Gunter Fandel· Thomas Gulledge

Albert Jones (Eds.)

Operations Research
in Production Planning
and Control
Proceedings of a Joint German/US Conference,
Hagen, Germany, June 25-26, 1992

Under the Auspices of

Deutsche Gesellschaft flir Operations Research (DGOR),
Operations Research Society of America (ORSA)

. With 152 Figures

Springer-Verlag
Berlin Heidelberg New York
London Paris Tokyo
Hong Kong Barcelona
Budapest
 Prof. Dr. Gunter Fandel, Institute of Production Management
Femuniversitat Hagen, FeithstraBe 140/AVZ II
D-5800 Hagen, FRG
Prof. Dr. Thomas Gulledge, The Institute of Public Policy
George Mason University, 4400 University Drive
Fairfax, Virginia 22030-444, USA
Dr. Albert Jones, AMFR, NIST B 124 Metrology,
Gaithersburg, Maryland 20899, USA

ISBN-13: 978-3-642-78065-3 e-ISBN- 13: 978-3-642-78063-9

001: I 0.1 007/978-3-642-78063-9
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broad-
casting, reproduction on microfilms or in other ways, and storage in data banks. Duplication of
this publication or parts thereof is only permitted under the provisions ofthe German Copyright
Law of September 9, 1965, in its version of June 24, 1985, and a copyright fee must always be paid.
Violations fall under the prosecution act of the German Copyright Law.
© Springer-Verlag Berlin· Heidelberg 1993
Softcover reprint of the hardcover I st edition 1993
The use of registered names, trademarks, etc. in this publication does not imply, even in the absence
of a specific statement, that such names are exempt from the relevant protective laws and regulati-
ons and therefore free for general use.
2142/7130-543210 - Printed on acid-free paper
 Preface

This proceedings volume contains selected and refereed contributions that were presented at
the conference on "Recent Developments and New Perspectives of Operations Research in
the Area of Production Planning and Control" in Hagen/Germany, 25. - 26. June 1992.

This conference was organized with the cooperation of the FernuniversiHit Hagen and was
jointly hosted by the "Deutsche Gesellschaft fur Operations Research (DGOR)" and the
"Manufacturing Special Interest Group of the Operations Research Society of America
(ORSA-SIGMA)". For the organization of the conference we received generous financial
support from the sponsors listed at the end of this volume. We wish to express our
appreciation to all supporters for their contributions.

This conference was the successor of the JOInt ORSA/DGOR-conference in

Gaithersburg/Maryland, USA, on the 30. and 31. July 1991. Both OR-societies committed
themselves in 1989 to host joint conferences on special topics of interest from the field of
operations research. This goal has been successfully realized in the area of production
management; and it should be an incentive to conduct similar joint conferences on other
topics of operations research in the years to come. The 36 contributions in this proceedings
volume deal with general and special problems in production planning as well as approaches
and algorithms for their solution. They cover a wide range of operations research within
product management and will therefore address a wide circle of interested readers among
OR-scientists and professionals alike.

Last, but not least, we would like to thank all who have shared efforts to make this conference
a success. In particular Lioba Lanwer, Peter Fran'$ois and Dr. Richard Lackes are recognized
for their help in organizing and conducting the conference. We thank also Prof. Dr. Wolfgang
Buhler and Prof. Dr. Lothar Streitferdt of the DGOR for initiating this joint venture of
DGOR and ORSA.

Gunter Fandel, General Chairperson, DGOR

Thomas Gulledge, Co-Program Chairperson, ORSA-SIGMA
AI Jones, Co-Program Chairperson, NIST
 TABLE OF CONTENTS

I. General Planning and Modelling Approaches

Planning Product Assortment Using Portfolio Optimization 3

Charles V. Trappey , Amy J . C. Trappey and Richard Feinberg

Rapid Modeling: The Use of Queueing Models to Support 21

Time-Based Competitive Manufacturing
Rajan Suri and Suzanne de Treville

Process Network Theory and Implementation for 31

Technology Assessment in Manufacturing
R . Lal Tummala, Bruce E. Koenig and
Herman E . Koenig

Is Lean Production Really Lean? 49

The Design of a Lean Production System
Joachim Reese

The Performance of Kanban Controlled Serial 71

Production Systems
John A. Buzacott , Suzanne M. Price and
J. George Shanthikumar

Capacity Oriented Production Control for a 89

Job Shop Production
Karl-Werner Hansmann

On Solving a Large-Scale Resource Allocation Problem 105

in Production Planning
Christof Dillenberger , Laureano F . Escudero ,
Artur Wollensak and Wu Zhang

Integrated Station Location and Flow Network Design 120

for Manufacturing Systems Layout
Dilip Chhajed. Timothy J . Lowe and Benoit Montreuil

Using Optimization Model to Control workpiece 138

Rigidity and Deformation in Workholding to Achieve
Precision Machining
Amy J . C. Trappey , Parag Gupta and C. Richard Liu

A New Approach to Determine the Pose and Dimension 151

of Polygonal Shapes
Jose A. Ventura and Jen-Ming Chen
 VIII

II. Flexible Manufacturing Systems

Problems of Modelling and FMS Control Implementation 169

Janez Dekleva and Matjaz Gaberc

Formal Models for Control of Flexible Manufacturing 184

Systems
Sanjay Joshi, Jeffrey S. Smith and Richard A. Wysk

Candidate Rule Selection to Develop Intelligent 201

Scheduling Aids for Flexible Manufacturing Systems (FMS)
Yuehwern Yih and Albert T. Jones

Tool Blocking in Flexible Manufacturing Systems 218

Ulrich A. W. Tetzlaff

III. PPC and CIM

Architectural Framework for Integrated Production 235

Control Systems
August-Wilhelm Scheer

A New Conceptual Framework for Hierarchically 246

Integrated PPS-Systems
Christoph SchneeweiB

Issues in Specifying Planning Horizons for Production 270

Planning within elM Environments
S. D. Thompson, J. A. Jewell and Wayne J. Davis

IV. Scheduling

A Closed Loop Control Heuristic for Scheduling a 293

Manufacturing System Subject to Random Setup Times
Mitchell H. Burman and Stanley B. Gershwin

Resource- and Time Window-Constraint Production 307

Scheduling with Alternative Process Plans:
An Artificial Intelligence Approach
Andreas Drexl and Arno Sprecher

Robust Scheduling and Game-Theoretic Control for 321

Short-Term Scheduling of Job-Shops
Victor Jorge Leon , S . David Wu and Robert H. Storer

The Two-Stage Hybrid-Flowshop Scheduling Problem 336

with Sequence-Dependent Setup Times
Stefan VoB
 IX

Shop- Floor Scheduling and Control: A Systems Approach 353

Richard A. Wysk, Richard J. Hayer, Hyuenbo Cho,
Sanjay Joshi, Jeffrey Smith, Walter Hoberecht,
C. Dennis Pegden and Albert L. Jones

Spatially-Oriented Production Planning and Scheduling 371

Uwe Petersen

V. Lot-Sizing and Inventory Control

Models for Integrated Inventory Control by EDI for 389

A Single Supplier and Multiple Buyers
Avijit Banerjee and Snehamay Banerjee

Optimal Inventory Policy in Assembly Systems with 405

Component Commonality
Srinivas Bollapragada and Ram Akella

Valuation of Leadtime Reduction in Multi-Stage 413

Production Systems
Karl Inderfurth

Nervousness and Reorder Policies in Rolling Horizon 428

Environments
Thomas Jensen

Some Valid Constraints for the Capacitated 444

Assembly Line Lotsizing Problem
Heinz D. Hathes

DLSP for Multi-Item Batch Production 459

Wolfgang Bruggemann and Hermann Jahnke

Information Production Functions in Dynamic Lot-Sizing 473

Gunter Fandel and Peter Fran90is

VI. Quality

How Much Quality Is Enough? 499

James R. Dorroh , Thomas R . Gulledge and
Norman K. Womer

Decision Methodology for Product Quality in Flow 511

Lines
Layek Abdel-Halek and Xiuli Chao
 x
VII. Neural Networks and Genetic Algorithms

Multiple-Step-Ahead Prediction by Hierarchical 529

Neural Networks
Tep Sastri

Integrating Neural Nets, Simulation, and Genetic 550

Algorithms for Real-Time Scheduling
Albert Jones and Luis Rabelo

A Genetic Algorithm for Scheduling with Resource 567

Consumption
Timothy Starkweather, Darrell Whitley and Bud Cookson

Genetic Algorithms in Problem Space for Sequencing 584

Problems
Robert H. Storer, S. David Wu and InKyoung Park

List of Sponsors 598

I. General Planning and Modelling Approaches
 Planning Product Assortment Using Portfolio Optimization

Charles V. Trappey
Department of Consumer Sciences and Retailing
Purdue University, W. Lafayette, In 47907, USA

Amy J. C. Trappey
Department of Industrial & Manufacturing Systems Engineering
Iowa State University, Ames, Ia 50011, USA

Richard Feinberg
Department of Consumer Sciences and Retailing
Purdue University, W. Lafayette, In 47907, USA

Abstract

The task of planning product assortments using portfolio optimization is developed and
applied to a case in apparel manufacturing and distribution. The planning process begins with the
retail buyer, the key individual in the organization that selects merchandise for resale to the
consumer. In order to maximize the profit and minimize the risk of the merchandise investment,
the retail buyer must simultaneously consider many different factors that impact the merchandise
mix decision. A merchandise planning model is developed for the retail buyer based on the
concept of stock portfolio theory. The investment returns and risks are considered to determine the
product assortment. A retrospective statistical test is conducted to verify the goodness of the
product assortment portfolio model. The methodology, developed for a large Midwest retailer, can
be extended to a wide variety of cases in manufacturing, especially cases involving the production
of consumer products.

Key Wonk Merchandise assortment planning, portfolio theory, product life cycle.

1. Introduction

This research introduces the use of investment portfolio theory to optimize two
merchandise investment planning decisions. The first decision is the merchandise item selection.
Given a set of n items available for investment, find m items to form a maximum profit, minimum
risk merchandise assortment (Figure 1). The selection criteria are based upon the buyer's purchase
objectives and evaluation of risk and returns. The second decision is the item percentage decision.
4

Given the set of m items, selected by the item selection model, as a portfolio, find the optimal
percentage of investment dollars for each item (Figure 2). That is, maximize the profit and
minimize the risk of the merchandise percentage investment portfolio based upon the buyer's
investment objectives and evaluation of risk and returns. Both analytical optimization models
depend upon the projected merchandise returns that can be derived from a merchandise investment
knowledge base. In this paper, the item percentage decision model is implemented and tested.

m items to be
choosen for
purchase

Figure 1. Merchandise item selection (Decision 1).

al% item 1

a2% item 2

a3% item 3

.
.
am% itemm

al % + a2% + ... + am% = 1

Figure 2. Item percentage decision (Decision 2).

The main objective is to assist the merchandise buyer in deciding which merchandise to buy
and in what quantities. The objective of applying portfolio theory to these decisiqns is to lower the
investment risk and increase the investment return. In Section 2, the background of the portfolio
theory and investment systems is reviewed. In Sections 3, the mathematical basis of two
merchandise portfolio decisions is discussed. Section 4 describes the formulation of the
optimization models incorporating portfolio theory to support merchandise assortment decisions.
In Section 5, an example of the merchandise investment planning scenario is presented to
demonstrate the model implementation. Section 6 introduces the retrospective test to compare the
decisions made based on the portfolio model, an expert buyer, and the existing textbook method.
Finally, Section 7 summarizes the contribution of this research and the future research directions.
 5

2. Background

Markowitz (1959) first introduced the portfolio selection concept to demonsttate how the
risk of a portfolio investment can be reduced, and how investment capital can be optimally
allocated between assets which have different risks and returns. Diversification of a portfolio is a
strategy whereby the investor attllmpts to reduce risk by investing in assets that provide different
yields under different economic conditions. By using a portfolio diversification approach, the
assets held in the portfolio are protected from extreme fluctuations in the economy. Portfolio
diversification is necessary because there is risk in investment, especially when high rates of return
are expected. Retailers in particular are continually investing in portfolios of merchandise that have
degrees of risk and rates of return highly correlated with fashion levels. Siegel (1979) introduces
this concept as the Fashion Triangle. At the top of the triangle, there are the "new fashion" items.
These items are the highest risk group, with almost a 100% chance of requiring a mark-down in
price. Buyers are advised never to invest more than 5% of their budget in these items. The
"fashion right" items occupy the next 15% of the triangle, and represent a high risk category of
merchandise that will require advertisement to educate the consumer and stimulate interest in the
new look. At the bottom of the triangle lies the "fashionable" merchandise, the 80% portion of the
investment that can be handled by assistant buyers and will never require a markdown. Siegel's
approach to merchandise investment is based upon "rules-of-thumb," a heuristic derived through
experience. The Fashion Triangle approach may be suitable for reducing the risk of merchandise
investments, but it is unknown whether or not the assortment chosen will yield the maximum
returns for the given level of risk. Neither is it known that the 5%-15%-80% assortment will
minimize the risk for the expected level of returns.
Markowitz portfolio theory uses four standard statistical measures to analyze the variability
of investments and the expected returns of those investments. As described by Baumol (1970), the
expected value (or mean) is used to measure the returns of the portfolio as well as the returns of
each component investment. The standard deviation measures the dispersion of the returns around
the expected value, and the covariance and correlation coefficient are used to measure the degree of
parallelism in the behavior of the investments. The principle assumption of Markowitz's approach
is that the investor's preference for a portfolio depends only upon the mean and standard deviation
of the components as the predictor of the risk of the portfolio. In order to minimize the risk of the
portfolio, Markowitz formulated a nonlinear optimization model with the formula for portfolio risk
as the objective function. The constraints consisted of a proportionality requirement to insure that
the investments added up to 100%, and a non negativity requirement to insure positive proportions.
Many extensions to this early formulation are discussed in Portfolio Management (Smith 1971).
Although the original intention of Markowitz's theory was to reduce the risk of portfolios
of securities, the theory has been extended to include investments in a number of different types of
assets. Several authors (Corstjens & Weinstein, 1982; Mahajan, Wind, & Bradford, 1982; Wind,
1974; Wind & Mahajan, 1981) provide both a theoretical and a mathematical development of
6

portfolio theory and apply the theory to product selection and analysis. Another group of authors
(Allaway, Mason, & Brown, 1987; Mason & Mayer, 1990; Rosenbloom, 1980) discuss the
application of portfolio theory in the retail organization and provide an overview of the product
portfolio concept. The research conducted by these authors establishes the theoretical foundation
for creating and utilizing merchandise portfolios in a retail setting.

3. Mathematical Basis of Merchandise Portfolio Risk and Returns

Given the forecasted states of apparel demand and the expected returns under the demand
scenarios, the expected merchandise rate of return I{Rj] is computed using the formula

(1)

where Ps is the probability of the demand scenario s occurring and Rjs is the investment return for
item j under the dem~nd scenario s. Referring to the example in Table 1, Rj is the weighted
average of the returns Rjs (or the geometric mean of the distribution) under the different levels of
consumer demand for each jacket. The variance and the standard deviation of the distribution are
computed using the formulas:

(2)

(3)

Table 1 Example data for the comparison of the expected returns and standard deviations of
four related merchandise items.

I uRVISBC ORVISCJ :CROADS] CROADCJ

EconomiC State
P~ Rl R2 R~ ~
Bad 0.2 0.05 0.15 -0.15 -0.15
Average 0.5 0.10 0.20 0.00 0.15
Good 0.3 0.20 0.05 0.50 0.35
0.120 0.145 0.120 0.150
Rj
OJ 0.056 0.065 0.255 0.173

ORVISBC - low risk Orvis Back Country jacket

ORVISCJ - low risk Orvis Corduroy jacket
CROADS] - high risk Country Road Storm jacket
CROADCJ - high risk Country Road Cashmere jacket
 7

0.6 . , . - - - - - - - - - - - - - - ,

0.5

Probability .. ORVISBC
... ORVISCJ
of 0.4 .. CROADSJ
... CROADCJ
Occurance

0.3

0.2 +-_.---.--O--.-........---.--....---.r---I
-0.2 0.0 0.2 0.4 0.6
Rate of Returns

Figure 3 Dispersion of merchandise risk for a portfolio of four jackets

The individual jackets are combined as an investment portfolio. Other investment

classifications could be used to build a merchandise portfolio. A portfolio of vendors, a portfolio
of merchandise lines, or a portfolio of department stores are examples of the different types of
merchandise portfolios that can be created. The simple portfolio of four jackets is used to illustrate
how the risk of a portfolio can be reduced while simultaneously maximizing the returns.
The risk of the merchandise portfolio can be reduced by building a collection of
investments that respond differently to the expected states of economy. Figure 3 illustrates the
differences in risk for four jackets. Notice that the Orvis Corduroy jacket (ORVISCJ) performs
best during bad and average economic conditions, since the jacket is a basic item with little fashion
risk. The high fashion items (Country Road jackets) and the more fashionable Orvis Back Country
jacket have curves which counter the investment returns of the Corduroy jacket. The jacket
scenario models a case where the state of the economy drives the consumer to switch from
expensive, high fashion jackets (e.g., Country Road Cashmere) to basic styles constructed out of
more affordable fabrics (e.g., Orvis Corduroy). The Orvis Corduroy jacket is negatively
correlated with the three other jackets since it is more profitable when the other jackets are less
profitable and vice versa.
By considering the interactions of returns within the portfolio, a formulation can be derived
which minimizes the risk. The covariance of the merchandise returns,

3
COVij = L Ps(Ris - RiXRjs - Rj)
8=1 (4)
8

is used to measure the portfolio risk when items i and j are selected. Notice that when i = j, then
COVij = Of, which is the variance of item j. Since the risk of the portfolio is a function of the
weighted variances and covariances of the merchandise returns (Baumol, 1970), the portfolio risk
may be substantially increased or decreased by the selection of items or by the percentage
investment in items. The equation for portfolio risk is presented below.

n m
I. I. Wi\\jOOVij
i=1 j=l (5)

For the decision depicted in Figure 1, Wi is 1 (0) when item i is selected (not selected).
Furthennore, Wi is the percentage invested in item i when the equation is used to model the
decision depicted in Figure 2. The merchandise investment planning objective is to minimize the
risk and maximize the expected returns. The expected returns of the portfolio are represented as

(6)

In order to maximize the returns of the portfolio while minimizing the risk, Equations 5 and 6 are
combined as an objective function. In the next section, mathematical optimization models are
derived to represent the merchandise portfolio concept.

4. Modeling Merchandise Portfolio Decisions

Two mathematical optimization models are provided to assist the retail merchandise buyer.
One is the item selection decision model (Decision 1), and the other is the item percentage decision
model (Decision 2). If the buyer is given a set of n items for purchase, the task is to find the best
m items to fonn a maximum profit, minimum risk portfolio. If the buyer is given a set of m items
as a portfolio, the task is to find the percent investment of each item that will maximize the returns
and minimize the risk of the portfolio. Each of the above tasks is fonnulated as a non-linear
programming problem. Only the first model (Section 4.1) requires that the variables be restricted
to integers. Section 4.2 presents the item percentage decision fonnulation.

4.1 Item Selection Decision

Given a set of n items (Xl' x2' x3' ... , Xn) available for purchase, find m items to fonn a
maximum profit, minimum risk portfolio. The selection criteria is based upon buyer's purchase
objectives and tolerance of risk 0\:).
 9

n n n
Max. L Xi~ - nc L L XiXjOOVij
i=1 i=l j=1 (7)
ST.
n
L Xi=m
i=l (8)

Xi = 0 or 1. i = 1.2.3 •...• n

where n is the number of Xi merchandise available fo~urchase and m is the desired

number of Xi items to carry in the merchandise portfolio. Ri is the expected returns for item
Xi. COVij is the covariance (or variance when i=j) for items Xi and '1. Other optional
constraints:

v
L Xj~y%m
i=1 (9)
(at least y% of m items selected have to be staple goods (or fashion goods. etc.) as
described by the set of v items (xl' x2. x3' ...• xv' ...• Xm).

v
LXi=1O
i=1 (10)
(at least 10 of the m items selected have to belong to the set of v items. v ~ 10)

".;=1 (11)
(item j must be chosen)

PIc is the risk adjustment factor used to reflect the buyer's investment philosophy. Three
levels of PIc are identified:

a. Risk Adverse - the buyer wishes to maintain a moderate level of returns but is willing to
tolemte only minimum levels of risk. Thus. PI is determined from

n _ In n
PI >>> L Xj~ L L XjXjCOVij
i=l i=l j=l (12)

b. Risk-Return Indifferent - risk and returns are equally weighed in the selection of items. P2
is determined from
10

n _ /n n
P.2 = L ~Rj L L ~"jroVij
i=1 i=1 j=1 (13)

c. Profit Seeker - the buyer wishes to maximize returns and is willing to tolerate high levels of
risk. P3 is determined from

n _ /n n
P3 <<< L ~Rj L L ~"jroVij
i=1 i=1 j=1 (14)

4.2 Item Percentage Decision

Given a set of m items (CI'~' c3 •...• em) as a portfolio. find the optimal percentage
investment in each item. That is. maximize the profit and minimize the risk of the portfolio
depending upon the buyer's investment objectives and tolerance of risk (<JIc).

m m m
Max L wiRj - CIk L L WiWjroVij
i=1 i=1 j=1 (15)
ST.

(16)

where m is the desired number of items to carry in the merchandise portfolio.

'tc is the risk adjustment factor that reflects the buyer's investment philosophy. Three levels of'tc;
risk adverse(Clt). risk-return indifferent (CJ2). and profit seeker (<11). are identified. The value of'tc
is set using the equations specified for 1\.

5. An Example of the Percentage Investment Decision

The sequence of steps required to solve the percentage investment decision model are
shown in Figure 4. The steps illustrate the optimization of the percentage investment decision.

Step 1 Select the items to be included in the portfolio. The item selection model could be applied
at this point. However. in order to simplify the problem. the selection is pre-determined
and is given as the four men's jackets listed in Table I.
 11

Step 1 Pre-select merchandise items

for men's sportswear portfolio

Step 2
----
Detennine investment returns
using a knowledge base MIR_3b
I
•
. .. -- ,..

Step 3 Compute variance I covariance

matrix using C program "Matrix"

Step 4
•
Fonnulate Markowitz quadratic
programming model to minimize
risk and maximize returns.
.....l.
Fonnulate constraints for the
Step 5 model.

•
Solve nonlinear model using
Gino, a generalized reduced
Step 6
gradient optimizer.

Figure 4. The automated decision making process for the merchandise investment plan

Step 2 A knowledge base program is used to estimate the Gross Profit on Sales (GPS) for each
jacket. Given the GPS value, the Merchandise Investment Returns (MIR) can be
estimated. For the jacket example, assume that a knowledge base MIR_3b (Trappey,
1992) predicted the following for the Orvis Back Country jacket:

Ec!;lDQwil; Sta~ Ps QPSTenn Q~S Yillu~ MIR

Bad Economy 0.2 Make Some 40% 5%
Average Economy 0.5 Make Some 42% 10%
Good Economy 0.3 MakeAlot 45% 20%

Using actual retail sales data from Younkers, a Mid-west chain of department stores, the
relationship between GPS and MIR (for men's sportswear) is modeled as a polynomial of
degree four. In order to use the polynomial equation to represent both positive and
negative investment returns (a GPS value of 38% represents a MIR of zero), a simple
transfonnation is used to shift the curve down the y-axis:

y" = - 0.567 - 0.574 x + 11.765 x2 - 26.054 ~ + 24.738 " (17)

12

Using Equation 17, the MIR for the Orvis Back Country jacket is computed for each
economic state. Also, the MIR_3b knowledge base is used to estimate the probability of
occurrence for each economic state.

Step 3 Given the Ps and MIR values for ORVISBC, ORVISCJ, CROADSJ, and CROADCJ, the
third step is to compute the variance I covariance matrix using Equation (4). The matrix
output is summarized in Table 2.

Table 2 Using a variance I covariance matrix to model the merchandise investment risks.

Step 4 Use the data from Step 3 to formulate the Markowitz quadratic programming model:

I Objective Function - Maximize portfolio returns for an average economy where the buyer
is risk adverse (Ck = 10.0). Using Equation 15,
1) MAX = ( 0.1200 • ORVISBC + 0.1450 • ORVISCJ + 0.1200 • CROADSJ +
0.1500· CROADCJ) - 10.0· (0.0031 • ORVISBC 1\ 2 + 0.0042 • ORVISCJ 1\ 2
+ 0.0651 • CROADSJ 1\ 2 + 0.0300 • CROADCJ 1\ 2 - 0.0058 • ORVISBC •
ORVISCJ + 0.0282· ORVISBC· CROADSJ + 0.0180· ORVISBC· CROADCJ -
0.0288 • ORVISCJ • CROADSJ - 0.0120 • ORVISCJ • CROADCJ + 0.0780 •
CROADSJ • CROADCJ );

Step 5 Formulate the constraints for the programming model:

I Constraint - All classifications should be included in the portfolio and the

summation of all investment percentages is 100%. Using Equation 16,
2) ORVISBC + ORVISCJ + CROADSJ + CROADCJ = 1;

! Constraints - Each classification should represent at least 15% of the portfolio

4) ORVISBC > .15;
5) ORVISCJ > .15;
6) CROADSJ > .15;
7) CROADCJ > .15;

I Constraints - No classification should represent more than 50% of the portfolio

8) ORVISBC < .50;
9) ORVISCJ < .50;
10) CROADSJ < .50;
11) CROADCJ < .50;
 13

Step 6 Solve the non-linear programming model using Gino software (Liebman, Lasdon,
Schrage, & Waren, 1986). Gino uses the generalized reduced gradient algorithm (GRG2).
The solution of the model, for three different investment strategies, is presented in Table 3.

Table 3 shows the optimal percentage investment for a risk adverse strategy, an indifferent
strategy, and a profit seeker strategy. For the risk adverse strategy, ORVISCJ (the Corduroy
jacket) has the least risk, so it has the highest investment percentage allocated. High risk items
such as CROADSJ and CROADCJ have the lowest suggested investment percentages. In order to
increase the returns of the portfolio, the buyer must be willing to assume more risk. The buyer of
merchandise assumes more risk for the risk indifferent strategy (the model risk adjustment factor
(<It) is set to 1.0). The portfolio returns increase to 13.5% as the risk is increased. The model
switches 5% of the investment from the low risk Orvis Back Country jacket to the high return,
high risk Country Road Cashmere jacket. When the merchandise buyer totally disregards the risk
of the portfolio (profit seeker strategy) and chooses only to max~mize the returns of the investment,
the model allocates 70% of the investment to the two most profitable items - the Cashmere and the
Corduroy jacket. Using a profit seeker strategy, the returns of the portfolio reach a maximum of
14%.

Table 3 Example Investment Returns for Three Investment Strategies

For each of the three investment strategies described above, the probabilities were based
upon the assumption of an average economy. If the probabilities had predicted a good or a bad
economy, then the model would have yielded an investment strategy tailored to satisfy either of
these conditions. In conclusion, the portfolio approach gives the retail buyer a decision making
tool which can be used to model (1) merchandise under different economic conditions, (2) different
risk tolerance strategies, and (3) a variety of different types of merchandise classifications. The
example used in this section illustrates the application of the model to a simple portfolio of four
jackets. The application can be extended to a number of different applications in retailing. Instead
of building a portfolio of jackets, the buyer could feasibly build a portfolio of vendors, or explore
the optimum allocation of goods between staple and fashion merchandise categories. The model
could also be used on a more aggregate level to plan merchandise investments across departments
in a department store, or across stores for a multi-store operation. In the next section, a
14

retrospective test is used to statistically compare the investment outcome of the model to the
investment outcome of a retail buying expert. The Younkers men's sportswear buyer typically
purchases over 300,000 items every six months, representing a seasonal investment of over 5
million dollars. Management confmned that the buyer providing data for the test was highly
capable and successful, and had a proven record of profitable investment planning.

6. Retrospective Test

A Merchandise Assortment Plan (MAP) was generated using the percentage investment
portfolio optimization model (Equations 15 & 16). A test methodology called a retrospective test
(Hillier & Liberman, 1974) was used to determine how well the MAP would have performed had it
been used. The objective of the test was to determine whether or not the optimized assortment plan
would yield as good of a profit as the expert buyer's best plan.

6.1 Test Methods

The net profit of a portfolio of men's sportswear items was used at the performance
measure to test the Hypothesis. Net profit was selected as the performance measure since
profitability provides "the hard numbers on which continuing merchandise assortments are built."
(Seegal, 1979). The MAP obtained from the expert buyer and the MAP obtained from the
optimization model provided the data from which the net profit values were computed. The
buyer's MAP was taken directly from the Younkers 1991 Fall season assortment plan, and
represents a sample of the buyer's investment strategy for 22 classifications of merchandise. The
model MAP is the output generated from the application of the percentage investment decision
optimization model (Equations 15 & 16) to the 22 classifications of merchandise. The net profits
are computed for each of the 22 classifications (Table 4). The research hypothesis is stated as
follows:

Hypothesis: The profit of the mathematical optimization model is significantly different

from the buyer in planning 22 classifications of merchandise.

The experiment for testing the Hypothesis is concerned with the differences in net profit that may
occur when a portfolio optimization model is used or not used to generate the MAP. The test of the
hypothesis is based upon the nonparametric sign test (Conover, 1971) of the net profits computed
for the MAPs. The sign test is used to determine if the net profit measures derived from the
expert's MAP are significantly smaller than the net profit measures derived from the optimized
MAP. The inputs to the optimization model included (1) a set of 22 basic classifications of
merchandise for the Fall planning season, (2) the expected returns for each classification (Rj), (3)
 15

the probability (Ps) of a given state of economy occurring during the planning period, and (4) the
risk adjustment factor <1\:).
During a field interview, the expert buyer stated that the plans for the 1991 Fall Season
were based upon expectations of an average economy. When the buyer was formulating the plan,
Younkers was experiencing an increase in store traffic, fewer sales, no change in markdowns, and
no change in gross margin. The national economy also seemed to be experiencing growth. Given
this data, the knowledge base created for the test stated that "there is weakly suggestive (-0.3)
evidence that the economy is not 'bad,' suggestive evidence that the economy is 'average,' and
weakly suggestive (0.3) evidence that the economy is 'good.' After converting the certainty values
to probabilities, the three economic states were set to 20% chance bad economy, 50% chance
average economy, and 30% chance good economy. The buyer's records provided information
about the minimum and maximum stock levels recommended by management, information which
was used to formulate the constraints of the optimization model. Since the buyer was observed to
be risk I return indifferent, the impact of risk and returns were equally weighed in the selection of
items (Le., the model risk adjustment factor CJk was set to 1.0). The net profit for each of the 22
classifications of merchandise was calculated using the standard formula for net income, earnings,
or profits (Davidson, Stickney, & Weil, 1982). Gross profit represents the excess of merchandise
sales revenues over the cost of the merchandise sold, whereas net profit represents the excess of
gross profit over the operating costs of the department store. The expert buyer reported that the
Operating Cost Index (OCI) for the department store typically ranged from 30% to 40% of the total
retail sales. Based upon an analysis of the Fall season sales results, the operating cost for the
men's sportswear department was set to 35.7% of the retail sales. The buyer's plan provided an
accurate count of the number of items received and the cost and retail of the items sold during the
1992 Fall season. Using the buyer's data, the net profit was computed for 22 classifications of
merchandise. The merchandise portfolio selected by the buyer (MAPl) was optimiZed using the
Gino software to form MAPl. Using a published merchandise planning heuristic called the
Fashion Triangle (Siegel, 1979), MAl'3 was created.
As an example of the retrospective net profits computed, consider the classification for
rugby knit shirts. Table 4 shows that the expert buyer's net profit for rugby shirts was $88,300.
Using the results of the Gino portfolio model, the net profit was determined to be $lO4,000.
Using the Fashion Triangle (FT) planning heuristic, the net profit in retrospect was determined to
be $11,900.

6.2 Test Results

The retrospective test results are presented in two parts. Section 6.2.1 provides a historical
basis for the buyer's investment strategy. Section 6.2.2 utilizes the non-parametric sign test to
evaluate the statistical significance of the differences between the buyer's plan and the
16

mathematically optimal plan. The differences between the portfolio approach, the expert buyer,
and the Ff approach are also tested.

Table 4 Comparison of net profit (in thousands of dollars) for three assortment plans

MAP)
Abbrev. Ex
or nswtr
cttnswtr -1.7
woolswtr -0.1
wlblswtr -4.2
-16.3

6.2.1 Historical Analysis

Sweaters
In retrospect, acrylic sweaters were poor sellers during the 1991 Christmas season. The
portfolio model correctly reduced the investment in acrylic sweaters and increased the investment in
orlon sweaters. The buyer also argued (in retrospect) that an increase in the supply of cotton
sweaters during the peak of Christmas sales would have substantially improved his departmental
performance. The model results however, do not support an increase in cotton sweaters.

Woven Shirts
The expert buyer invested the largest percentage of woven shirt money in high risk updates
and fashions. Since the model selected assortment percentages based upon a 30% probability of a
good economy, a 50% probability of an average economy, and a 20% probability of a bad
economy, the optimal allocation is somewhat different from the expert. Evidently the expert
selected a wovens investment strategy based upon expectations of a good economy. Furthermore,
 17

since the model parameters were set to avoid risk, the largest investment percentage went to the
low risk, regular weight wovens. The portfolio model, in retrospect, recommended that the buyer
focus on a basic stock of wovens instead of a fashion forward collection of wovens.

Knit Shirts
During the 1991 Fall season at Younkers, the best selling men's sportswear items were
knits. Rugby and fleece knit shirts were top performers in the stores located in Iowa, Nebraska,
and Illinois. The buyer's largest investment was in rugbys, followed by fleece wear. The
portfolio approach fairly matched the expert's rugby strategy, and invested the maximum allowed
by the model constraints. Since fleece items are considered by the buyer to be most profitable ina
good economy, and since Docker's knits are typically profitable regardless of the state of the
/
economy, the model shifted investment percentages away from polos, updated knits, and Bugle
Boy knits and increased the investment in rugbys and fleece. The portfolio model correctly
reduced the investment in Bugle Boy knits.

In order to maintain an advantageous sales relationship, Younkers committed a large

portion of its pants budget to Levi Dockers. Even though the expert buyer felt that Docker pants
were not strong performers in average and bad economies, he assumed that the purchase of larger
quantities of pants would provide a sound basis for negotiating good prices and delivery terms for
tops. Dockers knits and wovens are considered by the buyer to be strong performers regardless of
the state of the economy. The Gino model investment strategy did not account for this loss / profit
relationship and pursued a different goal. Since the model parameters were set for an average
economy, a larger percentage of low risk jeans and Bugle Boy pants were ordered. Investments
which perform poorly in an average economy, such as Docker pants and updated / fashion pants,
were reduced to lower bound levels.

6.2.2 Sign Tests

Si~n Test for MAP!. MAP}

A one tailed sign test (Conover, 1971) was used to test the null hypothesis that the buyer's
net profit values for MAPl are more likely to exceed the optimized net profit values for MAPl.
Rejecting the null hypothesis would indicate that the optimized net profit values tend to be larger
than the buyer's. The data in Table 4 show that 13 of the 22 net profit values planned by the model
exceeded the net profit values planned by the buyer. If 13 < (n - t), the null hypothesis cannot be
rejected, where (n - t) is the critical value, n =22 in this case study, and t =0.5 x (n + ou Yii).
When a = 0.05, ou =~.05 = -1.6449 and then (n - t) = 15. Thus, the results of the sign test fail
18

to reject the null hypothesis and provide no statistical evidence that the net profit values of the
optimization model tend to be larger than the buyer's (p > 0.05).

Sign Test for MAP!. M&3

A one tailed test was used to test the null hypothesis that the Ff net profit values for MAP3
are more likely to exceed the buyer's net profit values for MAP•. Rejecting the null hypothesis
would indicate that the buyer's net profit values tend to be larger than the values resulting from the
FI' approach. When the net profit values for the two plans were compared, 17 of the buyer's net
profit values exceeded the net profit values obtained via the Ff approach (fable 4). The results of
the sign test provide statistical evidence, i.e., 17> (n - t) = 15 at IX = 0.05 (or p ~ 0.05), that the
buyer's net profit values tend to be larger than those computed using the FI' approach. The null
hypothesis is rejected since there is a significant difference between the two sets of net profit
values.

Sign Test for MArl. M&3

A one tailed test was used to test the null hypothesis that the Ff net profit values for MAP3
are more likely to exceed the Gino model net profit values for MAIl2. Rejecting the null hypothesis
would indicate that the model's net profit values tend to be larger than the values resulting from the
FI' approach. When the net profit values for the two plans were compared, 17 of the model's net
profit values exceeded the net profit values obtained via the Ff approach (fable 4). The results of
the sign test provide statistical evidence (p ~ 0.05) that the model's net profit values tend to be
larger than those computed using the FI' approach. The null hypothesis is rejected since there is a
significant difference between the two sets of net profit values.

7. Conclusion

This paper demonstrate the applicability of the portfolio concept to the product assortment
planning problem. The traditional task of planning merchandise assortments, often regarded as an
art that cannot be taught (Mason & Mayer, 1990), can be quantitatively modeled using the
technique provided and tested in this paper. This research contributes to the retail automation by
presenting a scientifically sound methodology to assist product assortment decision making
processes.
The portfolio model developed for the percentage merchandise investment decision cannot
be shown to be significantly better than the expert buyer when net profit is used as the comparative
basis. The net profits resulting from the use of the portfolio model are at least as good as the
expert for the given economic condition. Even though there is no statistically significant difference
between the expert and the model, there is an observable difference in net profit. When the net
profit of the buyer's assortment ($123,600) is compared to the net profit of the portfolio model
($155,000), the portfolio model outperforms the buyer by at least $31,000.
 19

The plans for the expert and portfolio model were also compared to the plan derived using
the Fashion Triangle approach. Both the expert and the model were significantly better than the FT
approach. Both the expert and the model produced profitable plans, whereas the FT approach
produced a plan which yielded a net loss of $202,000. The retrospective analysis of the FT
approach indicates that there are a number of serious limitations to the heuristic. The most serious
limitation is that the FT approach failed to produce a profitable plan. The FT approach also
produces a plan inconsistent with the expected state of the economy and the buyer's risk tolerance
level.

8. References

Allaway, A., Mason, J.B., and Brown (1987). An optimal decision support model for department
-level promotion mix planning. Journal of Retailing, Vol. 63, No.3, pp. 215-242.

Baumol, W. J. (1970). Portfolio Theory: The Selection of Asset Combinations. New York:
McCaleb-Seiler Publishing Company.

Conover, J. (1971). Practical NOllparametric Statistics. New York, NY: John Wiley & Sons,
Inc.

Corstjens, M. and Weinstein, D. (1982). Optimal strategic business units portfolio analysis. In
A.A. Zoltner (ed.), Marketing Planning Models: T1MS Studies in the Management Sciences,
Vol. 18, pp. 141-160. New York: North-Holland Publishing Co.

Davidson, S., Stickney, C., & Weil, R. (1982). Financial Accounting: An Introduction to
Concepts, Methods, and Uses. Chicago, II: The Dryden Press.

Hillier, F. S., & Lieberman, G. J. (1974). Operations Research. San Franscisco, CA: Holden-
Day, Inc.

Liebman, J, Lasdon, L., Schrage, L., & Waren, A. (1986). Modeling and Optimization with
GINO. San Franscisco, CA: The Scientific Press.

Mahajan, V., Wind, Y., & Bradford, J. (1982). Stochastic dominance rules for product portfolio
analysis. In A.A. Zohner (Ed.), Marketing Planning Models: T1MS Studies in the
Management Sciences, Vol. 18, pp. 161-183. New York: North-Holland Publishing Co.

Markowitz, H. M. (1959). Portfolio Selection: Efficient Diversification of Investments. New

York: John Wiley & Sons.

Mason, J. B., & Mayer, M. L. (1989). Retail merchandise information systems for the 1980's.
Journal of Retailing, Vol. 56, No. I, pp.56-76.

Mason, J. B., & Mayer, M. L. (1990). Modern Retailing: Theory and Practice. Boston, MA:
BPI/Irwin.

Rosenbloom, B. (1980). Strategic planning in retailing: Prospects and problems. Journal of

Retailing, Vol. 56, No. I., pp. 107-120.

Smith, K. (1971). Portfolio Management. New York: Hoh, Rinehart and Winston.
20

Seegal, H. (1979). Keys to profit. In R. Patrick Cash (Ed.), The Buyers Manual, pp. 357-376.
New York, NY: National Retail Merchants Association.

Siegel, J. B. (1979). Managing the risk of fashion merchandising. In R. Patrick Cash (Ed.), The
Buyers Manual, pp. 287-298. New York, NY: National Retail Merchants Association.
Trappey, C. V. (1992). A Merchandise Investment Planning System Using Knowledge Based
Representation and Portfolio Theory. Ph.D. Thesis, Purdue University, W. Lafayette,
Indiana. .

Wind, Y. (1974). Product portfolio analysis: A new approach to the product mix decision. In
Ronald Curhan (ed.), American Marketing Association Combined Proceeding, Series No. 36,
pp. 460-464.

Wind, Y., and Mahajan, V. (1981). Designing product and business portfolios. Harvard
Business Reyiew, Vol. 59, No. I, pp. 155-165.
 Rapid Modeling:
The Use of Queueing Models to Support
Time-Based Competitive Manufacturing

Rajan Suri
University of Wisconsin-Madison
Dept of Industrial Engineering
1513 University Avenue, Madison, WI 53706, USA
and

Suzanne de Treville
Helsinki University of Technology
Institute of Industrial Automation
Espoo 02150, Finland

Limited Practical Application of Queueing Models

A recent paper in Interfaces laments that! "the mathematical and often

abstract nature of many articles on queueing suggests that queueing theory is just
that: only theory and difficult to apply in practice." It goes on to document the
"limited use of queueing theory" in practice. We agree wholeheartedly with those
comments, and in the present paper we will address this issue from the point of
view of manufacturing applications of queueing models. Despite the fact that the
analysis of queueing systems was formalized with Erlang's work at the beginning
of this century2, the use of queueing models in manufacturing and operations
management (OM) remains limited. There has actually been a lot of progress in
the research on queueing models in manufacturing3 , but queueing models are
seldom used in industry for real manufacturing decision-making. When
compared with other OR techniques such as LP or Simulation, we see that there
is also a scarcity of commercially supported software packages enabling the use of
queueing models for manufacturing.

Thus we see that an important OR technique, which -- according to papers

that are over two decades old -- should be very useful in manufacturing, is seldom
even considered by people in industry. While there are many reasons for this,
discussed at conferences and in the society publications, we will focus here not on
the reasons for failure, but rather, on rectifying this situation. Today we stand at
the brink of a new opportunity for queueing models in OM. The face of
manufacturing is changing and the new developments provide great potential for
22

successful application of queueing models. We discuss this change, and also

provide references to some recent applications papers that may signify a trend to
improve the "track record" of practical uses of queueing models. We will only
discuss a few of the main points here, so we refer the reader to a companion paper
by us which describes some additional concepts, has more detailed case studies,
and also discusses some resulting implications for the successful application of
ORIMS techniques in general 4 •

Time-Based Manufacturing

Much of the traditional OM work in manufacturing has focused on

scheduling problems, and often on deterministic scheduling. Key objectives
usually included utilization of labor and equipment, or meeting pre-specified due-
dates. The new paradigm of manufacturing is primarily based on reducing
overall lead time. While resources are not unduly wasted, utilization is of
secondary concern -- particularly in the case of non-bottleneck resources.
Although the questions raised by the old paradigm are not well addressed by
queueing theory, we will argue here that queueing theory may indeed be the best
approach to addressing the questions raised in the new paradigm. We will now
introduce this new paradigm.

'90s ""' Speed

'80s ""' Quality

l

'70s ""' Cost

l

'60s ""' Scale

...

Figure 1: The Evolution of Manufacturing Strategy from the 1960's on

The keyword for competitive manufacturing in the 1990's will be speed.

Some perspective will help here (see Figure 1). In the 60's the world was still in a
high growth phase due to the ongoing recovery from World War II, and so the
 23

more a firm produced, the more it could sell. U.S. firms dominated world
markets and the name of the game in manufacturing was Scale. As some of the
European and Japanese firms entered the world markets more effectively in the
70's, competition shifted to costJprice issues. Then in the 80's came the first major
"paradigm shift". Primarily through the results of the Japanese firms, the
manufacturing world realized that there does not have to be a tradeoff between
cost and quality: if one focused on improving quality, then cost competitiveness
would follow as well . Now in the 90's another paradigm shift has occurred,
leading to a focus on speed: this is also known as time-based competition. "Time is
becoming the main battlefield -- and weapon -- of competition".5 How exactly does
one compete on time? In essence, time-based competition means delivering
products or services faster than ones competitors 6. But there is more involved.
The very act of looking for ways to speed up existing procedures results in
manifold benefits. For instance, one company found that efforts to reduce
throughput time were closely linked to other improvements:7 "Everything we
wanted to do to improve operations had something to do with squeezing time out of
our processes. The deeper we probed, the more opportunity we saw."

For manufacturing firms , competing with time primarily means reducing

their lead times -- both, the time to bring new products to market, and the time to
manufacture an existing product from the raw materials -- and efforts to reduce
these can result in significant competitive advantage 8 . To take the quality
paradigm a step further, competing on speed usually results in both quality and
cost improvements as well. In other words, a firm that decides to be a time-based
competitor can, in principle, supply its customers with products that are of better
quality than other firms, at lower cost, and faster than anyone else -- a truly
formidable competitive advantage.

Rapid Modeling Teclmology

The shift to time-based manufacturing provides us with the opportunity to

make queueing models playa greater role in applications. However, we take our
old friend from the beginning of the century, namely queueing theory, and
reposition it in an effort to make greater industrial impact. We call our "applied"
version of queueing theory Rapid Modeling Technology (RMT).

The idea of building a model of a manufacturing process to aid in decision-

making is, of course, not new. In manufacturing situations where there exist
complex system dynamics, uncertainty, and other stochastic elements, (so that
24

LP cannot be readily used) the OR tool of choice traditionally has been

simulation9 • Given the power of simulation, why is an alternative modeling
technology required? The answer lies in the time required for both, building and
running, simulation models. Typically, model development requires weeks or
months, and run time is measured in hours (particularly when the model is run
on a PC).10 With a few recently available software tools based on queueing theory,
models of similar size can be created in hours, and run in seconds l l . Therefore,
we felt the name "Rapid Modeling Technology" (RMT) better highlighted the
advantages of queueing theory-based tools, and helped to differentiate them from
other approaches. We have also developed "industrial strength" descriptions of
RMT which are quite different from traditional descriptions of queueing theory
(see our companion paper12 ).

Capabilities of a State of the Art RMT Package

Many industrial engineers or OR analysts are under the impression that

the use of queueing models requires such restrictive assumptions as "Poisson
arrivals" or exponential service time distributions. However, many advances have
been made in the last two decades 13 . Approximate solutions can now be obtained
for a fairly general class of systems. With respect to the new manufacturing
paradigm, these approximations provide adequate estimates of performance to
guide decision making. Consider a company trying to reduce its lead time from 15
weeks to 10 days -- not uncommon in the world of time-based manufacturing.
Suppose an RMT model, one that might even be off by 20%, predicts that a set of
decisions will achieve a 10 day lead time. At worst, this means that the actual lead
time achieved may be 12 days, instead of 10. Looking at it from the standpoint of 15
weeks, 12 days looks great! The RMT model may be enough to justify an initial
decision. Instead of waiting weeks for detailed analyses, management can
proceed before the competition does.

To illustrate the class of manufacturing systems that can be modeled, we

describe the capabilities of an RMT tool frequently used by us in our consulting
(one of us was also part of the team that conceived of, and developed, the tool).
Figure 2 summarizes the inputs and outputs of MPXTM, a RMT software package
from Network Dynamics Inc., Burlington, Mass., USA14. This tool requires basic
information on the products to be made, and the resources available --
information that is easily entered into tables, guided by pull-down menus and
windows. Its RMT engine represents state-of-the-art queueing methods, with the
ability to model many manufacturing features such as: multiple products, each
 25

with its own routing, setup times, and cycle times; general service times;
equipment failures; labor shared across equipment (also known as "operator
interference"); lot sizes and transfer batches.

