
B(I,J)=W(J)*P(I,J)*Q(I,J)*BS(I)*(1-U(I)**2+<br />
& (1-Q(I,J))*(1-P(I,J)))<br />
S1=S1+P(I,J)*((1-Q(I,J))*P(I,J)+BS(I)*<br />
& Q(I,J)*U(I)**2*X(I))<br />
S1=S1<br />
IF(I.EQ.J) THEN<br />
260
MATRIX(I,J)=1.0-B(I,I)<br />
ELSE<br />
MATRIX(J,I)=-B(I,J)<br />
END IF<br />
write(15,*) MATRIX(I,J)<br />
340 CONTINUE<br />
A(J)=1.0+W(J)*((P0(J)*C02N(J)-1.0)+S1)<br />
350 CONTINUE<br />
C ---------------------------------------------------------------<br />
C Solve system of traffic variability equations<br />
DO 351 J=1,NODES<br />
WRITE(15,*) A(J)<br />
351 CONTINUE<br />
CLOSE(15)<br />
CALL SYSTEM('solveqs.out')<br />
OPEN(UNIT=15,FILE='lambdas')<br />
DO 355 I=1,NODES<br />
C Calculate utilization and server load<br />
READ(15,*) CA2N(I)<br />
355 CONTINUE<br />
CLOSE(15)<br />
C --------------------------------------------------------------<br />
C CALCULATE THE PERFORMANCE OF THE NETWORK (Congestion at the nodes)<br />
CTQ=0.0<br />
DO 800 I=1,NODES<br />
S1=2*(M(I)+1)<br />
S1=SQRT(S1)-1.0<br />
EW(I)=(CA2N(I)+CS2N(I))/2.0*TSN(I)*<br />
& U0(I)**S1/(M(I)*(1.0-U0(I)))<br />
EN(I)=ALFA(I)+FRN(I)*EW(I)<br />
C Return to the original measures (before eliminating feedback)<br />
FRN(I)=FRN(I)/(1.0-QAUX(I))<br />
TSN(I)=(1.0-QAUX(I))*TSN(I)<br />
CS2N(I)=(CS2N(I)-QAUX(I))/(1.0-QAUX(I))<br />
EW(I)=(1-QAUX(I))*EW(I)<br />
CTQ=CTQ+EW(I)<br />
800 CONTINUE<br />
CTtot=0.0<br />
DO 900 K=1,R<br />
CT(K)=0.0<br />
DO 850 I=1,NODES<br />
IF(I.EQ.2) TS(K,I)=TSN(I)<br />
CT(K)=TS(K,I)+CT(K)<br />
850 CONTINUE<br />
CT(K)=CT(K)+CTQ<br />
CTtot=CTtot+CT(K)*PCT(K)<br />
900 CONTINUE<br />
910 FORMAT(1X,F12.3,F12.3)<br />
920 FORMAT(1X,F12.3,F12.3,F12.3)<br />
C<br />
C -------------------------------------------------------------------<br />
261
C Estimate cost of the product<br />
MCSA1=27620.0<br />
MCSA2=18810.0<br />
MCSA3= 5905.0<br />
PCSA1=MCSA1<br />
PCSA2=MCSA2<br />
PCSA3=MCSA3<br />
LCSA1=17.0*2*TSA1<br />
LCSA2=17.0*2*TSA2<br />
LCSA3=17.0*2*TSA3<br />
COST=LCSA1+LCSA2+LCSA3<br />
ICSA1=WIPSA1*PCSA1*0.1/80.0<br />
ICSA2=WIPSA2*PCSA2*0.1/80.0<br />
ICSA3=WIPSA3*PCSA3*0.1/80.0<br />
COST=COST+ICSA1+ICSA2+ICSA3<br />
PCWTM1=PCSA1+LCSA1+ICSA1<br />
PCWTM2=PCSA2+LCSA2+ICSA2<br />
PCWTM3=PCSA3+LCSA3+ICSA3<br />
WIPWTM1=TSN(2)*ARRIVAL<br />
WIPWTM2=(TSN(2)+TSN(1)+CTQSA1-TSA2-CTQSA2)*ARRIVAL<br />
WIPWTM3=(TSN(2)+TSN(1)+CTQSA1-TSA3-CTQSA3)*ARRIVAL<br />