Products to be
Fabrical9d
Equipment
Resources
I ~=-I Indented Bill
of Materials
Part
Routings
Operation
Assignments

~\I//
RMT
ENGINE
Automated Reports and Color Graphics

//I\~~
Production
Equipment
Utiization
Labor
Utilization
Scrap &
Assembly
Work-in-Process r--::I
L::-1
Levels Inventory
and Queues and Queues Reports

Figure 2: Inputs and Outputs of a State-of·the-art Rapid Modeling Package

These features are a far cry from the "exponential assumptions" of the
early queueing models, and are critical to the industrial acceptance of the tooL
Outputs from MPX include reports on production levels, equipment and labor
utilizations and queue lengths, work-in-process, and product lead times. While
the only values reported are "steady state averages", use of RMT tools has shown
that most of the time these reports are sufficient for management to choose
among alternatives, or to obtain insight on critical parameters for a particular
manufacturing system 15_

Traditional Versus Modern Manufacturing

As one shifts from traditional manufacturing to time-based

manufacturing, fundamental changes are needed in manufacturing decision-
making (see Figure 3). The traditional view of capacity (as espoused in MRP and
other capacity planning systems) is simple -- you either have enough or you don't_
So the production schedule is either feasible or infeasible. The modern view
however, is that there is a gradually worsening tradeoff between capacity
utilization and lead times. Comparing the upper and lower graphs on the left side
26

of Figure 3 helps to explain many of the problems of large firms. Since

manufacturing managers were measured on how effectively they could use
capacity, they tended to operate their resources at close to 100% (or even higher!)
utilization -- with disastrous results such as very long lead times and high work-
in-process. They also needed frequent bouts of expensive "overtime" to catch up on
past due jobs (periodically bringing the utilization under 100% by such draconian
measures). The two graphs on the right show another piece of the same story. The
utilization measure just mentioned only gave credit for actual "run" time. Setup
was thus considered bad, and so manufacturing managers tried to run as large a
lot size as they could. Their efficiency appeared to increase with lot size (upper
right graph). The lower right graph shows, once again, the result of this
mistaken strategy -- very long lead times and high work-in-process. (The U-shape
of this graph is explained further below.)

Traditional View
Capacity Volume
Utilization Production

Production
Feasible

Utilization Lot Size

Modern View
Time-based Time-based
Competition 100% Competition
D.I

Utilization Lot Size

Figure 3: Traditional Versus Modem Views of Manufacturing

The two points we would like to emphasize now, are that (i) queueing
models can do a good job of illustrating the above behaviors which were previously
neglected, and (ii) queueing (or RMT) is arguably the "best" decision support
technology for time-based manufacturing. We now elaborate on these points.
 27

Rapid Modeling in Support of Time-Based Manufacturing

In traditional manufacturing situations, products can spend 80-95% of

their time waiting for resources. Thus, any attempt to quantify the impact of
alternative manufacturing strategies for lead time reduction, needs to have a
good estimate of waiting times ...and that's exactly what queueing theory is all
about! In fact, RMT is the only technology that can accurately predict the three
main performance measures -- utilizations, lead times, and work-in-process
(WIP) -- in only seconds of PC time.

It is worth discussing further the two graphs on the lower half of Figure 3.
The one on the left is the "classic" result to be found in any basic text on queueing
theory, that lead time and WIP grow unbounded as utilization approaches 100%.
The one on the right however, while fundamental for modern manufacturing, is
relatively unknown in practice, although it has been known in the OR literature
for many years16. We give here a brief explanation of it. The graph goes up on the
right because, as lot size increases, two things happen: (i) pieces spend more time
waiting for other pieces in their own lot; (ii) lots spend more time waiting for other
lots ahead of them -- while the first item can be estimated by a deterministic
formula, this second item can only be predicted by queueing theory. The graph
goes up on the left because as lot size decreases, setups increase, leading to higher
utilization and thus higher waiting times. Somewhere in-between is a "good" lot
size which leads to low lead times and WIP (and incidentally, this lot size bears
little relation to the now-obsolete "Economic Lot Size" formula). Thus there are
some very interesting dynamics and tradeoffs in a multi-product manufacturing
system that are well-captured by queueing theory. A sophisticated RMT tool can
capture these effects for a real factory, as illustrated in several articles 17 .

A Wisconsin manufacturer of landscaping equipment was under pressure

to reduce lead times. Not only was the WIP carrying cost high, but customers
were constantly complaining about delayed orders. One line that produced
crankshafts had severe problems. The entire processing time for a crankshaft
couldn't have been more than an hour, yet the 30 station line had a lead time of
three weeks! Using RMT to study the problem, we were able to set up an adequate
model of the line in under a day. Just an hour of playing "what-ifs" pointed
towards a very simple solution. The crankshafts were moved from station to
station in containers of substantial size. We found that by cutting this size down, a
significant reduction in lead time could be obtained. This somewhat obvious
28

conclusion had been previously proposed but not implemented because it was
thought that too much material handling would be required, and line
inefficiencies would result. The RMT model showed, however, that lead time
could be cut by one half or more, with little impact on overall line efficiency. The
beauty of this solution was that it could be done at a nominal cost -- the cost of the
new containers -- and had been obtained with little analysis effort.

In addition to guiding the reduction of lead times, RMT also helps to speed
up new product introductions. The ease and speed with which models can be built
encourage productJprocess designers to use RMT at early stages of a project. This
can provide valuable feedback on critical processes and equipment needs, which
removes potential sources of delay later on in the product introduction effort.
Brown I8 details the case of RMT helping to speed up a new product introduction
by IBM. As reported there, the firm was planning a large factory-of-the-future to
manufacture printed circuit boards. The entire factory was modeled in a matter of
days, and hundreds of what-ifs were answered in just over a month, including
many alternatives for implementing just-in-time, lot sizes for material handling,
and yields under various scenarios. The rapidity with which the analysis was
completed sped up the project implementation time; in addition, early feedback
influenced process choices and equipment purchases, and enhanced the
successful implementation of the factory.

As we have argued elsewhere, the implemention of time-based competition

in manufacturing firms is supported by two other modern developments -- Kaizen
and Concurrent Engineering l9 . Our companion article 20 discusses how RMT can
be key in supporting these two concepts as well.

The Potential for Success

As manufacturers strive to implement speed, the rapidity of analysis that

can support decision-making becomes equally vital. RMT captures enough of the
time-based elements of the manufacturing system to provide useful management
insight, and yet does so in a manner that is simple and fast -- it is thus usable by
most people in a manufacturing enterprise. No other technology offers this
combination of abilities. Queueing theory, in its new incarnation as Rapid
Modeling Technology, has the potential for successful application by supporting
management decision-making for time-based competition in manufacturing.
 29

FOOTNOTES
1 Buchanan, J. and J. Scott, "Vehicle Utilization at Bay of Plenty Electricity", Interfaces,
Vol.22 No.2, March-April 1992,28-35.

2 See, for example, the Preface in Queueing Systems, Vol. I, L. Kleinrock, John Wiley, 1975.

3 R. Suri, J.L Sanders and M. Kamath, "Performance Evaluation of Production Networks",

Chapter in Handbooks in Operations Research and Management Science, Vol.4, S.C .
Graves, A.H.G. Rinnooy Kan and P.H. Zipkin (Eds.), Elsevier, to appear in 1992.

4 R Suri and S. deTreville, "Full Speed Ahead: A Look at Rapid Modeling Technology in
Operations Management", OR/ MS Today, June 1991, 34-42.

5 Peters, T. "Tomorrow's Companies", The Economist, March 4, 1989, 19-22.

6 Stalk, G., Jr., "Time -- The Next Source of Competitive Advantage", Harvard Business
Review, July-August 1988, pages 41-51.

7 Merrills, R "How Northern Telecom Competes on Time", Harvard Business Review, July-
August 1989,108-114.

8 Suri, R, "Lead Time Reduction through Rapid Modeling", Manufacturing Systems, July
1989,66-68.

9 Law, A.M. and S.W. Haider. 1989. Selecting simulation software for manufacturing
applications: Practical guidelines and software survey. Ind. Eng. 31 33-46.

10 See the statistics in: Anderson, K.R 1987. A Method for Planning Analysis and Design
Simulation of CIM Systems. Proc. 1987 Winter Simulation Conference, 715-720; and Brown,
E. 1988. IBM combines rapid modeling technique and simulation to design PCB factory-of-
the-future. Industrial Eng. 20 (6) 23-26.

11 See the articles in the preceding footnote.

12 Suri and deTreville, 1991, op.cit.

13 Whitt, W, Approximations for the GI/G/m queue, Advances in Applied Probability (to
appear); Bitran, G.R and D. Tirupati, Multiproduct queueing networks with deterministic
routing: Decomposition approach and the notion of interference. Management Science 34 (1)
75-100,1988; also see the survey in Suri, Sanders and Kamath, op.cit.
30

FOOTNOTES
14 This tool was developed by Greg Diehl and Mike Tomsicek of Network Dynamics along with
one of us (Suri).

15 See the articles in footnotes 4, 8 and 10.

16 See, for example, Karmarkar, U.S., "Lot Sizes, Lead Times and In-Process Inventories",
Management Science, March 1987,409-418.

17 See several application articles listed in Suri and deTreville, 1991, op.cit.

18 Brown, E., 1988. IBM combines rapid modeling technique and simulation to design PCB
factory-of-the-future. Industrial Eng. 20 (6) 23-26.

19 S. deTreville and R. Suri, "Understanding and Implementing Time-Based Competition in

Manufacturing Firms", IE Dept Working Paper, University of Wisconsin-Madison, July
1992.

20 Suri and deTreville, 1991, op.cit.

 PROCESS NETWORK THEORY AND IMPLEMENTATION
FOR
TECHNOLOGY ASSESSMENT IN MANUFACTURING

R. Lal Tummala, Bruce E. Koenig and Herman E. Koenig

Department of Electrical Engineering
Michigan State University
East Lansing, Michigan, USA 48823-1226

Abstract- This paper describes a theoretical framework called Process Network Theory and
its implementation for evaluating the economic, technical and environmental performance of
manufacturing systems. Historically, performance of manufacturing enterprises has been
measured largely in monetary terms such as cash, value added and net return on investment.
Achieving and maintaining a long-term competitive position in a global economy requires
looking beyond monetary performance measures and financial accounting systems. It is
necessary to evaluate alternative technologies in the context of their economic performance,
technical performance and environmental loading. Technical performance, as defined here
is the physical performance of manufacturing systems in terms of physical flows of materials,
products and energy. Many companies recognize the crucial role of technical performance
measures such as defect and scrap rates, yields, throughput time and resource efficiencies in
retaining competitive position in the economy. They are experimenting with processes and
procedures that stress quality control (TQC), reduce throughput time (JIT), improve process
efficiencies, and others. Process network theory provides the ability to determine the
technical and economic performance of manufacturing systems by following a two step
procedure. As a first step, the physical part of the manufacturing system is represented as
a network of elemental processes whose basic function is to transform the technical state of
materials using physical energy and skill-specific human time. The models of the elemental
processes are parameterized by a set of feasible technologies which are selected based on
constraints and opportunities available for the exchange of materials and energies between the
manufacturing system and its economic and natural environment. Linear graph theory is used
to obtain the overall model of the manufacturing system as a function of the technologies and
the network structure. The model is then used to obtain technical, economic and
environmental performance as a function of the technologies involved.
32

INTRODUCTION
Historically, performance of the manufacturing enterprise has been measured largely in
monetary terms such as cash flow, value added, and net return on investment, Johnson and
Kapl<w [1987]. Contemporary management accounting systems and budgeting processes deal
almost exclusively in financial terms. Achieving and maintaining a long-term competitive
position in an environmentally sensitive global economy requires looking beyond internalized
monetary performance measures and financial accounting systems.
Today, many companies recognize the crucial role of technical pe1jormance measures
such as defect rates, yields, throughput time, and resource efficiencies, in retaining a
competitive position in the economy. They are experimenting with processes and procedures
that stress total quality control (TQC), reduce throughput time (JIT), improve manufacturing
cycle effectiveness (MCE), flexible manufacturing (FM), and material requirements planning
(MRP) , Kaplan [1989] and [1982], Veilleus and Petro [1988], Sethi and Sethi [1990],
Fogarty, Blackstone and Hoffmann [1991], Browne, Haren and Shivan [1989]. Some have
begun to experiment with new approaches to management accounting systems for operations
control and technical performance measurement, at both the enterprise and industry levels of
organization, i.e., in the context of their own manufacturing operations as articulated with
their industrial partners as suppliers and customers.
Many companies are also searching for processes and procedures for factoring
environmental load constraints such as air and water emission standards, alternative natural
resource supplies, and available human resource skills into the management of their
operations and in their long-term strategic plans for products, production processes, and plant
siting, Society of Environmental Toxicology and Chemistry [1991]. Environmental loads,
along with available human resource skills and worker expectations are already factors in
plant closings and relocation.
Strategic trade-offs must now be made between monetary pe1jormance, technical
pe1jormance and environmental factors at both the enterprise and industry levels of
organization and/or simultaneous improvements must be realized in all three. Articulation
and evaluation of these trade-offs and/or improvements can be viewed as a two-stage iterative
process;
1) Technology innovation --- synthesis of specialized knowledge and understanding to
identify alternative technologies at the process level.
 33

2) Technology assessment --- evaluating the relative merits of alternative processing

technologies in the context of the overall manufacturing operation. Improvements
and/or trade-off in monetized performance, technical performance, and environmental
loads, are the central issues of technology assessment. Selection of processing
technologies based on strategic choices among these trade-off issues is referred to as
technical design.
Technology innovation at the process level has a long-standing history of achievement in
the industrialized nations of the world. However, rigorous technology assessment as defined
above has only recently been recognized as a critical element of design and operations control
in manufacturing. A theoretical framework for systematically quantifying trade-offs in
monetized performance, technical performance and environmental loads is discussed in this
paper. The overall paradigm is given in Figure 3.

I. THE MANUFACTURING ENTERPRISE

A diagram of the form illustrated in Figure 1 is used to characterize manufacturing
enterprise as a network of production processes. The larger closed boundary (shown dashed)
identifies the boundary of the network as an enterprise or manufacturing sub-network. The
internal small closed boundaries (boxes) identify the boundaries of production processes. The
directed line segments identify the flow rates of technically-specific materials between
processes. The human resources and synthetic energy required to maintain the material flows
are represented as energetic costs associated with each technically-specific material.
The technically-specific outputs of each production process within the network are
considered to be the objective of the respective production process. They are always
identified in the network as bold-faced line segments with end points, and carry labels 01,
02, etc., as indicated. The resources required and the by-products produced by each
production process are always identified with normal line segment and their endpoints. If the
product of one production process is linked as a resource to another process within the
network, it is so indicated by connecting the end points of the corresponding line segments.
The point of connection is called a node. Theoretical considerations require one and only one
product to be incident to any given node. All resources and by products that are linked
internally to production process are labeled 4-1, 3-2, 5-3, etc. to identify the processes so
linked. All materials crossing the boundary of the network are consistently labeled rl, r2,
34

processing assembly
by products by products

5-4
Processing 04 Assembly LI~-----," Product
3-2 2 : 02 No. 2
4

feed Pre-
processing
stock r6
5
4-1
Product
Processing 3-1 Assembly '-'-,~---" No. 1
3 I-I...c}-..--I 1 101
· ·····1:·······. ·······. ····. ·· ....· . · . · ....······F····
~~~~ 03 ~--~~

··~ ·· ··

preprocessing processing assembly

by products by products by products

Fig. 1 illustrative process network diagram for a

manufacturing enterprise. [ Closed boundaries
(small boxes) identi!y' manufacturing process.
Line segments identity inter process material
flows. ]
 35

etc., and referred to as boundary resources. The unit energetic cost associated with
technically-specific boundary resources represent the skill-specific human resources and
technically-specific synthetic energy required to deliver the material to the process or to
remove it, as the case may be.
The network diagram provides an unambiguous reference frame for all processes and
observables to be used in the technical analysis of the manufacturing enterprise. But, it is
not unique. It is always associated with a particular level of organization. More detailed
representation of the same enterprise can be obtained, for example, by partitioning one or
more of the production processes illustrated in Figure 1 into a more detailed network of
processes, each of which might be associated with a particular manufacturing operation such
as docking, milling, grinding, etc., in durable goods manufacturing or cracking, stripping,
scrubbing, etc. in chemical and biochemical production operations.
The level of organization considered in any technical analysis depends primarily on a) the
alternative technologies and management details to be considered, b) the level of organization
at which the processing technologies can be parameterized, and c) the physical features of the
production processes. The guiding principle is to start at a level of organization no more
detailed than that required to address the design and management issues of interest.
Aggregation from lower levels of organization to higher levels of organization is carried out
analytically using concepts and tools being described herein.

ll. mE PROCESSING TECHNOWGIES

For each production process identified in the network diagram, the technically-specific
resource materials are transformed into a technically-specific output product and by-products
including defects and "scrap" and processing effluents. The flow rates of the technically-
specific materials, into and out of the process, are typically measured in terms of number of
units per day. However, other units and time scales may be used, such as tons/day or
number of units/year.
For any given set of measures, material balance is maintained at the boundary of each
component process, if and only if, the flow rate of each resource and each by-product is
proportional to the flow rate of the product. The constants, k, of proportionality are called
the coefficients of conversion. They are dimensionless constants if the same units, such as
tons/year are used for both the resource and the product. If different units of measure and/or
36

time scales are used, such as number of units/day and number of units/year, then the
coefficients of material conversion are dimensional as required to convert from one unit of
measure and/or time scale to another. The only requirement is that the same time scale and
units must be used between certain production processes so that resource requirements of one
process can be equated to the product of another at nodes in the network.
Synthetic energy in one or more of its various technical forms --- kWh of heat, work, or
electricity --- is almost always involved in these production processes. Human time in its
various skill-specific forms --- technicians, chemists, engineers, craftsman, equipment
operators, etc. --- is also usually involved. General purpose machine time such as
programmable milling machines, gestation vessels, separators, and robots may also be
involved. Finally, a production facility time is always involved in the production process.
As a class, these resources are called energetic resources as distinct from material resources,
because synthetic energy is dissipated and human and machine time as a resource are
irreversible. The time rate of resource utilization, respectively, of synthetic energy, labor
time, machine time, and facilities time are typically measured in terms of kWh/day, person
hours/day, machine hours/day, and facility hours/day. Other units and time scales can be
used for each energetic resource, provided only that the same units and time scales are used
for all transformation processes, so that energetic resources of the processes can be summed
for the network as a whole.
For any given set of measures, the rate at which energetic resources are utilized in a
given network process is proportional to the flow rate of the product. The coefficients, f, of
proportionality are called the energetic cost coefficients. The units of measure are energetic
units per unit of material produced. More specifically, if the material output rate is measured
in tons/day, then for each of the respective energetic resources, f has dimensions of kWh/ton,
person hours!ton, machine hours/ton, and facility hours/ton. However, unlike the coefficients
of material conversion, the energetic cost coefficients, f, may depend on the rate of
production. The manner in which it varies with production rate is, in fact, an important
characteristic of a processing technology.
The coefficients of conversion, k, and the energetic cost coefficients, f, are recorded as
entries in a matrix of the form illustrated in Table 1. The columns of this matrix have a one-
to-one correspondence with the products associated with bold-faced line segments of the
network diagram. Boundary (external) products are always listed as columns on the far right
 37

Table 1. Processing Technologies.

Material Products

Internal External
(ij
·c
Q)II)
Material rates
oo~
::E~ 05 04 03 02 01
3-1 k31
4-1 k41
3-2 k32
4-2 ~2
5-3 k53
5-4 ks4

r1 kr1
r2
(ij r3
....
C
Q)
X
r4 kr4
UJ
r5 kr5
r6 kr6
II)
Q)
....
()
:::J
0
labor el f15 f14 f13
II)
Q)
a: elect. e2 '24 122 f21
25 23
()
robots ' '
.~ e3 f35 f32 f31
e>
C
Q)

UJ
woe.
Ime
e4 145 f44 143 f42 f41
li5 155 f54 153 f52 151
()

~
e>1I)
Coo
Q)Q)

UJ ....

[ The columns corresponding to process technologies. Coefficients

k and f represent parameters of processing technologies as
determined from on-line measurements of energetic and material
input and output rates. ]
38

of the materials table as illustrated.

The rows of the matrix in Table 1 are divided into two categories;
a) material resources -- rows having a one-to-one correspondence with the material
resources and by-products associated with normal line segments of the network
diagram; external material resources being listed always as rows on the bottom of the
materials table as illustrated.
b) energetic resources -- rows having a one-to-one correspondence with the energetic
resource; synthetic energy, labor time, machine time, and facilities time.
The numerical values of k, and f in each column of the matrix represent the technical
parameters of a specific processing technology in the network diagram, i.e. each column
provides a quantitative measure of a technology utilized in the network. The technology is
characterized in terms of both the absolute and the relative magnitude of synthetic energy,
labor time, machine time, and facilities time required to produce one unit of a technically-
specific product. For example, a fully automated technology requires no labor. If there are
no reprogram mabie machines involved, automated technology is said to be fixed or
"inflexible." And, if no synthetic energy is involved, then the technology is a manual one.
The time cost coefficient (hours/unit of output) identifies the time required by the respective
technologies to convert input resources to output products, i.e. the processing time. This is
an important characteristic of the technologies, especially in custom-order batch process
operations. Base-line values for the technical parameters of the processing technologies are
obtained from on-line operating data. For each process, the units of material resources
utilized, the by-products produced (including defects and effluents), and the energetic
resources utilized over a given time interval are simply divided by the number of units of
product produced over the same time interval, i.e. divide recorded values of the variables
corresponding to the rows by recorded values of the variables corresponding to the columns.
For each component process and each specific energetic resource, the energetic
cost coefficient, f, is typically related to the production rate by a curve of the general form
illustrated in Figure 2. The intercept 0 is said to represent the intrinsic costs of the process
technology. In standard line-item production operations, the intrinsic costs are, in a large
measure, attributable to "set-up" time, idle work force and/or mechanical equipment and
other non value-added time. In custom-order batch operations they may be defined more
broadly to include time devoted to the design of special components, engineering change
 39

''U

5Q.
c
.S!
Q)
e>
(J)
c
UJ

intrinsic
losses, 0
-r~----------~------------------~
o typical
operating - - -...~~:
range

E
(J)
'0
:e
(J)
o
() /= -:
e
iii Yo
o
()
()
.~ .. ...... .. ....... ... .~.... ... .. ............ ......... .. ' ,. ... ~~.
e> . .
(J)
c
:. :.
UJ : :

o
Processing Rate, Yo

Fig. 2 Typical variation of Energetic Input Rates, e

and Energetic Cost Coefficient, f, with Processing
Rate,Yo
40

orders and other support activities involved in what is sometimes called activity-based
accounting (ABC), Kaplan [1989]. In principle, the energetic cost coefficient, f, is obtained
operationally by dividing measured values of the energetic resource input rates by the
production rate for at least two levels of operation. There are a number of specific
techniques for estimating and recording the variations of the energetic cost coefficients, f,
with production rate.
In practice, it may be necessary or desirable to represent k and/or f as random variables
of time, to account for indeterminate factors associated with the production processes. For
example, the mean and variance of the conversion coefficient for defects or effluents can be
evaluated by standard statistical control techniques at the process level. In general, the mean
and variance of k and f represent important technical performance parameters of processing
technologies.
Finally, alternative or modified processing technologies are introduced into Table 1 by
simply replacing the base-line parameters in the corresponding column by those of the
proposed technology. Of necessity, the parameters of the new technology must be determined
from a synthesis of information available from the vendor, from prototype studies, or
operating data from another system in which the technology has been, or is, in operation.
In principle, the vendor should be expected to characterize the production technology in terms
of its technical parameters so that its performance can be evaluated in the context of the
manufacturing network in which it is to function .

ID. THE PROCFSS INTERCONNECTIONS

For analytical purposes, the linkages between resources, by-products, and products
internal to the boundary of the network are recorded as entries + 1 in a matrix of the form
illustrated in Table 2. The rows of this matrix have a one-to-one correspondence with the
nodes of the network diagram, which are also endpoints of the bold-faced line segments. The
columns have a one-to-one correspondence with the resource and byproduct flows internal to
the boundary of the network. The entries 1 and 1 in the row 04 indicates, for example, that
product flow rate 04 must be identically equal to resource flow rate 4-1 plus 4-2.
 41

Internal Resources
(ij
0;:: Material
Q)I/)
«i~
rates
I/) ::E~ 3-1 4-1 3-2 4-2 5-3 5-4
t5
:J
"0 05 0 0 0 0
e
Cl..
(ij 04 0 0 0 0
c
~

03 0 0 0
c 0
Q)

Table 2 Process Interconnections. Material intercoMection as determined from network

diagram in Figure 1. Rows identify process outputs. Columns identify links to other
processes in the network.

IV. THE NETWORK TECHNOLOGY

The technical parameters for the processing technologies given in Table 1, and their
material interconnections given in Table 2, are mathematically transformed into a table of
technical parameters which characterize the processing network as a production technology
at its boundary. Two steps are involved. First, the coefficients of Table 3 are computed,
then the coefficients in Table 4. The mathematics involved in making these transformations
is given explicitly as equations (19), (21), and (24) in Koenig and Tummala [1991]. The
derivation and use of these equations, however, need be of no direct operational concern to
most users of these analytical procedures. The requisite transformations can be carried out
on desk-top computers using soft-ware packages adapted from commercial software. The
transformation is based on first principles of physics and meets the standards of logical
consistency, value independence, and repeatability required of all accounting systems.
The coefficients computed for Table 3 represent the number of units of internal products
required to produce one unit of output. It is referred to as the production schedule for the
network because it indicates the levels of internal production required to balance the system,
i.e. avoid inventories and shortages. The internal production rates are obtained by simply
42

Material Products
External
(ij
"C
(I)(/)
Material
-(I)
<'l!_ rates
(/) ~~ 02 01
U
:l
"C 05 k32ksa)
e
a.. (ij
(~2ks4+ (k3tks3+ ~tks4)

-
(ij
C
~
(I)
04 ~2 ~t
"C
(I) oS 03
"E k32 k3t
::E

Table 3 Process Schedule. Coefficients are computed from Tables 1 and 2 as a linear
matrix transformation. Production rates 02 and 01 are specified. Internal production rates
05,04, and 03 are computed.

multiplying the computed coefficients in the table by the target production rates, 01 and 02.
The coefficients computed for Table 4 represent the units of external resources, by-
products (including defects) and energetic resources utilized by the enterprise in the
production of one unit of outputs 01 and 02. The conversion coefficients between output
products and input resources represent material conversion coefficients of the network as
computed from the corresponding coefficients of the processes technologies. Careful
inspection will show that the technical parameters of the network as given in Table 4 are
linearly related to those technical parameters of the processes which appear below the
horizontal line of partitioning between internal and externally derived resources in Table 1.
This is a fundamental property of processing networks. Thus, it can be concluded that t~

statistical quality control parameters for t~ network are computable from t~ statistical
quality control parameters of t~ processes. Likewise, the statistical parameters of the
processing time for the network can always be computed from the statistical parameters of
the processing times for the technologies of the component process. The ration of the
 43

1
Table 4. Network Technologies.

Material Products

External

Exchange
Q) prices
0'1
c: (ij
ClSIIl ' C:
Material
.cal allll
00
X'C:
mal rates
III LU ::2 m....
Q) 02 01
~
::J
0
III (ij r1 kr1
al c:
....
a:
(ij x
al r2 kr2
'C LU
Q) r3 k31 kr3
ro
~
k32kr3
r4 ~2kr4 ~1kr4
r5 (~2k54+ k32ks3) krs (k31 k53+ ~1k54) krs
r6 (~2k54 + k32ks3) kr6 (k31 k53+ ~1k54) kr6

labor el
.
f 12
.
f 11
elect. e2 (22 (21
III
OQ)
:;:;~ robots e3 (32 (31
Q)::J
e>o
Q)1Il
c:al proc, e4 (42 (41
LUa:
time
e5 f S2 f 51
0
III
'$
al e>1Il
0 alQ)
'C C: m
a.. LU ....

where
t'i2 = ( ~2k54+ k32k53) fis + ~2fi4 + k32fi3 + fi2 , i = 1,2,3,4,5
t'i1 = ( k31 k53+ k41 k54) fiS + k41fi4 + k31fi3 + fi1 , i =1,2,3,4,5
1. Computed from tables I and 3 by master program. External production rates are specified by user. Exter-
nal material and energetic resources computed from table. Exchange prices and capital cost allocation func-
tion G specified by user. Monetized performance, technical performance and environmental loading
computed from table.
44

processing time to the actual throughput time is a measure of the manufacturing cycle
effectiveness (MCE).
The columns of coefficients in Table 4 represents the parameters of the network as
processing technologies, i.e. they represent the technologies of the network as computed from
the technologies of the component processes. All the irifonnation regarding monetary
performance, technical performance, environrnentalloading, and the specific human resource
requirements of the network is contained in this table of parameters. This information can
be used in two ways;
a) The network can be considered as a component process in a higher level network of
production processes; in which case, the above analysis can be repeated in technology
assessments and evaluations at a higher level of organization.
b) It forms the basis for the technical design and operations control of the enterprise.

V. APPLICATIONS IN TECHNICAL DFSIGN AND

TECHNOWGY ASSFSSMENT
As a practical expedient to the use of Table 4, exchange prices associated with the
boundary products of the network can be entered, by the user, on the horizontal axis of the
table in the space provided. Likewise, where applicable, the exchange prices of material
resources, by-products (including effluents), and energetic resources can be entered in the
space provided on the vertical axis of the table. Cash flow for the enterprise is obtained
simply as the algebraic sum of these exchange prices times their respective material and
energetic resource rates as computed from the table; recognizing, of course, the algebraic
distinction between purchases and sales, and including any monetary costs imposed on
environmental discharges as control mechanisms.
A similar summation appears in the value added equation. But an allocation function is
also required in the value added equation to amortize the monetary cost of developing and
maintaining each of the technologies involved in the processing network. Although the
magnitude of the capital costs, broadly defined to include safety, oversight, legal, and other
"indirect" costs may be technology specific, the technology itself does not and cannot
determine how these costs are to be allocated over time to the intermediate products and/or
to the final products. The rules and regulations regarding capital cost allocation are
promulgated by the accounting profession and the IRS.
 45

For single output, line-item production operations, the capital costs can be allocated
directly to the final product. However, for custom-order batch operations and/or multiple
output operations, such as that illustrated in Figure 1, it is necessary to allocate capital costs
at the process level to avoid distortions. Information contained in Table 3 is used to
transform process level capital cost allocations to the network level. The capital costs of the
product(s) so established, are based on their resource content and the processing technologies
used to produce them. Technical distortions in conventional allocation procedures are
frequently major issues in financial accounting for advanced technologies and/or custom-order
operations, where capital costs are frequently large compared to the cost of materials and
labor, Kaplan [1989].
Among other things, Table 4 establishes norms for defect rates, yields, resource
efficiencies, processing time and other technical performance measures, as well as
environmental load factors, and human resource requirements of the network as a whole.
The performance norms for the network become a basis for zero-based budgeting, risk
assessment, contingency planning, regulatory compliance, and other elements of tactical and
strategic planning. They also provide the base-line information from which to identify
improvements in technologies at the process level for new and improved environmental,
technical, and monetized performance at the network level. Deviations from the network
norms serve as a basis for corrective action in operations control.
In all but the very simple manufacturing operation it is desirable, if not necessary, to
partition the overall manufacturing network into sub-networks according to the specific
operational responsibilities of the members of the manufacturing team. In this context,
technical design and operations control take the form of a network of hierarchial desk-top
information and decision centers that are informationally interconnected in a very specific way
as illustrated in Figure 3. Specifically, output of Table 4 as computed at any given level of
organization literally becomes the input for Table 1 for the next higher level of organization.
Thus, each level of management has the information and the analytical tools to make
technology assessments and technical design and management decisions pertaining to their
respective sub-networks; within the context of higher-level technical and financial constraints.
Tables of processing technologies, rather than tables of specific material flow rates, or
monetary costs, serve as the accounting system and language of communication between
levels of organization and between enterprises. These tables contain all the information in
46

Ecological and
Environmental Social & Cultural
Factors Factors

t t
Resource
Con straints
I
Economic
Q)-
_ I/)
r--- Objectives u;
.o8
<IS
c
0
~ Q)
0
0
Q)
1/)- .- c Q.
.~.s
E·- 'H 4 <lSI/)
.s:;Q)
00
C/)
Q)
"O~g x ·c 0
_c
«()«
-
Monetized wa.. <IS <IS
'---- Performance .2E
c ...
.s:;0
o't:
Q)Q)
1-0..

Technical Performance Norms

3 I/)
Q)
Q) ·0
I/)
c
Q) .Q
Technical ~.Q
0 0 <ISO
cC
~Q) Performance ... .s:;
"t:l= Q)O
<ISO
«I-
~Q)
>()

2
c
01/)
.- Q.
Processing 1/)0
Q)o
Network I-- 0-

'.
-~
<ISo
I/)
Q. .2 <IS
coO
0 1 .s:;'
.Q 0"0
Q)Q)
~
1-$
~ Processing
.0
-6 Technologies I--
Q)
$ Organizational level s

Special Science and Technology

Fig. 3 Analytical Tools In Technical Design & Management

 47

a standard format required to quantify all the measures of technical performance,

environmental load and human resource requirements at all levels of organization. Monetary
pricing takes place only at the boundary of the processing network, where exchange prices
with other enterprises in the economy, including employees and governing bodies, are
negotiated to establish monetary values of the resources, products, and by-products, including
effluents.
Exchange agreements between enterprises can, and frequently do, include more than
prices, especially in advanced technology industries. The analytical tools and accounting
systems described herein can, in principle, quantify selected elements of technical
performance as well as environmental loading as metrics for extending the control of
enterprise beyond its boundary to both its supplies and its markets. Thus, technical
performance and environmental load factors can be viewed as currencies of exchange along
with monetary prices.

BIBLIOGRAPHY AND RELATED PUBLICATIONS

H.T. Johnson and R.S. Kaplan, Relevance Lost: The Rise and Fall of Management
Accounting. Boston, Mass.: Harvard Business School Press, 1987

R.S. Kaplan, "Management accounting for advanced technological environments," Science,

vol. 245, 25 August, 1989, pp. 819-823

R.S. Kaplan, Advanced Management Accounting, Englewood Cliffs, N.J.: Prentice Hall, 1982

R.F. Veilleus and L.W. Petro, eds. Tool and Manufacturing Engineering Handbook,
Manufacturing Management. vol. 5, Dearborn, Michigan: Society of Manufacturing
Engineers, 1988.

A.K. Sethi and S.P. Sethi, "Flexibility in manufacturing: A survey," The International
Journal of Flexible Manufacturing Systems, vol. 2, 1990, pp. 289-328

D.W. Fogarty, J.H. Blackstone, Jr., T.R. Hoffmann, Production & Inventory Management,
Cincinnati, Ohio: South-Western Publishing Co., 1991

J. Browne, J. Harhen, J. Shivnan, Production Management Systems, A elM Perspective.

Addison-Wesley, 1989

Society of Environmental Toxicology and Chemistry, A Technical Framework for Life-Cycle

Assessments. Washington D. C.: Society of Environmental Toxicology and Chemistry and
SETAC Foundation for Environmental Education, Inc, 1991.
48

B.E. Koenig and R.L. Tummala, "Enterprise Models for the Design and Management of
Manufacturing Systems", loint US/Gennan Conference on New Directions for Operations
Research in Manufacturing, National Institute for Science and Technology I National Science
Foundation Conference, July 30 -31, 1991. in press

H.E. Koenig and R.L. Tummala, "Principles of ecosystem design and management," IEEE
Trans. Syst. Man Cybem., vol. 4, pp. 449-459, Sept. 1972.

R.L. Tummala and L.J. Connor, "Mass-energy based economic models," IEEE Trans.
Syst.Man Cybem., vol. SMC-3, No.6, pp. 548-555, Nov. 1973

B.E. Koenig, "Economic systems as cybernetically controlled networks of physical and

biological processes," Proceedings ofthe 1990 IEEE Conference on Syst. Man Cybernetics.,
Paper No. AOO70, Nov. 1990

E.C. Alocilja, "Process network theory," Proceedings ofthe 1990 IEEE Conference on Syst.
Man Cybernetics., Paper No. AOO71, Nov. 1990

R.L. Tummala and B.E. Koenig, "Classes of processes models and their characteristics useful
for the design of manufacturing systems," Proceedings ofthe 1991 IEEE Conference on Syst.
Man Cybernetics., Oct. 1991

B.E. Koenig, "Enterprise Models for the Design and Management ofManufacturing Systems"
Ph.D. Dissertation, Department of Electrical Engineering, Michigan State University, 1992.
 Is Lean Production Really Lean?
The Design of a Lean Production System

by

Joachim Reese
Chair of Production Management
Department of Economics
University of Luneburg
0-2120 Luneburg, Federal Republic of Germany

1. The Problem

It is often argued that the success of Japanese manufacturing is

due to the basic requirements of lean production with a minimum
of fixed costs as concerns the production inputs. E.g. there
should be a minimum of investment in machine capacity, that
means the minimum of unproductive set-up and idle times, no
buffers for materials or finished products and personnel that is
universally trained and able to do teamwork. Or, in other words,
the general idea of lean production is to have less of every-
thing compared with mass production (WOMACK, JONES and ROOS,
1990, p. 13). The concept has been developed in the TOYOTA motor
company, where it gained excellent results during the last deca-
des. Therefore, it is first of all discussed for application in
the automobile industry.

The general - and perhaps striking - point behind this concept

is to install flexibility without accompanying slack. Fixed cost
factors are substituted by variable cost factors. This might be
a dominance directory as compared with concepts of mass produc-
tion, if productivity is not negatively influenced. But though
there exist lean solutions for single work factors, e.g.
flexible manufacturing systems in the factory, KANBAN principles
50

for the supply chain or methods of simultaneous engineering,

there remain two main topics to be further analyzed; the coordi-
nation of all these elements to a lean company as well as the
fact that there will always exist an environment of the firm,
which causes uncertainties and therefore demands flexibility,
which cannot be shifted towards other manufacturing cells.

The purpose of this paper is therefore to model a simple lean

production system with special reference to the managerial va-
riables, by means of which the lean process has to be control-
led. It is further examined, in how far the attribute "lean" has
to be restricted to the classical work factors. Finally, the
evaluation of a lean production system seems possible at a given
flexible system behavior.

2. Elements of a Lean Production System

The main components of lean production are the factory, the

product development, the supply chain, and the customer. All of
them can be described by several features, which characterize
the organizational solution within the single modules, but not
between them.

The heart of the production system is the factory, which is

organized by means of a group technology with team leaders. Each
group has the responsibility to improve performances. Quality
controls have to be done for each part wi thin the production
stage. This is in agreement with the premise of zero buffers
between two production stages. The orders from other stages come
in by means of KANBANS or similar informations just in time. The
automation of the production unit is not essential for high
productivity, but it enlarges the flexibility as concerns the
variety of products to be produced (KRAFCIK, 1989, p. 50). There
may be other supporting characteristics in a factory, which
produces on the lean principle, as e.g. the appropriate election
of a team leader by a democratic voting scheme or by delegation.
Nevertheless, the group appears as a single work factor, which
 51

is rewarded as a whole for its performance.

The product development in lean production requires durable

teams and a strong leadership. Members of the development team
do not return to the functional department after having finished
a project, as they use to do in other production systems. Due to
this large independency of the team and the strong position of
its leader, the so-called rugby team approach allows an anti-
cipation of necessary product design activities as well as the
simultaneous engineering. Empirical results show furthermore,
that the team oriented acceleration of the development process
costs less than in the traditional relay race process (WOMACK,
JONES and ROOS, p. Ill).

The supply chain is organized in functional tiers. This permits

a maximal decentralization of decisions, so that the factory
cell has only long-term contracts with some few main suppliers.
The community of suppliers is not competitive with each other,
though everyone is organized as a profit center. The suppliers
are informationally joined together, so that there is no need
for considerable inventories. Most of the suppliers are also
involved in the individual development of parts, which fit with
respect to the factory's production process.

The customer is integrated in a lean production system, because

a lifelong relationship between dealer and buyer is intended. By
way of door-to-door sales or similar marketing strategies of
active purchasing the supply chain is extended at this end, too.
The extensive informations on buyers' behavior give therefore
appropriate hints for just-in-time-production .

The process of lean production is visualized in Figure 1. The

information flow by means of orders is separated from the mate-
rial flow. The design and the evaluation of a lean production
system in the following chapters concentrates on the supply
chain between supplier, producer, and customer. The eventual
instabilities of this chain arise from the fact, that there is
no slack in and between the elements. Some of the most important
risks are marked in Figure 1.
52

S
U
P
Design
order How
Product Deve lopment
Design
order
I ~
Know-
How
C
U
S
P T
L Material BAN Customer
Team A 0

...
Team B
---
order order
I Product M
Material supply E
E supply
i Product _
(seml-
R supply -- R
I finished)

I
I

- Just-In-time - Strike - Lack of - Strike - Uncertain

deliveries - Machine inventory - Machine customer
disturbances disturbances demand
- Defective parts - Defective parts

Figure 1: Lean production system

3. System Design

3.1. General Performance Measures

In general, the performance of a production system can be regi-

stered by means of its productivity, flexibility, and quality
(KRAFCIK, 1988, p. 45). These are the main indicators though
there are several others for special reasons . The measurement of
system productivity may be done by counting the number of pro-
ducts, which leave the system in a given period. Slight modifi-
cations, but yet yielding the same information, are the time
elapsed for one product unit or simply the flowtime of the pro-
duct. The productivity measures for all products can be aggrega-
ted to a value added objective function z. Most common is the
objective function of cash-flow, which comprises the whole acti-
vities of the firm within a period.

While productivity is a unidirectional measure, flexibility is

not as simple. Flexibility is costly, but not immediately pro-
ductive, because it keeps resources available for future condi-
tions, which are not met today (REESE, 1991, p. 367). As uncer-
tainty increases with the planning period, the corresponding
 53

flexibility will become more and more expensive. An adequate

degree of flexibility is therefore not necessary equal to the
maximum degree. The current flexibility of a production system
may be defined as the set of future conditions, where the requi-
rements are met. If d denotes the demand, x the production volu-
me, and s the vector of the system parameters, the system flexi-
bility may be characterized as

{d I z[x(s)] ~ z(d)}. (1 )

I.e. the flexibility is given by the set of demand quantities,

which can be accepted with the given assets. The parameter vec-
tor s is not restricted to the simple machine and personnel
capacity and the inventory, because this would mean, that a lean
production system would not be flexible at all. At any growth of
demand, the production system would break down. Because practi-
cal evidence shows, that there is a flexibility in the lean
production system, s has to be interpreted as the total assets
available.

The quality of a production system is given e.g. by the defec-

tive parts, assembly defects or the time spent for finding the
defect and improving the situation.

As a rule, the productivity performance is of primary interest

for production systems, whereas flexibility and quality are
regarded as subgoals. But the lean system does not differ expli-
citly between the importance of each goal. Therefore, maximum
flexibility is as well required as zero defects. The refusal of
orders is therefore not discussed during the following steps of
system design.

3.2. The Master Problem

The real design phase for a lean production system is initiated

on the central layer of the firm. The problem is to minimize
fixed cost subject to given tasks. These tasks may include e.g.
a given market share, a production program or some special fu-
54

ture developments. Given the notations

t 1 to T periods under observation
n 1 to N products
q 1 to Q current and fixed assets
x periodical production quantity (N x 1)
d (minimum) periodical demand, i.e. d=min (d1, ... ,dT),(N x 1)
t

A matrix of production coefficients (Q x N)

b primary periodical capacity (Q x 1)
c = unit cost of capacity (1 x Q)
p contribution margin (1 x N),
a simple design problem of one stage production with continuous
material flow and no buffer at the end of the period may be
formulated as

min c b (2)
b

subject to

A x S b (3 )

x=d (4 )

x ~ o. (5)

The objective function can be reformulated as

max (px - cb) . (6 )

b

Of course, this standard model may be specified according to

practical needs (cf. e.g. BITRAN and CHANG, 1987, and ZAPFEL,
1991).

If we have am-stage, t-period design problem, where KANBAN

control is intended, let denote
m 1 to M work stations
F (m) product family, which is produced in work station m
production quality of product n in period t (KANBANs)
 55

Ynt free KANBANs of product n in period t (production or-

ders for period t + 1)
KANBANs of n not produced due to lack of capacity
in period t
Vnt full containers of n in buffer at the end of period t
dnt, n E F (M) = demand quantity of a finished product from
family F (M) in period t
production time per container of n (included set up
and quality control)
a *nsCn) quantity of product n per unit of s (n) (production
coefficient)
regular production capacity of work station m in each
period
C 1
n inventory cost per container of product n
C 2
III cost of regular capacity per hour over all periods.

Now the extended design model is (cf. also BITRAN and CHANG,
1987),

(7 )

subject to

~ an x nt S b m , m=l, ... ,M , t=l, ... ,T (8 )

"btml

X nt = d nt , n E F (M) , t =1, . . . , T (9 )

X nt + Vn,t_l 2: Ynt , n E F (m) , m=l, ... ,M-1 , t=l, ... ,T (10)

X nt S Un,t_l + Yn,t-l , n=l, ... ,N , t=l, ... ,T (11)

u nt = Un,t_l + Yn,t-l - x nt , n=l, ... ,N , t=l, .... ,T (12)

Vnt = Vn,t_l + Ynt - x nt ' n=l, ... ,N, t=l, ... ,T (13)
56

Yat = L a-a.(a) X.(a)t'

.(a)
nE F(m) , m=1, •.. ,M-1 , t=1, ..• ,T (14)

YaO = 0 , n=1, ... ,N (15)

vaO = V ~ 0 , n=1, ..• ,N (16)

uao=u~0,n=1, ... ,N (17)

Xat , uat , Yat , Vat EN, n=1, •.. ,N , t=1, ••• ,T. (18)

The conditions (8) and (9) are the generalizations of (3) and
(4). (10) guarantees the availability of semi-finished products
for further production stages. With (11) production must not
exceed the number of KANBANs in the production cycle. (12) is
the equation for determining the KANBANs, which remain in the
system for the next period due to lack of current capacity.
Analogously, (13) and (14) determine the buffers at the end of
period t and the periodical demands, respectively. (15) to (17)
give the starting conditions, i.e. the orders, which initiate
the production. ( 18 ) corresponds to the non-nega ti vi ty con-
straints of the original problem. The objective function (7)
prescribes the minimization of the cost of current and permanent
assets. Figure 2 describes once more the KANBAN and product flow
within a stage.

container { v nt

•
buffer

production!
capacity \L..-___ xn.. ;,t_ _-.J
production unit

Figure 2: KANBAN and product flow

 57

The master problem on the long-term layer may be interpreted as

the upper stage of a hierarchical process, For current planning
activities, the general problem has to be split up into periods
and planning units, It seems noteworthy, that all capacities of
current and fixed assets remain unaltered, as long as there is
no revised version of the master problem,

3,3, Short-Term Planning Activities

When the general design task is done, there remains the current
problem of controlling the periodical production process, In
periods t, where

(19)

an irregular (e,g, overtime) capacity is required to meet pro-

duction, With
xt production quantity in period t
dt demand in period t
b* t use of irregular capacity in period t
c * irregular unit cost (c * > c)
B* t irregular capacity,
the short-term problem of production runs as follows:

min (cb + c'b't) (2a)

b't

respectively

max (px t - cb - c'b't) (6a)

b't

subject to

(20)

b' t ~ B't (21)

(22)
58

(23)

As it can be easily seen from this model, the optimality would

not be guaranteed as compared to an appropriate case of slack
production. An example is given in Table 1 and Table 2 for a
single fixed asset. It is assumed that the other assets are
irrelevant. As the capacity requirements alter over the three
periods regarded, the minimization of b causes relatively large
overtime capacities. The relevant costs of production sum up to
60 (Table 1). In the case of slack production with a slightly
enlarged regular capacity (b = 8) total cost is only 56, though
there is a capacity slack remarkable in period 3.

t aXe c c* b b* cb c*b* Slack

1 8 2 4 6 2 12 8 -
2 10 2 4 6 4 12 16 -
3 6 2 4 6 - 12 - -
60
Table 1: Cost of lean production (example)

t aXe c c* b b* cb c*b* Slack

1 8 2 4 8 - 16 - -
2 10 2 4 8 2 16 8 -
3 6 2 4 8 - 16 - 2
56
Table 2: Cost of slack production (example)

Before continuing with the lean design - which is obviously

necessary, because the shown sub-optimalities must not be accep-
ted - let us reflect on the general short-term model. Supposed
that there is an optimal use of capacity, the objective function
 59

(2a) may be reformulated as

min [cAd + c'A (d t - d)]. (2b)

A

Furthermore, if a multiple product work station is regarded, the

processing times per unit have to be supplemented by eventual
waiting times. This can be expressed by

min [CFd + c'F (d t - d)], (2c)

F

where F denotes the flowtime matrix. Thus, the minimization of

overtime capacity demands for the solution of scheduling pro-
blems. For further simplification (2c) is eventually interpreted
as the minimization of flowtime. This is only correct, if the
unit cost for products of all stages are equal. But for practi-
cal reasons the partition of the master design problem is often
done in that way. Nevertheless, Figure 3 illustrates that the
reduction of a cost problem towards a time problem has its spe-
cial interpretations for the different areas of lean production.

: master r1em :

I. supply II. design III. product ion I V. customer

minimize minimize minimize minimize

delivery time development time flowtime delivery time
s.t. material s.t. customer s. t. customer S.t. (production)
requests request s requests capacity

Figure 3: Partition of the master design problem

60

4. Organizational Impacts

4.1. The Incentive System

It is a matter of organizational concern to prevent lean produc-

tion from sub-optimality, as has been shown in Table 1. General-
ly, in lean production there is no resource waiting for proces-
sing in future periods. All resources must be procured or stimu-
lated just in time. But if there have been taken no precautions
for unexpected demand situations, the procurement might become
dear. The other phenomenon of lean production is decentraliza-
tion. The short-term problem is split up in several stage-wise
sub-problems, which have to be solved by the production teams
and the external suppliers, respectively. This is why there are
adequate incentives necessary to lead the teams to the right
efforts in the sense of the central objective.

Simple incentive schemes are time wage and piece wage . If the
unit wage is denoted by w for regular work and w· for overtime
work, the time wage function is given by
~l = w b + w' b'
= wAx + w· A (x t - x)
(24)
(w -w') Ax +w'Ax t
= (w - w·) b + w' Ad t'

This function is to be maximized by each team. It can be impli-

citly written as ~l (a,d,b), i.e. it depends on the production
coefficient, the demand, and the regular capacity. Time wage has
to fail especially in the case of lean production, because there
is first of all no interest to enlarge efforts, but on the oppo-
site to reduce efforts. Enlarged productivity would diminish the
time wage. The interest of each team is obviously in conflict
with the central objective.

The piece wage function is simply

~2 = wx t (25)
= wd t •
 61

~2 can be written as ~2 (d) and is also to be maximized. As the

marginal utility of this wage may become less than the marginal
utility of the effort, which is necessary for producing another
unit, the time wage is likewise not treated for stimulating
efforts in the desired manner, although it does not contradict
automatically to the central objective.

A third alternative is the participation of each production team

in the periodical revenue or cash-flow, denoted here as z. But
the profit share function ~3 (z) has similar consequences as the
piece wage function, because z is first of all a transformation
of d (and b).

There remains the general incentive scheme ~ (z,d,b,l/a) as the

only promising alternative for vertical integration of the teams
into the central objective function. This scheme comprises in-
centive parts, which refer to the firm's profit, the exogenous
parameters demand and capacity, and the individual efforts. It
is universally suited to support optimal decentralized decisions
(cf. SHAVELL, 1979, and JENNERGREN, 1980). The complex shape of
the incentive system causes considerable problems of measure-
ment, which has to be done by the central instance. The measure-
ment of efforts is, if at all, only possible by means of an
expensive process control. If the upper management resigns from
this task for practical reasons, the teams have to be stimulated
in advance by a fixed charge, which should be congruent to the
importancy of efforts in the incentive system. Because the re-
sult is nevertheless not guaranteed in this case, and the wage
is paid for only a minimum performance of the team, this solu-
tion of the incentive problem is only second best (HARRIS and
RAVIV, 1979, and JENNERGREN, 1980). Those incentives given in
advance mean slack, which is installed in the hope that any
unexpected and uncontrollable development might be solved in the
best manner.

If it is not already evident, that efforts playa dominant role

in lean production systems and thereby cause a maximum of slack,
this shall be illustrated by the following example. Let only two
demand situations and two efforts be relevant for discussion: d 1
62

= 100, d 2 = 200; a 1 = 1, a 2 = 1.5. That will say that the demand

is either 100 or 200 in the next period under regard, and the
production time per unit is either 1 hour or 1 1/2 hour. Fur-
thermore, let p 4, c = 1, and c· = 1.2. The general incentive
function is given as

~ (z, d, b , ~) = ~l (z) + ~2 (d) + ~3 (b, ~ )

a a
= 0,02 z + 0,01 d + ~ • i. (26)

In an extreme case of slack production, the regular capacity

might be chosen as b = 300. The cash-flow function is defined as
1 = z (d,b,bO) = pdt - cb - cObo.
z (d,b,_) (27)
a

The net profit

Zo =z - ~ (28)

serves as the definite objective function of the firm, whereas

(29)

may be interpreted as the utility function of the team manage-

ment. (29) means that the incentive is not always decisive for
the team's action, because a great effort (short processing
time) reduces the objective value on the opposite side. The
single contributions to the firm's and team's objective
functions are calculated in Table 3. The interpretation of Table
3 is as follows: With a large amount of regular capacity over-
time production is not necessary in no case. The periodical
profit varies only with the demand. But there is yet an influen-
ce of the different efforts on the reward. Nevertheless, a glan-
ce on the utility function ZM shows that a low effort is dominant
for the management strategy. This corresponds to the objective
function of the owner of the firm, who is not interested in any
acceleration of the production process because of sufficient
regular capacities available. Thereby the incentive scheme is
optimal, as it brings both partners into harmony. From the
 63

owner's point of view it has to be checked, whether the reward

of efforts (~3) can be eliminated from the incentive function.
The result would not alter, although the net profit would in-
crease.

Z (d l , b, l/a l ) = 4*100 - 1*300 - 1.2*0 = 100 = Zll

Z (d l , b, 1/a 2 ) = 4*100 - 1*300 - 1.2*0 = 100 = Zl2

Z (d 2 , b, 1/a l ) = 4*200 - 1*300 - 1.2*0 = 500 = Z2l

Z (d 2 , b, 1/a2 ) = 4*200 - 1*300 - 1. 2*0 = 500 = Z22

~ (Zll, d l , b, l/a l ) = 0.02*100 + 0.01*100 + 1*1 = 4

~ (Zl2, d l , b, 1/a 2 ) = 0.02*100 + 0.01*100 + 1*2/3 = 3 2/3
~ ( Z2l , d2 , b, l/a l ) = 0.02*500 + 0.01*200 + 1*1 = 13
~ (Z22, d 2 , b, 1/a 2 ) = 0.02*500 + 0.01*200 + 1*2/3 = 12 2/3
II ZOll = 96 ZOl2 = 96 1/3 Z02l = 487 Z022 = 487 1/3
ZM ll = 4 ZM l2 = 4.49 zil = 13 zi2 = 15.51
Table 3: Incentives in slack production (b = 300)

Z (d l , b, l/a l ) = 4*100 - 1*100 - 1. 2*0 = 300

Z (d l , b, 1/a 2 ) = 4*100 - 1*100 - 1. 2*50 = 240
Z (d 2 , b, l/a l ) = 4*200 - 1*100 - 1. 2*100 = 580
Z (d 2 , b, 1/a 2 ) = 4*200 - 1*100 - 1.2*200 = 460
~ (Zll ,b,d l , 1/a 1 ) = 0.02*300 + 0.01*100 + 3*1 = 10 (7 )
~ ( Z 12 , b, d 1, 1/ a 2 ) = 0.02*240 + 0.01*100 + 3*2/3 = 7.8 (5.8)

~ (Z2l = 0.02*580 + 0.01*200 + 3*1 = 16.6 (13.6)

,b,d 2 , l/a l )
~ (Z22 ,b,d 2 , = 0.02*460 + 0.01*200 + 3*2/3=13.2 (11. 2)
1/a2 )
zoll = 290 (293) ZOl2 = 232.2 (234.2)
I Z02l = 563.4 (566.4) Z022 = 446.8 (448.8)
ZM ll = 10 (7 ) ZM l2 = 9.55 (7.1)
zil = 16.6 (13.6) ZM22 = 16.17 (13.72)

Table 4: Incentives in lean production (b = 100)

Lean production reduces the fixed assets to a minimum, i.e. b =

100. From Table 4 we can see, that overtime capacity is now
necessary for all but one parameter combinations. Although that
capacity is more costly, the profit increases as compared
against slack production with one exception. The most improve-
64

ments can be stated for cases, when the efforts are on the upper
level. The same is true for the incentive function. The evident
conclusion is that high efforts perform better for all partners.
The incentive scheme is again pareto optimal. The importance of
the team effort on the reward system ~3 is clear, when this part
is neglected in the calculations. These modified results are
given in Table 4 in brackets. The effort of the team would alter
towards the low level, although the firm is still interested in
a high effort.

4.2. The Coordination Mechanism

Decentralization causes another problem of a lean production

system, namely that of horizontal integration. Because there is
no buffer between the different stages of the supply chain, the
work stations are wholly dependent on each other. This requires
a special cooperation with respect to the minimization of the
flowtime . While (2c) minimizes the flowtime indirectly by mini-
mizing the product stays in each work station, the objective
function for a given period t may be now reformulated directly
as

(30)

Here, ~ denotes an element of the (M x N) flowtime matrix F and

c_ symbolizes the unit cost of time with special reference not
only to the stage of production, but also to the product type.
Besides the usual capacity constraints (20) - (23) it is of
additional importance, that given delivery dates Tn are not
exceeded. This can be expressed by the conditions
H
E
m.l
t=xntST n , n=l, ... ,N. (31)

Supposed that a maximum degree of decentralization should be

maintained in each case, the flowtime problem is split into
 65

several sub-problems. At each stage m there exists a problem of

the following form:

(32)

subject to

tjjjn x nt S Tn - ~ tin X nt , n=l, ... , N (33)

~
isH

and the capacity constraints. Applying this kind of decentrali-

zation means to come into severe difficulties. The last stage,
which initiates the production of the preceding stages, can
solve its problem without spectactular time restrictions. As a
rule there are no emergencies, which exclude some favoured sche-
dules. But the more the planning process proceeds to the basic
processes of the material supply, the more relevant time re-
strictions arise. I.e. that the first stages of a lean produc-
tion process work under strong pressure. They often have to
speed up production in order to realize the given delivery dates
Tn' Consequently, the optimum solution is not necessarily
reached.

An illustrated example shall again demonstrate the negative

consequences for lean production. Three products have to be
produced on two stages consecutively. Processing times and unit
cost of the production process are given in Table 5. The due
date for each product is T = 50.

product production production unit cost unit cost

time stage 1 time stage 2 stage 1 stage 2
product 1 all = 10 alZ = 20 c ll = 50 c l2 = 150
product 2 aZI = 20 aZZ = 5 CZI = 100 czz = 110
product 3 a 31 = 10 an = 15 C31 = 80 c n = 120
Table 5 : Basic dates for scheduling

The planning process starts with stage 2. There are six diffe-
66

rent sequencing orders, which bear also different cost conse-

quences. The best sequencing order on stage 2 is ( 1 , 3, 2).
Therewith product 2 leaves the system after its mere processing
time a22 = 5. The flowtimes of the products 3 and 1 are longer,
t32 = 20 and t12 = 40, because they cannot leave the system just
after processing, but must wait for the delivery date. For the
first stage there remains no other feasible schedule than also
(1, 3, 2). Due to the assumed cost and time dates the total cost
of product flow sums up to

= 50*10+150*40+100*20+110*5+80*15+120*20 (34)
= 12,650.

The resulting schedule is illustrated in Figure 4.

(stage)
1 3 2
1

2 . 1 3 2

(time) 10
.
20 30 40 50

Figure 4: Schedule in a lean production system

The decentral solution, which has been calculated in a lean

production system, is not optimal. The centralized planning
brings out a better solution with C 12,400. The corresponding
sequencing order on both stages is (3, 1, 2). The schedule is
given in Figure 5.

As against slack production the sub-optimality of the lean solu-

tion depends mainly on the lack of buffer and machine capacity.
If there had been more buffer or capacity available, the stages
would have become more independent to those time restrictions.
 67

A possible step towards such an independency is of course the

realization of the KANSAN-principle. The buffer is never empty
and thereby guarantees an optimal production process within the
single stages, assumed that the demand will not exceed the stock
quantity .

(stage)

3 1 2
1

3 1 2
2

(time) 10 20 30 40 50

Figure 5: Optimal schedule with centralized planning efforts

While a lean production really means minimum fixed cost and

therefore also a zero buffer, the decentralized planning proce-
dure has to work in a different way than to split up the problem
as described above . The general idea is, that all stages must be
in the same way interested in an optimal solution. The conse-
quence is a proposal of any kind of profit sharing (cf. GROVES
and LOES, 1979), which seeks to stimulate a participation on the
firm's general objective. E.g . the idea of profit sharing can be
installed here by the following decentralized problem structure
for each stage m:

(35)

subject to

(36)

With the solution of this problem stage mcares for everything,

which happens on the preceding stages. There is no more danger,
68

that an early stage is not able to produce in time or only at

considerable cost. The general problem behind the idea of profit
sharing is, that the informational flow activities arise im-
mensely. There is need of an extensive information system, which
transports all kinds of production and cost dates between the
stages and which stimulates great planning efforts. Altogether
considerable managerial slack may be created, because the use of
the managerial resources only depends upon the demand situation.
It seems questionnable, whether a decentralized planning proce-
dure remains superior to a centralized solution.

The described coordination problem is not only restricted to the

supply chain within the firm. Horizontal coordination is also
necessary with regard to the supplier and customer in order to
find economic agents, who deliver or demand products on call
just in time. Practical examples illustrate that there is spent
a lot of information cost e. g. for marketing researches and
skeleton contracts, before the first economic transactions are
realized. This is typical for informational slack. On the other
hand, if a supplier is not well prepared for immediate delive-
ries, we have the situation that has been described above as a
non-optimal time pressure on the first stages in the supply
chain. This can be by far more costly than the informational
slack.

5. Conclusions

The fundamental informations, how a lean production system

works, come from observations of the Japanese industry, and
there is some empirical evidence that the system works excellent
at least in the automobile industry. Although there is a wide-
spread discussion of the concept, the scientific analyses are
somewhat behind. A main reason may be that the concept is beyond
the theory of planning. It can be stated - and this is confirmed
by the discussions in chapter 3 - that the planning activities
are shifted towards the organizational design. Operative plan-
ning is rather trivial in lean production systems. The matter of
 69

fact here is to design a system and not only to find an optimal

solution within a known system (ZELENY, 1986). Although opera-
tions research has always concentrated on planning procedures,
it is also able to support the design process (cf. e.g. SHI and
YU, 1991).

The essential issues, which have to be taken into account during

the design process, have been worked out as follows: In contrast
with buffered production the lean system copes with certainty
instead of uncertainty. The necessary information is not cost-
less. Because the value of information can be only calculated
ex-post (REESE, 1989), the installation of an information system
is an economic decision between the surplus of assets and the
surplus of information. It appears in many forms in the decen-
tralization and coordination process of a lean system.

KRAFCIK (1988) maintains that a lean production management bears

higher risks, but allows higher potentials of return on invest-
ment. This would confine the application of a lean system to
those firms, which take a special risk into account. But it
would not give an explanation for the asymmetric applications
allover the world. Instead it seems more convincing that the
trade-off between information and incentive cost on the one hand
and the fixed costs of assets on the other hand is decisive for
the choice of the production system.

References

BITRAN, G.R., and CHANG, L., A Mathematical Programming

Approach to a Deterministic KANBAN System, Management
Science, Vol. 33 (1987), 427-441
GROVES, Th., and LOEB, M., Incentives in a Divisionalized Firm,
Management Science, Vol. 25 (1979), 221-230
HARRIS, M., and RAVIV, A., Optimal Incentive Contracts with
Imperfect Information, Journal of Economic Theory, Vol. 20
(1979), 231-259
70

JENNERGREN, L.P., On the Design of Incentives in Business Firms

- A Survey of Some Research, Management Science, Vol. 26
(1980), 180-201

KRAFCIK, J.F., Triumph of the Lean Production System, Sloan

Management Review, Vol. 30, No.1 (Fall 1988), 41-52
REESE, J., Theorie der Organisationsbewertung, Oldenbourg, Mu-
nich and Vienna 1989
REESE, J., Unternehmensflexibilitat, in: KISTNER, K.-P., and
SCHMIDT, R. (Eds.), Unternehmensdynamik, Gabler, Wiesbaden
1991, 361-387
SRAVELL, S., Risk Sharing and Incentives in the principle and
Agent Relationship, The Bell Journal of Economics, Vol. 10
(1979), 55-73
SHI, Y., and YU, T. -L., An Introduction to Selecting Optimal
Linear Systems and Their Contingency Plans, in: FANDEL,
G., and GEHRING, H. (Eds.), Operations Research, Springer,
Berlin et al. 1991, 57-77
WOMACK, J.P., JONES, D.T., and ROOS, D., The Machine that
Changed the World, MacMillan Publishing, New York 1990
ZXPFEL, G., Comparison to Two Production Logistic Concepts, in:
FANDEL, G., and ZAPFEL, G. (Eds.), Modern Production Con-
cepts - Theory and Applications, Springer, Berlin et al.
1991, 18-45
ZELENY, M., Optimal System Design with Multiple Criteria: De
Novo Programming Approach, Engineering Costs and Produc-
tion Economics, Vol. 10 (1986), 89-94
 The Performance of Kanban Controlled Serial
Production Systems
John A. Buzacott* Suzanne M. Price*
York University University of Waterloo
J. George Shanthikumar
University of California, Berkeley

Abstract
Manufacturing cells arranged in series with work flow controlled using Kanbans
are considered. A number of approaches for approximating the performance of the
system are described. Numerical tests of the approximations indicate the conditions
under which they give excellent performance predictions, and also the type of im·
balances between the cells when their accuracy is not so good. The approximations
are relatively simple and provide insights into Kanban system behavior.

1 Introduction
Kanban systems for the control and coordination of production have attracted wide in-
terest in recent years (see [4] [6]). More recently there have been a number of papers
about modelling Kanban systems with a view to using the models to determine the ap-
propriate number of Kanbans and the impact of the number of Kanbans on performance
(see [3], [7], [2], [5]). This paper is also concerned with the analysis of Kanban controlled
production systems. Our goal is to describe a simple approach to approximating Kanban
system performance which at the same time helps develop insight about the strengths
and weaknesses of Kanban systems.
We consider a production system that consists of a number, m, of Kanban controlled
production units or cells arranged in series. Cell 1 is supplied by raw material and there
is always ample raw material available. The output product of cell i becomes the input
product to cell i + 1, i = 1, ... , m - 1. Customers require the output product of cell m.
The production activity in a cell could be fabrication, inspection and test, or transport. In