PCS1=PCWTM1+WIPWTM1*PCWTM1*0.1/80.0+<br />
+ PCWTM2+WIPWTM2*PCWTM2*0.1/80.0+<br />
+ PCWTM3+WIPWTM3*PCWTM3*0.1/80.0<br />
COST=COST+WIPWTM1*PCWTM1*0.1/80.0+<br />
+ PCWTM2+WIPWTM2*PCWTM2*0.1/80.0+<br />
+ PCWTM3+WIPWTM3*PCWTM3*0.1/80.0<br />
LCS1=17.0*MS1*TSN(3)<br />
ICS1=PCS1*(ARRIVAL*EW(3))*0.1/80.0<br />
PCS2=PCS1+LCS1+ICS1<br />
LCS2=17.0*MS2*TSN(4)<br />
ICS2=PCS2*ARRIVAL*EW(4)*0.1/80.0<br />
PCS3=PCS2+ICS2+LCS2<br />
LCS3=17.0*MS3*TSN(5)<br />
ICS3=PCS2*ARRIVAL*EW(5)*0.1/80.0<br />
CALL CTUBES(PCTUBE,F,LT,NTE,NTA,ET,AT)<br />
COST=COST+LCS1+ICS1+LCS2+ICS2+LCS3+ICS3+PCTUBE<br />
END<br />
SUBROUTINE CTUBES(COST,F,LT,NET,NAT,ET,AT)<br />
INTEGER R,I,ET,AT<br />
REAL INVE,INVA,Q,F,S,TETA,CE(8),CA(8),D(8),NE(8),NRVAL,<br />
+ NA(8),F1(8),PCT(8),L(8),DEM,AUX1,AUX2,CAVG,COST,<br />
+ CE1(8),CA1(8),LT,NET(8),NAT(8),COSTE(4),COSTA(4)<br />
DEM=80.0<br />
DATA (PCT(I),I=1,8) /0.05,0.15,0.16,0.05,<br />
+ 0.15,0.12,0.25,0.07/<br />
DATA (COSTE(I),I=1,4) /6.0,6.5,7.0,7.5/,<br />
& (COSTA(I),I=1,4) /3.25,3.50,3.75,4.0/<br />
S=0.99<br />
INVE=0.0<br />
262
INVA=0.0<br />
CAVG=0.0<br />
DO 20 I=1,8<br />
D(I)=DEM*PCT(I)<br />
CE(I)=COSTE(ET)<br />
CA(I)=COSTA(AT)<br />
L(I)=(LT+2)*1.75+17<br />
NE(I)=(NET(I)+2)*115+440<br />
NA(I)=(NAT(I)+2)*115+440<br />
CE1(I)=L(I)*NE(I)*CE(I)<br />
CA1(I)=L(I)*NA(I)*CA(I)<br />
CAVG=CAVG+PCT(I)*(CE1(I)+CA1(I))<br />
F1(I)=F*PCT(I)<br />
TETA=D(I)/54.0*7.0<br />
AUX1=INT(D(I)/F1(I))<br />
AUX2=NINT(D(I)/F1(I))<br />
Q=MAX0(AUX1,AUX2)<br />
R=1<br />
CALL NR(R,TETA,NRVAL)<br />
AUX1=(1.0-S)*Q<br />
DO WHILE (NRVAL.GT.AUX1)<br />
CALL NR(R,TETA,NRVAL)<br />
R=R+1<br />
END DO<br />
INVE=INVE+CE1(I)*(Q/2.0+R-TETA)<br />
INVA=INVA+CA1(I)*(Q/2.0+R-TETA)<br />
WRITE(18,*) Q*NE(I), R*NE(I),Q*NA(I),R*NA(I),(R-TETA)*NE(I)<br />
20 CONTINUE<br />
COST=(INVE+INVA)*0.1/80.0<br />
RETURN<br />
END<br />
SUBROUTINE NR(R,TETA,NRVAL)<br />
INTEGER K,R<br />
REAL GR,NRVAL,PR,TETA<br />
GR=EXP(-TETA)<br />
DO 20 K=1,R<br />
CALL PROB(PR,K,TETA)<br />
GR=GR+PR<br />
20 CONTINUE<br />
CALL PROB(PR,R,TETA)<br />
NRVAL=TETA*PR+(TETA-R)*(1.0-GR)<br />
RETURN<br />
END<br />
SUBROUTINE PROB(PR,K,LAMBDA)<br />
INTEGER I,K<br />
REAL LNPK,PR,LAMBDA,AUX<br />
LNPK=-LAMBDA+K*LOG(LAMBDA)<br />
DO 10 I=1,K<br />
263
AUX=REAL(I)<br />
LNPK=LNPK-LOG(AUX)<br />
10 CONTINUE<br />
PR=EXP(LNPK)<br />
RETURN<br />
END<br />
264