between cell i and cell i+ 1 there will be space to store the output of cell i, i = 1, ... ,m-l,
and after cell m there will be a final store containing finished product inventory waiting
'Supported by Natural Sciences and Engineering Research Council of Canada
72

customer demands. We will label store i the store after cell i, i = 1, ... , m, and store 0
will be the input raw material store for cell 1 (assumed to never be empty).
The performance aspect which we focus on in this paper is the service level provided
to customers, where by service level we mean either PND, the probability that a customer
demand is met with no delay, or E[DJ, the average delay in meeting a customer demand
(including those demands met immediately and which then have zero delay).
The management set parameters in a Kanban system are the number of Kanbans
associated with cell i, ki' i = 1, ... , m, and the container or lot size represented by a
Kanban. We will make the assumption that this lot size is identical at all cells, otherwise it
is possible that remnant inventory may accumulate in between cells. A further assumption
that will be made for analytical convenience is that the size of all demands is equal to
this lot size, so we will be able to take a unit of demand as being equal to one lot. This
assumption can be relaxed but at the price of analytical complexity that obscures insight.
Each cell will be assumed to have limited capacity so only one unit or lot can be
processed at a time. This means we can represent the production activity in each cell by
a single server. Again for analytical simplicity we will assume that the processing time
for a unit (lot) in cell i is exponential with parameter {ti, i = 1, ... , m. However, it is
straightforward to modify the analysis to consider multiple servers in parallel or series or
more general networks of servers.

2 Kanban Control
Assume that at time zero there is no unmet demand and no cell has units in process. Then
store i will contain ki units. With Kanban control the occurrence of a customer demand
at store m will trigger the movement of a cell m Kanban from its output store to its input
store, store m - 1, as soon as the customer's demand is met. On arrival at store m - 1
this Kanban will trigger the release of a lot to production in cell m which in turn triggers
the movement of a cell m - 1 Kanban back to store m - 2 and so on. Cell i Kanbans
circulate in a loop from store i to store i - I to production in cell i and eventually, when
processing is completed, back to store i. Note, however, that the cell i Kanbans wait at
store i until they are triggered by the arrival of a cell i + 1 Kanban, wait at store i - I
until a unit completes processing by cell i - I and is available for release to production
in cell i, wait for the limited processing resources in the cell and then processing itself
will take a finite time. Furthermore, note that the information about the occurrence of a
final demand does not get passed immediately back through the system to cell 1. If there
are no Kanbans (and associated inventory) in store i when a cell i + 1 Kanban arrives at
the store then there will be a delay in transmitting the information back to store i - I
until cell i completes processing a unit and delivers that unit and its associated Kanban
 73

to store i.
It can be seen that a model of Kanban control has to describe all these delays. To do
so we develop a set of equations that describes the system dynamics, i.e. the movements
of Kanbans within the system. For n = 1,2, ... , let Dn be the time of occurrence of the
nth demand ; A~) be the time at which the nth cell i Kanban is triggered at store i and is
transmitted to store i - 1; R~), i = 1, . . . ,m, be the time at which the nth cell i Kanban
(and associated parts or raw material) is released to production in cell i; R~m+l) be the
time at which the nth customer demand is met from cell m; G~i), i = 1, .. . , m , be the
time at which the nth cell i Kanban (and the associated unit completes processing at cell
i and is delivered to store i; S~i), i = 1, . . . ,m, be the time required for processing the unit
accompanied by the n t h cell i Kanban on the production facility of cell i. For generality
we sometimes view the customers as forming another cell, cell m + 1. So we can write
A~m+l) == Dn.
From these times there are a number of quantities that can be determined:
~n = R~m+l) - Dn = the delay in meeting the nth customer demand, n = 1,2, ... ,
~~) = A~) - R~) = the delay of the nth cell i Kanban at store i - 1, i = 1, ... , m,
n = 1,2, . . .,
F~i) = GAi ) - R~) = the flow time of the nth cell i Kanban through the processing facility
at cell i including queueing time, i = 1, . . . , m, n = 1, 2, ...
h~) = A~) - A~+I) = the delay in information flow about the nth demand experienced at
= 1, ... ,m, n = 1,2, ....
store i, i
D I y we some t'Imes wn't e Un(m+J) =
ror genera l't - Un, n - 1, 2, . . ..
A A -

Dynamics of Kanban system A Kanban and its associated raw material and parts
can only be released to a cell for processing when there are parts available in the store.
Hence
R~) = A~), i = 1, ... , m + 1; n = 1,2, ... ,ki ;
and
R~i) = max {A~) , G~i~1L }, i = 1, . . . , m + 1; n = ki + 1, ... , (1)

where G~~ko == 0 for all n. Similarly, for a cell i Kanban at store i to be triggered there
has to be both a cell i + 1 Kanban arrive at the store and a cell i Kanban available. That
IS,

A~) = max{A~+l), G~i~d, i = 1, ... ,min = 1,2, ... ; (2)

(i) _ _
where Gn - k• = 0 for n - 1,2, ... , ki . It follows from equations (1) and (2) that

R (i+I)
n
= A(i)
- n'
.
Z=
1, .. . ,m+ 1;n= 1"2 .... (3)
74

and
8(i) A(i) _ A(i+I)
n n n

R(i+I) _ A(i+I)
n n

~~+I), i = 1, ... ,m;n = 1,2, ... (4)

Since the processing in each cell has been assumed to be equivalent to a single server,
the dynamics of a cell will be given by

(5)

where C~i) == O.
Given the customer demand process Dn and the processing times S~i), i = 1, ... , m,
it is thus possible to determine the dynamics of Kanban movement from equations (1),
(2) and (5).

Recursion for delay Note that

~(i+I) R(i+1) _ A(i+I)
n n n

= max{O C(i) _ A(i+I)}

'n-ki n

= max{O,F~2ki + R~~ki - A~+1)}

max{O F(i) . + ~(i) .. + 8(i) . + A(i+I) _ A(i+I)}
'n-k. n-k. n-k. n-k. n

= max{O, F~2ki + ~~~ki + 8~i~ki - T(i+1)(n - ki , n)}. (6)

Note that ~~+I) > 0 implies that C~i~ki > A~+I). Next, let Kji) = L~=j ku and define
the following quantities

U,i)
T n-Kj = ~ F(u)
L...J
n-Ku' J =
. 1 .. 1
, ... ,z;z = , ... ,m
u=j

and
i
L~,i) = L 8~u+I), j = 1, ... , i; i = 1, ... , m
u=j

T~~1j is the total flow time of a job through cells j, j + 1, ... , i given it never has to wait
at any store while Ln(j, i) is the total time information about the nth demand at store i
takes to reach cell j.
Now substituting for ~~~ki into (6) one can show

~~+I) = ~f~i{T~~1j + L~~kJ - T(i+I)(n - Kj, n)}+, 'i = 1, ... , m (7)

Property of Delay It is clear from equation (7) that the delay is decreasing as the kj
increase for j = 1, ... , i.
 75

2.1 Analysis of a Single Stage Kanban System

Consider a single stage Kanban system. We have

If demands are Poisson T(n -k1 , k) will have an Erlang- kl distribution, however, we need
to find the distribution of the other quantities in the equation. In the case of a single
stage system as demands arrive they will enter the cell (i.e. they trigger Kanbans to enter
the cell) until the cell contains kl Kanbans. After this demands wait at store 1 until a free
Kanban becomes available. It follows that the queue of Kanbans in the cell plus demands
waiting at the store is equivalent to a single server queueing system, so if demands are
Poisson and the cell is equivalent to a single server with exponential processing times then
the combined queue behaves as if it is a M / M /1 queue. That is, in the cell itself we would
have that the probability there are n jobs is given by

p(n) pn(l_p) n=O,I, ... ,k-l

/ n=k (8)

while the probability of n jobs waiting for Kanbans at the store will be

q(n) pn+k(1_ p) n = 1, . ..
1- /+1 n =0 (9)

Note that the average number of jobs in process in the cell will be N = p(1 -l)/(I- p)
while the average number of jobs waiting at the store will be F,\E[Dj = l+1 /(1 - p).
Also PND = 1 - l. The distribution of the delay time of information about demands
coming to cell 1 can be shown to be

h(t) (1 - p)/J1-e-I'(1-p)t, t >0

P{8 = O} 1-/.
2.2 Analysis of Two Stage Systems
Consider now a two stage Kanban system. In equation (6) set F!~k; + ~~~k; = J!~k;
where J~i) is the time between the nth stage i Kanban leaving store i and its next return
to the store. Then

~(3)
n
= max{J(2)
n-k2
+ ~(3)
n-k2
_ T(3)(n _ k2 ' n) , O} (10)

Equation (10) suggests that we view the processing system as equivalent to a service
facility in which no more than k2 customers can be in service (and with the service time
76

of the nth customer JA ») and an external queue in which the nth customer waits for a
2

time Do~3) . If we knew KA2) = JA2) + Do~3) then we could use the basic result for delays in
a single stage system with an output store of size z, Don = max{Kn _ z - T(n - z , n), O}, to
find the service level parameters of interest.

Equivalent closed queue if all Kanbans active. As long as Do~3) > 0 then there
are no Kanbans waiting at store 2. This suggests that the processing system can be
considered as a closed queueing system consisting of the two production facilities 1 and 2
(where service is provided at cells 1 and 2), a split point A where a customer splits into
two customers (corresponding to the separation of the stage 2 Kanban from the order
tag conveying information to stage 1) and the merge point B where two customers are
combined into a single customer (corresponding to the combination of the stage 2 Kanban
and a part from stage 1). kl Kanbans circulate in stage 1 and k2 Kanbans circulate in
stage 2. This means that there are four queues in the system, one at each production
facility and the two queues waiting at the merge point B. Denote by NI(t) and N 2(t) the
queue lengths at the production facilities 1 and 2 (including any jobs in service), N[(t)
the number of stage 2 Kanbans waiting at point Band N["(t) the number of parts from
cell 1 waiting at B. Note that

NI(t) + N["(t) = kl
N2(t) + NUt) = k2
N[(t).N;"(t) = 0
To analyse the closed split and match queue further we will assume that each facility 1
and 2 is equivalent to a single server with service times having an exponential distribution
with parameters ttl and tt2 respectively. Define p(nl,n2,ni,ni") = limt ..... ooP{NI(t)
nl, N2(t) = n2, N[(t) = n l , N["(t) = ni"lnl + ni" = kl , n2 + nl = kd· Let Sk"k, =
{(nl,n2,n l ,ni") : 0:::; nl :::; kl,O:::; n2 :::; k2' O:::; n l
+ n l = k2}. Then as has been shown in [1]
kl , n2
1 ( 1 ) n, +nl" ( 1 ) n,+n;"
p(nl, n2, n~, n~) = G(k l , k2 ) ttl tt2
where
G(kl,k2) = L
k, k,
L
(1) n,
-
+k,-n, (
-
1) n,+k,-n,

n,=On2=O ttl tt2

That is, the closed split and match queue has the same queue length distribution as a
simple cyclic queue consisting of two servers 1 and 2 and with queue length nl + nl at
server 1 and queue length n2 + ni" at server 2. The throughput of this closed queue is
k, k,
TH(kl ,k2) = tt2 L LP(nl,n2,kl - nb k2- n2)
n2=l n,=O
 77

G(kb k2 - 1)
(12)
G( k), k2 ) •

Furthermore, from the closed queue one can obtain the mean time E[JD] for a stage 2
Kanban to go from store 2 to store 1 to store 2 and also the distribution. Since there are
k2 Kanbans circulating in the loop the E[JD] will be given by
k2 k2 G( kI, k2 )
E[JD] == TH(k b k) (13)
2 G(kb k2 - 1)

Stage 2 equivalent closed queue if all Kanbans active. The simple equivalent
closed queue above represents both stages 1 and 2. However, there is an alternative
closed queue which represents stage 2 only and which models the circulation of stage
2 Kanbans . This closed queue has a server corresponding to cell 2 and another server,
denoted by lu, which represents the service of stage 2 Kanbans arriving at store 1, i.e.,
the merging of the Kanban with a part from stage 1. Now, if the Kanban arrives and finds
N{"(t-) > 0, (and hence Nr(t-) == 0) then the service time of this customer (Kanban)
will be zero while if a stage 2 Kanban arrives at store 1 at time t and N{"(t-) == 0 then
the service time of the customer (Kanban) at the head of queue lu will be determined by
the service process at stage 1 and will thus be exponential with parameter Ill. Now, if
N{"(t-) == 0 then it is possible for Nt(t-) to be zero while if Nr(t-) > 0 then N{"(t-)
must be zero. Define qlu(t) == P{N{"(t-) == OIN;(t-) == O} and qlu == limt-+ooqlu(t).
From equation (11) it is possible to obtain the queue length distributions at the instant
of an arrival at the split point A as the distribution p(nl' n2, n~, n;") in a system with the
number of customers circulating in loop 2 given by k2 - 1. Hence

L~\n=o(1/ Ild 1 - nj (1/ 1l2)k2 -1+ n j

(Ild Ilill
(14)
L:~o(1l2/ Ill)"
Let Sk 2 == {(n2,nn,0 :::; n2 :::; k2,0 :::; n~ :::; k2,n2 + n~ == k2}. It follows that the queue
length distribution in the closed queueing system corresponding to the circulation of stage
2 Kanbans will be given by

(15)

where
nI == 0
nI == 1,2, ... ,k2
and
78

Also, the throughput of the system will be given by

TH(k ) = G'(k2 -1) (16)

2 G'(k2 )
and an alternative expression for E[JDJ is

(17)

Equivalent closed queue if not all Kanbans active. If not all Kanbans are active,
i.e., D.~3) = 0, then again the system can be represented by a closed split and match
queue. The system consists of the two production facilities 1 and 2, the split point A
and the merge point B as above, but in addition there is a further server in the stage
2 loop where circulating customers (stage 2 Kanbans) receive service at a single server,
denoted by 2d, at rate A. Unfortunately, the addition of the extra server means that
this system no longer possesses a product form solution. However, suppose we consider
the effect of adding the additional server to the stage 2 equivalent closed queue. Let
n~ = limt-+oo Nf(t) where Nf(t) is the number of customers (stage 2 Kanbans) at server
2d and let Sk 2 = {(n2,ni:,n~): 0:::; n2:::; k2'0:::; ni: :::; k2'0:::; n~:::; k2,n2 + ni: + n~ = k2}.
Then it can be shown that the cyclic system possesses a product form solution.

(18)

where
ni: = 0
ni: = 1,2, ... , k2
and
2
"
G (k2) = Lk2 kL-n, (-1 )n, (-1 )k2-n,-n~ (l)n~
~ h(n
r
l )
n,=O n~=O III 112
From equation (18) it is possible to determine the throughput of the stage 2 equivalent
system, T H N D( kb k2 ) from
G"(k2 -1)
THND (k l ,k2 ) = G"(k2 ) (19)

The expected time for a Kanban to circulate around the system, E[TND(kb k2 )J = kdT H(k l , k2 )
However, to find E[JNDJ, the time from when a Kanban departs store 2 until its return,
it is necessary to subtract the expected time a Kanban is at store 2d, E[T2d J, i.e.,

(20)

E[T2dJ will be given by

1 k 2 -1
E[T2dJ = ~(1 + L rp-(n~ = r))
r=l
 79

where p- (n~ = r) is the probability an arriving customer at server 2d finds n~ = r. This

can be found from the marginal distribution of n~ in a system where k2 - 1 customers
circulate in stage 2.

Determination of qlu. While equation (18) indicates a product form solution, to

determine the probabilities we require the quantity qlu. This can be approximated by
observing that qlu = limt .... ooP{NI(t) = kIIN{(t) = O} and then assuming that the queue
length NI(t) behaves like an M/M/l/k l queue. It is clear that service in this queue is
exponential with parameter /11 while arrivals will be equal to AId, the arrival rate of Kan-
bans from server 2d. Equivalently, one can determine qlu from qlu = limt.... oo P{N{"(t) =
OIN{(t) = O} and assume that N{"(t) is determined by a birth and death process on
the integers 0,1 , . . . ,kl in which the birth rate is /11 if 0 :::; N{"(t) < kl and zero when
N{"(t) = kl and the death rate is AId if 0 < N{"(t) :::; kl and zero when N{"(t) = O. One
approach to finding AId would be to determine the departure rate from server 2d in the
stage 2 equivalent closed queue when the 1u queue is empty and so there are a total of k2
customers at either server 2 or server 2d. That is, we set

and

where Pld = Ald/ /11·

However, this approach to finding AId ignores the effect of arrivals at server lu from
server 2d when there is a queue of jobs outside the system and so service at 2d takes zero
time. Thus we need to consider a more complete closed queue equivalent to the movement
of stage 2 Kanbans for both ~~3) > 0 and ~~3) = o.

Equivalent closed queue for stage 2 Kanbans. In this closed queue there are, as
above, three servers: a server 2 corresponding to processing at cell 2, server 1u corre-
sponding to the merging of a stage 2 Kanban with a part from stage 1, and server 2d
corresponding to the triggering of a stage 2 Kanban by an external demand. As above,
service at stage 2 is exponential with rate /12 and service at luis, when a Kanban ar-
rives at lu at time t, exponential with rate /11 if N{"(t-) = 0 and takes zero time if
N{"(t-) > o. But now service at 2d for a Kanban arriving at 2d at time t is exponen-
tial with parameter A if N2(t-) = 0, where N2(t-) is the number of requisition tags
from external customers waiting at store 2, and takes zero time if N2(t-) > o. Define
80

q2d(t) = P{N;(t-) = O/Nnt-) == O} and qu = liml ..... ex> qu(t). Then this equivalent closed
queue has a product form solution

p(n2,n~,nD = G'~k2) (:J n2 (:J nr (~r~ h(nDh(nD, (21)

where
nj" =0
nj" = 1, 2, . . . ,k2 ,
n~ = 0
n~ = 1, 2, . . . , k2

E
and
G'(k2) = kE'r (2.)k2-nr-n~ (2.)nr (~)n~ fl(n~)h(n~)
nr=o n~=O JL2 JLI
In this closed queue we require Aid in order to find qlu. We have

AId = Ap( n~ > O/n~ = 0) + JL2(1 - q2d)p( n~ = O/n~ = 0)

G d(k2 - 1)
(22)
Gd(k 2 )
where

and
(23)

with Pld = Aldl JLI'

Determination of qu. However, to evaluate Gd(k 2) we now require q2d. Let N2'(t) be
the number of completed parts from stage 2 at store 2. Observe that qu = liml-+ex> P{N;(t-)n
N2'(t-) = O/N;(t-) > 0 U {N;(t-) n N2'(t-) = O}}. Then qu can be found from the
probability of an idle server in a queue where it is assumed that the queue length process
N;(t) behaves like an MIMll queue with arrival rate A and service rate JL2u, where JL2u
is the rate at which Kanbans arrive back at store 2 when Nnt-) = O. It follows that

q2d =1- P2u (24)

where P2u = AI JL2u and

(25)

where
G:(k2) = t (2.)k2-nr (2.)nl
nr=o JL2 JLI
fl(nD .

Equations (22) , (23), (24) and (25) can then be solved to determine the qlu and qu.
 81

PI
0.1 0.3 0.5 0.7 0.9
0.1 Approx 0.990 0.982 0.937 0.796 0.447
Sim 0.990 0.981 0.927 0.750 0.350
0.3 Approx 0.910 0.898 0.846 0.707 0.378
Sim 0.911 0.902 0.847 0.677 0.301
0.5 Approx 0.750 0.737 0.678 0.530 0.201
Sim 0.752 0.744 0.692 0.532 0.182
0.7 Approx 0.510 0.496 0.435 0.282 ***
Sim 0.515 0.507 0.454 0.300 ***
0.9 Approx 0.190 0.177 0.117
,.,.,. ,.,.,.
Sim 0.196 0.187 0.131 *** ***
Table 1: Comparison of approximation and simulation for PND for a Kanban system with
kl = kz = 2. (***: unstable)

Performance estimation The probability of no delay PND will be given by

= lim P{N;:(t-) > O}

t-+oo

p(n;' > 0)
1 - p(n~ = 0)
Using the stage 2 equivalent queue one obtains
k2
PND = 1- L p(kz - n;,n;,O) (26)

(27)

The expected delay E[D] can be found using Little's law from
00

AE[D] =L n~p(n;) (28)

n;=l
and since p(n;) = pzup(n;-l), n; = 2, . . ., p(n; = 1) = pzup(n;nn2' = 0) and p(n;nn2' =
0) + E::;=I p(n;) = p(n~ = 0)
'E[D] PZu (r n m 0)
A = (1 _ pzu)ZP n z n z =
~(l-PND) (29)
1 - PZu

Tables 1 and 2 compare the performance measures calculated using this approximation
for a two stage system with kl = k z = 2 for a variety of different values of PI and pz.
Note that the agreement is in general very good except when PI is large and pz is small.
82

PI
0.1 0.3 0.5 0.7 0.9
0.1 Approx 0.001 0.003 0.029 0.242 2.20
Sim 0.001 0.006 0.067 0.549 5.418
0.3 Approx 0.039 0.047 0.102 0.376 2.52
Sim 0.039 0.046 0.127 0.689 6.064
P2 0.5 Approx 0.251 0.276 0.419 1.02 6.73
Sim 0.251 0.266 0.402 1.23 10.5
0.7 Approx 1.14 1.22 1.68 3.90 ***
Sim 1.14 1.19 1.55 3.70 ***
0.9 Approx 7.30 7.97 13.49 ""PI'
***
Sim 6.94 7.45 11.4 *** ***
Table 2: Comparison of approximation and simulation for E[D] for a Kanban system with
kI = k2 = 2. (***: unstable)

Alternative approximation. Based on equation (10) it is possible to develop an alter-

native approximation which attempts to improve on the above approximation. Equation
(10) suggests that we could find PND from

PND = PND (1- Ji.~ p{J~2) - T(3)(n, n + k2) ~ 01~~3) = O})

+ (1 - PND )(l - Ji.~ p{J~2) + ~~3) - T(3)(n, n + k2) ~ Ol~~k) > O} (30)

We then assume that JND(t) = limt ..... ooP{J~2) > tl~~3) = O} is given by

where E[JNDJ is given by equation (20). Let Hk2 (t) = limn ..... ooP{T(n,n + k2) :::; t}.
Because of our assumption of a Poisson arrival process with parameter .x for demands
dHk2 (t) = .xk2tk2-Ie-Atdtl(k2 -I)!. Thus

Next, consider the case where ~~3) > O. Then, we require an initial approximation for
E[WJ = limn .....oo E[~~3)1~~3) > 0], the average time that customer requisitions spend
waiting in the queue at store 2 until they can trigger a stage 2 Kanban. Assuming that
job arrivals are Poisson with rate .x and service is exponential at rate TH(kt,k2) when
~~3) > 0, using standard results for a MIMl1 queue the average waiting time will be
given by
(31)
 83

and

Fw(t) Ji..~ p{~~3) > tl~~3) > O}

e-tjE[W]. (32)

Let K = JD+W. Then E[K] = E[JD+W] = kdTH(kI,k2)+1/(TH(kl,k2)->'). Again,

by analogy with the M / M /1 queue we could assume that K is exponentially distributed

1:
with mean E[JD + W]. Thus

li..~ p{J~3) + ~~3) - r(3)(n, n + k2) ~ 01~~3) > O} = K(t)dHk2(t)

>.E[K] ) k2
(
1 + >.E[K]
Then from equation (30)
1 - (>.E[K]/(1 + >'E[K])k2
PND = 1 + (>'E[JND ]/(1 + >'E[JND j)k2 _ (>.E[K]/(1 + >.E[Kj)k2 (33)
The distribution of the delay can also be found using equation (10)

Ft.(t) = PND 1:0 JND(t + u)dHk2 (U) + (1 - PND ) 1:0 K(t + u)dHk2 (U) (34)
The integrals in the above equation can be readily evaluated.

~oo=o JND(t + u)dHk2(U) (>'E[JND] )k2 e-tjE[JND]

i" 1 + >'E[JND ]

~oo=oK(t+u)dHk2(U) (>'E[K] )k2 e-tjE[K]

i" 1 + >.E[K]
Hence from (34 It follows that the expected delay in meeting a demand can be approxi-
mated by

E[~]

(35)

Table 3 shows the PND calculated using this approach (Approx2) and compares it with
simulation and the earlier approximation (Approxl). Note that the approximation is
generally worse than the first approximation. Note that for simplicity the approximation
assumed that JND and K have exponential distributions, but as shown by equation (45) J D
does not have an exponential distribution. Assuming W has an exponential distribution,
for kl = k2 = 2 it is possible to determine the distribution of K = W + JD and also
the distribution of JND . The resulting approximation for PND is also shown in table 3 as
Approx3. The resulting approximation tends to underestimate PND .
84

0.1 0.3 0.5 0.7 0.9

0.1 Approxl 0.990 0.982 0.937 0.796 0.447
Approx2 0.989 0.980 0.936 0.798 0.449
Approx3 0.989 0.977 0.914 0.718 0.286
Sim 0.990 0.981 0.927 0.750 0.350
0.3 Approxl 0.910 0.898 0.846 0.707 0.378
Approx2 0.910 0.893 0.830 0.687 0.366
Approx3 0.909 0.888 0.804 0.600 0.217
Sim 0.911 0.902 0.847 0.677 0.301
P2 0.5 Approxl 0.750 0.737 0.678 0.530 0.201
Approx2 0.776 0.760 0.699 0.564 0.240
Approx3 0.749 0.726 0.635 0.439 0.117
Sim 0.752 0.744 0.692 0.532 0.182
0.7 Approxl 0.510 0.496 0.435 0.282 *,.*
Approx2 0.604 0.591 0.533 0.385 ***
Approx3 0.509 0.488 0.405 0.231 ***
Sim 0.515 0.507 0.454 0.300 ***
0.9 Approxl 0.190 0.177 0.117 *,.* *,.*
Approx2 0.333 0.316 0.228 *** ***
Approx3 0.190 0.175 0.109 *** ***
Sim 0.196 0.187 0.131 *** ***
Table 3: Comparison of approximations for PND when kl = k2 = 2. (***: unstable)

3 Multiple Stage Systems

The two stage approximation can be extended to multiple stage series systems controlled
by Kanbans. Again the circulation of stage i, i = 1, ... , m, Kanbans from store i to store
i - I to cell i and back to store i is modelled by a closed cyclic queue with three servers,
id, i-I u and i. The service rate of Kanbans at server id is /Lid if the Kanban has to queue
at the server, however, there is a probability 1 - qid that if the Kanban arrives at the
server and finds no Kanbans waiting then it will require no service. Similarly, the service
rate of Kanbans at server i - lu is /Li-lu unless the Kanban arrives at an idle server and
then with probability 1 - qi-lu it will not require service. The system then has a product
form solution with

(36)

where
ni_l = 0
ni_l = 1,2, ... ,ki ,
 85

nf = 0
h(nf) ={ ~id nf = 1,2 , .. . ,ki
and

To determine the /1i,/S , the /1id's and the q's it is necessary to determine the rate at which
products are delivered to store i from cell i and the rate at which Kanbans are delivered
to store i from store i + 1. Define Aid( ni) as the rate at which Kanbans are delivered
to store i from store i + 1 when there are ni stage i + 1 Kanbans at store i. Obviously
Aid(ki+1 ) = O. Then

\.(n T )=G:i(ki +1 -n i -1)

A,d I G*(k
d i+1 - ni
T) ni
T
= 0, ... , ki+l- 1 (37)

where

Similarly, define Ai" (nf) as the rate at which products are delivered from cell i to store
i + 1 when there are nf stage i Kanbans at store i. Note Aiu(ki ) = O. Then

A ( P) = G: (ki -
IU n, G~ (k i
nfnf)
_
- 1) ni
P
= 0, ... , ki - 1 (38)

where

It follows that

/11" /11
/1iu .\;,,(0) i = 2, . . . ,m
/1id Aid(O) i=1, ... ,m-1
/1md A (39)

Since qi" = limt~oo p{Nr(t) = OIN[(t) = O} it can be found by analyzing the birth-
and-death process representing the change in the inventory level in store i. It follows
that
1
qi" = Ow A ()'
1 + "ki- 1 y=O iu v
i = 1, . .. , m (40)
L.."w=O Aid(O)W+l

where Al,,( v) = /11 for v = 0, .... kl - 1.

86

Similarly qid = limt_ooP{N[(t-) = 0IN!(t-) = O} and it can be found by analyzing

the birth- and-death process representing the way in which stage i+ 1 Kanbans accumulate
at store i. It follows that

(41)

Since Aid(v) = A for all v it follows that

(42)

The equations defining the q's, the Ai.. 'S, the Ai/S and hence the iLid'S and the iLiu'S
parameters can be solved using an iterative scheme.
The performance measures can then be found. Define Pm .. = AI Am .. (O). It can be
shown that
(43)

and
AE[D] = Pm.. (1 - PND) (44)
1- Pm ..
Preliminary computational results on three and four stage systems which are not
extremely imbalanced suggests that the approximation gives very good results for PND
(within .02) but the E[DJ approximation can be 10-15% in error. This is probably due
to the effective service times of Kanbans not being exponential.

A Distribution of Kanban active time

Consider the three server cyclic queue equivalent to the movement of stage 2 Kanbans
from store 2 to store 1 to cell 2 and back to store 2. We wish to find the distribution of
the time between departure of a Kanban from store 2 until its return to store 2. Suppose
a specific Kanban is marked as it leaves store 2 and the time of its departure set as time
zero. Assume that the Kanban leaves no Kanbans waiting at store 2. The time t at which
the marked Kanban returns to store 2 is observed. We wish to determine FJD(t), the
distribution of t.
Let Pn~,n2(t) = probability that, including itself, there are n2 Kanbans at cell 2 and
n~ Kanbans at store 1 in front of the marked Kanban at time t.

dPnr,k2-nr(t)
dt = -(iLl + iL2)Pn].k2-nr(t)
+ iLIPnr+1 ,k2-nr-l(t), n; = 1, .. . , k2 - 1,
 87

dPO,k2 (t)
dt
dPnr ,n2 (t)
dt
-(ttl + tt2)Pnr ,n2(t) + ttI Pnr+l,n2-I(t)
+ tt2Pn"n2+l(t), 1::; ni, 1 ::; n2 < k2 - ni
dPnr,o(t)
dt
dPO,n2(t)
dt
dFJD(t)
dt
The initial conditions are determined by the distribution of (n;:, n2) seen by the marked
Kanban on leaving store 2, i.e.

Let P:r,n2 be the Laplace transform of Pn"n2(t). Taking the Laplace transform of both
sides of the above equations and solving the equations gives

If ttl = tt2 = tt then

88

which means that

If a Kanban leaves n~ Kanbans behind at store 2 then the above distributions would
be modified with k2 replaced everywhere by k2 - n~.
In the case where all Kanbans are always active and qu is given by equation (14) then
the distribution of JD becomes

/-LI f- /-L2 ( 45)

References
[1] J. A. Buzacott. Queueing models of Kanban and MRP controlled production systems.
Engineering Costs and Production Economics, 17:3-20, 1989.

[2] J. L. Deleersnyder, T . J. Hodgson, H. Muller, and P. J. O'Grady. Kanban controlled

pull systems: an analytic approach. Management Science, 35(9):1079-1091, 1989.

[3] T. M. Kim. J.I.T. manufacturing systems: a periodic pull system. Int. J. Prod. Res.,
23(3):553-562, 1985.

[4] O. Kimura and H. Terada. Design and analysis of pull system: a method of multi-stage
production control. Int. J. Prod. Res., 19(3):241-253, 1981.

[5] D. Mitra and I. Mitrani. Analysis of a Kanban discipline for cell coordination in
production lines, 1. Management Science, 36(12):1548-1566, 1990.

[6] Y. Monden. Toyota Production System. Industrial Engineering and Management

Press, Atlanta, Ga., 1983.

[7] K. M. Rege. Approximate analysis of serial manufacturing lines with buffer control.
Information and Decision Technologies, 14(1):31-43, 1988.
 Capacity Oriented Production Control
for a Job Shop Production

by

Karl-Werner Hansmann
Professor of Industrial Management
Universitiit der Bundeswehr Hamburg
D-2000 Hamburg 70, FRG

1. The Structure of Modem Production Planning and Control

Systems

Modem production planning and control is based on computer aided information systems

called Manufacturing Resource Planning (MRP II) which deal with the main functions of

the production process in a hierarchical way as shown in Figure 1.

Practical experiences of the author with German business firms have shown that the

performance of standardized MRP II systems is not so good as expected with respect to

flow time and work-in-process inventories. The main reason for this is the fact that MRP

II systems plan the functions of the production process successively and not by a

simultaneous approach integrating the individual functions of Fig.1 as it should be. But this

global aim is hardly to be reached in practice because the tremendous volume of data

which have to be processed simultaneously still exceed the capacity of current main

frames .
90

Production Planning

Master Production Scheduling

Material Requirements Planning

Capacity Requirements Planning

Production Control

Order Release

Scheduling

Shop-Floor Data Collection

Fig; 1 Structure of a Production Planning and Control System

 91

As a fIrst step to simultaneous planning this paper presents an approach that integrates two

functions of production control: order release and scheduling. Based on the data of a real

world problem it can be demonstrated with a simulation model that the integration of order

release and scheduling results in lower mean flow time and tardiness compared with

standard approaches of production control.

The here presented method is called Capacity Oriented Production Control (COPC)

because it explicitly allows for the bottleneck machines on the shop-floor while

determining the orders to be released to the manufacturing area.

In the next section order release will be described and analyzed as a function of production

control. Subsequently the integration of order release and scheduling will be discussed.

2. Order Release as a Function of Production Control

2.1 Load Oriented Order Release of Wiendahl

As mentioned above a lot of manufacturing fIrms using MRP II systems complain about
high flow times and in-process inventories connected with a considerable tardiness of some

orders. To improve the performance of shop-floor control BECHTE (1980) and

WIENDAHL (1987) designed an approach, called Load Oriented Order Release, which

can be described as follows:

In contrast to MRP II an order is not yet released to the shop-floor when the required

materials are available but only when it is certain that the order does not exceed the load

limit of any work station of the shop-floor.

92

The procedure comprises four steps:

Step 1:

The "urgency" of the arriving orders is determined by subtracting the planned flow time

from the due date to get the starting date. Only those orders are urgent whose starting date

lies within a previously defined time horizon.

Step 2:

In this step the load limit of the work stations is determined. In order to preserve the work

stations from being idle it is intended to keep a certain level of in-process inventory in

front of each work station. This desired inventory level plus the capacity of the work

station yields the load limit. The ratio of the load limit to the capacity is called the load

percentage. This measure has the same numerical value for all work stations and serves

to control the utilization of the shop-floor machines.

Step 3:
The third step consists in determining the capacity requirements of the urgent orders,

considering the load percentage calculated in step two. The capacity of a work station

required by an individual order is its processing time Pij multiplied by the probability that

this order will reach the work station in question during the planning period.

Expected capacity requirement

= processing time (pjj) • probability (Prjj)

that job i is processed at work station j during the planning period.

To calculate the probability Prjj we start from a load percentage of 200 % (that means

twice as much as capacity is available) for all work stations. Then the chance of a job just
 93

passing work station 1 to be processed at work station 2 is 0.5 because half of the jobs -

on the average - have to wait in front of work station 2 due to the chosen inventory level.

The same job can be processed at work station 3 only if it passes work station 2

(probability =0.5) and reaches work station 3 without waiting in front of it (probability
=0.5) during the planning period. Thus the probability for being processed at work station
3 is 0.5 • 0.5 =0.25 (statistically independent events). The probability for reaching work

station 4 is 0,5 3 and so on.

Consequently the load percentage determines the probability Prij by its reciprocal

Step 4:

After determining the expected capacity requirements for each order on each work station

the order release is carried out according to the following rule:

An order is released if its expected capacity requirements do not exceed the load limit of

any work station during the planning period.

The method of Wiendahl has proved to be a suitable approach to shop-floor control and

has been implemented in about 50 manufacturing fIrms with good results (WIENDAHL

1988). In spite of this experience three theoretical drawbacks are to be mentioned:

- The calculation of the expected capacity requirements with one single load percentage

seems to be too simple for a manufacturing process with very different kinds of work

stations.
94

- The load limit is not determined by an optimization procedure but arbitrarily chosen.

- The approach does not distinguish between bottleneck machines and non-bottleneck

machines.

In the next section the Wiendahl approach will be extended to a capacity oriented order

release avoiding the just mentioned drawbacks.

2.2 Capacity Oriented Order Release

This approach has been developed by K. Kleeberg and the author for a job shop produc-

tion (HANSMANNIKLEEBERG 1989) and has also been extended to flexible manufactu-

ring systems (HANSMANNIKLEEBERG 1991).

The procedure has the following essential features:

I . The bottleneck machines are established by using information from MRP II

production and capacity planning.

2. Individt,lal load limits are determined for bottleneck and non-bottleneck machines

according to an optimization procedure discussed in section 3.2.

3. The probability that an order reaches a work station is not simply expressed by the

reciprocal of the load percentage but is calculated considering

- the number of work stations to be passed before the work station in question and

- the processing times of the order on these work stations

with the following heuristic formula:

 95

j-I

T - E RW mk
G- j- I
k • i
Prmij = ----:::::T:----
n -
n
i)
if E RW mk ~ T
k = i

j-I
Prmij = 0, if E RW mk >T
k = i

= I, ... , n-I
J = i+l, ..., n
m = I, ... , M

i = I, ... , n
m = I, ... , M

with:
Pr = probability
m = job
M = number of urgent jobs
= work station where the order is at the moment
j = work station to be reached
n = number of work stations
T = Length of the planning period
RW mk = Remaining work (processing time) of the job m on work station k

The second term of the probability Prmij stands for the "risk" that an order has to wait in

front of a work station. As we do not know the actual waiting time we use the heuristic

rule that the number of work stations still to be passed increases the expected waiting time

and reduces the probability that the order reaches the work station j .

The expected capacity requirement (ECR) of job m on work station j is

EC~j =processing timemj . Prmij

96

4. The orders are released to the shop-floor according to the following rule:
The orders whose expected capacity requirements do not exceed the load limits of the

work stations are released in the sequence of their starting dates.

This procedure of the capacity oriented order release results in clearly reduced flow time

and tardiness compared to the Wiendahl approach as Fig. 2 shows.

3. The Integration of Order Release and Scheduling within

Production Control

The elements of the Capacity Oriented Production Control (COPC)

are

- a suitable and measurable objective function,

- optimizing the load limits of the bottleneck and non-bottleneck machines,

- scheduling with a suitable combination of priority rules in connection with the optimal
load limits.

These elements will be discussed in detail in the next sections.

3.1 The Objective Function of Production Control

The global objective function of production control could be:

processing the orders with minimal cost in consideration of the due dates.

As the direct impact of sequences and schedules on the production cost is hardly
 97

measurable in practice we need an operational objective function and choose the mean

flow time (MF!') of the orders as an indicator for capital tie-up cost which make up a

great deal of production cost. A suitable criterion for meeting the due dates is minimizing

the mean tardiness (MT) of the orders.

Consequently the following combined objective function (COF) is an appropriate

approximation to the above mentioned global objective function (a determines the weight

of the two objectives MF!' and MT):

COF = ~~N [a· MFTUI. + (I-a) . MTUI. ] O<a<1.

with:

L E set of Load Limits

REset of Priority Rules

As the simulation model in section 4 will show the combined objective function (COF)

enables a cost-effective production control on the shop-floor.

3.2 Load Limit Optimization

To determine the optimal load limits we have gradually varied the load percentage for all

work stations from 100 % to 500 % to find a load percentage which minimizes the

combined objective function COF. Fig. 2 shows that both the mean flow time (MF!') and

the mean tardiness (MT) are dependent on the load percentage. The priority rule used for

scheduling is a combination of WINQ GYork in Queue of the next operation of the job)

and SPT ilihortest frocessing lime).

98

Combined Objective Function

WINQ+SPT
20

18 ------

16
0 14
a
y 12
S
10

4
100 150 200 250 300 350 400 450 500
Load Percentage
~ MFT MT -COF

Fig. 2 Relationship between Load Percentage and the components of the combined
objective function

Fig. 2 shows that the mean flow time (light grey curve) increases with a higher load

percentage but not monotonously because the points of the curve are averages of 30
simulation runs and thus contain a stochastic element.
 99

The theoretical reason for the increasing mean flow time is that more orders are released

when the load limit is growing. More orders on the shop-floor often cause higher

in-process inventories and thus tend to increase the mean flow time.

The mean tardiness (lower dark grey curve) is initially going down with higher load limit

and subsequently begins to rise. That can be explained as follows: Additional orders

which are released beyond the 100 % load percentage are at least given a chance to be

finished by their due date. This positive effect on the mean tardiness is more than compen-

sated beyond 250 % load percentage by the increasing in-process inventory and mean flow

time. Consequently the mean tardiness begins to rise beyond that load limit.

The combined objective function COF takes both goals into consideration. We have chosen

the weight (X = 0.6 according to the decision of the manufacturing firm, to give a little

higher weight to the mean flow time than to the mean tardiness.

According to Fig. 2 the combined objective function COF (dark grey curve) reaches a

minimum value of 11 days at a load percentage of just under 200 %. The structure of

the curve shows that optimal load limits do exist on the shop-floor and can be determined

by a simulation model as it will be described in detail in the next section.

4. A Simulation Model for Production Control

4.1 The Structure of the Simulation Model

The simulation model approximates to the real manufacturing firm as closely as possible.

The shop-floor comprises ten shops each containing one group of equal work stations

which can be considered the stages of production. We know by information from capacity
100

planning that one shop is a bottleneck while the other shops have free capacity reserves.

The manufacturing firm carries out both job order production and job lot production. The

customer orders are poisson-distributed and require five to ten operations.

The planning period of production control is one week, that means five workdays with

one shift. The simulation period includes 104 weeks (= 2 years), the first 12 weeks being

omitted in the evaluation.

The flow time of an customer order consists of the time for processing, transporting and

waiting while the tardiness is defmed by the formula

Max [0 ; release date + flow time - due date].

To calculate the mean tardiness only the actually tardy orders are taken into consideration.

The simulation runs were carried out on an mM compatible PC 80486/33 MHz. As one

single simulation run takes 30 minutes simulation had to be restricted to 30 runs which

seem to be sufficient for getting typical results.

4.2 Comparison of the Simulation Results

The performance of the here presented capacity oriented production control (COPC) can

be demonstrated with respect to two alternative approaches:

- The capacity oriented order release results in significantly lower values of the

objective function than Wiendahl's approach.

- The integration of capacity oriented order release and scheduling greatly improves

production control compared to classic scheduling.

 101

4.2.1 Order Release

Fig. 3 shows the direct comparison between Wiendahl's approach and the capacity oriented

order release developed by KLEEBERG and the author.

Combined Objective Function

35~ ___________________________

30

D
25
a
y
s
20

15

10~~~~__~~__~~__~~
100 150 200 250 300 350 400 450 500
Load Percentage
c=J COF(1)

Fig. 3 Comparison between Wiendahl's approach (1) and the capacity oriented order
release (2)
102

A look at Fig. 3 illustrates that the capacity oriented order release results in significantly

lower values of the objective function than Wiendabl's approach and this holds good for

aU load percentages. The reduction of mean flow time and tardiness makes up about 5

days i.e. 33 %. The structure of the light grey curve (Wiendabl) and the dark grey curve

(COOR) is similar (except the level) that we conclude the differences of the objective

function values actually reflect differing performance.

4.2.2 Integrated Production Control

The second comparison shall demonstrate the superiority of the capacity oriented produc-

tion control using different optimal load limits for bottleneck and non-bottleneck shops
over the traditional scheduling with priority rules.

Tab. 1 shows the values of the combined objective function, the mean flow time and the

mean tardiness including their standard deviations when production control is merely
conducted by scheduling without consideration of order release (that means all aniving

orders are released to the shop-floor immediately). The four priority rules are the best out
of a set of 25 rules tested which were also applied to the capacity oriented production
control shown in Tab. 2. (DO = Earliest Due Date; SL = Smallest Slack)

Tab. 1 Scheduling with priority rules (in days)

Rule Combined Mean Standard Mean Standard

Objective Flow T1Ille Deviation Tardiness Deviation
Function (MFT) ~Ff (MT) ~T

WINQ+SPT 13,2 15,4 1,8 10,0 15,1

DD+SPT 19,2 26,5 9,3 8,3 10,4

DD 19,8 27,4 9,4 8,4 10,8

SL 19,9 27,7 9,7 8,1 10,9

 103

Tab. 2 Capacity oriented production control (in days)

Load limit: bottleneck 185% non-bottlenecks 150%

Rule Combined Mean Standard Mean Standard

Objective Flow Time Dee on Tardiness Deeon
Function (MFT) (MT)

WINQ+SPT 10,3 14,2 1,0 4,5 4,9

DD+SPT 17,1 16,0 1,1 18,8 17,8

SL 17,4 16,6 1,1 18,7 17,7

DD 17,7 16,6 1,2 19,3 17,9

Comparing Tab. 1 and Tab. 2 allows us to draw the following conclusions:

- The capacity oriented production control is clearly superior to scheduling with priority

rules because its combined objective function yields lower values with all priority rules

applied. The best rule combination WINQ + SPT reduces the objective function by

roughly 28 %.

- The optimal rule combination WINQ + SPT dominates all other rules as it leads to

both the shortest mean flow time and the shortest mean tardiness. This means that

suitably assembled combinations of rules are more effective than simple priority rules.

- The standard deviation of the mean flow time is largely diminished by the capacity

oriented production control. This holds true for the mean tardiness as well if WINQ +

SPT is used. Small deviations reduce the risk of production planning and control

substantially.

- The optimal load limits are 185 % in the bottleneck shop and 150 % in the non-bott-

leneck shops, which means relatively more orders are loaded in the bottleneck shop. This

seems sensible because an idle time of the bottleneck shop would strongly affect mean

flow time and mean tardiness and thus has to be avoided as much as possible.
104

It can be summarized that the capacity oriented production control presented here has

shown encouraging findings for the job shop production which could be confirmed in a

similar approach with Flexible Manufacturing Systems (HANSMANNIKLEEBERG 1991).

Thus the procedure can well contribute to the integration of important functions of

production control within MRP II systems.

References

Hansmann, K.-W. (1992): Industrielles Management, MOnchen-Wien 1992.

Hansmann, K.-W.; Kleeberg, K. (1989): Comparison of New Approaches to Job Shop

Scheduling with an Interactive Simulation Programm on PC, in: Paper for the
ORSAfTIMS conference in Vancouver/Canada 1989.

Hansmann, K.-W.; Kleeberg, K. (1991): Extension of Capacity Oriented Scheduling to

Flexible Manufacturing Systems, in: Paper for the 2nd conference of POMS in New
York 1991.

Haupt, R. (1989): A Survey of Priority Rule-Based Scheduling, OR-Spektrum, Vol. 11,

Febr. 1989, S. 1-16.

Kettner, H.; Bechte, W. (1981): Neue Wege der Fertigungssteuerung durch

belastungsorientierte Auftragsfreigabe, VDI-Z, 123 Jg., Heft 11, 1981, S. 459-466 .

Wiendahl, H.-P. (1987): Belastungsorientierte Fertigungssteuerung, MOnchen 1987.

Wiendahl, H.-P. (1988): Fertigungssteuerung, in: Schriften zur Untemehmensfiihrung, Bd.

39, 1988, S. 51-87.
 On Solving a Large-Scale Resource Allocation Problem
in Production Planning

Christof Dillenberger(J), Laureano F. Escudero(2)

Artur Wollensak(l), Wu Zhang(l)

(I)IBM German Manufacturing Technology Center (GMTC)

Dept. 0817: Modeling Methods
P.O. Box 266, 0-7032 Sindelfingen, Germany

(2)IBM T.J. Watson Research Center

Dept. 432G: Manufacturing Modelling and Logistics
P.O. Box 218, Yorktown Heights, NY 10598, USA
currently: Mathematical Sciences, University of Madrid
Ciudad Universitaria, Madrid, Spain

Abstract

Various production planning problems can be described and solved with Linear Pro-
gramming (LP) models. In manufacturing applications setup often plays an important
role. To solve a setup problem 0-1 variables must be introduced in the model, which
leads to an increase in the model size and computational effort. This paper discusses
several models to consider setup times in production planning. A decomposition meth-
odology that does not consider the integrality of all 0-1 variables at once, but succes-
sively, is outlined. Results for an application in the modeling of the IBM Sindelfingen
Multi-layer-ceramics (MLC) Punching area, are discussed .

1. Introduction

Production planning is concerned with determining production and allocation resources

to meet fluctuating demand requirements. If resources can be acquired as needed and
plant capacity is infinitely expandable and reductable at no cost, the optimal production
schedule consists of producing end products according to the demand schedule.
However, in many real production systems, the pro(!uction capacity may vary due to dif-
ferent reasons even for short planning horizons and the demand for products fluctuates,
both in total volume and in product mix. As a result, just-in-time production is not
usually feasible, and when feasible, may result in poor utilization of resources. In these
circumstances, the production planning consists of deciding on the best utilization of the
available resources given the current forecast of demand requirements.

There is a broad body of literature on production planning systems. See for example
hierarchical approaches in [6-8]. Here, we restrict ourselves to production allocation to
machines and time periods in a single-level, multi-item capacitated production environ-
106

ment with start setups as well as setups that are partially dependent on production
sequence and independent of time periods in a flexible manufacturing system.

Resource allocation to determine optimal programming of lot-sizes is a topic that has

attracted increasing attention over the last 25 years since the work of Dzielinski and
Gomory [5]. We are more interested on manufacturing systems where machine setups
are required. Typically, start setups (Le., time period dependent setups) are only
studied, see [1-3, 10-14]. In this work, we benefit on the notion of setup that allows an
uninterrupted sequence of periods on which production for a given item takes places
and, then, the setup is invariant to time periods change.

Let us define the following problem:

A production volume of various part types is to be processed over a given planning

horizon (e.g., one week). The planning horizon is divided in (integer and consecutive)
time periods (e.g., days or shifts). We assume that the availability of time units in each
time period may be different from one to another (e.g., Monday through Friday 3 shifts,
Saturday 2 shifts). The set of parts can be decomposed into subsets of identical parts;
let us name any of these subsets, part type. For the sake of simplicity, let us assume
that only one operation is to be performed on any part. The set of parts can also be
decomposed in families, such that a part type must belong to one and only one family.
A part type can be processed by more than one MG; only one machine is required to
process a given part, preemption is not allowed and no more than one operation can be
performed simultaneously by any machine.

The number of time units that are required to perform a given setup is a function of the
machine, the family and the part type that is to be processed. If two consecutively proc-
essed parts belong to the same part type (and family, of course) then there is no setup
for the second part. If they belong to the same family but not to the same part type,
then a minor setup is necessary, otherwise it is a major setup. Assume that different
MGs may have different efficiencies in the processing of the same· part type. Then, the
processing time is a function of the part type and the MG.

Each part type requires a number of resources that mayor may not be available at a
given time. These resources include tools, manpower, robots and others. We distin-
guish the resources by their availability mode and their consumption mode. It results in
the following four resource types:

~ non-transferable resources consumed per part type (e.g., operator time),

~ non-transferable resources consumed per process time unit (e.g., energy),
~ transferable resources consumed per part type (e.g., components), and
~ transferable resources consumed per process time unit (e.g., tools lifetime)

Note: Non-transferable resources that are not used at a given period cannot be used in
(transferred to) the next period, they are lost.
We allow the backlog of some part types to be strictly positive and upper bounded (i.e.,
it may not be required to satisfy the entire d(~mand of all part types). The production per
period, as well as the cumulated production up to the given period per part type may be
lower and upper bounded. There are also logical constraints related to the utilization of
the machines. Some examples are as follows: maximum number of machines where a
 107

part type may be assigned at any period, maximum number of part types that one
machine can handle at any period, etc.

The goal consists of obtaining a feasible deployment of the MGs and resources that opti-
mizes some of the objectives stated below. The deployment of the resources is a func-
tion of the main unknown in the problem, that is, the production volume of each part
type that is assigned to each machine at each time period. The possible objective func-
tions are

~ maximizing the shipment value along the planning horizon,

~ maximizing the weighted serviceability,
~ minimizing the weighted production backlog,
~ minimizing total processing time,
~ minimizing total cost of the resources, and
~ minimizing total setup time.

As a by-product the following results should also be provided:

~ part type backlog,

~ machine utilization,
~ resource utilization,
~ resource cost,

~ net availability of transferable resources per time period,

~ number of major and minor setups per machine and period,
~ total setup time per period,
~ number of part types assigned to machine and period,
~ number of machines and periods assigned to part types, etc.

The main difference between our approach and those published in the literature is that
we are interested on large scale instances, so optimality is not our main goal. We look
for practical solutions where the optimality gap, hopefully small, is also provided and the
computational effort is still affordable. On the other hand, we allow for a variety of deag-
gregated capacity constraints that are not frequent in the literature.

Finally, we are dealing with so-called start setups (i.e .• a fixed cost or machine utiliza-
tion is incurred if the production of an item is allocated for a given period~;_ see refer-
ences above , but our approach also deals with other (minor and major) machine setups.
This type of setups is partially dependent on the sequence of the parts and is inde-
pendent of the time periods on which the planning horizon is segmented. The
modelization and the so-called Fix-and-Relax algorithmic methodology that we propose
are new in the open literature. See the companion paper [4] for a detailed description
of the full model and the algorithmic framework.

In this work we focus on the comparison of different approaches for modelling the setup
requirements. The paper is organized as follows. Section 2 presents a base LP model
where the setup requirements are not considered. Section 3 discusses different
modelizations of the setup requirements. Section 4 outlines our Fix-and-Relax algo-
rithmic r:1ethodology. Section 5 reports on our computational experience with a large-
scale real-life problem. Finally, we offer some conclusions.
108

2. A Base lP Model

In this section we present the model for the production planning problem given above,
but setup requirements are not considered. First we introduce the notation.

T, set of time periods {I} in the planning horizon to study.

I, set of part types {i} .

M, set of machine groups, MGs {m}.

0 ;, demand of part type i to be satisfied along the planning horizon .

P" length of period I.

P;.m, processing time required for producing one unit of part type i in MG m .

Nole: P;.m = 0 indicates that part type i cannot be processed on machine m.

vm ,,, availability of machine m in period I

am, number of identical machines in MG m.

Wi. weight factor for backlog of part type i.

A" weight factor for period I.

The model must provide a value for the following decision variables:

X I.m . . production volume of part type i that is assigned to machine m at period I.

dl ,,, production volume of part type i at period I (it is an instrumental variable).

y;, backlog of part type i.

The target is to obtain the values for Xi.m.t Vi € I, m € M, I € T and Yi Vi e I, such that the
backlog and processing time along the planning horizon is minimized (1.1) and the fol-
lowing constraints are satisfied:

1. The workload assigned to a machine group does not require more time units than
the ava ilability for a specific period (1 .2)

2. The maximum allowed backlog for each part type is not violated (1 .3)-(1.4)

The LP model is as follows .

(1.1)

subject to:

b;,m"X;,m,t:S CTm·Vm.t"pt VmeM . teT (1.2)

;e I
 109

I
mEM
X; ,m,t = d ;,t ViE I, tET (1 .3)

2:>;,t+ y;= D; Vi E I (1.4)

tE T

X;,m,t~O,y;~O ViE/ , mEM , tET (1.5)

Table 1 shows a typical matrix Pim for the processing time in a job-shop environment
with parallel machines, and where a part type can be processed by more than one
mach ine.
Table 1: Processing Time Matrix

Machine groups

PT MGl MG2 MG3 .. . MG13 r~G14

1 1.892 1.853 O. 1. 703 1.853

2 0.782 0. 743 O. 0.703 0. 767
3 0. 723 0.723 0.723 0. 777 0. 777
4 0.477 0.492 0.477 0.492 0.492
5 0. 480 0.480 0.480 0.490 0.490
6 0. 735 0.735 0. 735 0.750 0.750
7 0.408 0.403 0.443 0.443 0. 443
8 0. 408 0. 408 0.443 0.443 0. 443
9 O. O. O. 0.798 0. 798
10 1. 383 O. O. 1.383 O.
11 1.062 O. O. 1.062 O.
12 1.1317 e. o. 1.('J17 O.
13 O. O. 1.340 1. 367 1. 367
14 0. 115 0. 443 0.315 0.443 0.443
15 O. O. 0.433 0. 442 El . 442
16 O. O. O. 1. 310 O.
17 O. O. O. 0.867 O.
18 O. O. O. 0. 952 O.

Solving the LP model (1.1)-(1 .5) leads to a solution as shown in table 2. From a setup
po int of view such a production plan is not very encouraging . Probably the LP solution
has generated unnecessary setups for the machines. The modelling approach that we
describe below avoids all unnecessary setups.
110

Table 2: Production Plan without considering Setup

Periods
MG t~ONDAY TUESDAY WED DAY THURSDAY /I Setup

1 10= 475 12=1091 12= 856 12=1091 5

14=3924 14=2076

2 7=2749 8=2715 7=2749 3

3 3= 480 5= 855 4= 977 6

15=1298 6= 480 13= 480
5= 415 4= 725

4 7=2749 8=2715 5=1656 3

8= 769

... . ..
11 4=2970 4= 32 4=2970 12=1091 4
12=1079

12 16= 846 16= 174 18=1165 17=1279 4

17= 157
18= 783

13 10= 802 18=1165 18=1165 12=1091 3

14 9=1389 9= 611 15= 496 3

15=1406 1= 480

Total 45

3. Setup Approaches

The modelling of setup requirements leads to the introduction of 0-1 variables since we
have to detect if a setup is necessary or not. In fact this is a typical situation for the
usage of 0-1 variables. We compared 4 different cases with respect to the modeling
accuracy and resulting problem size.

1. One approach is to split the user defined periods into so-called micro-periods of
equal length (see figure 1). Per each micro-period we allow only one single part
type (A, B, or C). From the modelling point of view this approach is very easy to
handle, but it has some drawbacks. The major one is the estimation of the optimal
number of micro-periods. If this number is too large, the size of the 0-1 model will
explode, since the number of 0-1 variables is proportional to the number of micro-
periods (and machines and part types). If the number is too small, we may waste
capacity, since the production volumes do not fit exactly into the micro-periods. The
waste of capacity is especially high for cases with short processing times or small
production quantities.
 111

Figure 1. Case 1. Two user periods split into 4 fixed length micro-periods. Only t part type
(setup and/or processing) per micro-period is allowed.

2. The drawback of wasted capacity in approach 1 can be lightened by allowing vari-

able length micro-periods (see figure 2). But even in this case the determination of
the number of micro-periods is still unresolved.

Figure 2. Case 2. Two user periods with up to 3 variable length micro-periods. Only t part type
(setup and/or processing) per micro-period is allowed.

3. An alternative is to stay with the user defined periods, but allow multiple part types.
A simple approach is to define always a setup for each part type per period (see
figure 3) . From the modeling point of view this is very easy to handle, but it creates
setups where in practice no setup is required (e.g., part types 8 and C in figure 3) .
This results in wasted capacity, too.

Figure 3. Case 3. User periods, multiple part types allowed per period and always setup for each
part type assigned to a period.

4. Our best approach allows a part type to be assigned to the same machine in two
consecutive periods without requiring machine setup for thE! second period. In
figure 4 the same production quantities are scheduled as in figure 3, but only one
setup is considered for part types Band C.
112

Figure 4. Case 4. User periods, multiple part types allowed per period and setup only when nec-
essary.

This approach (hereafter called detailed approach) requires more 0-1 variables than any
other, but it is very close to reality and the resulting model size is still acceptable.

Before we present the model for the detailed approach, we introduce the following addi-
tional notation.

Note: For the simplification, here we only consider minor setups, but it is easy to
extend the model to consider major setups, which are actually implemented (see [4]).

s;"", number of time units, that are required to perform a (minor) setup for any part type i
in any machine of MG m.

b;"",h 0-1 variable such that the value 1 means that part type i can be assigned to
machine m at period t, but it has not to be assigned and, otherwise, it is zero.

"';"",. 0-1 variable such that the value 1 means that a (minor) setup is required for part
type i in machine m at period t and, otherwise, it is zero.

Note: We can relax the integrality constraint for the variables ''';.m.• since any feasible
solution automatically satisfies this type of constraints .

U;"".h 0-1 variable such that the value 1 means that part type i is assigned to machine m
to be processed as the 'last' part type at period t and, otherwise, it is zero.

0-1 variable such that the value 1 means that two part types at least are assigned to
Zm,h

machine m at period t and, otherwise, it is zero.

Now we are able to model the additional constraints (3.1 )-(3.9) to add to the basic LP
model (1.1)-(1.5):

Force b;,rn,. to have the value 1 whenever part type i is assigned to machine m and period
t: (Constraint (3.1) is the most frequent usage of 0-1 variables in production planning .)

X;,m,t ~ M ob;,m,t Vie/,meM,teT (3.1 )

Prevent violation of the reduced capacity because of setup time: (This constraint
replaces constraint (1.2) in the base LP modef.)

L(S;,mOt//;,m,t + P;,mOXt,m,,) ~ (Jmovm"op, VmeM,teT (3.2)

; E I
 113

Identify t!;c part type that causes a (minor) partially sequence dependent setup in a
machine and a given period, if any:

o;,m,t :-:::; VI"m,t + U;,m,t - 1 ViE/,mEM,tET (3.3)

Define the 'last' part type to be processed in a machine at a given period:

~U
~ I,m, t= 1 Vm EM, t E T (SOS type 1) (3.4)
; E I

U;,m,t:-:::; O;,m,t ViE/,mEM,teT (VUB) (3.5)

Note: Constraints (3.4) define the so-called special ordered sets (SOS) of type 1. One
member of a SOS may and must take the value 1 and the other members must take the
value zero. Constraints (3.5) define the so-called Variable Upper Bounds (VUB).

Define the number of part types that are assigned to a machine at a given period, such
that variable Zm .• will take the value 1 whenever more than one part type is assigned to
machine m and period t:

LtJ;,m,t:-:::; 1 + M ·zm,t VmeM,teT (3.6)

;e I

where M is a big enough number that we should keep as small as possible.

Prevent that a part type can be assigned the 'last' part type to be processed in a
machine at a given period, provided that the same part type was assigned the 'last' part
at the previous period and at least some other part type has also been assigned to the
machine at the current period:

U;,m,t + U',m,t -1 :-:::; 2 - Zm,t Vi e I, m EM, t E T (3.7)

b;,m,t' U;,m,t' Zm,t E {O,l} Vi E I, m E M, tE T (3.8)

0:-:::; "';,m,t:-:::; 1 Vie/,meM.teT (3.9)

The model above requires thousands of constraints, continuous and 0-1 variables for
real life cases: For example, a problem with 18 part types (in 7 families). 14 machines,
and 4 periods expands to a 0-1 model with 3691 constraints, 3385 continuous and 0-1
variables, 12537 elements, and 12950-1 variables with 43 SOS sets.

Solving this problem until optimality exceeds the tolerable time, even when using
mainframes and state of the art optimizers as OSLIVM [9]. It still takes around 6
minutes of CPU time to obtain the first integer solution in a IBM 3090-400E.
114

4. Fix-and-Relax Algorithmic Methodology

Given the large-scale nature of the problem. it is unrealistic to seek for the optimal sol-
ution in our environment. Instead of providing optimal solutions. we favor a decompos-
ition methodology that does not consider the integrality of all 0-1 variables at once. but
successively. Basically. it consists of analyzing the characteristics of the instance and
considering iteratively the integrality of some subset of variables. Here. the first iter-
ation optimizes the original model. where the integrality constraint of all variables is
relaxed. except the 0-1 variables that belong to the first time period. After solving the
subproblem associated with a given iteration. the 0-1 variables related to the given
period will be permanently fixed to the optimal solution for the subproblem except when
an infeasibility occurs in later periods. The optimal solution of the last subproblem
should lead to a near-optimal solution of the original problem.

The above approach can be interpreted as the following branch-and-bound scheme for
exploiting promising subtrees:

~ The order to branch the 0-1 variables is given by the periods sequence.
~ No nodes are created by branching on 0-1 variables that belong to period T while
nodes with fractional values for 0-1 variables that belong to period t are still in the
waiting list for T> t.
~ The nodes related to a given period are discarded once the best 0-1 solution is
obtained for that period. but the node with the incumbent solution.

The implementation logic for this approach can be described by the following frame-
work.

~ Define variables (j;.m.t for period t = 1 to be integer. otherwise continuous.

~ Solve MIP problem.

~ For periods t = 2•...• I rl.

• Define variables (ji.m,t for period t to be integer. otherwise continuous.
• Fix all variables (ji,m" to the (optimal) values assigned by previous iterations (for
periods T = 1.... , t - 1).
• Decide (heuristically. see below) which part type (family) should be produced the
'last' one in the previous period t - 1.

The number of 0-1 variables is reduced drastically since we can omit all con-
straints related to integer variables U; ,m t and Zm ,f'
• Fix minor (major) setup times to 0 for part types (families) selected to be
produced the 'last' one in previous period t - 1.

Note: Here the 'last' part type (family) is not forced to be the first one in the
current period. it is only encouraged.
• Solve MIP problem.
The constraints (3.4)-(3.9) detect the 'last' part type to be assigned to a given machine
and period. say m and t. respectively. In our period-by-period approach. a heuristic
decides the part type (and family. of course). say J (resp .• f). to be assigned the 'last'
 115

one for machine m in period t. If i (resp., f) is the part type (resp., family) to be
assigned to machine m tile 'first ' one for period t + 1 then tile minor (resp., major) setup
is not necessary for period t + 1.

We have to distinguish th e following cases for machine m and period t:

1. There is only one part type (say, 1) assigned to the machine and period. Then, by
definition , it is the 'last ' part type and f(l) is the 'last ' family.

2. There are two part types (say, J and 1) that have been assigned to the machine and
period, and 1 was assigned the 'last ' part for period t -1. Then, obviously, 1 is the
'last' part type for period t. (A similar mechanism works for selecting the 'last' part
family.)

3. Otherwise, let j denote the part family among all families remaining with the largest
major and minor setup. (Note that this setup can be avoided, provided that j is
selected the 'first' part family for period t + 1.)

4. Any way, let j denote the 'last' part family selected by our heuristic for machine m
and period t. Assume that there is still room to select the 'last' part type . It will be
the part type with the largest remaining workload . Alternatively, J can be the part
type with the largest minor setup to be avoided.

This approach has been implemented as the optimization kernel of an IBM internal deci-
sion support system so-called DACAPO (Demand Allocation for Capacity Optimization).
The tool has been designed in such a manner that the model generator and the algo-
rithm are completely transparent to the user. No knowledge in the field of mathemat-
ical programming is required to use the tool.

The user interface allows the parameter specification and start of the optimization run. It
gives access to the results, too. The results can be represented either in graphical or
numerical form. The parameters and results are stored in a relational database. There-
fore it is easy to exchange data with other applications and to build user defined reports.
According to the parameters the optimization kernel generates and solves the math-
ematical model using OSL as basic software.

The user interface and relational database are implemented on an IBM PS/2 using OS/2
EE 1.3. The optimization kernel is currently available on VM/XA and AIX/6000. The
implementation of all components was done with the programming language C.
116

User Interface
(ace. CUA) ..

OS/2 Present.
'~anager Optimization -
Kernel

'1
VM/XA

ReI. Database I AIX/6000

.. t
OS/2 Database
Manager

Figure 5. DACAPO components. The three major components o( DACAPO are user interface,
relational database, and optimization kernel.

5. Results

Elsewhere [4] we report an extensive computational experience by using DACAPO in

some real-life problems. This section presents results from a reduced version of a real
application of DACAPO for the IBM Sindelfingen MLC punch ing area.

The input parameters for the model are as follows:

~ number of servers per machine group (here, 1),

~ machine group ava ilability per period (here, 77%) ,
~ period length (here, 1440)
~ processing time per part type and machine group (see table 1),
~ major setup time per family and machine group (here, 20)
~ minor setup time per part type and machine group (here, 5)
~ demand per part type (total 85520, varying from 480 up to 12400).

Table 3 gives the solution when setup restrictions are enforced. As expected the
number of necessary setups is reduced significaj,tly. Without modelling setup con-
stra ints the total number of setups was 45 (see table 2).
 117

Table 3: Production pl:," considering Setup

Periods
MG MONDAY TUESDAY ~IED DAY THURSDAY /I Setup

1 11=1021 11=1044 11=1044 11= 91 4

1= 49
10= 126
14=6000

2 5=2258 5=1576 7=1478 4

3= 480 8=1182

.. . ...
11 12=1066 12=1091 12= 663 4=2970 2
4=1098

12 18=1139 18=1165 18=1165 18= 419 2

17= 807

13 10= 783 12=1066 12=1091 12= 575 3

11= 464

14 9=1356 9= 642 1= 431 2

15=1293 13= 201

Total 39

Finally. let us compare the period-by-period model (using our Fix-and-Relax approach)
with the global model (see Table 4).
Table 4: Model Comparison

Global model Period-by-period

Model size 3691 rows 1562 rows

3385 columns 1798 columns
12537 elements aV9. 5368 elements
1295 0-1 variables 177 0-1 variables
43 SOS sets o SOS sets
LP sol' n 126264.1

1st MIP 128496.6 128514.3

solution after after
219 nodes and 192 nodes and
14107 iterations 4771 iterat ions

GAP 1.76% 1. 76%

CPU time 362 . 3 seconds 93.8 seconds

The average size of the submodel solved per period (especially the number of 0-1 vari-
ables) is significantly smaller than the size of the global model. This results in 93.8
seconds of CPU time for all 4 submodel optimizations. compared with 362.3 seconds for
118

the global model. In both approaches, the optimization was stopped when the first
integer solution was found. The LP solution for the global model is 126264.1. It is a
lower bound for the optimal solution . The optimality gap of the first integer solution for
the global model is only 1.76% . The optimality gap of the solution for the period-by-
period approach is only 0.02% worse than the gap for the detailed model, but reached
much faster. Especially for practical problems it is more important to provide near-
optimal, but fast solutions instead of optimal ones.

6. Conclusion

A practical approach for resource allocation and single-level multi-period lotsizing pro-
duction planning and scheduling with setups has been proposed. This framework can
be applied to production planning as well as machine scheduling and lotsizing prob-
lems.

We have presented a decomposition methodology to get solutions that will be afford-

able. For practical purposes the quality of the solutions is more than enough. Our
approach exploits the multi-period nature of the problem and our decomposition philos-
ophy can be framed in a partial exploration of the branch-and-bound tree. No matter the
performance of the approach, its applicability needs to embed the algorithm in a system
that automatically generates the submodels and is completely transparent to the user.
For this purpose we have developed the system called DACAP6 for IBM Manufacturing
plants.

References

[1] Chand, S., and S. P. Sethi, .. A dynamic lot sizing model with learning in setups,"
Operations Research 38 ( 1990 ) 644-655.

[2] de Mata, R., .. On solving production scheduling problems with changeover costs
using Lagrangean relaxation," Working paper 89-12-04, The Wharton School, Uni-
versity of Pennsylvania, Philadelphia, 1989.

[3] de Mata, R., and M. Guignard, .. Production scheduling with sequence-

independent changeover cost," ORSA I TIMS meeting, Denver (Colorado), 1988.

[4] Dillenberger, C., L. F. Escudero, A. Wollensak, and W. Zhang, "On Practical

Resource Allocation for Production Planning and Scheduling with Different Setup
Types," European Journal of Operational Research submitted ( 1992 ).

[5] Dzielinski, B. P., and R. E. Gomory, "Optimal programming of lot sizes, inventory
and lot size allocations," Management Science 11 ( 1965) 874-890.
 119

[6J Escudero, L. F., "An inexact algorithm for part input sequencing and scheduling
with side constraints in FMS," International Journal of Flexible Manufacturing
Systems 1 ( 1989) 143-174.

[7J Graves, S. C., " Using Lagrangean techniques to solve hierarchical production
planning problems," Management Science 28 ( 1982) 260-275.

[8] Graves, S. C., "A tactical planning model for a shop job," Operations Research 34
( 1986) 522-533.

[9] IBM , Optimization Subroutine Library Guide and Reference, Kingston (NY). 1991 .
I BM Form SC23-0519

[10] Lozano, S., J. Larraneta , and L. Oniera, "Primal-dual approach to single-level

capacitated lot-sizing problem," European Journal of Operational Research 51 (
1991 ) 354-366.

[11J Magnanti, T. l., and R. Vachani, "A strong cutting plane algorithm for production
scheduling with changeover costs," Operations Research 38 ( 1990) 456-473.

[12J Pocket, Y. •and l. A. Wolsey, "Solving multi-item lot sizing problems using strong
cutting planes," Management Science 37 ( 1991 ) 53-67.

[13] Trigeiro, W. W., L. J. Thomas, and J. O. McClain, "Capacitated lot sizing with
setup times, " Management Science 35 ( 1989) 353-366.

[14J Wolsey. L. A., "Incapacitated lot-sizing problems with start-up costs ," Operations
Research 37 ( 1989) 741-747.
 INTEGRATED STATION LOCATION AND FLOW NETWORK
DESIGN FOR MANUFACTURING SYSTEMS LAYOUT

Dilip Chhajed
Depanment of Business Administtation
University of Dlinois at Urbana-Champaign
1206 South Sixth Street, Champaign, n.. 61820.

Timothy J. Lowe
~ntof~gementScience
University ofIowa, Iowa City,lA 52242.

Benoit Montreuil
Groupe de Recherche en Gestion de la Logistique
Department of Operations and Decision Systems
Laval University, Quebec, Canada GlK 7P4.

Abstract

In this paper the component approach to designing a detailed layout is considered. The component
approach has four distinct phases of layout design: (a) block layout, (b) input/output station location, (c)
flow network design, and (d) aisle netting-ouL We look at the problem of simultaneously designing the
station locations and flow network. Two different topologies are considered. Formulation and solution
procedures for these topologies are discussed.

1. Introduction

Recently, renewed interest in the ~ design of facilities has developed, in part, because of
increased global competition in manufacturing and an increased consciousness toward reducing
manufacturing costs. Tompkins and White (1984) suggest that about 20% to 50% of the total operating
expenses within manufacturing are attributed to material handling. Efficient facilities design can reduce
these costs by at least 10% to 30% and thus increase productivity. Facilities design involves three
interrelated components: structure, layout, and the determination of a (network) system to suppon material
 121

flow interaction between facilities, e.g. material handling systems. For best results, various components
offacilities design must not be designed in isolation. However, at present, a decomposition approach is
usually taken, e.g. the material handling system (network) is determined aflcr the structure (size and
shape) and layout (orientation of facilities) has been determined.
Classical layout design approaches have primarily focused on laying out the block plan . Foulds
(1983) and Kusiak: and Heragu (1987) give comprehensive swveys of various research effons to develop
the block layout. Note that in the block plan, operational details such as circulation regions, aisle structure
and the location of departmental input/output stations are generally not modeled. Recently, several
researchers have come to recognize that considering aisle travel explicitly in the major layout design phase
provides significant potential for improvement in flow travel (via aisles) and space devoted to the aisle
system. O'Brien and Abdul Barr (1980) proposed a layout improvement algorithm named S-ZAKY
which computes the expected flow distance savings based on the location of the input/output stations. The
INTALA interactive layout software of Warnecke and Dangelmaier (1983), which uses a construction
approach, permit three internal configurations (shape with input/output locations) for each department
Instead of directly generating a layout with aisles using a myopic approach, Montreuil and Ratliff
(1988) and Montreuil (1987) have proposed a component approach to layout design where the components
are a) block plan, b) input/output station location, c) flow network design, and d) aisle netting-out (defmed
below). In this approach, initially a block plan is generated (Figure I). Given a block plan, departmental
input/output station locations can then be determined. All material flow between departments occurs
through these stations. Given the block layout and the location of input/output stations, an optimum flow
network must be designed. This network is the basic structure on which the flow of material and people
will take place. Finally, this network is transformed to a set of aisles and circulation areas (aisles are
netted-out).
The block design can be generated by using any of the classical approaches mentioned earlier (Foulds,
1983). A polynomially solvable model for optimally locating the input/output stations, when all the
departments are rectangular, is given in Montreuil and Ratliff (1988). Chhajed, Montreuil, and Lowe
(1991) have given a Lagrangian based heuristic to develop one type of flow network design. Montreuil
and Venkatadri (1988) have developed a comprehensive linear programming model generating a net layout
from an aggregate layout, coupled with a flow network.
All of the above papers tend to design one component at a time. It is clear that such a sequential
optimization approach could lead to a suboptimal overall design, but solving the overall problem as a
single integrated problem is a fonnidable task. However, an attempt to integrate as many of these
components as possible must be made in order to truly gain the benefits of optimization and to design a
system which is as close to ideal as possible. In this paper we present a step forward in the direction of
integration. The two components we have chosen to integrate are the station location and the flow network
design; we denote this problem as ISLAND (Integrated Station Location And flow Network Design).
Since there are a number of topologies possible for the flow network design, the same topologies are
possible for the ISLAND problem (Chhajed et al., 1991). Thus the ISLAND problem involvesdlOQsing
the location of a station for each department and designing a network which permits each flow between the
chosen stations subject to the topological requirements.
122

I J
(LA) (DR (LA) (DR

(PR) (PR) ~

(A ) (A! )

-
(S'I)
(S'I)
r- /PI \
(OR)
.,
ror
(OR)
(WR
(RE) (MI: (SH) ! /WE) (RE)
~ (SH) (WE) (WR

(a) A Block Layout (b) Block Layout with Input/Output Stations

[LA]
[DR]

[PR]

(RE)
-I

.....~~
(<:In
' l L _ - I WR

(c) A Block Layout with (d) A Block Layout with

110 Stations and Flow Network I/O Stations and Aisles

Figure 1. The Component Approach To Detailed Layout

We now discuss previous work related to our approach. Usher, Evans, and Wilhelm (1988)
consider simultaneous determination of flow paths and locations for load transfer stations within an
automated guided vehicle system Their two-phase heuristic first assumes the locations for the load
transfer points and solves for network design using the zero-one integer program formulation developed
by Gaskin and Tanchoco (1987). In the second phase, the location of load transfer points is improved by
solving a series of one-median problems. This process is repeated until no further improvement is found in
the total weighted distance traveled by loaded vehicles. The objective function in Usher, Evans, and
 123

Wilhelm does not consider the network cost. Note that the phase 1 problem is in general difficult to solve
and Gaskin and Tanchoco did not give any special property of this formulation which could be exploited in
a solution procedure. Furthermore, their procedure is iterative and works in two phases, since only one
component is solved at a time.
In this paper we discuss two different topologies for the ISLAND problem. In Section 2 we
discuss the case where the flow network is not restricted to be a subnetwork of any prespecified network.
an
Such assumption is called afreeflow assumption. We show that the problem can be posed as a fixed
charge network design problem and suggest ways to solve the problem and generate lower bounds. In
Section 3 we assume that the flow network has to be a subset of a prespecified network. This prespecified
network, for example, may be the network defmed by the block layout. We present a path-based
formulation for this case and identify special cases for which the problem is relatively easy to solve.
Section 4 has some concluding remarks. In the Appendix we provide proofs of results stated in the paper.

2. Free Flow

In this section we consider the case when flows are permitted to occur anywhere (free flow)
through the building enclosing the departments. In this situation, the actual flow network which is to be
designed, may pass through the interior of some departments. Such an assumption is called afreeflow
assumption. The free flow assumption is )lseful during the design stage to obtain the best (unconstrained)
material handling network. Flow network design with a free flow assumption is considered in Chhajed,
Montreuil, and Lowe (1991). We also constrain the network to have only horizontal and vertical edges and
each station must be located on the boundary of its respective department. Figure I (c) shows an example
of such a network. We call the ISLAND problem with these assumptions, F-ISLAND. The objective is to
determine the station locations and the network so that sum of the flow cost and the network cost is
minimized.

2.1 Formulation

Let D be the set of departments with cardinality n. We assume that only one station per department
is to be located and thus the set of stations has one member corresponding to each department. Let F be a
set consisting of pairs of departments such that if (u,v) e F, then there is a flow between departments u
and v. We consider only undirected flows and a flow (u,v) is said to be in F if either (u,v) or (v,u) is in
the set F. Occasionally we will denote a generic flow (u,v) by the letter f.
The block layout is assumed to be given with each block (department contour) an isothetic simple
polygon. Free flow is assumed and only travel parallel to one of the axes is allowed. Given a set of station
locations and the flow network (to be determined), each flow is assume to take the shonest path on the
network. The flow cost for flow f is computed by taking the length of the path for the flow f and
124

multiplying it by the annual flow cost per unit length of flow f, which we denote by Or (also denoted as
Q(u,v) for flow (u,v». Without loss of generality we assume that the network cost per unit length is 1. We.
now define two graphs:
Definition: Given a block layout, the department contours (boundaries) can be thought of as a graph with
nodes at the points where contour lines intersect or make a 9()0 tum. We will call this graph the COf/lOUT
Graph and denote it by C. V(C) and E(C) represent the nodes and edges, respectively, of the contour
graph C. The graph formed by solid lines and solid nodes in Figure 2 is an example of Contour graph.

Definition: Given a set of points x on the plane a grid graph of a set x is defined as: Draw a horizontal and
a venicalline through each point of 7t. The intersection of each horizontal line and venicalline defines a
grid poinL Take the intersection of these lines with the smallest rectangle containing x with sides parallel
to the axes. The union of the grid points, the set x, the rectangle, and the collection of horizontal and
vertical lines contained in this rectangle is the grid graph, n(lt). Note that the contour graph is a subgraph
of n(V (C». The union of all edges and vertices in Figure 2 is a grid graph.
Let Nu denote the set of those nodes of n(V(C» that are on the boundary of department u. We
now show that there is an optimal solution to the F-ISLAND problem where the stations are located on the
nodes of the grid graph n(V(C».

lA
DR

PR PL AS

ST ..·
!P-- ···..·······11··············· ......•.... ...........................•.....
~

GR WR

MI SH WE

Figure 2. Contour Graph C and Grid Graph n(V(C»

Theorem 1· There exists an optimal solution to the F-ISLAND problem in which the station of department
u belongs to Nu for all u e D, and the network connecting them is a subnetwork of n(V(C».
fr.QQf;. The proof of this theorem is given in the Appendix.«»
 125

This Theorem is the key to the following formulation. We first note that some of the edges of the
grid graph can be eliminated without affecting the optimal solution. For example. the dotted edges not in
the block layout of Figure 2 can be eliminated. For simplicity we will continue to call the graph after
elimination of these edges as TI(V(C». We now formulate an optimization problem. To the graph
TI(V(C». for each department ue D. append a node nu and add arcs between no and each node in Nu. The
length of each appended arc is set to be '1'. a number at least as large as the sum of all the arcs in TI(V(C».
DenOIC by R the resulting graph. Create another graph H by making a copy of R and replacing each arc in
the copy of R by two directed arcs. directed in opposite directions and with length identical to the replaced
arc. For an arc ae E(R). the corresponding directed arcs in E(H) are denoted by a and a'.
For each flow f=(u,v)e F. arbitrarily designate nu as the source node and nv as the sink node for
flow f. For each fe F.let t>f be a column vector with -I in the position of the source node and +1 in the
position of the sink node and a 0 in all other positions corresponding to the nodes of H. Let the node-arc
incidence matrix ofH be M. We now have the following fixed-charge network design problem
(Balakrishnan, Magnanti, and Wong, 1989).

f f
(Fl) min r Caxa + r Qr( r CaY a + r Ca'Ya' ) - e (1)
aeE(R) feF aeE(H) a'eE(H)
SUbject to: Myf=»f V fe F (2)
f
Ya ~x. Vfe F,ae E(R) (3)
f
Ya' ~xa Vfe F,ae E(R) (3')

xa = (0,1) and integer (4)

f
y. ~O (5)

where Xa are the arc variables, y~ are the flow variables. and Ca is the length of arc a, e is a constant with
value 'PIDI + 2 'P l:teF Qt. Here constraints (2) are the flow balance equations. Constraints (3) and (3')
permit flow on an arc if and only if the arc is selected. Constraints (4) force an arc either to be in the
solution or out of the solution and (5) are the non-negativity constraints on the flow variables. In the
objective function the first term includes the network cost while the second term includes the flow cost
These two terms also include the fixed cost and the flow cost due to the artificial edges. We note that the
length of the appended arcs are such that in any optimal solution to (Fl), for every u eD, exactly one arc
of R, from among the set of arcs between nu and members of Nu• will be chosen, Thus the formulation
ensures that exactly one station is chosen for each department.
Since each artificial arc has a fIXed cost 'P, the totaljixed cost of selected artificial arcs is IDI'I'. The
totalflow cost on the selected artificial arcs is 2'1' rfeF Qt, since each flow uses two artificial arcs. Thus
the total cost contribution of the arcs in an optimal solution to (Fl) is e, and so e is subtracted in (1).
above.
Note that certain additional constraints on the path of a flow can be incorporated in (Fl). For
example. if certain flow should not pass through a set of departments, the corresponding flow variables
can be set to O.
126

2.2 Lamn~an Relaxation

We form a Lagrangian relaxation of (FI) by multiplying (3) and (3') by multipliers v! ' v!, and

adding them to the objective function to get,

. f f f ff ff
(FIR)mmZl(V)= I (Ca-Iv a - Iv.,)x.+ I ( I vaYa+ I va'Ya')
aeE(R) feF. feF.' aeE(R) fe F. fe F.'
f f
+ I Qf ( I CaY a + I Ca'y a, ) - 9 (7)
feF aeE(H) a'eE(H)
Subject to: (2), (4), (5), and (6),
where Fa is the set of flows which can use arc a

This can be simplified to,

min Zl(Vf ) = I (Ca - I v! - I v~,) x.

aeE(R) feF. feF.'

+ I Qc( I (Ca+V~)Y~ + I (Ca'+v~,)y~,) - 9 (8)

feF aeE(H) a'eE(H)
SUbject to: (2), (4), (5), and (6).

The value ofxacan be set based on the sign of its coefficient, i.e., if (Ca -IfeF.V! -IfeFa'V~') is
~ 0, we set Xa to zero, otherwise set Xa to one.The remaining problem decomposes into IF! problems of the

following form - one for each flow f:

f f f f
(FIRf) min I maYa + I ma,ya·
aeE(H) a'eE(H)
Subject to: Myf=bf
1. ~O, y~. ~O
where m! = Ca+v!. FlRf is a shortest path problem which can be computed in 0(IV(H)12) time. Thus,
with a fixed value of the multipliers, (FIR) can be solved in 0(1F1*IV(H)12) time.
Finally, we want to find the multipliers such that Z*l(vf) is maximized to get the tightest lower
bound. This is given by the dual problem:
Z*l = Max f Z*l(V f ).
v
(9)
The multipliers can be adjusted in moving from one iteration to the next following the subgradient
approach (Fisher, 1981). Alternatively, the dual ascent method of Balakrishnan et al, (1989) can be used
to adjust the multipliers.
An upper bound during each Lagrangian iteration (for a fixed value of the multipliers) can be
obtained by taking the network formed by those xa's which are set to 1 and adding to it arcs from
TI(V(C» so that there is a path for each flow. The dual ascent method of Balakrishnan et al, (l~9) can be
used to obtain a lower bound to (FI).
 127

3. Contour Flow

We now consider the topology when the flow is restricted to travel along department contours or
along any other prespecified network. Again one station per department is to be located on the contour of
each depanment We call an ISLAND problem with such topology a C-ISLAND problem. As in section 2,
the total cost which we minimize consists of the cost of the network, which we assume to be linear in the
length of the network, and the flow cost which is proportional to the amount of flow times the distance
traveled.

3.1 Eonnulation

Let D be the set of departments and let C be the contour graph with nodes V(C) and arcs E(C). We
assume that there is one station per department Theorem 2 shows that the integrated problem with contour
flow has a solution in which all the stations are located at nodes of C.

Theorem 2· For the C-ISLAND problem in which each flow takes the shortest path along the contour
graph between the stations of the two departments defining the flow, there exists an optimal solution
where the stations are located at the nodes of C.
fmQt The proof is given in the Appendix.

By the virtue of Theorem 2, we can formulate the problem with only a finite set of points as
candidates for station locations. We now introduce the remainder of our notation. For every flow (u,v) e
F we flnd a shortest path, P~ , in C between node k e Nu and node I e Nv • To keep our notation simple
we assume that there is only one shortest path between any two such nodes of interest. Multiple shortest
paths can be easily incorporated within the framework of our model and solution methodology. Instead of
shortest paths, one can also proceed with any (finite) number of acceptable paths for each flow. Let ~ be
the length of path P~ multiplied by the cost of annual flow per unit length, between departments u and v,
and let Ca denote the (annualized) cost of arc a of E(C). P~~ is a discrete variable which takes on the value
1 if path P~ is chosen for flow between departments u and v, and 0 otherwise. Let ~a = 1 if P~ uses
arc a, 0 otherwise; Yuk = 1 if node k is selected for department u, 0 otherwise; and Xa = 1 if arc a is used, 0
otherwise. Here P~~, Yuk, and Xa are decision variables.
We now give our formulation, denoted as problem (I):

(I) min I. CaXa + I. I. I. s~P~ (10)

aeE(C) (u,v)eF keNu leNv

I. rp~ =1 V (u,v) e F (11)

keNu leNv
uv V (u,v)e F, ke Nu, Ie Nv
Pkl S Yuk (12)
uv
Pkl S Yvl V (u,v)eF, keNu,leN v (13)
128

I Yuk = 1 'I:IueD (14)

keNu
uv uv
PkI Ckla S xa '1:1 (u,v)e F, ke Nu, Ie Nv, ae E(C) (15)

Xa, Yuk, p~ e {O,l} (16).

Constraint (11) ensures that one path between department u and department v is chosen. Constraints (12)
and (13) allow a path to be chosen only if the end nodes of the path are chosen. Constraint (14) forces the
selection of exactly one node for each department Constraint (15) ensures all the arcs which are on a path
P~ are chosen (corresponding x variables are set to one), if the path is chosen.
The following Lemma may help in reducing the number of integer variables in (I), which is helpful
while solving (I) by an integer programming code.

I.emmn 1 : For a relaxed version of problem (I), where each Xa and each ~ is a continuous variable
between zero and one, there exists an optimal solution to the relaxed problem where each Xa and each p~~
is integer.
ftQQf: For each u, let Yuk· be the y variable which has value 1 in the set {Yuk: ke Nu}. For department u
and v, the right hand side (RHS) of (12) is 1 if and only if k = Ie" and the RHS of (13) is 1 if and only if I
= 1°, otherwise the RHS is zero. Hence only P~~lo can be positive and all other pr.;, k'e Nu\ko, Ie Nv\lo
will be zero. In order to satisfy (11), we have P~~lo = 1. Since we are minimizing, and each Ca is positive,
from (15) all { Xa: ~~Io =1} will be 1, and all other xa's can be set to zero. «»
By the above Lemma, the xa's and P~'s can be relaxed to continuous variables between zero and
one. With this relaxation there will be Iue D INul integer variables in (I). The following properties may
help in reducing the problem size:

Property 1: For any (u,v)e F, if all the paths between departments u and v use some arc a, then arc a can
be flXed, i.e. we can set Xa = 1 .

Property 2: For any ue D and k e Nu if all shonest paths, originating at k, to every department v, where
(u,v) e F, pass through some node I e Nu, then node k of department u, along with all paths starting from
node k of department u can be dropped, i.e. set Nu +- Nu\k and P~~ +- 0 '1:1 v and '1:1 I e Nv •

We now show how the C-ISLAND problem can be formulated as a flow network design problem
(without station location). For the pwpose of this alternate formulation, we introduce, in the contour graph
C, a new node us· for each department u. We also introduce arcs (us·,k) '1:1 ke Nu each with a very large
identical cost (length), denoted by '1'. Let C denote this modified contour graph. We concatenate every
path P~~ with two additional links, (us·,k) and (vs·,I). Let the variable P~ now refer to this concatenated
path. Also ~ is set to 1 when a = (us·,k) or (vs·,I), and 0 for all other new arcs. Consider the
formulation.

(I') min 1: Caxa + 1: 1: 1: suvpUv -IDI*'I'

aeE(C,) (u,v)eF lteNuleNv Itl kI
 129

:E :E P~ C! I \f (u,v) e F (17)
keNuleNv
P: ~ S xa \f (u,v)e F, ke Nu,le Nv, a e E(C') (18)

Xa, P: e {O,I}

Lemma 2 . Formulation (n and (I') are equivalent.

fmQt Since the cost of the arcs joining the dummy station uso with any node ke Nu is very large, in the
optimal solution only one of these arcs, say (us°,k°), will be selected. This arc determines the choice of the
station location for department u and corresponds to the variable Yuk0 taking value 1 in <no The cost of each
such arc is 'P, and so after subtracting the term IDI*'P as we do in (1'), the remaining terms in the objective
function of (f) are identical to the objective function of <n.«»

In the above formulation the dummy station takes the role of the given station for a department.
Chhajed (1989) gives a dual ascent procedure to solve the formulation (I') (with no flow cost). We now
look at an alternate approach by forming a Lagrangian relaxation of formulation <no

3.2 Lagrangian Relaxation

If we relax (15) in <n by multiplying each such constraint by ~~ and adding the result to the

objective function, we get problem (IL(u»:

(IL(u» min :E CaXa +:E :E:E s~P:

aeE(C) (u,v)eF keNuleNv
+ :E :E :E :E (P~ ~ - x.) ~
aeE(C)(u,v)eF keNu leNv
Subject to: (12), (13), (14), and (16).

.
=mrn L ( Ca L L L u uv
kla )xa
aeE(C) (u,v)eF keNu leNv
+ L L L (s~+ :E e~ ~)p~r (19)
(u,v)eF keNuleNv aeE(C)
Subject to: (12), (13), (14), and (16).

Since there are no constraints involving x, the value of any Xa can be set based on the modified
coefficient. If (ca - L(u,v)e F :Eke Nu Lie Nv ~ ) < 0, then we set Xa to one, otherwise set Xa to

zero. The remainder of problem (lL(u» reduces to:

(IL'(u» min :E :E:E C!)~ p~r (20)

(u,v)eF keNuleNv
Subject to: (12), (B), (14) and Yuk, pU~ e {O,I},
130

where~ =S:;+ 1: e~ ~ .
aeE(C)

Our interest is in finding the multipliers which solve the following problem:
maxu Z*(ll..'(u».
Problem (ll..'(u» is still a mixed integer programming problem which in the absence of special
structure will be difficult to solve. We next discuss some special cases which are easy to solve.

3.3 Special Cases of the Relaxed Problem

IT we choose one node in Nu for each depanment u and select, from the path set P, those paths
with both their end nodes in the set of nodes chosen, then a feasible solution to (ll..'(u» is obtained. The
value of the objective function of (ll,..'(u» corresponding to this feasible solution is given by summing the
modified path-cost, ~, for all the selected paths. This path-cost depends on the nodes selected for hmh
departments, thus suggesting an alternative formulation of (ll..'(u». Before describing this, we need some
definitions.

Definition: A flow graph G is a graph with a node for each department and an arc (u, v) e E(G) if and
only if (u, v) e F. Figure 3 has an example of a flow graph.

MI

The number
represents the
amount of flow
between pairs of
departments

Figure 3. A Flow Graph

Definition: Given a flow graph G with a set Nv (e.g. a candidate set of input/output points for department
 131

v) associated with each vertex v e V(G). a G-Partite Graph, GO. is formed as follows: Corresponding to
each node v e V(G) we create a super-node in GO which contains INy I new nodes (vt: k = 1• ...•INyl}.
We will denote the node family for department v by av. Two nodes vt and Ul (wu) are connected in GO if
and only if arc (v.u) e E(G). Arc (vt.Ul) in GO is assigned weight~. Thus if (v.u) exists in G. nodes
of node-families u and v fonn a complete bi-partite subgraph of GO.
Given a G-partite graph GO. let S(GO) be an induced subgraph of GO with one node of each node-
family and Z(S(GO» be the sum of the weights on the arcs in S(GO). We now define a problem on the G-
partite graph.

(NSP) Node Selection Problem: Given a graph G and the corresponding G-partite graph. GO. with arc
weights ro. fmd an S(GO) such that Z(S(GO» is minimum. We will denote an optimum solution of NSP by
S*(GO).
Now we show that problem (n...'(u» is reducible to NSP.

Theorem 3. Problem (IL'(u» is reducible to NSP.

frQQt Suppose we have an instance of (IL'(u» with weights~. Create a G-partite graph with a node-
family for each department and consider a fea~ible solution S(GO). We form a solution to (IL'(u» by
setting Yuk = 1 if and only if in S(GO). node Uk is selected from node family au. This will satisfy (14).
Next, set P~ = 1 if and only if both Yuk and yyl are set to 1 thus satisfying (11), (12) and (13). With the
values of the variables set as described, the value of (20) is (1: ~ : both Yuk and Yvl are set to I} =
{r.ro~~ : nodes Uk and VI are selected in S(GO)}=Z(S(GO» .«»

Although NSP is NP-hard, certain special cases of NSP are solvable in polynomial time. These
special cases are characterized by the structure of the flow graph. When the flow graph is a k-tree. NSP
can be solved in polynomial time for a fixed k. A k-tree is recursively defined as: a complete graph on k
nodes (a k-clique) is a k-tree. Given a k-tree and a subgraph of the k-tree which is a k-clique, the graph
obtained by introducing a new venex and connecting it to every vertex of the k-clique is again a k-tree.
Polynomial algorithms to solve NSP when the flow graph is a 2-tree or a k-tree are given in Chhajed and
Lowe (1991) and Fernandez-Baca (1989). Series-parallel graphs (partial2-trees) are particularly
interesting because they are very easy to identify (Wald and Colbourn, 1983). Moreover every planar
graph can be edge partitioned into two outerplanar graphs, which are special series-parallel graphs (Heath,
1991). Another defmition of a series-parallel graph is a graph which can be reduced to a single edge by
repeated use of (i) replace a node of degree 2 and the edges adjacent to it by a single edge, (ii) replace two
parallel edges by a single edge, (iii) eliminate a node with degree one. In the next section, we will restrict
our discussion to series-parallel graphs.

3.4 Heuristic for the General Case

In general the flow graph may not be a k-tree and thus even the relaxed sub-problem (n...'(u» will
be difficult to solve. Nevertheless, the results of the previous section may be exploited in an heuristic way.
132

In the rest of the section we will discuss only series-parallel graphs because of their simplicity and easy
characterization.
Given a flow graph G', we first select a maximal series-parallel subgraph of G', with V(G') =
V(G), and solve the subproblem for this restricted flow graph. Then we preform local improvement to get
a locally optimal solution to the (D..'(u» for flow graph G'. The multipliers are adjusted and the process
can be repeated. Algorithm SOLVE_C-ISLAND has the details.

ALGORI1HM SOLVE_C-ISLAND
Given: A flow graph G', flows, and a block layouL
Step 1: Develop formulation (1)
Step 2: Use Properties 1 and 2 to reduce the problem size
Step 3: Select a 'good' series-parallel graph G as subgraph of G'.
Step 4: Select a set of multipliers for constraints (15) and form problem (D..'(u».
Step 5: Set the values ofx variables and construct a G'-partite graph and G-partite graph. Note that the G-
partite graph will be a subgraph of the G'-partite graph.
Step 6: Solve the (NSP) on the G-partite graph.
Step 7: Using the nodes selected in Step 6, construct a feasible solution to (NSP) in G'-partite graph
which is an induced subgraph of G'-partite graph with one node of each color. Call it J.
Step 8: Pick a department u and for each ke Nu• compute the objective function value of G' which
corresponds to the solution ke Nu and all other nodes in J constant. IT the best solution, over all ke Nu• is
bener than the current solution, retain the new solution and set J to this new solution.
Step 9: Repeat Step 8 until no more improvement occurs.
Step 10: Construct an upper-bound to the original problem (I) (with flow graph G') by selecting all the
arcs in the paths selected in Step 9 and evaluating the objective function of (I) for these paths.
Step 11: IT the current solution is satisfactory then Stop. Else update the multipliers and go to Step 5.

In order to get a solution to (NSP) for a series-parallel graph G close to a solution to (NSP) on
G', it seems reasonable to have as many edges in G as possible (while maintaining G as series-parallel).
Furthennore, we may choose these edges so that the total flow represented by these edges is maximal.
Thus G must be a maximal (in terms of total flow weights) series-parallel graph. A series-parallel graph is
characterized by the absence of a complete graph on 4 nodes. Finding a maximal graph characterized by
the absence of certain sub-graphs is NP-hard (yannakiks, 1979). However we can heuristically select a
'good' series-parallel graph as follows:
Convert the graph G' to a complete graph Q by adding edges with zero flow. Since series-parallel
graphs are sub-graphs of a 2-tree. the maximal series-parallel sub-graph of a complete graph will be a 2-
tree. We generate a maxima12-tree in a manner similar to the recursive definition of a 2-tree. First we
initialize the 2-tree by selecting a triangle with vertices (x,y,z) such that the sum of the flows on arcs in the
set (x,y),(x,z),(y,z») is maximum. This is our initial2-tree, 2T. Now we give the recursive step. Given a
2-tree, 2T, for all edges (x,y)e E(2n, find a vertex s e V(Q)\v(2n such that the sum of the weights on
arcs (x,s),(y,s») is maximum. This vertex s along with the two edges (x,s) and (y,s) are added to 2T and
 133

the recursion continues until all vertices of G are in 2T. For the flow graph in Figure 3, a 'good' 2-tree
generated by this procedure is shown in Figure 4. The zero length edges are not present in the original
graph. This graph represents a flow of 505 which is more than 82% of the total flow of 610.

4. Conclusion

In this paper we have considered the problem of designing detailed layout for which a component
approach is suggested. We concentrated on integrated design of the station locations and the flow network
design. Two topologies were considered and the problems were modeled as integer programs. Lagrangian
based solution procedures were proposed. For one topology, special cases were identified for which each
Lagrangian subproblem is easy to solve.
This paper is an attempt to apply optimization-based techniques to designing a detailed layout For
efficiencies in material handling, all components of the facilities must be carefully coordinated. We hope
that this paper will generate more interest on designing detailed layouts.

MI New edges
with zero flow

Figure 4. A 2-tree subgraph of the Flow Graph in Figure 3

References

Balakrishnan, A., T. L. Magnanti, and R. T. Wong (1989), "A Dual-Ascent Procedure for Large-Scale
Uncapacitated Network Design," Operations Research, 37, 716-740.
Chhajed, D.(1989), "Manufacturing Systems Layout, Station Location, and Flow Network Design," Ph.
D. Thesis, Purdue University, West Lafayette, Indiana.
134

Chhajed, D., B. Monteuil and T. J. Lowe (1991), ''Flow NetwOIk Design for Manufacturing Systems
Layout," to appear in the Eumoean Journal of Operations t ~.
Chhajed, D. and T. J. Lowe (1991), "M-Median and M-Center Problems with Mutual Communication:
Solvable Special Cases," to appear in OJ>erations Research
Femandez-Baca, D. (1989),"Allocating Modules to Processors in a Distributed System," IEEE Trans.
Software En, 15,1427-1436.
Fisher, M L. (1981), "The Lagrangian Relaxation Method for Solving Integer Programs," Manaument
~27, 1-18.
Foulds, L. R.(1983), 'Techniques for Facilities Layout: Deciding which Pairs of Activities Should be
Adjacent," Mana~ement Science, 29, 1414-1426.
Francis, R. L. and J. A. White, "Facilitv Layout and Location: an Analytical Approach," Prentice-Hall,
(1974).
Gaskins, R. J. and J. M. A. Tanchoco (1987). ''Flow Path Design for Automated Guided Vehicle
Systems," International Journal of Production ReSearch. 25. 667-676
Heath, L. S. (1991), "Edge Coloring Planar Graphs with Two Outerplanar Subgraphs," Proc. of the 2nd
ACM-SIAM Symposium on Discrete Algorithms.
Kusiak, A. and S. S. Heragu (1987). "The Facility Layout Problem," EJll'Oj>ean Journal of Operations
Research. 29, 229-251
Montreuil, B. (1987), "Integrating Design of Cell Layout, InputlOutput Station Configuration. and flow
Network of Manufacturing Systems", in C. R. Liu. A. Requicha, and S. Chandrasekar (eds.). Intellj~ent
and Intewted Manufacturin~ Analysis and Synthesis, ASME Ped-Vol. 25, 315-326.
Montreuil. B and H. D. Ratliff (1988). "Optimizing the Location of InputlOutput Stations Within Facilities
Layout", En~neerin~ Costs and Production Economics, 14, 177-187.
Montreuil. Band U. Venkatadri (1988). "From Gross to Net Layouts: An Efficient Design Model",
Document No. 88-56. Operations and Decision Systems Department, Laval University, Quebec. Canada.
O'Brien. C. and S. E. Z. Abdul Barr (1980), "An Interactive Approach to Computer Aided Facility
Layout. " International Journal of Production Research. 18, 201-211.
Tompkins, J. A. and J. A. White, "Facilities Plannin~," John Wiley & Sons, (1984).
Usher, J. S., G. W. Evans, and M. Wilhelm (1988), "AGV Flow Path Design and Load Transfer Point
Location. "fiE Conference Proceedin~, 174-179.
Warnecke, H. J. and W. Dangelmaier (1983), "Progress in Computer Aided Plant Layout," Technical
Repon, Institute for Manufacturing Engineering and Automation - Fraunhofer, Stuttgan, Fed. Rep. of
Germany.
Wald, J. A. and C. J. Colbourn. (1983), "Steiner Trees. Panial2-Trees. and Minimum IFI Networks,"
Networks, 13, 159-167.
Yannakakis, M. (1979). 'The Effect of Connectivity Requirement on the Complexity of Maximum
Subgraph Problem," J of the ACM, 26, 618-630.

Appendix

Proof of Theorem 1.
Consider some arbitrary imbedded network R" which is a feasible solution to the F-ISLAND
problem. Since R" is feasible, for every ue D there exists some point ~ in R" on the boundary of
depanment u where the point ~ represents the station location for depanment u. Also, since R" is feasible.
there exists a path in R" between ~ and $v for every (u,v)eF. For every (u,v)eF,let P(u,v;R") be the set
 135

of edges in a shortest path in R" between ~ and ~ and define R' as the subgraph of R" which is the union
of this collection of shonest paths. Clearly R' is also a feasible solution to the F-ISLAND problem, with
objective function value no larger than that of R".
We define the node-set ofR' as follows: Let S={~: ueD). and let M denote those points (distinct
from S) of R' which have degree 3 or more (note that M is a subset of the nodes of R"). The node-set
V(R') ofR' is taken to be (M)u{S). The edge set E(R') is defined by ((~.$q»). where ~ and $q are
consecutive members of V(R') in some simple path in R'.
Recalling that R' is an imbedded graph. for every edge (~.$q) . if the length of this edge is greater
than d($p.$q) (the rectilinear distance between the embedded points ~ and $q). then replace the edge with
an edge of length d($p,$q)' After replacement of all such edges. we denote the resulting graph by R. with
edge set E(R). Let P(u.v;R) denote the set of edges on a shonest path between ~ and~, (u.v) e F.
Again we assume that each edge of R is in some P(u.v;R). otherwise the edge is removed from R. With all
such edges removed (if any) we redefme the resulting graph to be R.
With Z(.) the objective function value of F-ISLAND, we note by construction that Z(R) ~ Z(R') ~
Z(R"). Funhermore. it is easy to verify that

Z(R) = L d(<Pp,$q) + L Q(u,v) L d(~.$q).

(p,q)e E(R) (u,v)e F (~,~q)e P(u,v;R)

For every edge ($p,$q) e R. if we defme

wpq = (1 + L(Q(u,v): (~,$q) e P(u,v;R). (u.v) eF) .
it follows that
Z(R) (AI)

We note that (AI) expresses Z(R) as a sum of weighted rectilinear distances between pairs of points in the
plane (imbedded nodes of R), where the weights correspond to unit construction costs (=1) along with the
sum of all flows using the corresponding edge in R.
We now construct an optimization problem for which the set of imbedded nodes of R is feasible.
and where Z(R) (via AI) is equal to the objective function value of this optimization problem evaluated at
the embedded nodes.
Since ~ is on the contour of department u for every ue D, then either ~ue N u• or ~u is on some
horizontal or vertical edge between two members, (Xul and <lu2' of Nu. Let D' cD be the set of
department indices where the latter holds, and S' the subset of vertices of S where $ue S' if and only if
ueD' .
Let 9 be a large number, and for each pe MuS. define a variable yp (a location in the plane). Using
the weights (wpq : p,qeMuS) from (AI). defme the following optimization problem (0):

Z·o = miny Z(Y) = L L wpqd(yp,yq) + L L wpu d(yp,yu)+ 9{ L d(Yu,$u) )

pe M qe M pe M ue S ue 5\5'

+9{ L (d(Yu,(XuI) + d(Yu,(XU2) - d«Xul'(XU2»)' (A2)

ueS'
In (0), if (~.$q) «~ ,~» is not an edge of R, then wpq (wpU> is defmed to be zero.
We note that by setting yp~, pe MuS. in (0). Zo( (~) = Z(R). This follows since the last two
136

sums in (A2) are zero, i.e., for ue S'S', d(fu,+u) =0 and for ue S', d(~u,«uI) + d(fu,au2) = d(«UI,<<U2)'
e is large enough so that in an optimal solution to (0), Yu* = fu, ue S'S', and d(yu*,au I ) +
In addition,
d(Yu*,au2) - d(aul'~) =0, ue S', i.e., Yu* will be on the horizontal or vertical line between au l and
«U2'
We also note that (0) is equivalent to a rectilinear multifacility location problem in which the
variables (Y.,. peMuS) are new facilities and the fIXed points, EF = (fu: ue S'S'}u(<<Uj: ue S', i=I,2)
are existing facilities. We now make use of the following lemma from Francis and White (1974):

Lemma AI: There exists a solution to the rectilinear multifacility location problem in which the new
facilities are located at the nodes of the grid graph defined by the existing facilities.«»

Letting TI' be the grid graph defined by the points in EF, we see that TI' is a subgraph ofTI(V(C»
and that there exists an optimal solution y* in (0) where each Yq* is located at a node of TI' (and hence a
node ofTI(V(C))). Furthennore, Zo(Y*) S Zo({~}) = Z(R).
Using the points Y*, it is now easy to connect each pair (Yp* ,yq*) where (~~q)e E(R) using
edges of TI' so that the length of the path between yp* and Yq* is equal to d(yp*,Yq*). The constructed
graph, R*, satisfies Z(R*) = Zo(y*).
In conclusion, we have shown that given any feasible solution, R", to F-ISLAND, there exists
another feasible solution, R*, where Z(R*) S Z(R"), and the nodes and edges of R* are in the grid graph
TI(V(C». Thus an optimal solution to F-ISLAND must satisfy the latter property.
This completes the proof of Theorem 1.«»

Proof of Theorem 2
Suppose we have a solution R in which there is at least one station, say u, which is not in V(C).
Let (b1,h2) be the arc of C on which u is located in R. At least one of the nodes bl or b2 must be in R. If
only one of the nodes bl or h2 is in R, then shifting the location of station u to this node in R will result in
a better solution. Now we consider the case when both bi and h2 are in R. If we move the location of
station u to any point on the arc between nodes bl and hl, only the flow component of the total cost of the
current solution will be affected. We now show that keeping aU other station locations fixed, we can move
u to one of the nodes (bl,h2) without increasing the flow component of the total cost
Let ~ be the length of the arc (bl,hl), and for an arbitrary v such that (u,v) e F, let dbl(v) denote
the shonest distance on R from node bl to station v. Similarly we define db2(V). The shonest distance
from station u (located on arc (bl,b2» to station v on R will be either through node bl or node b2 and is
given by du(v) = min (dbl(v) +x, db2(v) +~-x} where x is the distance between nodes bl and u on the arc
(bl,hl). The cost of flows associated with node u is given by Zu =Iv:(u,v)E F Q(u,v)Ciu(v) where Q(u,v) is the
flow cost per unit length of the flow between stations u and v. We now make the location of station u a
variable, thus making x a variable, and constrain it to be on arc (bl,h2). With x as variable, the distance
function du(v) is a concave function of x since it is minimum of two linear functions. Zu also becomes a
concave function of x since it is the sum of positively weighted concave functions. Moreover, there exists
at least one minimum of a concave function at an extreme point. This implies that at least one minimum of
 137

Zu as a function of x will occur at an end point, i.e. either at x=O or x=j3. This minimum will certainly be
no more than the value of Zu with the current location of u. Thus station u can be relocated to a node
without increasing the flow cost This relocation does not change the network cost and thus the total cost
after relocation is no more than the total cost before relocation of station u.
Such a process can be repeated sequentially for all those stations not located at nodes of C; finally
obtaining a solution satisfying the Theorem with a solution value no larger than the initial solution. This
completes the proof of Theorem 2.0
 Using Optimization Model to Control Workpiece Rigidity and Deformation
in Workholding to Achieve Precision Machining!

Amy J. C. Trappey
Parag Gupta
Department of Industrial and Manufacturing Systems Engineering
Iowa State University
Ames, Iowa SOOII, USA

C. Richard Liu
School of Industrial Engineering
Purdue University
West Lafayette, Indiana 47907, USA

Abstract

The basic criterion in precision machining is to machine a workpiece that satisfies

dimensional accuracies and low tolerance variations. Precise machining demands the work-piece
to be rigidly fixed, which in turn requires high clamping forces. The clamping forces, as well as
cutting forces, result in deformation of the workpiece. This deformation of the work-piece hinders
the fmal goal of the machinist, which is to machine within low dimensional and geometric tolerance
band-widths.
In this paper an analytical, yet practical, nonlinear optimization model is developed which
ensures the rigidity of work-holding and guarantees the precision and accuracy of machining
results by proper fixturing. Principles of statics, kinematics, stress-strain, and geometric
constraints are applied in developing the model. The model consists of constraints in order to (1)
minimize the deformation of the workpiece, (2) compute the fixturing and cutting forces under
which there is no slippage, and (3) ensure the applicability of Coulomb's law of friction. A
computer program is developed to demonstrate the capability of this model. A given workholding
configuration for a specific part geometry is verified when a solution of the model is found.

This research is partially supported by the Society of Manufacturing Engineering Education

Foundation and the Iowa Center for Emerging Manufacturing Technology.
 139

1• Background

Several verification approaches have been developed in the past. A brief literature review
of each approach is described as are some issues that were not addressed and will be discussed in
this paper. First, the kinematic modeling approach is developed by Asada and By (1985). In this
model, the characteristics of the workholding, such as the deterministic position, accessibility,
detachability, and complete restriction, are formulated as geometric constraints. The constraints are
structured in terms of the boundary Jacobian matrix and the fixture displacement vector. Asada
and By solely consider the relationship between the workpiece boundary and the fixture elements,
the possible force impact of the machining process is not included in the workholding analysis.
Further more, due to the need of the boundary Jacobian matrix, the workpiece boundary domain
are restricted to differentiable equations.
Screw theory (Ohwovoriol and Roth, 1981) is applied by Chou, Chandru, and Barash
(1987) to model the fixturing equilibrium problem for a prismatic workpiece. A linear
programming model is used to solve the optimal clamping force magnitudes given fixturing and
cutting wrenches. There are some issues that were not emphasized by Chou et al. First, it is
presumed that all of the unit clamping and locating wrenches are under a friction free condition. In
contrast, the locating wrenches only occur dynamically to counter balance the clamping wrenches
with respect to the clamping force magnitudes and directions as well as the friction coefficient.
However, most of previous researches do not consider the frictional effect. Due to the simplified
fixturing force direction definition, only prismatic workpiece fixturing can be verified.
The naive kinematic approach (Shoham, 1985) is applied by Mani and Wilson (1988) to
construct the constraints and to ensure the validity of the fixture configuration. The counter
clockwise and the clockwise rotational triangles are used to model and prevent the rotational
movement. Although this approach is rather easy to implement in a knowledge based expert

system, the application domain is limited to a 2!-D prismatic workpiece. The friction factor
between the workpiece and the fixture is not considered, i.e., a friction free surface is assumed.
Further, the counter acting concept is not modeled in the analysis so the system may be too
restrictive in some situations. The force magnitude is not defined for each force; therefore, the
allowable fixturing force limits are not considered. In the next section the theory of workholding
verification model is presented. The mechanical and geometrical constraints of the workholding
configuration are discussed in detail.

2. Theory of Workholding Verification

This research treats the analytical model as the combination of several constraints which
cannot be violated by any workholding element. These mechanical and geometrical constraints are
the theoretical bases of the workholding verification.
140

2.1. Equilibrium

The effective force and moment resultants that occur on the workpiece have to be zero
(Malvern, 1976) to ensure the stability of the workpiece. A free body diagram is applied to
represent all possible forces acting on the workpiece from time to time. Either during the
workholding process along (without any effective outside force) or during the machining process
(with outside effective force), it is necessary to maintain the eqUilibrium of the 3-D noncoplanar,
nonconcurrent, and nonparallel force system. The three force equilibrium equations with respect to
X, Y, and Z directions are

(1)

and the three moment equilibrium equations about X, Y, and Z axes are

LMx =O.LMy=O. LMz=O (2)

can be applied to verify the workpiece stability for the most general case. If the equilibrium system
is a parallel force system. e.g.• along the Z axis. only forces in the Z direction and moments about
the X and Y axes need to be considered:

(3)

If the equilibrium system is a 2-D coplanar force system. e.g.• on the Xp-YP plane (the fixturing
plane of the base plate in fixture design). only forces in the X and Y directions and moments about
Z axis need to be considered (Higdon and Stiles. 1968):

(4)

2.2. Force Directions and Limits

For the completeness of the system. the geometrical constraint for force directions has to be
modeled as a verification condition. Assuming the workholding operation is a grasping (not
magnetic pulling) action. the force vector-(Fx.i • Fy.i • Fz.i ) has to be directed toward the
workpiece. For instance. the workpiece boundary face has a surface normal with direction cosines
Ix. my. and nz. Then the force vector constraint is

(5)
 141

This constraint is even applicable to the machining force direction, when the machining process
creates the force impact toward the workpiece.

2.3. Frictional Force vs. Normal Force

According to the coefficient of the static friction (J.1.) between the workpiece surface and the
contacting holding element, the magnitude of the frictional force element (IF~) has to be less than
the fractional normal force magnitude, i.e., J.1.1 Fn I (Malvern, 1976). Therefore, if there is no (or
little) friction between the workpiece and the holding element, the fixturing force direction is nearly
perpendicular to the boundary face (this is implicitly assumed by most of the previous research as
shown in the examples of the previous section). If friction exists as it does in most of machining
cases, the fixturing force has to be analyzed as the components of the force, the frictional force and
the normal force. The proportional constraint of the frictional force and the normal force is
formulated regarding the frictional coefficient. The fixturing force (F) acting on a workpiece at any
point can be divided into normal force component (Po) and frictional force component (FC), i.e.,
I Fj2 =I~l + IFR. Fo is nothing but a dot product of fixturing force and unit surface normal vector
(N) or Fn = IF·NI. The ratio of magnitudes of the frictional force (IFf!) and normal force (IFni)
should be less than the coefficient of friction to avoid slippage. This can be expressed
mathematically as

J.1.<
(6)

Further, the system allows for frictional (J.1. >0) as well as frictionless (J.1. =0) conditions.

2.4. Center of Gravity

The weight of a workpiece has to be considered as one of the force elements in the
workholding force system. Vertical supports can be applied to hold the workpiece against the
gravitational attraction of the earth (Hoffman, 1985) and create the counter acting forces to balance
the weight of the workpiece. The vertical supports (usually three supports) have to be located in
such a way that the center of gravity of a workpiece must be inside the supporting polygonal region
defined by the vertical supporting points.

2.5. Stress and Strain in Workpiece

It is essential to know the stress and strain developed in the workpiece due to fixturing and
cutting forces in order to know the deformation in the workpiece. If Fx.j, Fy.i, and Fz.i are the
force components in the x, y, and z directions and Ax, Ay, and Az are the normalized equivalent
142

areas in the x, y, and z directions, by using the spatial occupancy enumeration approach the
workpiece can be discretely represented and, then, the normalized equivalent areas in x, y, and z
directions can be computed as follows.

r2 L L (Xmax.jk - Xmin.jk)
Ar..= 'ltj 'ltk
(Xmax - Xmin) (7)

r2 L L (Ymax.jk - Ymin.jk)
At= 'ltj'ltk
(8)
(Ymax - Ymin)

r2 L L (Zmax.jk - Zrnin.jk)
At= 'ltj'ltk
(Zrnax - Zmin) (9)

where xmax.jk and Xmin.jk are the maximum and minimum x values at cell address jk while Xmax
and Xmin are the overall maximum and minimum x values of the workpiece. The Yand z notations
are interpreted the same way as the x's. Thus, the compression stresses in the x, Y, and z
directions are derived as

=Fx.i cr .=Fy.i and"--.·=I'z.i

r<. ..
"X.I Ar..' y.l At' ~L..l At (10)

Ifv is the Poisson's ratio and E is the elastic modulus, the strains in the x, y, and z directions are

n n n
L O"x.i L Oy.i L Oz.i
~=_i=I +vi=I +v.o.;i=~1_ (11)
E E E
n n
L Oy.i L Oz.i
i-I +v-'----'i-"---'I~_ (12)
E E

(13)

3• Nonlinear Optimization Model

Following the discussion of fixturing criteria above, we can generally represent the
flXturing procedure in a nonlinear optimization model. The fixturing configuration, the allowable
 143

maximum holding force limits, and the machining force are given through a user interface. If a
configuration is valid, the nonlinear model computes the unknown fixturing forces that produce
minimum workpiece deformation and ensure workholding rigidity. The model is summarized as
follows.

o The objective function is to minimize the deformation (described in Section 2.5) caused by all
the forces, which is mathematically expressed as

Min. {[Ex (Xmax - "min)f + [fy (Ymax - Ymin)f + [ez (7max - Zmin)f}1/2 (14)

o This objective function is subjected to the following constraints:

(a) Zero resultants of total forces and moments as discussed in Section 2.1.

({; X 3' F3 X 1 = A(j X 1 (15)

(b) A force can be represented in three force elements with respect to the X, Y, and Z axes.
Thus, the sum of the squares of X, Y, and Z force elements is equivalent to the square
of the force magnitude. In order to ensure that there is no yielding due to localized
stress, there should be a limitation on the maximum amount of force acting on a locator
or a clamp. No yielding in turn would ensure the applicability of Coulomb's law of
friction. If Fmax.i is the maximum force which is allowed then the constraint is:

FX.I.2 + Fy.1.2 + FZ.I.2 < Fmax I.2 (16)

(c) The force directions are constrained toward the workpiece according to the assumption
defined in Section 2.2.

(17)

(d) The frictional force magnitude (IFfjl) is limited by the fractional normal force
magnitude (1.1. IFnjl). (Ff.i and Fn.i are functions of Fx.i, Fy.i. and Fz.iJ

IFil2 -IFi .N il2

IFnP (18)

The matrices in constraint (a) (Eqn. 15) can be expressed as

144

1 0 0 0
0 0 0
0 0 0 0 1
C 6X3 =
0 0 -Zl -Zn Yn
Zl Zn 0 0 -X n
-y 1 -Yn Xl Xn 0

F3X} =[Fx.} .. Fx.n Fy.} .. Fy.n Fz .} .. Fz.n ]

A 6Xl = [-fex -fey -fe (~feY'Ycfcz) (Xcfcz-~cfcx) (ycfcx-xcfcy)]'f

Equation (15) represents three force equilibrium equations and three moment equilibrium
equations. In the next section, a numerical example is discussed to demonstrate the use of this
model for a fixture design verification purpose.

4• Numerical Example

The quadratic verification model described in the previous section is coded by using a
nonlinear programming package GINO. A numerical example (shown in Figure 1) is given for a
workpiece configured with three vertical supports, three horizontal locators, one vertical clamp,
and one horizontal clamp for fixturing. The fixture configuration data are listed in Table 1.
The friction coefficient for the contact of the workpiece and the fixture element is 1/3
(Higdon and Stiles, 1968). The maximum allowable locating force which a locator can take is
conservatively set to be 500 pounds (lbs.), and the maximum allowable clamping force which a
clamp can take is set to be 350 lbs. (Colbert, 1985). The workpiece weighs 30 lbs. The center of
weight of the workpiece is at the point (3.5, 3.5, 2.5). The maximum cutting force is at (4, 7, 5),
and the maximum cutting force elements are (-200 lbs., -200 lbs., -200 lbs.) with respect to the X,
Y, and Z directions.
The model is employed to check the feasibility of the fixture design. The complete results
are shown in Table 2. These forces ensure the rigidity of the workpiece and in addition under the
given constraints produce a minimum deformation (as defined by the objective function) of 22
micro-inches.
The complete resultant of the force system acting on the workpiece is zero. The results
demonstrate that the workpiece is completely restricted, the fixture elements are assigned to the
right locations, and the allowable force magnitudes and directions are verifiably counter-acted.
 145

Cutter

F 1, F2, F3: vertical locators (supports). Xp

F4, F5, F6: horizontal locators (stoppers).
F7: vertical clamp (e.g., swing clamp).
F8: horizontal clamp (e.g., side clamp).
Fi's tixturing force is represented by the basic force elements
(Fx.i, Fy.i, Fz.i) with respect to the Xp, Yp, and Zp directions.

Figure 1. 3-D fixturing illustration of the numerical example.

Table 1. Example of fixture configuration data.

Fixture Type of Fixture Element (x,y,z)

f1 vertical locator (1,1,0)
12 vertical locator (6,1,0)
f3 vertical locator (3,6,0)
f4 horizontal locator (1,0,1)
f5 horizontal locator (6,0,1)
f6 horizontal locator (0,4,1)
f7 vertical clamp (4,1,5)
f8 horizontal clamp (6,6,1)
146

Table 2. Results of the example model.

Fixture Computed Forces (Fx, Fy, Fz)

f1 ( -1.94, 19.62, 75.88 )
f2 (-2.99, -0.92, 9.39 )
f3 (74.08, 23.66, 393.62 )
f4 (-7.17, 167.44, -53.26)
f5 (-3.04, 174.91, -48.09 )
f6 (213.3, 16.43, 69.05 )
f7 (30.127, -6.84 ,-242.31 )
f8 (-102.35, -194.31, 25.71 )

5. Conclusion

This research establishes a verification procedure that can be applied with the computer-
aided workholding interface to check the correctness of the workholding configuration in places
such as fixture design or robot hand grasping. Moreover, it can minimize deformation to meet low
tolerance band-width required for precision parts. The model can be further embedded into the
automatic fixture design system (Trappey and Liu, 1990) and used as an analytical tool to
determine the proper fixture positions. The direction of future research is to develop a complete
automatic workholding design system utilizing heuristic decision rules as well as the mechanical
and geometrical knowledge developed in this paper.

6. References

Asada, H., and By, A.B., 1985, "Kinematics of workpart fixturing," IEEE Con! on Robotics and
Automation, pp. 337-345.
Chou, Y.C., Chandru, V., and Barash, M.M., 1987, "A mathematical approach to automatic
design of fixtures: Analysis and synthesis," Proceedings, ASME Winter Annual Meeting, Boston,
Mass., pp. 11-27.

Colbert, J.L., (Major Prof: DeVries, W.), 1985, "Workholding technology in flexible
manufacturing and the design of a modular fixturing system," M .S. Thesis, Rensselaer
Polytechnic Institute, Troy, NY.

Higdon, A., and Stiles, W.B., 1968, Engineering Mechanics, 3rd Edition, Prentice-Hall, Inc.,
Englewood Cliffs, New Jersey.

Hoffman, E.G., 1985, Jig and Fixture Design, Second Edition, Delmar Publishers Inc., Albany,
NY.
 147

Malvern, L.E., 1976, Engineering Mechanics, Volume I: Statics, Prentice-Hall, Inc., Englewood
Cliffs, New Jersey.

Mani, M. , and Wilson, W.R.D., 1988, "Automated design of workholding fixtures using
kinematic constraint synthesis," Proceedings, 19th North American Mfg. Research Conference,
SME, University of Illinois, Urbana, Illinois, pp. 437-444.

Ohwovoriol, M.S., and Roth, B., 1981, "An extension of screw theory," Journal of Mechanical
Design, Vol. 103, pp. 725-735.
Shoham, Y., 1985, "Naive kinematics," Proceedings, International Joint Conference on Artificial
Intelligence, pp. 436-442.
Trappey, J.C., and Liu, C.R., 1990, "Automatic generation of configuration for fixturing an
arbitrary workpiece using projective spatial occupancy enumeration approach," Proceedings,
Symposium on Advances in Integrated Product Design & Manufacturing, ASME Winter Annual
Meeting, Dallas, TX, November 25-30, pp. 191-202.

7. Appendix

Example Model Formulated in GINO:

MODEL:
1) FXl + FX2 + FX3 + FX4 + FX5 + FX6 + FX7 + FX8 - 200 = 0;
2) FYI + FY2 +FY3 +FY4 +FY5 + FY6 + FY7 +FY8 - 200=0;
3) FZl + FZ2 + FZ3 + FZ4 + FZ5 + FZ6 + FZ7 + FZ8 - 200 - 30 = 0 ;
4) - FY4 - FY5 - FY6 - 5*FY7 - FY8 + FZl+ FZ2 + 6*FZ3 + 4*FZ6 + FZ7 + 6*FZ8
- 5*200 - 7*200 - 3.5 *30 = 0;
5) FX4 + FX5 + FX6 + 5*FX7 + FX8 - FZl - 6*FZ2 - 3*FZ3 - FZ4 - 6*FZ5 -
4*FZ7 - 6*FZ8 - 5*200 + 4*200 + 3.5*30 =0 ;
6) - FXl -FX2 - 6*FX3 -4*FX6 - FX7 -6*FX8 + FYI + 6*FY2 + 3*FY3 + FY4 +
6*FY5 + 4*FY7 + 6*FY8 + 7*200 - 4*200 = 0 ;
7) - FZl <0;
8) -FZ2<0 ;
9) -FZ3<0;
10) -FY4<0;
11) -FY5 <0;
12) - FX6<0;
13) FZ7 <0;
14) 0.707*FX8 + 0.707*FY8 < 0 ;
15) (FXl A 2 + FYI A 2 ) / ( FZl A 2 ) < 1/9 ;
16) ( FX2 A 2 + FY2 A 2 ) / ( FZ2 A 2 ) < 1/9 ;
17) ( FX3 A 2 + FY3 A 2 ) / ( FZ3 A 2 ) < 1/9 ;
18) ( FX4 A 2 + FZ4 A 2 ) / ( FY 4 A 2 ) < 1/9 ;
19) ( FX5 A 2 + FZ5 A 2 ) / ( FY5 A 2 ) < 1/9 ;
148

20) (FY61\ 2 + FZ61\ 2) 1 (FX61\ 2) < 1/9 ;

21) ( FX7 1\ 2 + FY7 1\ 2 ) 1 ( FZ7 1\ 2 ) < 1/9 ;
22) (FX8 1\ 2 + FY8 1\ 2 + 2* (FZ8 1\ 2) - 2*FX8*FY8 ) 1 (FX8 1\ 2 + FY8 1\ 2 +
2*FX8*FY8 ) < 1/9 ;
23) FXll\ 2 + FYI 1\ 2 + FZll\ 2 < 250000 ;
24) FX2 1\ 2 + FY2 1\ 2 + FZ2 1\ 2 < 250000 ;
25) FX3 1\ 2 + FY3 1\ 2 + FZ3 1\ 2 < 250000 ;
26) FX41\ 2 + FY41\ 2 + FZ41\ 2 < 250000;
27) FX5 1\ 2 + FY5 1\ 2 + FZ5 1\ 2 < 250000 ;
28) FX61\ 2 + FY6 1\ 2 + FZ6 1\ 2 < 250000 ;
29) FX7 1\ 2 + FY7 1\ 2 + FZ7 1\ 2 < 122500 ;
30) FX81\ 2 + FY8 1\ 2 + FZ8 1\ 2 < 122500;
31) XTERM = ABS( FXl ) + ABS( FX2 ) + ABS( FX3 ) + ABS( FX4) + ABS(
FX5 ) + ABS( FX6) + ABS( FX7 ) + ABS( FX8) + 200;
32) YTERM = ABS( FYI) + ABS( FY2) + ABS( FY3 ) + ABS( FY4 ) + ABS( FY5 )
+ ABS( FY6) + ABS( FY7 ) + ABS( FY8) + 200;
33) ZTERM = ABS( FZl ) + ABS( FZ2 ) + ABS( FZ3 ) + ABS( FZ4 ) + ABS( FZ5 )
+ ABS( FZ6) + ABS( FZ7 ) + ABS( FZ8) + 230;
34) MIN = ABS( (l/( 10000000 1\ 2) )*(49*(-( XTERM 133.57) + 0.05*( YTERM /
33.57 ) + 0.05*( ZTERM / 47 ) ) 1\ 2 + 49*( 0.05 * ( XTERM /33.57) - ( YTERM
/33.57) + 0.05*( ZTERM 147) ) 1\ 2 + 25*( 0.05 * ( XTERM / 33.57 ) + 0.05 * (
YTERM / 33.57) - ( ZTERM 147) ) 1\ 2 ) ) 1\ 0.5 ;
END

Example Output from GINO:

SOLUTION STATUS: OPTIMAL TO TOLERANCES. DUAL CONDmONS: SATISFIED.

OBJECI1VE FUNCflON VALUE

34) .000022

VARIABLE VALUE REDUCED COST

FXl -1.940474 0.000000
FX2 -2.994801 .000000
FX3 74.080698 .000000
FX4 -7.171431 0.000000
FX5 -3.048294 0.000000
FX6 213.302969 0.000000
FX7 30.127334 0.000000
FX8 -102.356001 .000000
 149

FYI 19.623131 0.000000

FY2 -.921439 0.000000
FY3 23.664606 0.000000
FY4 167.441511 0.000000
FY5 174.916865 0.000000
FY6 16.437538 .000000
FY7 -6.847188 0.000000
FY8 -194.315026 .000000
FZl 75.882110 0.000000
FZ2 9.397702 0.000000
FZ3 393.626052 0.000000
FZ4 -53.268695 0.000000
FZ5 -48.094998 .000000
FZ6 69.053000 .000000
FZ7 -242.313570 0.000000
FZ8 25.718399 .000000
XTERM 635.022002 .000000
YIERM 804.167304 .000000
ZTERM 1147.354527 .000000

ROW SLACK OR SURPLUS PRICE

1) -.000000 0.000000
2) o. ()()()()()() 0.000000
3) o. ()()()()()() 0.000000
4) O. ()()()()()() 0.000000
5) o. ()()()()()() 0.000000
6) 0.000000 0.000000
7) 75.882110 .000000
8) 9.397702 .000000
9) 393.626052 .000000
10) 167.441511 .000000
11) 174.916865 .000000
12) 213.302969 .000000
13) 242.313570 .000000
14) 209.746421 .000000
15) .043583 .000000
16) -.000056 0.000000
17) .072077 .000000
18) .008068 .000000
150

19) .035205 .000000

20) .000370 .000000
21) .094854 .000000
22) 0.000000 .000007
23) 243853.072698 .000000
24) 249901.865310 .000000
25) 89010.567631 .000000
26) 219074.357061 .000000
27) 217081.669340 .000000
28) 199463.333984 .000000
29) 62829.593453 .000000
30) 73603.483869 .000000
31) -.000000 0.000000
32) -.000000 0.000000
33) -.000000 0.000000
 A New Approach to Determine the Pose
and Dimension of Polygonal Shapes

by

Jose A. Ventura and Jen-Ming Chen

Dept. of Industrial and Mgmt. Syst. Engineering
The Pennsylvania State University
University Park, PA 16802, USA

Abstract

Machine vision has the potential to significantly impact both quality and
productivity in computer integrated manufacturing, due to its versatili-
ty, flexibility and relative speed. Unfortunately, algorithmic development
has not kept pace with advances in vision hardware technology, particu-
larly in the areas of analysis and decision making. The specific subject
of this investigation is the development of machine vision algorithms for
the dimensional checking, pose estimation, and overall shape verification
for general polygonal objects. Algorithmically, the image boundary da-
ta is partitioned into n segments, and then the best-fitting polygon is
found using a non-ordinary least-squares technique. The algorithm is
well-suited for on-line implementation in an automated environment due
to its flexibility and demonstrated speed.

1. Introduction

Machine vision has the potential to effect significant improvements with

respect to both quality and productivity in automated manufacture. This
is due in large part to the versatility of machine vision, which can perform
functions such as guidance and control, product identification, inspection
and measurement. The successful implementation of a computer integrat-
ed manufacturing facility, for example, would not be possible without the
reliable real-time performance of these activities. Historically, inspection
activities have been performed by human inspectors. However, it is well
documented that human inspectors are unreliable, even when multiple
152

inspectors are employed to check each others work (Schaffer 1984) . Ma-
chine vision is often preferable to comparable automated contact systems
(e.g., a coordinate measuring machine), due to factors such as its greater
relative speed and inherent nondestructive nature. In addition, an in-
tegrated machine vision system not only has the capability to identify
a part and subsequently qualify (or classify) it, but can provide timely
process feedback as well.

Machine vision systems are well-suited for data acquisition from prod-
ucts with complex and diverse geometric features, as they are less re-
stricted (than are contact measuring systems) by the configuration of
the part. However, the analysis of these complicated shapes and subse-
quent decision processes are not straightforward. There are two general
methodologies for shape analysis and decision making: referential and
non-referential (design-rule) methods. Essentially, referential methods
"match" the features of an item's captured image with that of an ideal
stored image; a common example is the well-known template matching
method. Non-referential methods, on the other hand, employ a paramet-
ric approach; that is, certain parameters are estimated from the image
and compared with ideal values. These parameters are selected such that
they completely describe the geometric properties (of the ideal shape)
and their spatial relationships.

The subject of this investigation is the dimensional checking, pose estima-

tion, and qualification of objects (esp. production parts) with polygonal
profiles. Polygonal shapes, in general, warrant special attention due to
the number of products with this particular profile, including keyboard
cut-outs, sheet-metal formed cases and boxes, integrated circuit chips
and capacitors, fastener heads, surface mounted electronic components,
printed circuit boards, microfilm cards, and labels.

The existing approaches regarding the polygonal shape representation,

recognition and analysis are numerous. Among them, the vertex-based
approach, or point pattern matching, is one of the most commonly used
methods (e.g., Atallah et al. 1991, Cox et al. 1989, Han et al. 1989,
Koch and Kashyap 1989, etc.) in which the boundary of a polygon is
represented by a set of cyclic ordered list of vertices, expressing each ver-
tex as a coordinate pair. The major difference between the vertex-based
method and the proposed edge-based method is that the latter approach
 153

utilizes all the extracted edge points for the shape matching, instead of
using only the vertex points; that is, it preserves all the available infor-
mation from the input image. Hence, the edge-based shape model is to
describe boundary contours by a set of cyclic ordered edges, expressing
each edge by a set of correlated parameters. Using the edge-based model,
the subsequent contour fitting procedure has two significant advantages:
1) it is more precise and yields a global optimal solution with respect to
the available contour pixels, and 2) it can provide more functions for the
image analysis, such as dimensional checking, pose determination, and
detection of the edge flaws.

In current polygonal shape analysis, a distance function must be defined

to compute the similarity between the input shape and the reference
shape. For example, Atallah et al. (1991) used the Hausdorff distance
function:

DH(P, Q) = DH(Q, P)
= max{ d(p, Q), d(q, P) I pEP, and q E Q}, (1)

i.e., the maximum Euclidean distance from any point of a polygon (P

or Q) to the other polygon (Q or P), where P and Q are the reference
shape and input shape, respectively. Cox et al. (1989) used the sum of
squares of the Euclidean distances from each vertex of each polygon to
the convex hull of the other polygon:

Dc(P,Q) = Dc(Q,P)
L d2 (p,Q) + L d2 (q,P), (2)
pEV(P) qEV(Q)

where V(P) (resp. V(Q)) denotes the set of vertices of polygon P (resp.
Q). Koch and Kashyap (1989) used the mean of squares of the Euclidean
distances from each vertex of a polygon (fragments) to the corresponding
vertex of the other polygon (fragments):

DK(P,Q) = DK(Q,P)

= !:.
n
t
i=l
{(Xi(P) - Xi(Q))2 + (Yi(P) - Yi(Q))2}, (3)
154

where n is the total number of vertices of P (resp. Q), and (Xi(P), Yi(P))
are the coordinates of the ith vertex of P (resp. Q). Many other distance
functions have been developed for the analysis of polygonal shapes, such
as the L2 distance between the turning functions of the two polygons
(Arkin et al. 1991), the sum of squares of the distances between each
pair of corresponding points along the polygonal arcs (Kamgar-parsi et
al. 1991), etc.

In this paper, a least-squares contour fitting procedure will be developed

herein for the analysis of polygonal objects. Although this algorithm
will be restricted to specific shapes, it can be preferable to other non-
referential techniques with respect to reliability and speed. In a related
study, Ventura and Dung (1991) have developed a global optimization
procedure for analyzing rectangular and square objects. The procedure
proposed here is a generalization of their approach to general polygons.

The remainder of the paper will be organized as follows. In Sections 2

and 3, data segmentation and shape analysis procedures will be detailed.
Experimental studies supporting the model's validity will be presented in
Section 4. Issues related to the speed and accuracy of this algorithm are
also discussed in this section. Finally, concluding remarks and directions
for further study are summarized in Section 5.

2. Segmentation

Image segmentation is fundamental to the success of subsequent shape

analysis. For dimensional inspection, the boundary contour of the object
to be inspected is detected, extracted, and then split into segments. For
a fixed number of segments, n, the objective is to minimize the error
norm of the boundary data with respect to the piecewise approximation
segments. The boundary data

(4)

is separated into n data subsets

 155

in which
(6)
(7)
and
Sk n Sl = 0, \;f (k, I), k =1= l. (8)
In Eqs. (4-5), m is the total number of boundary data points, and mi is
the number of points in Si-the subset of boundary data approximated
by the ith segment. Eqs. (6-7) indicate that the union of all the subsets
of data is equal to the set of all boundary data, and Eq. (8) shows that
the subsets are mutually exclusive.

The segmentation algorithm presented below is a hybrid of a boundary

split-and-merge method and an end-point adjustment procedure. The
error norm used is based on the integral square error.

Segmentation Algorithm (Ventura and Chen 1991):

Let E z (a, b) denote the integral square error of a segment with end-points
a and b, and let Eoo( a, b) represent the maximum deviation between the
data and the segment which connects the two end-points of the data set.

Step 1. The initial segments are obtained by equally dividing the data
set S into n subsets Si, i = 1,2"", n. Let the first point of
Si, denoted as Ui, i = 1,2"", n, represent the break point of
segment i.

Step 2. For i = 1,2"", n, compute EZ(Ui, Ui+l - 1). Choose the seg-
ment with the maximum Ez among all segments; this segment
will be designated with a "k". Let v represent the point having
maximum deviation error Eoo(Uk,Uk+l -1).

Step 3. Let el = (EZ(Uk-l,Uk -1) +(EZ(Uk,Uk+l-1)) - (EZ(Uk-l,V-

1) + Ez(v, Uk+l - 1)), and let ez = (E2(Uk, Uk+l - 1) +
(E Z(Uk+l,Uk+2 - 1)) - (EZ(Uk,V - 1) + E 2(v,Uk+2 - 1)). If
max{ el, ez} ~ et, a preset small quantity, go to step 4. If
el ~ ez, split the segment k at v and merge the two segments
adjacent to Uk (i.e., substitute v for Uk); go to step 2. If el < e2,
156

split the segment k at v and merge the two segments adjacen.t

to Uk+l (i.e., substitute v for uk+d; go to step 2.

Step 4. Perform the end-point adjustment procedure sequentially for

each segment i, i = 1,2"", n. Given i, the end-point ui IS
adjusted by the following procedure:

4a. Let e~ = (E2 (Ui-l, Ui - 1) + E 2 (Ui, Ui+l - 1)), e~ -

(E2(Ui-l,Ui - 2) + E 2(Ui - 1,ui+1 - 1)), and e~
(E2(Ui-l,Ui) + E 2(Ui + 1,Ui+l -1)).

4b. If ma.x{ e~ - e~, e~ - e~} ~ 0, stop; else if e~ ~ e~, let

Ui = Ui + 1, and go to step 4a; else let Ui = Ui -1, and
go to step 4a.

3. Shape Analysis

General shape analysis consists of two steps: 1) shape modeling, and 2)

matching algorithm. Shape modeling is to represent a planar object in
terms of a set of primitives. Then, the matching procedure is to specify
the optimal transformation and overall least squares error between the
segmented scene data and the model.

3.1. Polygonal Model

An arbitrary polygon can be defined by n straight lines, where n is the

number of edges. Naturally, the edge segments are the most prominent
primitives for the edge-based shape representation. An equation exists
for each edge:

Li : X cos8i +Y sin8i = di , i = 1,2"", n, (9)

which is characterized by the ith normal angle (8 i ) between the normal

line to the edge Li and the X-axis, and the ith normal distance (d i ) be-
tween edge Li and the origin. Hence, there are 2n parameters associated
with this system of n linear equations. Each pair of parameters, however,
cannot be independent from each other. They must satisfy the object's
spatial and angular relationships among the edges. Therefore, we want
 157

to develop a parametric description for each edge primitive, so that the

geometric relationships are encoded, and the overall shape can be repre-
sented in terms of the dimension and pose features of the polygon. The
shape modelling starts with the a priori knowledge. First, let the dimen-
sion of the segment Li be ii, i = 1,2, ... ,n. Then, the a priori knowledge
for a given polygonal shape can be represented by:

"'i: dimensional scale of Li with respect to Ll (i.e, idid, and

ai: internal angle of the polygon between L i - 1 and L i ,

where "'i represents the spatial relationship between edges Li and L 1 ,

and ai represents the angular relationship between two consecutive edges
arranged in counterclockwise direction. The dimensional relationship be-
tween the edges is fixed by

(10)

In addition, let /3i be the turning angle between L i - 1 and L i , which is

determined by
/3i = 7r - ai, i = 1,2,· .. , n. (11)
It is worth noting that the turning angle /31: is positive if the edge Li turns
counterclockwise; otherwise it is negative. The sum of the turning angles
for a closed boundary contour is 27r.

Fundamental to shape modelling and representation is the selection of

the characteristic parameters; that is, the parameters that will be used
as model variables so that the geometric configuration of the shape can
be completely determined. Four parameters are required to characterize
a general planar polygon: 1) the dimension is represented by iI, 2) the
orientation is defined by ()1, and 3) the object position is represented by
the intersection of Ln and L l , denoted by (Xl, Yl). Using the underlying
four parameters, the modelling procedure is divided into two major steps:
Initially, we set the pose parameters (Xl, Yl) = (0,0), and ()l = 0, so that a
simple preliminary model can be derived based on the information about
the shape, "'i and ai, i = 1, 2, ... , n, and the scale parameter II. Finally,
the pose parameters (()l, Xl, Yd will eventually be encoded into the model
through an affine transformation.
158

In the first step, since the first vertex of the polygon is set to the origin,
and the orientation angle is zero, the coordinates of the remaining vertices
of the polygon, (Xi, Yi), i = 2,3"", n, are expressed as a function of
II and (Ki, ai), for i = 1,2"", n, as summarized in Lemma 1. The
prehIninary model is then derived in Lemma 2 based on the equations of
the vertices. In the second step, the shape model described in terms of
the underlying parameters is obtained and summarized in Theorem 1.

Lemma 1 (Ventura and Chen 1992). Given that (XllYl) = (0,0) and
81 = 0, the coordinates of each vertex of the polygon, (Xi, Yi), 2 :s: i :s: n,
can be obtained as follows:
i-I
xi=IlLKjCOSPj, i=2,3,···,n, (12)
j=l

and
i-I
Yi = II LKj sinpj, i = 2,3, ··· ,n, (13)
j=l

where
i
Pi =L f3j + 7r /2, i = 2,3, .. . , n, (14)
j=2
and
PI = 7r/2, (15)
in which Pi is the inclined angle of Li with respect to the X-axis.

Proof. The proof is straightforward. Since we know the inclined angle

(Pi) and length (Ii) of Li, for i = 1,2"", n, the coordinates of each
vertex (Xi, Yi) can be computed from the coordinates of the previous
vertex (Xi-l,Yi-l), and Ii and Pi:

Xi = Xi-l + (Ki-l1d cos Pi-l

= Xl + KIll COSPI + K2ll COSP2 + ... + Ki-lll cos Pi-l
i-I
= II L Kj cos Pj, i = 2,3, ... , n,
j=l
 159

and

Yi = Yi-l + (Ki-l 1l) sin Pi-l

= Yl + Klh sin PI + K211 sinp2 + ... + Ki-l 1l sin Pi-l
i-I
= II 2:= Kj sin Pj, i = 2,3, ... , n.
j=1

Since Pj is a function of i3i, and i3i is a function of ai, the derived vertices
(Xi, Yi), i = 2,3" . " n, are functions of 11 and (Ki' ai), i = 1,2", " n. 0

The next step of the modelling process is to derive the normal distance
and normal angle for each edge using the information about the vertices.

Lemma 2 (Ventura and Chen 1992). Given (Xi, Yi), i = 1,2·· " n, the
preliminary model of the polygon can be represented by the following
system of equations:

Li : X cos e~ +Y sin e~ = d~, i = 1,2" . " n, (16)

where
2 ",i-I. . ",i-l
I -1 cos PiL.;j=I K jsmpj - smpicosPiL.;j=IKjCOSPj (17)
0i = tan . 2 ",i-l . ",i-I.'
sm Pi L.;j=l Kj COSPj - smpi COSPi L.;j=1 Kj smpj
d~ = Ci1l, (18)

and
i-I i-I
Ci = [( sin 2Pi L Kj COSPj - sin Pi cos Pi L Kj sinpj)2
j=l j=l
1/2
+ ( cos 2Pi L L
i-I i-I
Kj sin Pj - sin Pi cos Pi Kj cos pj) 2] .(19)
j=1 j=1

Proof. Given the two end points of a segment L i , (Xi, Yi) and
(Xi+ 1, Yi+ I), the intersection between the segment and its normal line
160

from the origin, denoted by (X~, yi), can be obtained by

I Yi - Xi(Yi+l - Yi)/(Xi+l - Xi)

Yi = 2
1 + ((Yi+l - Yi)/(Xi+l - Xi))
I 1 ("i-l . "i-l )
L..Jj=1 "'j sm Pj - L..Jj=1 "'j cos Pj tan Pi
1 + tan 2Pi
i-I i-I
= II ( COS 2Pi L "'j sinpj - sinpi COSPi L "'j COSPj), i = 1,2,···, n,
j=1 j=1

and
I I (Yi+l - Yi)
x · = _yo
t t (Xi+l - Xi)
i-I i-I
= -II ( cos 2 Pi L "'j sin Pj - sin Pi cos Pi L "'j cos pj) tan Pi,
j=1 j=1
i-I i-I
= II (sin 2Pi L "'j COSPj - sin Pi cos Pi L "'j sinpj), i = 1,2,···, n.
j=1 j=l

The normal distances and normal angles can be derived based on (x~, yi).
o

It should be noted that d~ is a function of only the scale variable II, and 8~
is a constant. Therefore, the preliminary model has only one parameter
h; that is:
Li : X cos8~ +Y sin8~ = d~
=Ci11, i=1,2,···,n. (20)

The encoding of the pose features into the primitives is accomplished by

performing a transformation of the preliminary model (20).

Theorem 1 (Ventura and Chen 1992). The dimension, orientation and

position of an n-sided polygon can be completely characterized in terms
of II, 81 , Xl and YI by the following system of equations:

Li : (XT - Xl) cos(8~ +( 1 ) + (yT - yd sin(8~ +( 1 ) = cilI, i = 1,2,···, n,

(21)
 161

where p:-T, }-T) are the coordinates of the edge points after the object is
rotated by an angle ()l counterclockwise, and translated by a distance of
(xl , yd·
Proof. The translation and rotation of the coordinates of the preliminary
Model (20) can be specified below:

Then, by determining X and Y from Eq. (22) , and substituting them in

Eq. (20), Model (21) can be obtained. <>

3.2. Shape Matching

The shape matching is an optimization procedure, which objective is to

minimize the sum of the squares of the errors between the input data and
the reference shape. The reference shape is represented by Model (21) ,
and the input data for the matching procedure is obtained from Eq. (5) .
To match the input data with the model, the edge points are first rotated
so that each edge is aligned with the reference edge (Ld :

I
Xi ,j = Xi ,j cos ( ) I ' ()I
i + Yi,j sm i, V (i,j), (23)

and
I ()I . ()I
Yij = Yi ,j cos i - Xi ,j sm i, V (i ,j), (24)
where ()i is computed from Eq. (17). Therefore, Model (21) is simplified
to

where (XI, yl) represent the edge points after the rotation.

Next, the best-fitting polygon is obtained using a least-squares procedure.

The associated model to obtain the best fitted polygon is represented by:
162

where
Ci,j = I(x~,j - Xl) cos 01 + (Y:,j - yd sin 01 - Ci id. (27)
The error Ci,j is the normal distance between the data point (x~,j' yi,j)
and the fitted straight line L i . Estimates for the four parameters of the
regression model are obtained by minimizing the sum of squares of the
errors:

( Min SSE(i1 ,X11Y110d

,\,n ,\,mi 2
Min L.Ji=l L.Jj=l Ci,j
Min E~=l E~l ((x~,j - xd COS01 + (Yi ,j - Y1) sin 0 1 - Ci i1)2 .
(28)

The optimal matching is accomplished by finding a solution of the sys-

tem of nonlinear equations generated by taking the partial derivatives of
SSE(i1,x1,th,Od with respect to iI, Xl, Y1 and 01,

3.3. Solution Procedure

The Newton method (Mathews 1987) has 'been employed to solve the
proposed least squares Problem (28). This particular method was chosen
because it generally converges faster (quadratically) than alternative pro-
cedures, such as the steepest ascent, conjugate gradient, or quasi-Newton
methods. The Newton method includes the initialization step and the
iterative step. The initialization step is to compute a good starting solu-
tion which is crucial to the subsequent computational effort. The starting
solution is found in the following manner: 1) the initial scale variable l~
is estimated by computing the distance between the two end points of
Sl1 2) the initial position (x~, y~) is set to the intersection of Li and L~,
which are the independent least squares lines for the points in subsets Sl
and Sn, respectively, and 3) the initial orientation angle (e~) is estimated
by computing the normal angle of Li. Each iterative step incorporates a
second-order Taylor polynomial and, as such, requires the computation
of both the gradient ([G]) and the Hessian matrix ([H]) at the current
iterate. The descent direction for the quadratic approximation is found
by premultiplying the gradient (i.e., the steepest descent direction) with
 163

the inverse of the Hessian matrix. Thus, the estimators are updated in
the following manner:

(29)

where
(30)
The algorithm is terminated when the relative error norm of the gradient
satisfies a predetermined stopping criterion; that is, when

(31)

4. An Example

A profile of a spindle was used to test the proposed algorithm. The edge
points were generated equally spaced along a defect-free polygonal con-
tour. Then, the points were perturbed by adding a normally distributed
noise along the normal direction of the object contour. The coordinates
of each edge point, (Xi,j,Yi,j), were updated as follows:

(32)

and
(33)

where f(€) is the magnitude function of the error, and ()i is the normal
angle along the edge. After adding the noise, all the sampled points were
rounded into the nearest integer so that the data resembles the error due
to the spatial resolution of the image plane. The magnitude function of
the error is
f(€) = aN(O, 1), (34)
where N(O, 1) is the standardized normally distributed random variable
with mean zero and variance one, and a is the variance (magnitude factor)
of the noise. The set of generated data is shown in Figure l(a), where
a = 3, n = 12 and m = 203.
164

The proposed shape matching procedure was implemented on a VAX 8550

computer using the single precision C compiler. The CPU time to per-
form both the shape modelling and matching procedure, excluding input
and output, is about 0.02 seconds, and takes 14 iterations to reach the
stopping criterion (the relative norm of the gradient less than 10- 6 ). The
optimal matching between the model and the input data is shown in Fig-
ure l(b) with a solution (it\xt\y}4,O}4) = (149.8,382.7,248.9,0.5013).
The convergent behavior of the algorithm is illustrated -in Figure 2, in
which the X-axis represents the number of iterations, and Y-axis repre-
sents the relative norm of the gradient.

5. Conclusions

A new method for the analysis of general polygonal objects has been pro-
posed. This procedure includes data segmentation and shape analysis.
The shape analysis consists of two steps: shape modelling and matching
algorithm. The shape modelling step is to describe a defect-free reference
shape in terms of a set of edge primitives. The primitives are represented
in a parametric form, as opposed to the structural description. Based on
the shape model, the matching algorithm uses the least-squares criterion
to find the best fitting between a set of scene data and the set of primi-
tives. The developed algorithm is well suited for on-line implementation
in an automated environment due to its speed and accuracy.

References

Arkin, E. M., L. P. Chew, D. P. Huttenlocher, K. Kedem and J. S. B.

Mitchell, An efficiently computable metric for comparing polygonal
shapes, IEEE Trans. Pattern Anal. Mach. Intell. PAMI-13,
1991, 209-216.
At allah , M. J., C. C. Ribeiro and S. Lifschitz, Computing some distance
functions between polygons, Pattern Recognition 24, 1991, 775-781.
Cox, P., H. Maitre, M. Minoux and C. Ribeiro, Optimal matching of
convex polygons, Pattern Recognition Letters 9, 1989, 327-334.
Han, M-H, D. Jang and J. Foster, Inspection of 2-D objects using pattern
matching method, Pattern Recognition 22, 1989, 567-575.
 165

500 500

400 400

300 300

200 200

100 100

O ~,--~~--------~~ O+-~~~~~~~~
o 100 200 300 400 500 o 100 200 300 400 500

(a) (b)

Figure 1. Test problem of a nonconvex polygon:

(a) Input data, and (b) Optimal matching.

o 5 10 15

Iteration

Figure 2. Algorithm performance.

166

Kamgar-parsi, B., A. Margalit and A. Rosenfeld, Matching general polyg-

onal arcs, CVGIP: Image Understanding 53, 1991, 227-234.
Koch, M. W. and R. L. Kashyap, Matching polygon fragments, Pattern
Recognition Letters 10, 1989, 297-308.
Mathews, J. H., Numerical Methods for Computer Science, Engineering,
and Mathematics, Prentice-Hall, Englewood Cliffs, NJ, 1987.
Schaffer, G. H., Machine vision: a sense for CIM, American Machinist,
pp. 101-120, McGraw-Hill, New York, June 1984.
Ventura, J. A. and J. M. Chen, Se~mentation of 2-D curve contours,
Pattern Recognition (to appear).
Ventura, J. A. and J. M. Chen, Optimal matching of non-convex poly-
gons, IMSE Working Paper, Dept. of Ind. and Mgmt. Syst. En-
gineering, the Pennsylvania State University, University Park, PA,
1992 (to appear).
Ventura, J. A. and Z. A. Dung, Automated inspection of rectangular
and square production parts using vision systems, European J. of
Operational Research (to appear).
II. Flexible Manufacturing Systems
 Problems of Modelling and FMS Control Implementation
by

Janez Dekleva
Matja~ Gaberc

Ljubljana University
Faculty of Mechanical Engineering
Murnikova 2, 61000 Ljubljana
Slovenia

1. Introduction

The design of the FMS as proposed in literature [1] and which is

based on the Gantt chart model of the system represents the
work load of each resource (i. e . of each element of the FMS)
obtained by use of Modified ~lgorithm for generating Mon-Qelay
§chedules (MAGNDS) .

The Gantt chart of this so-called two-phase method is based on

the maximization of the utilization of machines and on the
minimization of the through-put time of individual parts . The
first goal is achieved by the very nature of the non-deiay
scheduling and the realization of the second goal is left to the
priority ruTes incorporated into the MAGNDS . The final result of
the procedure designed is a satisfactory FMS with the proposed
schedule of operations of parts on individual elements of the
system.

The question of the usability of the near to optimal schedule

arises specially in connection with the operation of the FMS.
Unfortunately, the deterministic processing, transportation and
170

manipulation times used for schedule generation, render this

option useless for controlling the operation of the FMS, although
(even though) it may be quite useful for testing the reality of
master schedule. It is known by experience that processing times
of jobs have, in addition to their basically deterministic
portion of time interval, small random components caused by non-
ideal material, tools, processing procedure etc. The persistence
in the usage of the schedule above for system operation would
certainly lower the obtained high utilization of the NC
machines/centres.

On the other hand, the operat i on of the FMS symbol i zes the
concurrent flow of parts within the FMS, which all compete for
the same set of final resources. Out of this concurrent competi-
tion a classical deadlock situation may occur .

In this paper, we first modify the already known model of a job-

shop system based on Timed Event-Graphs for the purpose of better
understanding of deadlock situations in the FMS. Then we define
two command circuits: the so-called external logistic circuit and
the internal one, we formulate the problem of deadlock avoidance
and f i na 11 y, on the bas is of the fact that the two command
circuits represent two time axes of the improved Gantt chart
model of the FMS, the movements and transportation between
different resources are scheduled with the accordingly adapted
MAGNDS.

2. FMS Using Timed Event-Graphs

The mode 11 i ng of the FMS by use of Petr i Nets (PN) is a we 11

known method. PN are a special kind of bipartite, weighted,
directed graphs consisting of two types of nodes, cal1e"tl places
and transitions.
PN can be defined as a 5-tuple, PN=(P, T,F,W,MO)' where [2]:
P = {p" P2' • •• , Pn} is a finite set of places,
 171

T = {t : , t ; , ... , t~} is a finite set of transitions,

F !: (Px T)u( TxP) is a set of arcs from places to
transitions and from transitions
to places,
W: F - {1, 2, 3 ... } is a weight function, where 3
weighted arcs stand for the 3
parallel arcs,
MD:P - {O, 1, 2 ••• } is the initial marking which
assigns a non- negative number of
tokens to each place of the PN.
,en T = 0 and PuT ~ 0

A Timed PN is a 6-tuple (PN, ~), where

Jl: T - {TI' T2' ••• , Tm} is a finite set of firing times
which associates to each transi-
t i on a fir i ng time ( pos it i ve
rational number).

A Timed Event-Graph called Decision Firing PN is a PN with the

characteristic of each place being connected to exactly one input
and one output transition with firing times defined as above.

The PN model of the FMS is very similar to the model of a job

shop [3] with some necessary extensions. There are at least four
additional resources (elements) in the FMS: the material handling
system (HMS) responsible for carrying parts, pallets and jigs
into the system (FMS); the L!U station, i.e. the resource
responsible for loading, reloading and unloading the parts; the
transportation vehicles (TV) which perform all transportation
movements within the system; waiting stations (WS) for parts
moving within the system . In modelling these four resources with
transitions and places, the transitions represent tasks or jobs
(i.e. moving, loading, moving, waiting), the input places
represent need for resources, and the output places represent
released resources.
172

The typical closed loop model of manufacturing process is sh:w~

in Figure 1 .

MHS resource MHS

a)

L/U TV NCcentre TV Wcentre TV L/U

. ...
.
............................. .............................. ,

b) WS TV NC TV WS~

Figure 1: a) The closed loop of manufacturing process for

the part IO 7 where the input buffers of work-
ing centres are avai lable (Wcentre - washing
.
centre) ,
b) The closed loop model of manufacturing process
where the input buffers of working centres are
busy and a pallet with the part is release: on
waiting stations before entering work'ng
centres.

The sequence of parts at the individual resources (e.g . NC

centres, transportation vehicle TV, I/O station etc.) i s
determined by the sequence of operations on individual ax i s ( " .e.
resource) as shown in the Gantt chart (Figure 2). The obta ; ned
sequence represents the order of the transitions in new circ~ ' ts
which are called command circuits. Two of these command circu~:s,
the MHS circuit called the external logistic circuit and the TV
command circuit called the internal logistic circuit, are of
special importance and they should be extended. The extensior of
theexterna 1 logistic circuits and of the internal ones is
 173

required because both resources physically stop in the posit~:~

where they are released as resources. In order to join the

required position in the next processing circuit to perform
either the job of movement or that of transportation, they have
to be transferred to the places from their waiting positions. The
internal logistic circuit then looks as shown in Figure 3.

"
VU

Net
NC2
tIC3

IIC4

UC

1\1

8 8.5 t 1.5

Figure 2: The original Gantt chart used in the design pr:)-

cedure of the FMS

The basic function of both logistic circuits is the movement ~.

individual parts between the sequence of resources defined in the
manufacturing process (routing) of each part. But these mcvemer~s
are reasonaele only if the resources are avai lable with each
movement, i.e. with each resource - allocation decision. We can
express this type of situation with the statement that the par~s
inside the FMS compete for the shared resources. It is the
phenomena of dead lock and u 1t i mate 1y the p rob 1em of dead lock
prevention and deadlock avoidance which are studied in the F~S
using Petri Nets. In this connection two articles should be
mentioned [4], [5] and their definitions of deadlock avoidance
policy and of markings in which processes are blocked or beeing
174

in deadlock are accepted in our reasoning.

interna.l lOgistiC Circuit

.. _-----------------.----------------------" I

k-th processing
circuit

<k+l)-th processing
circuit

Figure 3: The internal logistic circuit (= TV - transportation

vehicle internal command circuit)

3. Modification of the Gantt Chart Model

The study of modelling the FMS by use of Timed Event-Graphs, the

survey of the investigation of deadloc~ prevention by reachabi-
lity graphs, and finally, the acceptance of definitions proposed
for deadlock avoi dance, none the 1ess, promoted some necessary
simpl ification in our original Gantt chart model. Instead of
modelling each element (resource) of the FMS by the separate time
axes (see Figure 2), only two of them should be modelled: MHS's
represented as MHS command c i rcu its (i. e. external log i st i c
c i rcu its) and TV's represented as TV command c i rcu i ts (i. e .
internal logistic circuits). The activities carried out by these
two command circuits are connected with all the possible
resources which are either held or needed. by a closed chain of
 175

processes (i.e . processing circuits ) and where the avoidance _

circular wait among the parts is of prime importance.

At this point it is necessary to emphasize the benefits of the

proposed modified model. Firstly this model simplifies the
understanding of the behaviour of the system (e.g . in the case
presented in the next paragraph there are only two time axes in
the Gantt chart model) . Secondly the model clearly points to two
basic decisions connected with the material (i.e. parts) flow in
the system . The decisions are associated with:
1 . the optimal entrance of parts into the system, i . e. with the
sequencing entrances requested by a higher level of the
production system (i . e . the fourth level),
2. the selection of opt i mal transportation activities in the
system.

The first decision is connected with the requirement of minimiz-

ing the through-put time in the system and with the problem of
deadlock avoidance . It requires the precise allocation of the l/U
station (including reservation!) either to the outer command
ci r cuit or to the inner one and the definition of priorit i es ~or

each decision maker .

The selection of optimal transportation activities is responsible

for optimal utilization of NC centres, for minimization of
through-put time of parts and for deadlock avoicance in -:he
system . The first goal is carried out by the MAGNOS, but the next
two goals may be confl i cting . The MAGNOS generates schedul~s in
the class of dispatching procedures which means tha~ scheduling
decisions need not be made in advance, but only when they are
requested . This is essential for the real time control of the
FMS. The requests either for entrance into the system or "'or
transportation among resources of the system are solved at the
time of occurrence. The request for entrance into the system i s
admitted into the decision making procedure only i'" the MHS 1S
free and the L/U station is available (i.e. not busy and not
176

reserved). The requests for transportation within the system at

the stage t are first investigated for availability of resources
and only then included into the set of schedulable transportati-
ons at stage t. The request with the highest priority at the
earliest time at which the selected transportation could start
is determined for executi9n . With a greater number of waiting
stations we could certainly prevent blocking situations or even
deadlocks, but at the same time the through-put time might
greatly increase, which is exactly what we have tried to avoid
in the first place . In order to cope with this conflicting
s i tuat i on successfu 11 y , we have to introduce some necessary
restrictions regarding the number of parts waiting for the NC
centre, the number of occupied waiting centres etc .

To accomplish the outlined requirements the control procedure of

the FMS ought to be designed according to hierarchical rules with
three levels [6] :
FMS control system,
Integrated fhysical ~roupings ~ontrol ~ystem (IPGCS),
equipment control system .

The emphasis should be on the third level , i.e. on the decision

mak i ng system of the FMS, respons i b 1e for real time schedu 1e
generation . The second level should be responsible for the
presentation of the status of the system regarding either the
motion of parts or the availability and the functioning of its
resources (elements). Another function of the second level should
be to report all requests of the first level and/or to execute
all requirements decided at the third level and to convey the
necessary commands to the first (lower) level.

These rules take into account the following facts . The movement
of individual parts within the system, 'i.e. among resources, is
guided by corresponding routings. Only minor changes may be made
(to these movements) when non-ava i 1ab i 1 i ty of some resou rces
requires detours to the waiting stations. The movement within the
 177

system as a who 1e is schedu 1ed accord i ng to the MAGNDS as

described above and can be imagined as a queuing situation where
the part upon arrival at the resource (i.e. NC centre) may be
processed immediately or may have to experience waiting time at
the input buffer until the working position becomes available.
Operating characteristics of this queuing situations and/or at
least general grasp of them would greatly enable the analyst to
make inferences concerning the operation of the system, specially
in devising the correct priority rules for the MAGNDS.

Let us make a brief survey of some general rules connected with

the resource allocation in respect to accepted requests :
1/ The L/U station may be required:
¢ by the outer logistic circuit when the entrance of the
part into the system is under consideration , the
requirement has the highest priority after the time t=Q
when the FMS starts to operate; the L/U station is
reserved in order to avoid deadlock;
when the entrance of parts is temporarily stopped by
the fact that all input buffers of NC centres desig-
nated for the first operation of parts are occupied
(busy) and that the working positions are busy too; the
entrance of parts is resumed at the time t, = tc - t:
after input buffer(s) become(s) available; the critical
time ~ is defined as the sum of all possible waiting
and executing times needed for loading of the part and
for its transportation to the input buffer without
jeopardizing the utilization of working position of the
NC-centre, t: being the deterministic operation time of
the part in the process;
when the entrance of parts to some NC centre is brought
to a standstill in the case when the NC centre desig-

The entrances of parts included in the order for manufacturing

execution represent the permanent request to the system. These
requests should be compared with other requests generated with the
system and corresponding decisions should be made according to
priority rules.
178

nated to the second , third . .. operation of the same

part has a busy input buffer and the working position
as we 11 and when the first wa it i ng stat i on becomes
occupied by the incoming part; the entrance of parts
into the system is resumed after the waiting station
becomes available;
o by the inner logistic circuit for reloading and unloading
of parts. These parts, usually waiting at output buffers
of NC centres (machines), may exceptionally occupy one
waiting station per NC centre. These parts may reserve the
L/U station, i.e. they have the priority of transportation
to the L/U station. For all the other parts the priority
of movement to the L/U station is defined by the order of
entrances into the system .
2/ The transportation vehicle is requested for the transfer of
parts from the output buffer of one NC centre (machine) to
the input buffer of the next NC centre (mach i ne). The
priorities of movement are defined according to the order of
entrances into the system . If the request is blocked by the
busy input buffer, · the alternative solution, i.e. the
movement to the wa it i ng stat ion is offered. For each NC
centre (machine) only one single waiting station is allowed.
The busy waiting place in front of the NC centre with the
traffic intensity of p=l/~ < 1 (l = mean arrival rate of
parts to the centre, ~ = mean processing rate per NC centre)
releases (triggers) the request for transportation to the
input buffer (with the same priority!), but no changes are
foreseen at the entrance of the waiting station. In the case
where p-approaches 1 (one) or becomes greater than 1 (>1) ,
the entrance has to be stopped (see paragraph 1/).
3/ In addition to priorities described in paragraphs 1/ and 2/
some additional rules may be added , e.g. minimum slack time,
earliest due dates etc . [7] [8]
 179

4. Case Study

The usefulness of the proposed modification will be explained in

the following case study. The FMS under consideration (Figure 4)
is assembled of : 4 (four) NC-centres, 1 (one) NC-washing machine,
(one) L/U station, (one) transportation vehicle, 4 (four)
waiting stat i ons and (one) mater i al handling system (MHS)

IV. LEVEL

III. LEVEL IWU'ACTURING DATA

lAst.
MIIIU. rDR R£Al
TIM[ SCHtJaE
GENERA TIOI <MAGHD$)

II. LEVEL IOfITORING stotus r.ports

of the rMS
IPGCS
DIAUIlSTIC

... sso~s COfIfIGnds

rf1~sts

I. LEVEL

B
VAR£HOUS£
FMS

Figure 4 : Flexible manufacturing system used in the case study

with the scheme of the control levels
180

needed for the movement of parts, pallets and jigs from / ~o ~r.e

warehouse to/from the L/U station.

There are fourteen different parts that are produced in this

system; the manufacturing process (routing) of each part is
uniquely defined, together with the deterministic times spent on
each active element (centres, L/U, TV). Each part has its own
pallet; all pallets, the number of which is fourteen, are equal.
Waiting times of parts spent on waiting places and on input and
output buffers of work i ng (washi ng) centres are mon i tored and
processed by the d i agnost i c system . A11 data on starting and
terminating times of current activities (operations, loading,
transportation) can be obtained. The same diagnostic systerr: keeps
track of availability (or non-availability) of resources, waiting
places, buffers etc .

The described FMS was designed by use of two-phase method and we

are going to investigate performances of the system when the
modified Gantt chart model is applied for its control (Figure 5).
Two axes are shown in this figure, one for the MHS called the
outer command circuits and the other for the TV called the ;nr,er
command circuit. The activities of both circuits in the ~ig~re
represent the first twelve minutes and the corresponding even~s
are indicated when the appropriate messages (M), requests (?) and
commands (C) are generated and communicated between the first a~d
the second level. Decisions regarding commands, i.e. sche:;,;les,
are made on the third level and they are modified and se~t v:a
the Integrated Physical Groupings Control System (see Fi~~re 5)
to the fi rst- 1eve 1.

The meanings of the events indicated by arrows in Figure 5 a~e

as follows: M1: the L/U is available; M2: the TV is ava;:able;
C1: command for transportation of the part IDS from ware~o~se ~o
the L/U station; M3: the MHS is available; ell: comma~d fer
loading of the part IDS; M4 : loading of the part IDS is ccY!-
pleted; R1 : the transportation of the part IDS from the L!U
 181

station to the input buffer of the NC centre is requested; C2:

command for transportation of the part lOB from the L/U station
to the input buffer of the NC centre 1, etc.

RI R2 R3
M4 M6 148 "1\ "12 Ml3 H15
t t
MHS i
C3 Ct2 3 C7 4
M5 H7 M9 HIO HI4 HI6

TV ::::::~::::::
........

C2 C4 C6

tlt2 t:J t5 tiMe

Figure 5: The modified Gantt chart for controlling the FMS;

activities on two time axes for the first twelve
minutes and corresponding events such as: messages
(M), requests (R) and commands (C) are shown

The performances of the FMS were tested wi th the product i on

program defined by six repetitions of a pack cf 14 (fo~rteen)

parts. The following utilization of resources was obtained:

L/U station 58.33%
NC centre 89.08%
NC centre 2 93.71%
NC centre 3 100.00%
NC centre 4 92.41%
NC washing machine 44.73%
TV 44.08%
Waiting station 41.88%
Waiting station 2 27.92%
Waiting station 3 19.33%
Waiting station 4 14.64%
182

The ratio of the average through-put time and the average net
processing time equals 1.65. In order to interpret the calculated
number, the average through-put time has to be explained. This
is obtained by adding the appropriate waiting times to the net
processing times of individual parts. (Added waiting times are
the lowest estimates and they can not be omitted.) By doing this
we obtained an estimated value of 1.62 which is close enough to
the above calculated ratio of 1.65.

The order completed at the time t = 1368 minutes has to be

compared with the longest processing time of 1312.5 minutes
scheduled at the NC centre No.3.

5. Conclusions

The paper deals wi th the mode 11 i ng of the FMS and wi th the

proposal of the Gantt chart model based on two different time
axes; one for the MHS and the other for the transportation system
within the FMS. The model is suitable for on-line control of the
FMS. By the case study we tried to demonstrate the performances
obtained within the system where 84 items were manufactured. The
avoidance of deadlock should be counted as one of the achieve-
ments along with the high utilization of centres and with the
satisfactory through-put time.

6. References

[1] Oek 1 eva J.: Non Oe 1 ay-Rev is i ted, Modern Product jon Concepts 1
Theory and Applications, FB.ndel G., Zapfel G. (Eds.), pp.
601-616, Springer-Verlag, Berlin, 1991.

[2] Murata T.: Petri Nets: Properties, Analysis and Applica-

tions, Proceedings of the IEEE, Vol. 77, No.4, pp. 541-580,
1989.
 183

[3J Hillion H. P., Proth J.-M.: Performance Evaluation of Job-

Shop Systems Using Timed Event-Graphs, IEEE Transactions on
Automatic Control, Vol. 34, No.1, pp. 3-9, 1989.

[4] Viswanadham N., Narahari Y., Johnson T. L.: Deadlock

Prevention and Deadlock Avoidance in Flexible Manufacturing
Systems Using Petri Net Models, IEEE Transactions on
Robotics and Automation, Vol. 6, No.6, pp. 713-723, 1990.

[5J Banaszak Z., Krogh G. H.: Deadlock Avoidance in Flexible

Manufacturing Systems with Concurrently Competing Process
Flows, IEEE Transactions on Robotics and Automation, Vol. 6,
No.6, pp. 724-734, 1990.

[6] Jones A. T., McLean C. R.: A Proposed Hierarchical Control

Model for Automated
Manufacturing Systems, Journal of
Manufacturing Systems, Vol. 5, No.1, pp. 15-25, 1986.

[7] Conway R. W., Maxwell W. L., Miller L. W.: Theory of

Schedu ling, Add i son-Wes 1ey Pub 1 ish i ng Company, Read i ng,
Massachusetts, 1967.

[8] Baker K. R.: Introduction to Sequencing and Scheduling, John

Willey & Sons Inc., New York, London, 1974.
 FORMAL MODELS FOR CONTROL OF FLEXIBLE
MANUFACTURING SYSTEMS

Sanjay Joshi
&
Jeffrey S. Smith

Department of Industrial & Management Systems Engineering

Pennsylvania State University
University Park, PA 16802

Richard A. Wysk

Royce Wisenbaker Chair In Innovation

Department of Industrial Engineering
Texas A & M University
College Station, Texas 77843-3131

Abstract

Explicit formal models, theorems and abstractions form a large segment of the science base
in many fields. These models are considered as necessary precursors to the development of pragmatic
knowledge and tools, and provide the basic understanding of generic issues related to various classes
of problems. The need for a solid science base in manufacturing has been articulated by several
authors . One area of manufacturing that could benefit significantly from the development of formal
models is the control of flexible manufacturing cells. Applications of formal models and abstractions
could further the science and understanding of the control of manufacturing cells. The development
of formal models often leads to the discovery of ambiguities, omissions, and contradictions, the formal
model may lead to semi-automated (or even automated) system development methods, the formal
model can be verified for correctness by mathematical methods, a formally verified subsystem can be
incorporated into a large system with greater confidence that it behaves as specified, and different
designs can be compared. Two approaches to the development of formal models namely the use of
context free grammars for control and a new set of formal models called Message-based Part State
Graphs (MPSG) will be presented. It is shown that, through the use of these models, a significant
amount of the shop floor control software can be automatically generated based on a high level
description of the system.
 185

Introduction
Explicit formal models, theorems and abstractions form a large segment of the science base
in many fields. These models are considered as necessary precursors to the development of pragmatic
knowledge and tools, and provide the basic understanding of generic issues related to various classes
of problems. The need for a solid science base in manufacturing has been articulated by several
authors [Suh 1984, Sohlenius 1984]. One area of manufacturing that could benefit significantly from
the development of formal models is the control of flexible manufacturing cells. A formal model of
a system is an unambiguous description of the system presented in an implementation-independent
language with precisely defined syntax and semantics. Applications of formal models and
abstractions could further the science and understanding of the control of manufacturing cells. The
following are other benefits of a formal framework [Sharbach 1984, Ostroff, 1990]:

in the process of formalizing informal requirements, ambiguities, omissions, and

contradictions will often be discovered hence reducing the possibility that errors are made
during system development, and the correctness of the specification can be considered
before detailed implementation is undertaken.

provides the vehicles for unambiguous communication of design decisions, and is supported
by the expressive power of good formal languages which improve conciseness.

the formal model may lead to semi-automated (or even automated) tools for system
development, simulation, verification, and implementations of specifications, and testing of
implementations.

the formal model can be verified for correctness by mathematical methods,

a formally verified subsystem can be incorporated into a large system with greater
confidence that it behaves as specified, and

different designs can be compared.

Formal methods and grammars have been used recently in the development of various aspects
of manufacturing system theory. Wu, et al (1986) apply grammars to the cell formation problem,
using a syntactic pattern recognition approach. Williams and Upton [1989] examine the use of
stochastic grammars for the measurement of flexibility. Graham and Saridis [1982] describe the
186

application of formal grammars to the control of a small robot, incorporating feedback and learning
into their scheme. More importantly their approach recognizes the usefulness of formal languages in
modeling the hierarchical nature of complex control systems. The use of syntactic models in
manufacturing processing and control is also presented by Upton and Barash [1988]. Tam [1989]
presents an approach using linguistic models to develop models for flow line production. The model
is used to check production feasibility. Petri nets have also been applied to the design, testing and
development of manufacturing control systems [Narahari and Viswanadham, 1985], Merabet [1985],
Kasturia et al. [1988], Xou et. al [1989]. Bruno and Marchetto [1987] have developed extended petri
nets called PROT net (PROcess Translatable net) where the basic structure of the processes and
synchronization can be generated automatically from the net, via translation into Ada program
structures which will provide a rapid prototype of the system.

Several references on the use of formal models for automatic generation of software and
compilers can be found in the computer literature [Aho et aI1985]. The use of formal models of the
factory floor as a precursor for software generation has been discussed by Naylor and Maletz [1986]
and Naylor and Volz [1987]. Unfortunately, methods for systematically extending their formal models
into running software have not yet evolved. One problem with their approach is that scheduling forms
a large component of control, making the task more difficult.

Chaar [1990] extends the model of Naylor and Volz by developing a methodology for
implementing real-time control software for FMS. This methodology is based on a component-
oriented rule-based specification language. Chaar uses the modeling formalism of Naylor and Volz
to model the softwarelhardware components of the manufacturing system and the part process plans
as a set of first-order logic-based rules. These rules are then coded in the specification language which
can be converted to executable code. Although some possibilities for automating this conversion are
mentioned, none are described. Instead, the conversion is done manually by the software engineer.
However, Chaar does present an interesting method for generating a cyclic schedule for the
manufacturing system. An example implementation for a prismatic machining cell is presented in
[Hadj-Alouane, 1990]. A detailed description of the implementation of a material handling system
as a softwarelhardware component is given in [Hadj-Alouane, 1988].

Biemans and Blonk [1987] present a formal specification technique based on LOTOS (Language
for Temporal Ordering of Specifications) for specifying CIM systems such as transport systems and
work cells. Other authors [Maimon and Fisher, 1988] have developed rapid software prototyping
systems that provide only for the structural specifications of cell control software, still demanding
 187

enonnous development of hand crafted code to realize effective systems.

Brooks et al. (1990] examine the use of fonnal models for implementing and verifying media
access control (MAC) protocols. They support the use of fonnal description languages which can be
automatically or semi-automatically converted into running code. Their specification language, LIFP
(language for implementing fast protocols), is based on the state machine model. Ostroff (1990]
presents an overview of the use of temporal logic for control of real-time discrete event systems.

Architecture
A shop floor control architecture serves as a blueprint for the development and implementation
of a shop floor control system. Therefore, a control architecture should completely and
unambiguously describe the structure of a system as well as the relationships between the system
inputs and the system outputs. Although there have been several shop floor control architectures
described in the literature, these are not well suited for use as a basis for software development. This
is because these architectures (1) are ambiguous, and (2) lack sufficient detail. In other words, these
architectures are not defined fonnally, but rather, they are described in prose.

We use a three level hierarchical control architecture in which each level has been fonnally
defined (Smith et aI., 1992]. This architecture is an extension of the "sealeable" architecture presented
by Joshi et al. (1990]. A similar architecture is described by Jones and Saleh (1990]. The three levels
in this architecture are the equipment level, the workstation level, and the shop level (see Figure 1).
The following paragraphs provide an introduction to the architecture.

Figure 1 - Shop floor control architecture.

188

The equipment level is the lowest level in the hierarchy. It represents a logical view of a
machine (complete with its machine level controller) and an equipment level controller. An equipment
level controller has responsibility for the piece of machinery under its control. To this end it makes
decisions regarding local part sequencing, keeps track of part locations, and is responsible for
monitoring the operation of the device. Shop floor equipment has been partitioned into four classes.
Material processors (MP) autonomously process parts in some way according to a set of processing
instructions. NC machines, inspection devices, and assembly devices are examples of material
processors. Material handlers (MH) load and unload parts from other equipment. Industrial robots
and pick and place machines are examples of material handlers. Material transporters (MT) transport
parts from one location to another within the shop. The distinction between MH equipment and MT
equipment is that MH equipment performs intra-workstation material movement and MT equipment
performs inter-workstation material movement. Examples of material transport devices include AGVs,
conveyors, and fork trucks. Automated storage (AS) devices store parts, raw materials, tools, and
fixtures. Automated storage and retrieval systems (AS/RS) are examples of automated storage
equipment.

The workstation level is the second level in the control hierarchy. A workstation is made up of
one or more pieces of equipment under the control of a workstation level controller. Workstations
are defined using the physical layout of the equipment and are generally configured so that multiple
material processing devices can share the services of a single material handling device. The
workstation level has been partitioned into processing workstations, transport workstations, and storage
workstations. A processing workstation contains one or more material processors along with
(optionally) a set of material handlers and buffer storage units. A storage workstation contains one
or more automated storage devices along with (optionally) a set of material handlers. A transport
workstation contains one or more material transport devices along with (optionally) a set of material
handlers. Additionally, the Resource Manager is a special workstation level entity which provides
centralized access to shared resources.

The shop level is the highest level in the control hierarchy. The shop controller is responsible
for controlling the activities of the workstations and the Resource Manager. The shop level is also
responsible for interacting with the user (e.g. accepting orders and order/part status requests).

The functionality of the controllers at each level has been partitioned into planning, scheduling,
and execution activities. Planning determines which controller tasks will be performed. Scheduling
sets start and end times for the planned tasks. Execution is responsible for performing the scheduled
 189

Feedback

Feedback

Feedback

System Operation System Development

Figure 2 - System infonnation/control flow ..

tasks (i.e., interacting with the subordinate systems). Table 1 shows examples of the planning,
scheduling, and execution tasks for the equipment, workstation, and shop levels. During system
operation, scheduling depends on planning, execution depends on scheduling, and the physical system
depends on execution. (see Figure 2). During system development, however, the dependencies are
reversed (see Figure 2). Execution depends on the physical system (or a model of the physical
system) and planning and scheduling depend on execution. Additionally, planning and scheduling also
depend on the production requirements. In the target environment, the production requirements change
much more frequently than the physical system. Consequently, the planning and scheduling modules
of a controller will require modification much more frequently than the execution module.
Furthermore, since the execution module is dependent only on the physical system, it is conjectured
that the execution software can be automatically generated based on a formal model of the physical
system.

Fonnal Models
We present two formal models of the execution portion of shop floor controllers. These fonnal
models are used as bases for the automatic generation of the shop floor controllers (the execution
module of the shop floor controllers). The first approach uses context free grammars to model the
control problems for each controller. This approach is based on a system state space model. The
second approach uses a new model called a message-based part state graph (MPSG). This approach
is based on a part state model rather than a system state model.
190

Context Free Grammar Approach

Formal
Our earlier work [Joshi et al.,1992, and Model

Mettala 1989] presents an approach to automate

the development of control software for flexible
manufacturing cells both at the workstation and
!
Grammar
Conslruction
cell controller level (under the new architecture
and definitions, these are the equipment and
workstation controllers, respectively), using high
~
Definition &
level systems specifications. The approach is Conslruction of
Semantic Actions

!
illustrated in Figure 3. The automatic code
generation process begins with the user creating an
informal definition of the system specifications, Parser
Generation
describing the basic inputs, outputs, and

~
relationships, for the system processes in
application specific terms. This is then converted
Automatic
into a formal system specification. The formal Code Generation

model provides the basis for syntactic and

semantic correctness checking, and provides the t
Distributed Control
Software
necessary foundation for automatic transformation
Figure 3 - Code generatIOn process usmg CFGs.
into an executable program which is
computationally equivalent to the design specifications. A state based approach using the "parts in
contact" to define system state is used to create graph models of the controllers, describing state
transitions as a function of the physical actions that take place in an FMS. These graph models are
then converted into Context-free grammars (CFG) which are used as the formal basis. Each controller
is modelled as a parser. The FMS is controlled as a by-product of the process of recognizing the
CFGs. The semantic actions associated with the production rules of the CFG provide the necessary
control actions for both the physical part flow and system information flow. The formal model is then
converted algorithmically into a skeletal program defining the relationship between the data structures,
objectives and algorithms. The code generator converts this skeletal program into actual useable code.
 191

Table 1. Planning, scheduling, and execution activities for each level in the SFCS control
architecture.

Planning Scheduling Execution

Equip. Operations-level Determining the start/finish Interacting with the

planning (e.g. tool path times for the individual machine controller to
planning). tasks. Determining the initiate and monitor part
sequence of part processing processing.
when multiple parts are
allowed.

Wkstn. Determining the part Determining the start/finish Interacting with the
routes through the times for each part on each equipment level
workstation (e.g. processing machine in the controllers to assign and
selection of processing workstation. remove parts and to
equipment). Includes synchronize the activities
replanning in response of the devices (e.g. as
to machine breakdowns. required when using a
robot to load a part on a
machine tool).

Shop Determining part routes Determining the start/finish Interacting with the
through the shop. times for part batches at workstation controllers
Splitting part orders into each workstation. and the Resource Manager
batches to match to deliver/pick-up parts.
material transport and
workstation capacity
constraints.

Message-based Part State Graph Approach

In a distributed control environment, control is exercised by passing messages between controllers
and by performing controller tasks. Controller tasks are actions which are performed by the controller.
For example, closing the robot gripper would be a controller task for an equipment level robot
controller. For a particular controller, the set of controller events consists of all messages coming into
and going out of the controller along with the set of controller tasks. We define the processing
protocol as the sequence of controller events required for a controller to process a part. For example,
Table 2 shows the sequence of controller events required for a robot to load a part on an NC machine
(under the control of a workstation controller).

There are three distinct processing protocols associated with the dialogue in Table 2. The first
is a workstation processing protocol, which includes events associated with the workstation controller
(events 1,3,4,6,7,9, 10). The second is the machine tool processing protocol which includes all
events associated with the machine tool equipment controller (events 4, 5, 6, 10). The last protocol
192

is the robot processing protocol which includes all events associated with the robot (events 1, 2, 3,
7, 8, 9). The individual controllers are responsible for implementing the associated processing
protocols. A new formal model of the processing protocol called a message-based part state graph
(MPSG) has been developed. The MPSG model can be used to automatically generate the execution
portion of a controller based on the processing protocol (Smith, 1992).

Table 2 - Sequence of controller events for processing a part.

Event Source Messageffask Destination

1. workstation put part robot controller

2. robot controller put task robot

3. robot controller put done workstation

4. workstation grasp part machine controller

5. machine controller grasp part task machine

6. machine controller grasp done workstation

7. workstation clear robot controller

8. robot controller clear task robot

9. robot controller clear done workstation

10. workstation clear ok machine controller

An MPSG describes the behavior of a controller from the parts' point of view. An MPSG is a
deterministic finite automaton which has been modified in two ways. First, the MPSG accepts input
from two separate streams. The first stream accepts incoming messages from other controllers. The
second stream is the list of tasks specified by the controller's scheduler. Second, the transition
function has been modified so that the controller tasks can be performed during state transitions. This
is accomplished by introducing a set of controller actions and a corresponding controller action
transition function.

An MPSG M is defined formally as M = (Q, q() F, I, A, P, 6, y), where Q is a finite set of

states. qo E Q is an initial or start state, F ~ Q is a set of final or accepting states. I = (II U Io
U IT) is a finite set of controller events which is partitioned into a set of input messages (II)' a set
of output messages (1: 0 ), and a set of controller tasks (1: T). A is a finite set of controller actions,
where a E A is an executable function which performs some controller action. P is a finite set of
 193

physical preconditions for controller actions, partitioned so that for each a E A there is a
corresponding Pu E P, where Pu is a boolean valued function which returns either true or false. The
individual Pu are evaluated using the physical configuration and the system status. b: Q x l: -.
Q is a state transition function (next state function). y: Q x (l:o U ~r) -. A is a controller action
transition function.

In a standard DFA the input symbols are read from a single stream (e.g. a source code file). In
a MPSG, however, the input symbols are read from two different streams: the controller
communications module, and the task list. The communications module receives incoming messages
from other controllers in the system. The task list is the communications medium between the
execution module and the planning and scheduling modules within the controller. As described above,
the input alphabet (l:) consists of the partitioned set of controller events. ~J is the set of valid input
messages which are received from other controllers. For example, in the machine tool' controller, the
grasp part event specified in Table 4.1 is an input message received from the workstation controller.
l:o is the set of valid output messages which are sent to other controllers in the shop. Output
messages are specified by the scheduler placing an output message task on the task list. l:r is the set
of controller tasks. Controller tasks are executable functions which implement the controller actions
or perform database updates. Controller tasks are specified by the scheduler placing the tasks on the
task list.

The set of controller actions A, is a set of executable functions which perform the controller
actions. For example, the a E A which corresponds to the grasp part task as used above, is an
executable function which closes the machine tool 's part fixture. The implementation of these
functions will be device specific. The controller task transition function y, maps state-task pairs (Q
x (l:O U l:r)) onto A. The interpretation of the controller action transition function,

y(q, b) = a, q E Q, bE l:o U ~p and a E A,

is that the MPSG M, in state q and scanning the event b, executes function a. Upon successful
completion of a, M undergoes a state transition specified by b(q, b).

The feasibility of many controller tasks depends on the current system status. For example,
putting a part on a fixture is only feasible if there is no other part currently on the fixture (assuming
a single part fixture) . The physical preconditions are used to verify the system status prior to
executing controller tasks. P is a set of physical preconditions which is partitioned according to the
194

set A such that, associated with each controller action a there is a physical precondition, Pa which
is a boolean valued function that must evaluate to true in order for the corresponding controller action
a to be called. For example, if the controller task is to start (cycle _start) the machine tool, the
following physical precondition must hold:

Pcycle_S1Qrl = (PI II P2 II P3 II p.J where:

P1 - The part must be in the chuck
P2 - the chuck must be closed
P3 - the part NC program must be loaded
P4 - the machine tool work volume must be clear

If this conditions does not evaluate to true, a cycle start can not be sent to the machine tool, and the
specified task is not consumed by M and no state change occurs. Furthermore, the controller action
must also successfully execute in order for the task to be removed and a state change (identified by
b) to occur. Therefore, given a MPSG M is state q, scanning task b, the following pseudocode
represents the controller operation.

a = y (q, b) II determine the controller action

if ( Pa evaluates true) II verify physical preconditions
if( aO ) II if the action executes
q' = b(q, b) II change states
remove b from the task list II remove task from task list
endif
endif

Verifying the successful execution of the controller task facilitates error recovery when an unexpected
event occurs. For example, if a robot drops a part during a move task, the task will not successfully
complete (assuming that the robot can detect that the part has been dropped). The last valid state will
correspond to the most recently completed action or message. Optionally, M could transition to an
error state based on the message received from the robot.

Figure 4 shows the transition diagram and the corresponding formal representation of an example
MPSG. Note that the set P of physical preconditions is not included in this example. The events in
this example have been prepended with 'i_' for incoming messages, '0_' for outgoing messages, and
't_' for controller tasks. This example gives a simplified view of processing a part on a MP class
machine. The part is loaded by receiving a load message from the workstation controller and by
subsequently executing a load task. Once loaded, the machine is started by executing a start task.
 195

Start -.@) ® t_release

7

i load i error
0 i unload

3 5 6
t start i finish o_done

M = (Q, qo, F, l:, A, P, /), y) where:

Q = {O, 1, 2, 3, 4,5, 6, 7, 8}

qo = °
F = {4, 8}
l: = (l:( U l:o U l:T ) where:
l:( = {Uoad, i_finish, i_error, i_unload}
l:o = {o_done}
l:T = {tJoad, t_start, t_release}

A = {loadO, start(), doneO, releaseO}

/) : Qx L -- Q y : Q x(Lo U L T ) -- A

6(0, Uoad) --1

6(1, tJoad) --2 y(l, tJoad) -- loadO
6(2, U tart) --3 y(2, t_start) -- startO
6(3, i_error) --4
6(3, i_finish) --5
6(5,o_d one) --6 y(5,o_done) -- doneO
6(6, i_unload) --7
6(7, tJelease ) --8 y(7, tJelease ) -- releaseO

Figure 4 - Example MPSG.

196

At some time in the future, the machine will either report that it has finished processing (finish
message) or will report that an error has occurred (error message). If the machine successfully
finishes processing, the controller sends a done message to the workstation signifying that processing
has been completed on the part. Finally, the part is unloaded by receiving an unload message and
subsequently executing an unload task.

In order to generate an equipment controller based on this MPSG, a mechanism to send and
receive messages and the set of controller task functions (A) would be required. Furthermore, the set
of physical preconditions (P) would have to be defined and the corresponding boolean functions
provided. MPSG provide enormous flexibility in controller development and modification. For
example, to add an elementary error recovery function to the controller, a new output message
transition could be added to the MPSG. This transition would send a processing error message to the
workstation and transition to state 6. To make this modification, the following line would be added
to the transition table in Figure 4:

y(4, error) .... error().

0_

Additionally, state 6 would replace state 4 in F, the executable function errorO would be added to A,
°
and the output message _error would be added to kO' This modification would allow the part to
be removed from the machine tool if an error was detected (e.g. the tool breaks, or the part is
damaged).

Figure 5 shows how the MPSG Builder is used to automatically generate the C++ code for a
shop floor controller. The input file (input.m)
is a textual description of the MPSG. The
output from the MPSG builder includes the
implementation of the transition functions
(MPSG.cpp), the header file (MPSG.h), and
the task action file (MPSG.a). The task action
file contains the function prototypes for the
controller actions and the physical
precondition functions. These functions are
implementation specific and must be
developed for each individual MPSG.

Figure 5 - MPSG Builder structure.

 197

A set of MPSG for generic controllers has also been developed. This set includes MPSG for
generic MP, MH, MT, and AS class equipment, processing, storage, and transport workstation, and
shop controllers (Smith, 1992). These MPSG include the implementation of the majority of the
controller tasks and the physical precondition functions. If the generic MPSG can be used, the user
only needs to supply the device specific controller action functions (e.g. the device specific routines
to close a robot gripper, move the arm, etc. or the routines to download an NC file to an NC machine
tool). Furthermore, the generic MPSG can by modified by adding or removing arcs, and the MPSG
Builder can be used to regenerate the transition functions.

A shop floor control system for a full -scale flexible manufacturing system at the Pennsylvania
State University's Computer Integrated Manufacturing Laboratory has been implemented using both
of the techniques described in this paper. This FMS includes three processing workstations, a storage
workstation, and a transport workstation (see Figure 6). The primary distinction between the CFG
approach and the MPSG approach is that, in the CFG approach, the system states are modeled and,
in the MPSG approach, the part states are modeled. Modeling part states rather than system states

Penn State elM Lab

Storage Workstation
Rotational Machining
Workstation
Material
Transport
Cart Horizon V
Vertical Mill

~
IBM 7535 D Part Inverter
Material
Transport

Prismatic Machining
Cart
U
Fanuc Ml·L
Daewoo
Puma
Turning
Center
Workstation Buffer

Material
AssemblY

~
Transport
Cart Q Workstation
Material
Transport
Cart

IBM 7545
Cartrac Unit Conveyor
Transport System

Figure 6 - PSU CIM Lab.

198

appears to eliminate the state space explosion problems associated with modeling system states [Smith,
1992]. Since both the CFG approach and the MPSG approach deal with the development of only the
execution portion of the shop floor controllers, work dealing with the development of planning and
scheduling systems which work with the execution systems is currently under way.

Acknowledgements
This work was partially supported through National Science Foundation (NSF) awards DDM-
9009270 and DDM-9158042, both awarded to Dr. Sanjay Joshi.

References
Biemans, F., and Blonk, P., "On the Formal Specification and Verification of CIM Architectures Using
LOTOS," Computers in Industry 7, pp. 491-504, 1986.

Brooks, C. A., "A Method for Specifying, Implementing, and Verifying Media Access Protocols,"
IEEE Control Systems Magazine, Vol. 10, No.4, pp. 87-94, June 1990.

Bruno, G. and Marchetto, "Process Translatable Petri Nets for the Rapid Prototyping of Control
Systems," IEEE Transactions on Software Engineering, Vol. SE-12, No.2, pp. 346-357, 1987.

Chaar,1. K., A Methodology for Developing Real-Time Control Software for Efficient and Dependable
Manufacturing Systems, Ph.D. Thesis, University of Michigan, 1990.

Graham, 1.H., and Saridis, G.N., "Linguistic Decision Structures for Hierarchical Systems," IEEE
Transactions on Systems, Man and Cybernetics, SMC-12, pp. 325-333, 1982.

Hadj-AIouane, N. B., Chaar, 1., Naylor, A., and Volz, R. A., Material Handling Systems as Software
Components: An Implementation, Technical Paper, Center for Research on Integrated Manufacturing,
The University of Michigan, May 1988.

Hadj-Alouane, N. B., Chaar, J .• and Naylor, A., The Design and Implementation of the Control
Software of the Prismatic Machining Cell, Technical Paper. Center for Research on Integrated
Manufacturing, The University of Michigan, March 1990.

Jones, A. and Saleh, A., "A MuIti-level/MuIti-layer Architecture for Intelligent Shopfloor Control,"
International Journal of Computer Integrated Manufacturing, Vol. 3, No.1, 1992, pp. 60-70.

Joshi, S. B, Mettala, E. G., and Wysk, R. A., CIMGEN - A Computer Aided Software Engineering
Tool for Development of FMS Control Software, lIE Transactions, Vol 24, No. 3, 1992, pp. 84-97.

Joshi, S. B., Wysk. R. A., and Jones, A., A Scaleable Architecture for CIM Shop Floor Control,
Proceedings of CIMCON '90, A. Jones Ed .• National Institute of Standards and Technology, May
1990, pp. 21-33.

Kasturia. E .• DiCesare, P. and Desrochers. A.. "Real Time Control of Multilevel Manufacturing
Systems Using Colored Petri Nets," Proceedings of IEEE International Conference on Robotics and
Automation, pp. 1114-1119, 1988.
 199

Maimon, O.Z., and Fisher, E.L., "An Object-based Representation method for a Manufacturing Cell
Controller," Arti/iciallntelligence in Engineering, Vol. 3,No. I, 1988.

Merabet, A.A., "Synchronization of Operations in a Flexible Manufacturing Cell: The Petri Net
Approach," Journal of Manufacturing Systems, Vol. 5, No.3, pp. 161-196, 1985 .

Mettala, E. G., Automatic Generation of Control Software in Computer Integrated Manufacturing,

Ph.D. Thesis, The Pennsylvania State University, 1989.

Narahari, Y. and Viswanadham, N., "A Petri Net Approach to the Modeling and Analysis of Flexible
Manufacturing Systems," Annals of Operations Research, Vol. 3, pp. 449-472,1985.

Naylor, A. W. and Maletz, M. C, The Manufacturing Game: A Formal Approach to Manufacturing

Software, IEEE Transactions on Systems, Man, and CybernetiCS, Vol. SMC-16, No. 3, May/June 1986,
pp. 321-334.

Naylor, A.W., and Volz, R.A., "Design ofIntegrated Manufacturing System Control Software," IEEE
Trans on Systems , Man, and Cybernetics, Vol. SMC-17, No.6, pp. 881-897, NovlDec 1987.

Ostroff,1. S., "A Logic for Real-Time Discrete Event Processes," IEEE Control Systems Magazine,
Vol. 10, No.4, pp. 95-102, June 1990.

Sharbach, P. "Formal Methods and the Specification and Design of Computer Integrated
Manufacturing Systems," Proceedings or the International Conference on the Development of Flexible
Automation Systems, July 1984.

Smith, J. S., A Formal Design and Development Methodology for Shop Floor Control in Computer
Integrated Manufacturing, Ph.D. Thesis, The Pennsylvania State University, 1992.

Smith,1. S., Hoberecht W. C, and Joshi, S. B., "A Shop Floor Control Architecture for Computer
Integrated Manufacturing," Working Paper Series, Department of Industrial and Management Systems
Engineering, The Pennsylvania State University, 1992.

Sohlenius, G., "Scientific and Structural Base of Manufacturing," Robotics and Computer Integrated
Manufacturing, Vol. I, No. 3/4, pp. 389-396, 1984.

Suh, N. P., "Development of the Science Base for the Manufacturing Field Through the Axiomatic
Approach," Robotics and Computer Integrated Manufacturing, Vol. I, No. 3/4, pp. 397-415, 1984.

Tam, KY., "Linguistic Modeling of Flexible Manufacturing Systems", Journal of Manufacturing

Systems, Vol. 8, No. 2, pp. 127-137, 1989.

Upton, D.M. and Barash, M.M., "A Grammatical Approach to Routing Flexibility in Large
Manufacturing Systems," Journal of Manufacturing Systems, vol. 7, pp. 209-221, 1988.

Wiliams, DJ. and Upton, D.M., "Syntactic Models of Manufacturing Processing and Control,"
International Journal of Computer Integrated Manufacturing, vol. 2, no. 4, pp. 229-237, 1989.

Wu, H.L., Venugopal, R., and Barash, M.M., "Design of a Cellular Manufacturing System: A
Syntactic Pattern Recognition Approach," Journal of Manufacturing Systems, vol. 5, pp. 81-87, 1986.
200

Xou, Mo, DiCesare, F. and Desrochers, Ao, "A Top-down Approach to Systematic Synthesis of Petri
Net Models for Manufacturing Systems," Proceedings of IEEE International Conference on Robotics
and Automation, ppo 534-539, 19890
 Candidate Rule Selection to Develop Intelligent Scheduling Aids for
Flexible Manufacturing Systems (FMS)

by

Yuehwern Yih Albert T. Jones

School of Industrial Engineering National Institute of Standards
Purdue University and Technology (NIST)
1287 Grissom Hall, Room 228·A Building 220, Room B310
West Lafayette, IN 47907, USA Gaithersburg, MD 20899, USA

Abstract

In this paper we propose an intelligent, real-time, computer-based decision aid to effectively

schedule jobs through a flexible manufacturing system (FMS). One way to develop such an
aid is to use a simulation technique to collect information about the impact of heuristic rules
on the system performance. The resulting information forms the basis of an inductive
learning process for choosing a given rule in a given situation. Since there are more than 100
commonly used heuristic rules, this is an impractical approach for analyzing all the rules and
their combinations. This paper advocates a three step systematic approach to real-time
scheduling: quick analysis to select a small number of candidate heuristics, followed by a
more thorough analysis of those candidates to generate a schedule, and the development of an
automated scheduling system. In this paper we concentrate on steps one and three. We utilize
neural networks for candidate rule selector and trace-driven knowledge acquisition (IDKA)
for developing the automated scheduling system. An example of training data for the rule
selection network is given. Also, an example is provided to demonstrate the learning process
through TDKA.

1. Introduction

In the past two decades, manufacturing systems have moved toward more and more
automated production, a larger variety of products, and smaller lot sizes. To efficiently and
effectively control such manufacturing systems, intelligent, computer-based decision aids are
needed. The approaches to develop such aids include optimization (YIH and THESEN 1991),
rule-based systems (THESEN and LEI 1986, THESEN et al. 1987, YIH 1988, 1990b), genetic
202

algorithm (AYTUG et al. 1992), neural networks (CHRYSSOLOURIS et al. 1990, 1991, WANG
and YIH 1992), and their hybrids (YIH 1990a, 1992, YIH et al. 1991, 1992). Typically,
decisions (such schedules) are generated from the knowledge contained in a knowledge base.

The ftrst step in developing the required knowledge base is knowledge acquisition. This itself
is a two step process: get the knowledge from knowledge sources and store that knowledge in
a digital form. Much work has been done in the area of knowledge acquisition, such as
protocol analysis, interactive editing, and so on (WATERMAN and NEWELL 1971, NISBETI
and WILSON 1977, BAINBRIDGE 1979, GEVARTER 1983, HAWKINS 1983, ERICSSON and
SIMON 1984, HART 1985). In most cases, knowledge sources are "human experts" who
describe or explain the process of decision making. There are two problems with this
approach. First, "domain experts" cannot accurately describe the process of decision making.
Second, "experts" do not exist for every domain. We assert that scheduling falls into the latter
category. Because of this, we often resort to a "computer-based expert" to help us analyze the
impact of scheduling rules on system performance - discrete-event simulation. In this
approach, simulation results can be used to form the knowledge base through inductive
learning processes. However, with more than 100 commonly used heuristic rules, it is
impractical to study all the rules and their combinations through simulation.

To shorten the lead time for developing intelligent scheduling aids, it is imperative that the
knowledge acquisition process be quick and easily reftned as new knowledge becomes
available. In addition, this process must fit into an overall framework for real-time
scheduling. We adopt the three phase framework described in (JONES and RABELO 1992). In
phase one, a small set of candidate scheduling heuristics is selected based on a quick analysis
of the "system". In phase two, a more detailed analysis of the selected rules is performed
using discrete-event simulation. In phase three, a genetic algorithm is used to improve the
schedules from phase two. We propose a fourth phase which employs inductive learning
techniques to "learn" from those decisions. During the inductive learning process, decisions
that perform well will be added to the knowledge base, thereby increasing the "intelligence"
of the scheduler. The goal is to eventually develop an automated, real-time scheduler.

In this paper, we will concentrate on phase one and phase four, that is, the initial rule
selection and inductive learning. The remainder of this paper is organized as follows. A brief
review of the generic framework of JONES and RABELO (1992) is presented in next section.
In section three, the framework of rule selection and improvement is proposed. Section four
describes the development of the rule selectors. Section ftve discusses the learning algorithm
that is used in phase four. This is followed by a summary and references.
 203

2. Generic Intelligent Controller

This research is built upon the generic framework of intelligent controller proposed by
(DAVIS et al. 1991) as shown in Figure 1. This framework has its foundations in (ALBUS et
al. 1989, JONES and MCLEAN 1985, JONES and SALEH 1990). It performs four production
management functions: assessment, optimization, execution, and monitoring.

Tasks Dnd times

from superulsor

Problem
formulDtion

Selected

Current Control Schedul

I NTERFRCE
MODULES 1-----...
Figure 1. Generic Control Architecture (DAVIS et al. 1991)

The Assessment Function (AF) formulates the planning and scheduling problems facing the
controller based on three sets of inputs. (These can be thought of as optimization problems,
although they are not solved using formal mathematical programming techniques.) The first
set is specified by the supervisor and includes assigned tasks, priorities, due dates, and any
other constraints or performance measures. The second set is the feedback from the
Monitoring Function (MF) that describe the current system state and the projected system
response with current control rules. The third input comes from the Optimization Function
(OF) indicating the current status of its attempts to generate "feasible/optimal" solutions. The
204

AF specifies the perfonnance criteria for each optimization problem to be considered by the
OF. In addition, the AF constructs constraints to be imposed as part of those optimization
problems.

The Optimization Function is responsible for selecting a run-time production plan and
scheduling rule. The production plan is selected from those feasible process plans passed
down by the Assessment Function. In general, there are no a priori limitations on the choice
of scheduling rules, but the framework allows for this if it is desirable. The selection is made
which optimizes the current set of perfonnance measures. In addition, the Optimization
Function must deal with violations of the scheduled start/finish times and any other
constraints imposed by the Assessment Function. In this paper, we concentrate on a method
for implementing the Optimization Function which we describe in the next section.

The Execution Function implements the decision rules selected by the Optimization Function.
Currently, it is envisioned to work as follows. Using the current state of the system, it does a
single pass simulation using the selected production plan and scheduling rule to compute the
starting and finishing times for each task. These become the "due dates" which are passed
down to subordinates. In addition, when minor deviations from these times are envisioned by
the subordinates, it can try to restore feasibility using techniques such as perturbation analysis
or match-up scheduling (JONES and SALEH 1990).

Lastly, the Monitoring Function updates the system state using feedback from subordinates,
and evaluates proposed subordinate responses against the current set of imposed (from the
Assessment Function) and computed (by the Optimization Function) constraints. During error
recovery, the Monitoring Function acts as an expediter. That is, the Monitoring Function
determines when violations occur, their severity, and who should deal with them.

The remainder of this paper describes a strategy for implementing the scheduling perfonn~
by the Optimization Function.

3. Framework of Rule Selection and Improvement

The learning-based framework consists of four phases as shown in Figure 2. Phase one deals
with initial candidate rule selection. Phase two conducts simulation based on the selected rule
candidates. Phase three analyzes the results from simulation, and applies Genetic Algorithm
(GA) for further improvement. These are described in detail in JONES and RABELO (1992).
We have added a phase four in which knowledge is extracted from the resulting schedules by
 205

an inductive learning algorithm. The outcome is then used to improve future invocations of
phase one. This paper focuses on the initial selection of candidate rules and the learning
process.

After receiving the production constraints and performance criteria from the Assessment
Function, an initial candidate rule selector will be activated. Based on the desired system
performance, the rule selector will select candidates that are more likely to fulfill the
requirement and pass them to the simulation phase.

Desired Production
Heuristic Rules Requirements

Candidate Rule
Selector

Resulted Schedule

Inductiue Learning

Figure 2. Proposed Learning-Based Approach for Optimization Function

The simulation phase uses the candidates selected in the first phase and any additional
constraints from the Assessment Function to generate feasible schedules. These simulations
are initialized to the current state of the system. The performance of each schedule is
evaluated against the imposed criteria. Those schedules performing well based on the criteria
will be selected as the basis for further improvement in the third phase.

Phase three is responsible for improving the quality of schedules when possible. This phase is
needed because it is very unlikely to have a single scheduling heuristic that can optimize all
the objectives (some of which can be negatively correlated) simultaneously. In this phase,
206

new schedules are generated based on those good schedules passed from the second phase by
using Genetic Algorithms (GA). Among those schedules generated, good ones will be
"genetically altered" to generate new and better schedules. This process of regeneration
iterates until no significant improvement is found. This is the schedule that gets implemented
by the Execution Function.

At the conclusion of the improvement phase, one "good" schedule has been generated for the
current system and performance measures. We would like to learn from this experience to
avoid a repeat of the process the next time similar conditions arise. However, we cannot
easily derive a simple "strategy" form which to generate this schedule (like FIFO, SPT etc.)
because the schedule is a complex combination of such rules. There are, then, two questions
to be answered. How do we determine the relationship between the schedule we derived and
the circurpstances which led to that schedule? How do we efficiently represent this
information for later use?

To extract the scheduling strategy, we propose to use an inductive learning algorithm- Irace-
I!riven Knowledge Acquisition (TDKA) (YIH 1988, 1990b). This algorithm results in the
representation of the strategy as a rule base. The result will be a new, albeit somewhat
complicated, heuristic which can be placed in the knowledge base and which can be used the
next time the candidate rule selection module is invoked.

In the following sections, we will discuss the development of the rule selector in the first
phase and the learning algorithm used in the phase four.

4. Development of Rule Selector

In JONES and RABELO (1992), the authors propose to use a (possibly more than one) neural
network to perform the initial selection of candidate rules. This neural network has three
inputs: scheduling rules, system configuration, and performance measures. The output of the
network is the matching score assigned to each heuristic rule. Those rules with higher scores
will be selected as rule candidates for further simulation study and probably improvement.

Neural networks, also called connectionist or distributed parallel processing, have been
studied for many years in an attempt to mirror the learning and prediction abilities of human
beings. Neural network models are distinguished by network topology, node characteristics,
and training or learning rules. An example of a three-layer feed-forward neural network is
shown in Figure 3.
 207

Input Layer

Hidden Layer

••• Output Layer

Figure 3. An example of a three-layer feed-forward neural network

Through exposure to "training patterns", neural networks attempt to capture the salient
information. PARKER (1985) and RUMELHART et al. (1986) showed how to learn by using
back-propagation. WERBOS (1974) developed the back-propagation training algorithm in his
doctoral dissertation, but his work was unknown for almost 10 years. Back-propagation
applies the gradient-descent technique in the feed-forward network to change a collection of
weights so that the cost function can be minimized. The cost function, which is only
dependent on weights and training patterns, is defined by:

C(W) = 1 / 2 L (Tip - Oip) (1)

where the T is the target value, 0 is the output of network, i is the output nodes, and p
is the training patterns.

After the network propagates the input values to the output layer, the error between the
desired output and actual output will be "back-propagated" to the previous layer. In the
hidden layers, the error for each node is computed by the weighted sum of errors in the next
layer's nodes. In a three-layered network, the next layer means the output layer. If the
activation function is sigmoid, the weights are modified according to

(2)
or
(3)
where Wij is weight from node i to node j, 11 is the learning rate, Xj is the output of
node j, Tj is the target value of node j, 8k is the error function of node k.
208

If j is in the output layer, relation (2) is used, relation (3) is for the nodes in the hidden layers.
The weights can be updated to reduce the cost function at each step. The back-propagation
has been successfully applied in many situations. But, we note that the problem of training a
three-layer neural network is NP-complete (BLUM and RIVEST 1988, JUDD 1987).

Processing Time Matching Score

Variation

Congestion Level FIFO

PTIMH Ratio
SPT

Average
Tardiness

LIFO
Maximal
Tardiness

Mean Flow EDD

Time

Maximal
Waiting Time

Figure 4. A neural network for rule candidate selector

To train the neural network for candidate rule selection, we need to collect training data set
from simulation studies. Three factors will be used to describe the system configuration:
processing time variation, congestion level, and ratio of processing times and material
handling (PT/MH). The performance measures for production requirements will be selected
from related literatures. An example of the rule selector network is shown in Figure 4. The
jobs in the system will be randomly generated to be scheduled through the system on a real-
time basis. Simulation will be conducted based on fractional factorial designs. The values of
performance measures will be compared with those of First-In-First-Out (FIFO) rule and the
ratio values will be used as inputs to the network. The output values of the rule candidate
selector will be the matching score of each heuristic rule. A threshold will be set by the user
to select the promising rule candidates for further improvement.
 209

For instance, using the simulation results listed in Table 1, the indices of four performance
measures are (1.4, 1.79,0.9, 1.67) and (1.1, 2.14, 1.04,3.54) for SPT and LIFO respectively.
Similar simulation studies could be conducted based on different levels of processing time
variance, congestion level, and PT/MH ratio for a given rule by using fractional factorial
design.

T a ble 1 Sampl e resuIts f rom SimuIatlon

.
System Configuration Performance Measures
Rule Proc. T Congo PT/MH Avg. Max Mean Flow Max Waiting
Variation Level Ratio Tardiness Tardiness Time Time
FIFO 1.0 0.7 10.0 5.0 14.0 80.0 24.0
SPT 1.0 0.7 10.0 7.0 25.0 72.0 40.0
LIFO 1.0 0.7 10.0 5.5 30.0 83.0 85.0
EDD 1.0 0.7 10.0 1.2 12.0 90.0 45.0
where SPT = Shortest Processmg TIme; LIFO = Last In FIrst Out; EDD = Earhest Due Date

The matching score (M-score) of rule k compared with the performances of rule r is
calculated by the following formula.

M-score(k, r) =400 - 100*Lj max {(Kj - Rj), o}

where Kj = the index of i th performance measure of rule k
Rj = the index of i th performance measure of rule r
If we use the simulation results of FIFO as an example, M-score(FIFO, FIFO) will be 400
(perfect match) because the performance measures are resulted from FIFO. For the same
performance of (1.0, 1.0, 1.0, 1.0), M-score(SPT, FIFO) is equal to

400-100* [max( 1.4-1.0, O)+max( 1.79-1.0, 0)+max(0.9-1.0, 0)+rnax(1.67 -1.0, 0)]

= 400 - 100*[0.4+0.79+0+0.67] =400 - 186 = 214.
The training data transformed from the simulation results in Table 1 are listed in Table 2.

T able 2 Trammg
.. d ata f or ru e se ector network
Input Vector Output Vector
(1.0,0.7, 10.0, 1.0, 1.0, 1.0, 1.0) (400,214, 18,299)
(1.0,0.7, 10.0, 1.4, 1.79,0.9, 1.67) (390,400,164,356)
(1.0,0.7,10.0,1.1,2.14,1.04,3.54) (400,370,400,391)
(1.0,0.7, 10.0,0.24,0.86, 1.13, 1.88) (310, 191,20,400)
 210

5. Inductive Learning Algorithm - TDKA

Trace-driven knowledge acquisition (TDKA) (YIH 1988, 1990b) is a method that extracts
knowledge from a trace of the effect of decisions instead statements or explanations of
decisions. There are three steps in the process of trace-driven knowledge acquisition: data
collection, data analysis, and rule evaluation. An overview of the model of trace-driven
knowledge acquisition is given in Figure 5.

Figure 5. Model of trace-driven knowledge acquisition

In step one, data is collected through simulation, historical records, human inputs, actual
physical experiments, etc. In the proposed framework, the data (trace) is simply the schedule
generated from the Genetic Algorithm. In step two, the trace is analyzed to determine
whether or not any production rules (If ... Then ... ) are embedded in the trace. In step three,
simulation is used to compare the performance of the resulting rules with the trace (the
schedule). The process returns to step two to refine the rules if the resulted rules do not
perform as well as the original schedule.

The core of the TDKA lies in the step two- data analysis. There are two types of rules
involved -- decision rules and class assignment rules. If the state space is viewed as a hyper
plane, class assignment rules draw the boundaries on the hyper plane to define the areas of
classes as shown in Figure 6. For each class, a single decision rule is used to determine what
to do next. The resulting scheduling rules are in the form of
"If [state E class i] then apply Ri* "
 211

State Space

R2
Figure 6. State space hyper plane(Rm is the decision rule applied in class m)

In data analysis, the records collected are used to define classes and to select a single rule for
each class. The rule formation algorithm can be described by three steps as follows. (The
terms used in the algorithm are defined in Appendix A.)

I. Initialize
Detennine state variables Vi (i=1,2, ...,Ny )
Detennine decision rules Rj (i=1,2, ...,Nr)
Set initial value for L
Detennine initial classes Cj (i=1,2, ... ,Nc)
Detennine initial class assignment function F
Initialize class frequency counter Yj(i=1,2, ... ,Nc)
Initialize rule frequency counter Zik (i=1,2, ...,Nc, k=1,2, ... ,Nr)
Obtain a trace (Sj, aj) G=1,2, ...,Ns)
II. Determine the decision rules used by the user.
For each record (Sj, aj)
A. Determine what class Sj belongs to.
F(S-) = i
B. Find all decision rules RK that yield ajfrom Sj-
RK(Sj) = Aj S.t. aj E Aj
C. Update class and rule frequency counters.
Yj =Yj + 1
Zjk = Zik + 1 for all k E K
III. Evaluate the decision rules in each class.
For each class Cj
A. Determine the score for each decision rule.
Zik
X ik = Y. for k=1,2, ... ,Nr
1
B. Select the decision rule Rk* with the highest score.
Xik* = maxk {Xik}
C. Check whether the selected decision rule satisfies the minimal score
requirement.
If Xjk* ~ L then goto step Ill.G
212

D. Split this class into two classes.

1. Detennine the critical level for dividing this class into two classes.
a. Set up the testing rule with all possible variables Vp.
If [Vp .op. Dip] then apply Rk*
else apply oilier decision rules.
where Dip is a value of Vp
.op. E {~,<}
P E {l,2,... ,Nv }
b. Detennine the number of records (U~ that cannot be explained by
the testing rule for each possible value {Vip) of each variable (Vp).
c. Select the value D*il?* with the minimal Uip as the critical level.
U*ip*= minp {UipJ
2. Split class Ci into two classes based on the selected critical level D*ip*'
a. Add two class assignment rules.
IF [SjECi] and [Vp*<D*ip*] 1HEN [SjECil]
IF [SjECjJ and [Vp*~*ip*] 1HEN [SjECi2]
b. Update the class aSSIgnment function.

°°
E. Reset class and rulejrequency counters.
Yi =
Zuc =
for i=I,2, ... ,Nc
for i=I,2, ... ,Nc, k=I,2, ... ,Nr
F. Goto step II.
G. Form the scheduling rulejor this class.
If [SjE Ci] then apply Rk*
H. Stop.

This iterative process will stop whenever the rule formation is completed. That is, the state
space has been properly divided into several classes and a single decision rule is selected for
each class. If the performance of the extracted rules is worse than the trace, then it is
necessary to return to data analysis and refine the rules by rising the value of acceptance level
(L). Otherwise, the process stops.

For example, consider a flowshop problem in a cell with five workstations and one material
handling robot (Figure 7). Individual workstations have no buffers. There is, however, one
input and one output buffer for the cell. The robot is responsible for transporting jobs from
the input buffer to workstations sequentially and eventually to the output buffer. Jobs arrive
the cell randomly and the batch size is one. The problem is to select the job from the input
buffer or from the workstations for transportation and process so that certain production
requirements are fulfilled. If we represent the schedule by the identification number of the
workstation, a schedule of 3 -> 5 -> 2 -> 1 -> ° -> 4 means that the robot will transport the

°
job in workstation 3 to workstation 4, and then from workstation 5 to 6, etc. In this
representation, stands for the input buffer and 6 the output buffer. If the resulting schedule
from Genetic Algorithm is 3 -> 5 -> 2 -> 1 -> °-> 4, we can obtain the trace from simulation
in the following format.
(R, W_0, W_1, RPT_l, W _2, RPT_2, W _3, RPT_3, W _4, RPT_4, W _5, RPT_5) ---> Action
 213

where
°
R: current location of robot, range = {0... 6} where is the input buffer, 6 is the output
buffer, and the others are the respective workstation numbers
W _0 : feasibility of new entry

°
= 1 if new entry is feasible
= otherwise
W _i : current workstation status (for i = 1,2,... ,5)
=0 if empty
= job identification number if occupied by a job
RP'Ci : (for i = 1,2, ... ,5)

°
= remaining process time if the workstation is occupied by a job
= otherwise
Action: the identification number of the workstation where the job will be selected for
transportation next.

Workstation

Robot 1 1#411

m I/O
Buffer
Figure 7. A five-workstation cell

For instance, the trace may look like the following.

(2, 0, 124, 12, 123, 0, 122, 0, 0, 0, 120, 30) --> 3

(4, 0, 124, 8, 123, 0, 0, 0, 122, 112, 120, 26) --> 5
(6, 0, 124, 0, 123, 0, 0, 0, 122, 47, 0, 0) --> 2
(3, 0, 124, 0, 0, 0, 123, 85, 122, 26, 0, 0) --> 1
(2, 1, 0, 0, 124, 50, 123, 55, 122, 0, 0, 0) --> 0
(1, 0, 125, 40, 124, 20, 123, 25, 122, 0, 0, 0) --> 4

Each record in the trace will vote for the rules that could yield the same decision as in the
schedule. Of course, the feasibility of the movement will also be considered. In other works,
if the workstation can process one job at a time, due to no buffer among workstations, the job
on workstation #2 cannot be moved to #3 until #3 is empty. The first record, (2,0, 124, 12,
123,0, 122,0,0,0, 120,30), there are only two feasible movements: #3 to #4 and #5 to #6.
214

Based on the schedule, the job on workstation #3 is selected for transportation and this record
will vote for the rules such as, SPT (Shortest Processing Time), CB (Closest to the I/O
Buffer), and CR (Closest to the Robot), etc. The fifth record will vote for the rules such as
CB, CR, FR (Farthest from the Robot), and N (New job first). The summary of votes is listed
in the following table.

where LPT =Longest Processing Time; FB =Farthest from the I/O Buffer

We start with one class, called Class 1. If we are satisfied with the accuracy of 83.33%, we
may arbitrarily choose CB or CR and form the following rule.
"If [state E Class 1] then apply CB" (1)
If we would like to obtain higher accuracy, the variable W_5 can be used to split Class 1 to
Class 11 and 12 where Class 11 includes the states with W_5=0, and Class 12 has the
remainder. The following class assignment rules will be formed.
"If [W_5 *" 0] then state E Class II" (2)
"If [W_5 = 0] then state E Class 12" (3)
After splitting into two classes, the voting process repeats within each class. The results are
listed in Tables 4 and 5. The following two rules could be formed with 100% accuracy. The
process stops and the rule base contains four rules: (2), (3), (4), and (5).

"If [state E Class 11] then apply CR" (4)

"If [state E Class 12] then apply CB" (5)

Table 4. Summary 0 f the record votes (CIass 11)

Rule Candidate # of Votes Percentage
LPT 1 50%
SPT 1 50%
CB 1 50%
FB 1 50%
CR 2 100%
FR 0 0%
N 0 0%
 215

Table 5. Summaryof the reco~d votes (CIass 12)

Rule (andidate # of Votes Percentage
LPT 1 1?~
:sPT j 75~
CB 4 lUU%
FB ~ 50%
~R ~ 75%
FR 4 ~
N 1 ~5%

6. Remarks

In this paper, we proposed a strategy for implementing rule selection and inductive learning
for an intelligent scheduling aid for FMSs. Trace Driven Knowledge Acquisition and neural
networks were proposed Currently, the proposed framework is under implementation. We
will be able to verify the feasibility of the framework in the near future.

Reference

ALBUS, J., MCCAIN, H. and LUMIA, R., "NASA/NBS Standard Reference Mode for
Telerobot Control System Architecture," NIST Technical Note 1235, National Institute
of Standards and Technology, Gaithersburg, MD, April 1989.
AYTUG, H., KOElll.ER, G. J. and SNOWDON, J. L., "Genetic Learning of Dynamic Scheduling
within a Simulation Environment," Workshop on Learning Approaches to Scheduling,
Gainsville, FL, January 17-18, 1992.
BAINBRIDGE, L., "Verbal Reports As Evidence of the Process Operator's Knowledge,"
International Journal of Man-Machine Studies, 11,411-436,1979.
BLUM, A., and RIVEST, R. L., "Training a 3-Node Neural Network is NP-complete,"
Advances in Neural Information Processing Systems I, ed. by D. S. Touretzky, San
Mateo: Morgan Kaufmann, 494-501,1988.
CHRYSSOLOURIS, G., LEE, M. and DOMROESE, M., "The Use of Neural Networks in
Determining Operational Policies for Manufacturing Systems," Journal of
Manufacturing Systems 10,166-175,1991.
CHRYSSOLOURIS, G., LEE, M., PIERCE, J. and DOMROESE, M., "Neural Networks for Design
of Manufacturing Systems, " Manufacturing Review 3, 187-194, 1990.
DAVIS, W., JONES, A. and SALEH, A., "A generic architecture for intelligent control
systems," Computer Integrated Manufacturing Systems, to appear.
ERICSSON, K. A. and SIMON, H. A., Protocol Analysis: Verbal reports as Data
,Massachusetts MA: The MIT Press, 1984.
GEVARTER, W. B., "Expert Systems: Limited But Powerful," IEEE Spectrum, Aug., 39-45,
1983.
HART, A., "The Role of Induction in Knowledge Elicitation," Expert Systems, 2, 1,24-28,
1985.
HAWKINS, D., "An Analysis of Expert Thinking," International Journal of Man-Machine
Studies, 18, 1-47, 1983.
216

JONES, A. and MCLEAN, C., "A Production Control Module for the AMRF," Proceedings of
the International Symposium on Computers in Engineering, Boston, MA, August 1985.
JONES, A. and RABELO, L., "Real-Time Decision Making Using Neural Networks,
Simulation, and Genetic Algorithm," Proceedings of the International FAIM'92
Conference, Fairfax, VA, 1992.
JONES, A. and SALEH, A., "A Multi-levellMulti-layer Architecture for Intelligent Shop Floor
Control," I nternational Journal of Computerlntegrated Manufacturing Special Issue on
Intelligent Control, 3,1,60-70, 1990.
JuDD, J. S., "Learning in Networks is Hard," Proceedings of the First International
Conference on Neural Networks (San Diego 1987), II, New York: IEEE, 685-692,
1987.
NISBETI', R. E. and WILSON, T. D., "Telling More Than We Can Know: Verbal Reports on
Mental Processes," Psychological Review, 84, 231-259, 1977.
PARKER, D. B., "Learning Logic," Technical Report TR-47, Center for Computational
Research in Economics and Management Science, Massachusetts Institute of
Technology, Cambridge, MA, 1985.
RUMELHART, D. E., HINTON, G. E. and WILLIAMS, R. J., "Leaming Internal Representations
by Error Propagation," in Parallel Distributed Processing, 1, Cambridge, MA: MIT
Press, 318-362, 1986. •
THESEN, A. and LEI, L., "An Expert System for Scheduling Robots in a Flexible
Electroplating System with Dynamically Changing Workloads," Proceedings of the
Second ORSAITIMS Conference on FMS: Operations Research Models and
Applications, edited by K. E. Stecke and R. Suri (Amsterdam: Elsevier), 555-566, 1986.
THESEN, A., LEI, L. and YIH, Y. "Knowledge Acquisition Methods for Expert Scheduling
Systems," Proceedings of the 1987 Winter Simulation Conference, 709-714, December,
1987.
WANG, J. and YIH, Y. "Using Neural Networks to Select Control Strategy For Automated
Storage and Retrieval Systems (AS/RS)," Technical Report, School of Industrial
Engineering, Purdue University, West Lafayette, IN, 1992.
WATERMAN, D. A. and NEWELL, A., "Protocol Analysis as a Task for Artificial
Intelligence," Artificial Intelligence, 2, 285-318, 1971.
WERBOS, P., Beyond Regression: New Tools for Prediction and Analysis in the Behavioral
Sciences. Ph.D. Thesis, Harvard University, Cambridge, MA, 1974.
YIH, Y., Trace Driven Knowledge Acquisition for Expert Scheduling System, Ph.D.
dissertation, University of Wisconsin-Madison, December, 1988.
YIH, Y. "Learning Scheduling Rules for FMS from the Optimal Policy of User-Based Semi-
Markov Decision Processes," The Fourth International Conference on Expert Systems
in Production and Operations Management, Hilton Head Island, South Carolina, 175-
183, May 14-16, 1990a.
YIH, Y., "Trace-Driven Knowledge Acquisition (TDKA) for Rule-Based Real-Time
Scheduling Systems," Journal of Intelligent Manufacturing, 1,4,217-230, 1990b.
YIH, Y., "Leaming Real-Time Scheduling Rules from Optimal Policy of Semi-Markov
Decision Processes," International Journal of Computer Integrated Manufacturing, to
appear, 1992.
YIH, Y., LIANG, T. P. and MOSKOWITZ, H. "A Hybrid Approach for Crane Scheduling
Problems," Proceedings of Artificial Neural Networks in Engineering (ANNIE)
Conference, St. Louis, MO, 867-872, November 10-13, 1991.
YIH, Y., LIANG, T. P. and MOSKOWITZ, H. "Robot Scheduling in A Circuit Board Production
Line- A Hybrid OR/ANN Approach", lIE Transactions, to appear, 1992.
YIH, Y. and THESEN, A., "Semi-Markov Decision Models for Real-Time Scheduling,"
International Journal of Production Research, 29, 11,2331-2346, 1991.
 217

Appendix A. Definition of terms in TDKA

(1) ~ «Sj, aj), j=I,2, ...,Ns) - Records contains the data recorded during the problem-
solving process. Each record describes a system state (Sj) and the action taken by the
subject (aj). For instance, if the system state is defined as a vector of robot location,
status of machine#1 and status of machine#2, and the action is defined as the index of
machine to be served next, record «1,1 ,0), 2) indicates that machine#2 is selected and
that the system state is (1,1 ,0).
(2) ~ - The trace is a sequence of records collected from the subject during the problem-
solving process in phase one.
(3) State Variable (Vj, i=I ,2,... ,N v ) - State variables are the variables defined in the state
vector, e.g. robot location, machine status.
(4) Decision Rule (Rj, i=1,2,... ,Nr ) - Decision rules are the rules to select the alternative based
on the rank order of state variables, e.g. closest to the output first, shortest remaining
process time first. Each decision rule Rj will yield a set of possible actions (Aj) from a
system state (Sj), that is, Rj(Sj) = Aj.
(5) Class (Cj, i=I,2, ... ,Nc) - A class is a subset of the state space. For any Cj and Cj, if i~j then
CjrlCj = 0 . The union of all classes is the state space.
(6) Class Assignment Rule - Class assignment rules are rules defining the areas of classes,
e.g. If [robot location=2] then SjE C2.
(7) Class Assignment Function (F) - The class assignment function computes a class index
from a system state, e.g. (F(Sj)=i) ~ (SjE Cj).
(8) ~ (Xi]c, i=1,2,... ,Nc, k= 1,2,... ,Nr ) - «F(Sj)=i) and (Rk(Sj)=Aj) S.t. ajE Aj) ~ (Sj,aj) can
be explained by Rk in Cj. The score (Xjk) is the percentage of records which can be
explained by decision rule Rk in Cj.
(9) Acceptance Level (L) - The acceptance level is the percentage indicating the minimal
score that a rule must have before it is accepted as the rule that explains the subject's
decision.
(10) Critical Level (D*jj*' i = 1,2, ... , Nc, j*E (I ,2, ... ,Nv }) - The critical level (D*ij*) is the
value of a state variable Vj* used to split class Cj into two classes, Cil and Cj2. A
member of Cj will be classified into Cjl if Vj* is below D*ij*, and will be classified into
Cj2 if!j* is above or equal to D*jj*'
(11) Class Frequency Counter (Yj,i=I ,2,... ,Nc) - The class frequency counter keeps track of
the number of records (Sj,aj) with SjE Cj.
(12) Rule Frequency Counter (Zjk, i=1,2,...,Nc, k=1,2, ... ,Nr ) - The rule frequency counter
(Zjk) keeps track of the number of records (Sj,aj) which can be explained by Rkand
SjECj.
 Tool Blocking in Flexible Manufacturing Systems
Ulrich A.W. Tetzlaff
George Mason University
Department of Decision Sciences & MIS
Fairfax, Virginia 22030-4444

Abstract
An analytical procedure to evaluate the performance of a Flexible Manufacturing
System (FMS) with a separate tool transportation system is presented. Of particular
interest is the tool blocking time, which is defined as the time a machine is idle
because it has to wait for the required tools. In order to evaluate this time, both
systems of the FMS, the parts transportation system and the tool transportation
system, are modelled as two interacting closed queueing networks.

1 Introduction
Flexible manufacturing systems obtain to a large extent their flexibility through Comput-
erized Numerical Controlled Machines (CNC machines), which are capable of manufac-
turing different part types. This ability is due to a computerized controlled tool exchange
mechanism as well as to an efficiently working tool supply system. They enable the per-
formance of a large variety of different operations at one single machine. Sometimes up to
100 or more different tools are stored locally at the machines. Because the part mix in the
system can change or the tools are worn down or broken, the local tool storage must be
supplied with new tools. For the designer of a flexible manufacturing system the question
arises as to how to specify this tool supply system in order to deliver the necessary tools
in time without acting as a bottleneck on the performance of the whole system. Thus,
the system should be designed in such a way that a CNC machine does not need to stop
processing because the required tool is not yet available, i.e. without tool blocking.
Previous work considered only the influence of the tool blocking time on the perfor-
mance of a FMS in the form of a single, constant parameter. Thus, important design
issues involved with an automated centralized tool supply system were neglected (see
Vinod and Solberg, 1984, Vinod and Altiok, 1986, Widmer and Solot, 1990).
 219

2 Problem Description
The flexible manufacturing system under consideration consists of M -1 machine groups.
Each group m consists of one or several identical machines Sm. These machines can be
either CNC machines, load/unload stations, or washing machines, etc. Of this set of
M - 1 machine groups, only a subset $ of S machine groups is required to be supplied
with tools in order to operate. For this subset $, the number of machines Sm is equal to
one at each group. The tool transportation system T M (TT System) with one or several
transportation vehicles STM provides only those m E $ machine groups with tools. The
parts transportation system M (PT System) comprises one or several automated guided
vehicles (AGVs) SM supplying the machines with parts.
There is a fixed number of pallets N circulating in the system. Without loss of
generality it IS assumed that there is only one part on each pallet. There are furthermore
a total of L tool types available for the system and the number of different tool types
at a specific station m E $ is given by Lm . The number of identical tools of type I at
station m is Kim. Each type has an expected lifetime of WI. The average time between
tool breakages is exponentially distributed with mean a.
There are two different types of tool supply orders delivered to the TT System by each
machine group m E $. The first type, called a production supply order, is sent to the
control system of the TT System rm time units (with rm 2:: 0) before the actual operation
with the required tool(s) starts. A production supply order is issued in case of part mix
changes or when a tool has been already used for the length of its expected lifetime. The
second type, called a breakage supply order, is issued in case of tool breakage. Since the
exact point in time for a tool breakage cannot be forseen, this order is issued at the same
time when the actual requirement arises, i.e. when the breakage is detected. For each
operation of type 0 with tool type I at station m the operating time tmol is required.

Notation:
a mean time between tool breakages
em number of tool types at station m which are exchanged
per transportation sequence of the tool supply system
d~ total tool blocking time for all tool breakages during a part's visit
at station m
tm tool blocking time caused by each tool breakage at station m
d~m delay at breakage-station m within the TT System T representing
the intermediate time between tool requests due to breakage
~ tool blocking time at station m related to a production supply order
4m delay at production supply station m within the TT System representing
the intermediate time between tool requests caused by production
supply orders
9m average number of tools to be exchanged at station m if the
220

part mix in the system changes

Kim number of identical tools of type I at station m
Lm number of tool types at station m
N number of pallets
nm average number of part visits to station m between part mix changes
Pm probability that a pallet visits station m
P~ probability of tool breakage(s) during a part's visit at station m
~ probability of a tool breakage at station m during the processing
of a part
ri!n probability of a production supply order caused by a part mix change
at station m
v,:. probability of a production supply order during a part's visit
at station m
P~ probability of a production supply order at station m when
a tool's lifetime is over
qm queueing time at station m
q~m queueing time at the central server (transportation system)
of the TT System T for a breakage supply order for station m
q~m queueing time at the central server (transportation system)
of the TT System T for a production supply order for station m
rm preorder time, i.e. time between ordering the tool and the
moment when it is required
Sm number of machines at machine group m
tm approximated mean operating time at station m
lm operating time at station m
t;;, mean processing time at station m
tmol operating time at station m for operation 0 with tool type I
tTm average transportation time for a tool supply to station m
Um utilization per machine at station m
Vm average number of operations per visit at station m
WI : expected lifetime of tool type I

3 The Model
To evaluate the performance of our flexible manufacturing system a model consisting
of two interacting closed queueing networks is used (see Fig.I). One queueing network
models the PT System which supplies the CNC machines with parts, the other queueing
network models the TT System which supplies the CNC machines with the required tools.
 221

PT System TT System

VTL

I e;:~up I

JProa-.~
Mill

~ Mill

rrrii ns
sy.t.
for

T
too Is

"

Fig. I . : Two interacting closed queueing networks

3.1 PT System
The PT System is modelled as a central server queueing network model with one customer
class. The machine groups are modelled as queueing stations and the transportation sys-
tem acts as the central server. The machine groups, i.e. stations m f¢ S are modelled
as single stage stations with an exponentially distributed mean processing time t~. The
remaining stations m E S requiring tool supply are approximated by a three stage queue-
ing station as shown in Fig.2. Below, these three stage stations and their parameters are
described.
 222

Fig.2: Station with three stages

The first stage is reached with probability p.:. when a tool supply is required for the
next part to be processed and a production supply order must be issued. The machine is
idle during the tool blocking time d!:. when the station is waiting for the fullfillment of the
production supply order. The second stage of the station consists of the approximated
mean operating time tm at station m. The third stage is reached with probability of p~
in case of tool breakage(s) during the processing of a part. The tool blocking time d~ is
the total blocking time for tool breakage(s) during a part's visit at station m.
The probability p.:. for a production supply order consists of the probability F'!n for a
tool supply order caused by a part mix change and the probability p::; for replacement of
a tool at the end of its expected lifetime WI. For both of the latter probabilities, rough
approximations are given in the following .

~ =~+p::; VmES (1)

The probability ~ is approximated by dividing the average number of tools changed
per part mix change gm at station m by the average number of visits nm between those
changes and the number of tools Cm exchanged per production supply order for station
m. Note that the number of tools exchanged per order should be greater than or equal to
the required average number of tools to be changed per part mix change, i.e. Cm 2: gm,
in order to avoid more than one supply order at a time.

~ =~ VmES (2)
m nmCm
The probability p::; is approximated by the following expression:
 223

,,/~
W L- wlKlm U S (
Pm= LmCm Vm vmE . 3)
It captures the probability that during a visit of an average part with the operating
time lm' a production supply order is required. The total processing time performed wi th
one tool type, i.e.
(4)
is divided by the available tool performance time w/K/m of tool type I at station m.
However, as there are Lm tool types at one station, this term exists with multiplicity Lm.
An average value is given by the summation over all tool types I and dividing it by the
number of tool types Lm. Thus the expression

(5)

gives the average probability of a tool supply of a single tool per operation. To obtain
the probability ~ of a production supply request per visit, one has to include the ratio
of the average number of operations per visit Vm to the number of tools exchanged per
request Cm, resulting in the formula in equation (3). Note that ~ can be greater than
one, if more than one tool is required per visit and the number of tools exchanged per
request Cm is smaller than this demand. Here an increase of Cm to at least the number
of tools requested per visit is necessary.
The probability P~ is obtained by modelling the tool breakage(s) by a two-stage Coxian
distribution (see Altiok and Stidham 1983). The approximate average operating time tm
is given by
(6)
and the total tool blocking time for breakage d~ by

"1m ES. (7)

The approximate probability P~ is evaluated by:

P~ = d; (l/tm + l/lm + l/a) -

-b [
1- "1m E S. (8)
224

The probability density function fm(t) for a three-stage station has the following form:

(9)

In order to obtain the mean processing time t~ , the Laplace-Stieltjes transform .em (s)
of the probaqility density function fm(t) is derived.

( 10)

(11 )

The mean processing time t~ for a part staying at a station of type m E S is then
given by:

t*m = ds
(12)

= ~~ + tm + p~d~ VmES. (13)

3.2 TT System
To estimate the tool blocking times ~ and d~, a second queueing network is used mod-
elling the TT System. It consists of 2S stations which act as pure delay stations without
any queue. Thus, there are two stations corresponding to each station m E S of the PT
System, one for each type of supply order. Furthermore, it consists of a central server
which models the tool transportation system. This server has a queue with STM servers,
i.e. vehicles . Within the TT System 2S classes of customers circulate. Thus, the TT
System is modelled as a multi-class queueing network. Each customer class has only one
customer which represents a tool supply order for one of the 2S stations of the PT Sys-
tem. An order can be either at the queue of the central server, being processed by the
 225

central server or at one of the 2S delay stations. If the order is at the queue of the tool
transportation system, it is waiting for a transportation vehicle in order to be completed
by delivering tools to one of the 2S stations. Thus, the tool blocking times d:;' and d~
can be estimated by the queueing times qtm in case of production supply orders or by
the queueing times q~m in case of breakage supply orders at the central server, and the
tool transportation time tTm to one of the two delay stations m of the network. It follows
that the the tool blocking time for breakage is:

(14)

If the fact is taken into consideration that the control system is capable of placing the
production supply order for the tools before they are actually required with preorder time
rm, the tool blocking time of the production supply order is:

cP,;, = max{O, tTm + qtm - rm} 'rim E S . (15 )

Each customer class of the TT System, i.e. each kind of tool supply order, visits only its
specific delay station. The "processing" at the delay station means the tool supply order
is actually not processed but rather at rest. The delay time represents the intermediate
time between tool requests 4m or d~m' Thus, when this time is over, the customer
comes to the central server queue and a new order is delivered to the tool transportation
system. To estimate the intermediate time between tool requests the performance of the
PT System must be known. Assuming for a moment that only one pallet is in the PT
System and that at each visit a production supply order is issued, i.e. N = 1 and p:;' = 1,
then 4m is given by:

(16)

This is the throughput time of a single pallet through the system less the tool blocking
time d:;', since, by definition, no tool blocking takes place for a production supply order
at station m during dTm when no order is issued. Note, however, that it may happen
at other stations i =f m. Thus, here the mean times ti are used. If there is more than
one palJet in the system the given term in (16) is divided by N. Since a tool supply is
necessary only every liP!. times per visit, multiplication with this factor is done to obtain
the intermediate time between tool requests dtm for production supply orders:

~ _ Li=l i~m [~ (qi + tn] + (qm + tm + p~d~)

'rim E S. (17)
Tm - N Pm
The intermediate time between tool requests d~m is derived analogously:
226

"1m ES (18)

3.3 System Evaluation

The given model consists of two networks. The first network models the PT System as
a single class closed queueing network of the Jackson type (Gordon and Newell, 1967,
Solberg 1977). The three stage stations m E S are "exponentialized" in distribution by
using their mean processing times t:, within the Jackson network as the processing times
for a single stage station m (for similar approximations see Yao and Buzacott 1986, Vinod
and Altiok, 1986). The second network models the TT System and is a multiple class
network. For its evaluation a MVA approximation is used (Shalev-Oren et al., 1985).
Each of the two networks requires the results of the other one as input for its evaluation.
To overcome this problem, an iterative approach is suggested. At the beginning, the PT
System is evaluated by assuming that the queueing times of the tool supply orders are
zero, i.e. qfm = 0 and frm = 0 "1m. The resulting throughput times at the stations of the
PT System are then used as input for the TT System. After evaluating the TT System,
the resulting qTm are then used to evaluate again the PT System. This switching between
queueing network models and adjusting of the intermediate times between tool supply
orders d~m' 4m and tool blocking times d~, d~ continues until the queueing times qfm'
q~m reach a value within a certain tolerance level.
Convergence is ensured by the fact that qTm cannot increase infinitely since if qTm
increases, the tool blocking times d~ and ~ must also increase. Consequently, the inter-
mediate times between tool requests d~m' 4m and thus the delay in the TT System will
increase. This however has a decreasing impact on qfm and q~m'

The Algorithm:

STEP 1: Analyze the PT System with a single class closed queueing network model
of the Jackson type. Use the exponential approximation approach of
Vinod and Altiok (1986). In order to estimate the mean processing times
t:, "1m E S based on (13) use (I), (15) with frm = 0, (8) with tm = tTm,
(6), and (7).

STEP 2: Analyze the TT System with a multi class closed queueing network model.
There are 2S customer classes circulating, each one consisting of a single
customer. In order to evaluate the delays by equations (17) and (18) at
the 2S delay stations use the last results of the PT System.
 227

STEP 3: Analyze the PT System by using the results of STEP 2 to evaluate

equation (14) and (15). Update the approximate probability p~ , I.e.
as follows:
p~(n) -+ p~(n+l),

1 - p~(n+l) = (19)

and the approximated mean operating time tm

(20)

Calculate the mean processing times t~ Vm E S based on (13).

STEP 4: Perform STEP 2 and STEP 3 until the differences between the queue-
ing times of two successive iterations are less than a given value E, i.e.
b (n+l) _ qb (n) < E
qTm qP (n+l) _ qP (n) < E Vm with E being a small
Tm - , Tm Tm-
number and E > o.

4 Example
4.1 Input Data
The above algorithm is demonstrated by an example based on real life data. It considers
a FMS consisting of one load/unload station (l/u station), one vertical turret lathe (vti),
one milling machine (mill), one transportation vehicle for parts (transp. parts) and one
for tools (transp. tools) (see Table I).

I station I Sm I rm I gm I nm I tooll I tool2 I tool3 I

llu station 1 - - - - - -
vtl 1 0 20 30 1 1 1
mill 1 0 20 30 1 1 1
transp. parts 1 - - - - - -
transp. tools 1 - - - - - -
Table I: machine data
228

I no. I station I total time I tooll I tool2 I tool3 I

1 l/u station 90 - - -
2 transp. 10 - - -
3 mill 428 100 200 125
4 transp. 10 - - -
5 l/u station 90 - - -
6 transp. 10 - - -
7 mill 57 55 0 0
8 transp: 10 - - -
9 l/u station 90 - - -
10 transp. 10 - - -
11 mill 57 55 0 0
12 transp. 10 - - -
13 l/u station 90 - - -
14 transp. 10 - - -
15 mill 57 55 0 0
16 transp. 10 - - -
17 l/u station 90 - - -
18 transp. 10 - - -
19 mill 351 100 100 150
20 transp. 10 - - -
21 l/u station 90 - - -
22 transp. 10 - - -
23 vtl 229 25 100 100
24 transp. 10 - - -
Table II: Processing times for part 1 in minutes

I no. I station I total time I tooll I tool2 I tool3 I

1 l/u station 90 - - -
2 transp. 10 - - -
3 mill 470 100 200 170
4 transp. 10 - - -
5 l/u station 90 - - -
6 transp. 10 - - -
7 vtl 639 135 300 200
8 transp. 10 - - -
Table III: Processing times for part 2 in minutes
 229

I no. I station I total time I tool! I tool2 I tool3 I

1 l/u station 90 - - -
2 transp. 10 - - -
3 mill 185 0 100 85
4 transp. 10 - - -
5 l/u station 90 - - -
6 transp. 10 - - -
7 mill 130 0 0 130
8 transp. 10 - - -
Table IV: Processing times for part 3 in minutes

I no. I station I total time I tooll I tool2 I tool3 I

1 l/u station 90 - - -
2 transp. 10 - - -
3 mill 47 0 0 45
4 transp. 10 - - -
5 l/u station 90 - - -
6 transp. 10 - - -
7 mill 223 0 100 120
8 transp. 10 - - -
9 l/u station 90 - - -
10 transp. 10 - - -
11 vtl 19 0 19 0
12 transp. 10 - - -

Table V: Processing times for part 4 in minutes

 230

Four different part types can be manufactured on the system (part1, part2, part3,
part4). The production ratios between the part types are partl : part2 : part3 : part 4
= 0.2 : 0.2 : 0.4 : 0.2. There are four pallets circulating in the system, i.e. N = 4. Three
different tool types are in the system (L = 3) to perform the required operations. Each
tool has a life time of WI = 1000 minutes VI. The mean time between tool breakages is
a = 3000 minutes.
The set S of machines which requires tool supplies consists of S = {vtl, mill}. The
transportation vehicle can transfer one tool at a time, i.e. em
= 1 Vm. Furthermore,
Table I provides the number of tools Kim of each type which is available at a station,
the preorder time r m , the average number of tools to be exchanged if the part mix in the
system changes grn, and the average number of part visits to a station between part mix
changes n rn .
Table II-IV give the process plans for each part type with the operating times of the
required operations and the required tools.

4.2 Results and Sensitivity Analysis

Using the above input data, first the system is analyzed by neglecting tool blocking
completely. Afterwards, the system is evaluated by the algorithm given in section 3.3.
A comparison with results obtained from computer simulation is shown in Table VI in
order to analyze the accuracy of the suggested approach. The simulation is based on the
underlying modelling assumptions.

I paramo I station I no block " with block I simul. I

Urn l/u station .51 .48 .49
vtl .34 .33 .34
mill .88 .89 .92
transp. parts .11 .11 .11
transp. tools - .07 .07
t*rn l/u station 90.0 90.0 89.9
vtl 295.7 307.9 308.2
mill 193.3 207.3 207.2%
transp. parts 10.0 10.0 10.0
transp. tools 15.0 15.0 14.9

Table VI: Results

The throughput of the system without blocking is 1.898 10- 4 parts/minute. With tool
blocking, the algorithm evaluated a decrease in throughput by roughly 6% to 1.790 10- 3
parts/minute. The result by computer simulation is 1.857 10- 3 . A comparison of some
important parameters in Table VI underlines the accuracy of the model.
 231

parameter new change in chge of t;" d~ +d~ I

value throughput vtl mill vtl mill J
STM transportation vehicles 2 0.04% 0.14% 0.04% 30.01 30.01 I
3 0.04% 0.14% 0.04% 30.00 30.00 I
tTm transportation time 10.0 1.93% 1.35% 2.25% 20.54 20.33 I
5.0 3.92% 2.67% 4.49% 10.14 10.0S I
Tm preorder time 1.0 0.3S% 0.26% 0.45% 30.20 29 .74 I
2.0 0.77% 0.52% 0.90% 29.20 2S.76 I
em exch. tools per transp. 2 I.S5% 2.06% 3.42% 30.62 30.3S I
3 3.97% 2.73% 4.55% 30.42 30.25 I
Table VII: Sensitivity analysis of important design parameters

A system designer is interested in changing design parameters in order to improve

the performance of the system, i.e. to reduce tool blocking. In Table VII the number
of transportation vehicles STM of the tool transporation system is increased from one to
values of two and three. Alternatively, the transportation time tTm for tools is decreased.
Furthermore, a preorder time Tm can be introduced and, finally, the number of tools em
exchanged per tool supply order is increased.
Since in the given example the queueing times q~m and qtm at the tool supply system
are quite small (e.g. at the vtl 0.S7 minutes for a production supply order and LOS minutes
for a breakage supply order), a change in the number of transportation vehicles STm does
not have a significant influence, i.e. the change in throughput is small. A reduction in the
transportation time tTm is more beneficial, since it has a direct influence on the blocking
times given in equations (14) and (15). Introducing preorder times Tm has some influence,
however, it only influences production supply orders. The same fact holds for an increase
of the number of tools supplied em per supply order. However, the influence is stronger
here, since, e.g., doubling em cuts the probability ~ in half.
Based on these results the designer must now assess the technical feasibility of these
possible design changes. The outcome of this analysis can then be used to decide on the
best alternative to be implemented.

5 Conclusions
The presented model allows to evaluate the system performance of a FMS with a sepa-
rated, automated tool supply system. The tool blocking times, i.e. the times a machine
is idle because it is waiting for the required tooling, and their relationship to important
design parameters are modelled. Two different types of blocking times are considered:
those due to changes in the part mix and "normal" usage and those due to tool break-
age. The design parameters considered include the number of transportation vehicles, the
232

transportation times to machines, possible preorder times, and the number of tools to be
transported to a machine per tool supply order.

References
Altiok, T. and Stidham, S. (1983) "The Allocation of Interstage Buffer Capacities in
Production Lines," IIE Transactions, Vo1.15, No.4, pp.292-299.
Gordon, W.J. and Newell, G.F. (1967) "Closed Queueing Systems with Exponential
Servers," Operations Research, Vo1.15, No.2, pp.254-265.
Shalev-Oren, S., Seidman, A., Schweitzer, P.J. (1985) "Analysis of Flexible Manufacturing
Systems with Priority Scheduling: PMVA," Annals of Operations Research, Vol.3 ,
pp.115-139.
Solberg, J.J. (1977) "A Mathematical Model of Computerized Manufacturing Systems,"
Proc. 4th Intern. Conf. on Production Research, Tokyo, pp.1265-1275.
Vinod, B. and Altiok, T. (1986) "Approximating Unreliable Queueing Networks Under
the Assumption of Exponentiality," J . Opl. Res. Soc., Vo1.37, No.3, pp.309-316.
Vinod, B. and Solberg, J.J. (1984) " Performance Models for Unreliable Flexible Manu-
facturing Systems," OMEGA, Vo1.12, No.3, pp.299-308.
Widmer, M. and Solot, P. (1990) "Do not forget the breakdown and the maintenance
operations in FMS design problems," International Journal of Production Research,
Vo1.28, No.2, pp.421-430.
Yao, D.D. and Buzacott, J.A. (1986) "The exponentialization approach to flexible man-
ufacturing system models with general processing times," European Journal of Oper-
ational Research, Vo1.24, No.3, pp.410-416.
III. PPC and CIM
ARCHITECTURAL FRAMEWORK FOR INTEGRATED PRODUCTION
CONTROL SYSTEMS

Prof. Dr. August-Wilhelm Scheer

Institut fUr Wirtschaftsinformatik

1m Stadtwald, 6600 Saarbriicken, FRG

1. Introduction

The architecture of integrated information systems (ARIS), as described in the

following chapter, represents a framework that reduces the complexity of the
software engineering process by considering different views (functions, data,
organisation and control) and different phases (requirements definition, design
specification and implementation description) [1]. Chapter 3 demonstrates the
advantages of this approach for the development of production control systems
with emphasis on the domain-dependent characteristics.

2. Architecture for Integrated Information Systems (ARIS)

In order to reduce the complexity of developing information systems two steps

are undertaken:

1. the amalgamation of elements into more general descriptive views .

2. the reduction of relationships using a phased, or procedural, model.

First, an information system is broken down into individual views, which are
then described independently, and thus with less redundancy.
236

Events and environmental conditions are represented by data. They are

represented as infonnation objects using a unifonn data view. The recording of
events as part of a data view is common to many software development methods
[2].
The description of process rules and the process structure provides the process
or function view. The term "function" is therefore often used interchangeably
with the tenns "process" or "process chain", because it is often used in the
literature in association with functionally-oriented system design.
Because of their close connections, the two components user and organizational
unit are aggregated into a single element. Users are assigned to organizational
units, and these are constructed on the basis of criteria such as "same function"
or "same work object". This view is referred to as the organizational view.
In creating these views, however, the links between the views are lost. For this
reason, an additional view, referred to as control, is introduced to record the
relationships between the viewpoints.
In addition to reducing complexity and redundancy within the descriptive
subject matter, the formation of views also has the advantage that one
component (that is, one view) can already be developed without the other views
being available. This means, for example, that later the data model for a system
can be created without the functions being comprehensively defined. Insofar as
details from one view are absolutely essential for the description of another view
these can be established in general fonn by this view and then further specified
by the relevant view.

The process of transfonning the business requirements into the EDP-technical

implementation is often described by differentiated phase models [3]. The
following treatment adopts a five stage layering approach to implementing and
operating a business applicatlons system as shown in Figure 1.
 237

2nd Phase

Requirements
definition

3rd Phase

Design
specification

4th Phase 5th Phase

Operation
Implementation and
description maintenance

Fig. 1: Description phases dependent on the proximity to infonnation

technology
238

In the first step, an BDP-oriented semantic starting position is created. This is

the result of an actual analysis of the process chains with planned concepts built
upon it. This process chain analysis should make the fundamental benefits of the
information system visible. For this reason, at this stage all the views are
examined together.
In the second step, a requirements definition models the individual viewpoints of
the applications system independent of implementation considerations. In the
process, descriptive languages are chosen which are sufficiently formalized to
provide a starting point for a consistent BDP-technical implementation.
In the third step, the creation of the design specification, the requirements
definitions are adapted to the demands of the user interfaces of the
implementation tools (e.g. database systems, network architectures or
programming languages). However, at this point there is still no reference to
specific products.
In the course of the fourth step, the technical implementation description, the
concrete translation of the requirements into physical data structures, hardware
components and programming systems is undertaken.
These four phases describe the creation of an information system and are
therefore referred to as "build time". Thereafter, the completed system is
released for operation, so that the operation and maintenance step can be added,
which is referred to as "run time". This run time version of the information
system and its tools environment is not further considered below.
The requirements definition is very closely linked with the business applications
sphere, as is shown by the width of the arrow. It should, however, be created
largely independent of the implementation considerations, as is represented by
the width of the arrow to the design specification. Both technical
implementation and operation and maintenance are, in contrast, closely linked
with the "device and product level" of information technology. Changes in
information technology have an immediate impact on the kind of
implementation and the operation of a system.
The phases cannot always be exactly separated from each other. As a result, the
assignment of methods, representations and results of the software design
process is not always unambiguous. The phase concept should certainly not be
interpreted as a strict sequencing of the development process according to the
"waterfall" principle. Rather, a prototyping approach is also explicitly
incorporated. But even in the case of evolutionary software development the
descriptive levels are given.
 239

Initially , in the first two steps, each component is represented solely from the
conceptual viewpoint without implementation restrictions. Thereafter, the
factual content is further specified in the framework of the design specification
and in the implementation stage it is implemented using specific data processing
techniques. Only in the third and fourth phase information technology resource
considerations do have an effect.

As a result, each component can in the first instance be described independently,

the relationships to other components are then handled at the control level.
Within control, the three descriptive levels are also established with respect to
their proximity to the information technology resources. In this way, the links to
the other components can be created at every descriptive level.
Figure 2 represents the information system architecture. It consists of the
following elements: functions, organization, data, and control. All the
components are broken down in terms of their proximity to the information
technology resources into the descriptive levels: requirements definition, design
specification and implementation description.

equire-
..--L_-:mc:.:.;:.en:..:.:t:.::.s--=d=-=e.:..:;fin:. :.:i.:.:io::.:n~....:::........
.t Organization view

Design specification

Implementation description

Requirements
definition
Requirements definition
Design
Design specification specification

Implementation Implementation Implementation

description description description

Data view Control view Function view

Fig. 2: ARIS architecture

240

3. Architectural Framework for Integrated Production and

Control Systems

The tendency towards decentralization in manufacturing (e.g. the creation of

Flexible Manufacturing Systems) strongly influences the design of production
control systems. The major problem faced by a decentralized production control
approach is to cope with the constraint-based assignment of production orders to
the resources available.
The information technology in the area of production control is dominated by
so-called "Leitstand"-systems. A Leitstand system is a computer-based,
interactive decision support system for production control and -monitoring. At
the moment, Leitstand systems are used in many enterprises. Different reports
show their hard- and software concepts, functionality and design of the user
interface [4].
In the following chapters, the architecture of integrated information systems
(ARIS) is used to show the design of a Leitstand system focussing on the phase
of requirements definition.

3.1. Functions

A function can be defined as an operation carried out on an object with the aim
of supporting one or more goals. Functions can be described at various levels of
granularity which can be further broken down according to the level of detail of
the analysis.
The main functions of a Leitstand system are [5]:

production data administration,

short term scheduling,
production order release,
production monitoring.

Characteristic features of Leitstand systems are the window-oriented graphical

user interface and the interactive design of the user functions. Figure 3 shows an
example of a detailed functional structure of a Leitstand system.
 241

Fig. 3: Function hierarchy diagram of a Leitstand system

3.2. Organization

Organization units are the units responsible for the tasks to be supported by the
information system. Normally, organization structures are represented by
organizational charts. Another possibility is the definition of planning levels to
regulate planning and decision responsibilities. The organizational view for the
use of connected Leitstand systems on two different planning and control levels
is shown in figure 4.
 242

NC-
program-
ming

Plant
buffer
Plant

Tool
buffer

Production Production
area A area 8

Parts
production Area
buffer

Fig. 4: Organizational view of connected Leitstand systems

Here each production area (parts production, assembly) has its own Leitstand for
planning and production monitoring. A Master-Leitstand is responsible for the
coordination of the area specific, decentralized Leitstand systems; it is related to
NC-programming and different inventory control systems.

3.3. Data

Data structures, that describe mainly static and structural aspects, may origin
both from physlcal and logical objects of the manufacturing system [6]_ They
include personnel, materials, machines, tools, transportation devices as well as
data about the order spectrum such as type of order, quantities to be produced,
due-dates of orders, bill-of-materials and work schedules.
For modelling of data structures, an extended version of the well-known ERM-
approach [7] is utilized. This modelling technique is commonly used in the
databased design phase to depict conceptual schemas of the data base in a
graphical, user-oriented manner. Major modelling constructs are the entity- and
relationship type. The concept of an entity type denotes an abstract or real world
object that is distinguishable and carries meaning in itself. An entity type is
 243

graphically represented by a rectangle. A relationship type is used to model links

between entity types and is depicted by a rhombus. A cardinality constraint is
attached to the relationship type to further specify how many times an entity
participates in the relationship. To allow modelling of different logical
relationships between entity types, various specific relationship types are
introduced, of which the generalization/specialization relationship is the most
important one. Both entity types an relationship types may be assigned with
attributes that describe their characteristical features. The so-called key attribute
is a special attribute used to identify each individual entity.
The information objects needed in a Leitstand system are represented in area-
specific data bases of which a sample conceptual schema is depicted in figure 5.

O,n O,n O,n O,n

AREA PART

O,n O,n

O,n

O, n
OPERAllON DISTURBANCE
lYPE

O,n
RESOURCE
O,n

O,n

Fig. 5: Conceptual data structure of production control area

244

3.4. Control

The task of the control modelling is to reunite the separately considered views
(function, organization, and . data). In addition, dynamic aspects of the
information architecture are considered, since the linking up of the views allows
the description of changes of state and hence the dynamic behaviour of the
system. To model the control view, either pairwise relationships between the
architectural blocks (function-organization, function-data, data-organization) can
be considered, or all three views can be linked together.
The control view of a Leitstand system is exemplified by the link between
functions and organization. In the function level model shown in figure 6, the
main functions are allocated to the organizational units responsible.

NC-
program-
.....
0nI0< _ . arrong
ming

-
Aeprooerutlon 01...,.,.

""'"
~lunctlons
Gtop/IfcII_kNolions

...
PIamil; PIaIww>g
S""""*'O ,_Ion pion SITOOIhInQ .............. plan
Ro........-y_ R.Ie ........-y_
Procl.c:IIon .."..,.,. ProGIabn coni.,.
~ ......uor. s.. _ .......1ons
Grap/Qj
~

Production
area B

Fig. 6: Function level model for connected Leitstand systems

 245

4. Concluding Remarks

Due to recent advances in infonnation processing technology, production

infonnation systems are becoming more and more important for the
management and control of the manufacturing process. The architectural
framework presented allows a detailed description of Leitstand system
components. The division into the views function, organization, data and control
leads to a reduction in the complexity and avoids redundancies. In this way, the
architecture of integrated infonnation systems CARIS) is a well established
approach to develop infonnation systems for production control.

Literature
[1] See Scheer, A.-W.: Architektur integrierter Informationssysteme, Berlin et al. 1991, S.
1 If.
[2] See Olle, T.W., Hagelstein, J., MacDonald, I.G. et al.: Information Systems 1988, p.
43.
[3] See for example the 12 stage model in: Olle, T. W., Hagelstein, 1., MacDonald, I.G. et
al.: Information Systems 1988, p. 37.
[4] See Hars, A.,Scheer, A.-W.: Entwicklungsstand von Leitstanden, in VDI-Z
132(1990)3, pp. 20-26; Adelsberger, H.: Ein Vergleich von Leitstanden, in: Mertens,
P., Wiendahl, H.-P., Wildemann, H. (Ed.): Proceedings of PPS im Wandel, Mtinchen
1990, pp. 363-380; a detailed report about the main Leitstand systems can be found in:
Hemmerich, G. et al.: Fertigungsleitstand-Report der Fa. Ploenzke Informatik,
Wiesbaden 1990; another overview is given by: Hoff, H., Liebrand, T.: Elektronische
Leitstlinde - Auswah1 und Einflihrung elektronischer Leitstande, eine Marktstudie, in:
FBIE 39(1990)6, pp. 208-287.
[5] See Kurbel, K., Meynert, 1.: Flexibilitat in der Fertigung durch Einsatz eines
elektronischen Leitstands, in: ZwF 83(1988)12, pp. 58 If; Schmidt, G., Frenzel, B.:
Anforderungen an Leitstande flir die flexible Fertigung, in: CIM Management
6(1990)4, p. 34; Kruppke, H.: Problematik bei der organisatorischen und technischen
Integration von Fertigungsleitstlinden in die Untemehmenspraxis, in: Scheer, A.-W.
(Ed.): Fertigungssteuerung - Expertenwissen flir die Praxis, Mlinchen, Wien 1991, pp.
279f.
[6] See Zell, M.: Datenmanagement simulationsgestiitzter Entscheidungsprozesse am
Beispiel der Fertigungssteuerung, in: Scheer, A.-W. (Ed.) Veroffentlichungen des
Instituts flir Wirtschaftsinformatik, Heft 72, Saarbrucken 1990, p. 5.
[7] See Chen, P.S.: The Entity-Relationship Model - Towards a Unified View of Data, in:
ACM Transactions on Database Systems, 1(1976)1, pp. 12f.
 A New Conceptual Framework for
Hierarchically Integrated PPS-Systems

by

Christoph Schneeweiss
Chair of Operations Research
Faculty of Business Administration
University of Mannheim
Schloss, D-6800 Mannheim 1

1. Introduction

Traditional Production Planning and Scheduling (PPS)-Systems have

usually been designed to treat rather involved situations like that of a job
shop production of the make-to-stock or make-to-order type (e.g. VOLL-
MANN et al.). Generally, these systems consist of the Master Production
Schedule (MPS), the Material Requirements Planning (MRP)-module,
the Capacity Requirements Planning (CRP)-module, and the short term
job scheduling (sequencing) module (s. Fig. 1).
----------------.------.-----
Production Planning : medium term level

work order
planning level

job .hop
control level

Fig. 1 Traditional PPS-Systems

Within the MRP-module three main tasks have to be performed:

 247

(1) Requirements planning for components, (2) lot sizing, and (3) de-
termination of lead times. Typically, lot sizing is performed only for
single products not taking into account capacity constraints and stocha-
stic elements. Only subsequently the CRP-module performs some kind
of post-optimal calculation to balance demand and supply of capacity. In
addition, lead times are only taken to be rough estimates which do not
explicitly consider capacities, lot sizes, and possible stochastics.
In principle a production planning and scheduling problem could be sol-
ved in using a stochastic dynamic programming (DP) approach. The
first stages of such an approach would build up production facilities and
capacities which would then be used for the short term production and
scheduling task. This creates a hierarchical structure which, in using
the backward recursion procedure of DP, leads to an optimal solution
(SCHNEEWEISS). Obviously, however, such a calculation can only be
performed for extremely simple systems. Therefore, various approxima-
tions have been suggested. Depending on the nature of the production
system the following three approaches may be considered (s. Fig. 2).
1) The traditional PPS-system approach,
2) the deterministic lot sizing approach, and
3) the stationary queueing theoretical approach (QT-Approach).

Stochastic
DP

Traditional
PPS

Deterministic
Approach aT-Approach

Fig. 2 Three approximations to the general production planning pro-

blem

(1) PPS-Systems often represent rather poor approximations of the

248

original stochastic DP-model. They consist essentially of a two level

hierarchy: The MPS-MRP-CRP-Ievel (also called "work order plan-
ning module") and the short term scheduling (or job shop control)
level (s. also Fig. 1). One starts (in the MPS step) with a (one-
product) non-capacitated deterministic lot sizing procedure which is
then, in the MRP step, extended to all production levels. Only after-
wards capacities are incorporated. This is achieved within the CRP
step using a network procedure which extensively exploits the given
estimates of production lead times. These estimates are rather com-
plex quantities. They have to take into account the capacity situation
of multiple product bottlenecks, lot sizes, and possible stochastic di-
sturbances. Obviously, it is one of the main drawbacks of
existing PPS-systems that production lead times are not optimized
but rely on rough exogenous estimations. As a consequence, short
term scheduling has to tackle with an overwhelming amount of un-
solved problems.

(2) The second way to approximate the original stochastic DP- model is a
deterministic approach (s. Fig. 2) . One reduces the stochastic DP
to a deterministic DP resulting in a capacitated multi-product multi-
stage and multi-period deterministic production-inventory problem.
It is well known, that such an approach is again only possible for
very simple situations. In fact, only for the non-capacitated multi-
stage case and for the capacitated one-stage case are results available
(e.g. AXSATER et al.) . Assuming, however, that this problem could
be solved, then, of course, lead times could be replaced by setup
dates. As is always the case with inventory problems, lot sizes and
setup dates are calculated simultaneously. Since lot sizes depend
on capacities and capacities have usually to be determined well in
advance to setup dates, the multi-stage lot sizing approach is only
reasonable in completely deterministic situations. If, however, at the
point of time when capacities can still be changed, one is not sure of
the dates of future setups, deterministic multi-stage lot sizing is only
of limited value. Of course, one can try to use a rather course time
grid, e.g. weeks, but then the coupling to the short term scheduling
level with its refined time scale is rather difficult. Hence, in non-stable
situations multi-stage lot sizing cannot be recommended.
 249

(3) The so-called Queueing Theoretic (QT)-Approach which will be dis-

cussed in this paper represents the other extreme. It uses a stationary
(stochastic) approximation of the stochastic DP-model. It considers
production as a network of queueing models and minimizes in a first
step production lead times with respect to capacities, lot sizes, and
stochastics. In a second step, given these lead time-optimal capacities
and lot sizes, the setup dates are then determined according to a due
date rule. It is obvious that such an approach is particularly suited
for a situation where stochastics plays a dominant role and where
production can control the size of the lots, that is, in make-to-stock
or make-to-assembly job shops.
The QT-Approach is a new way of incorporating stochastics and
capacities into production planning. Compared with existing PPS-
Systems it uses different hierarchical levels which are strongly inter-
related with each other. In fact, the QT-Approach may be considered
as an extension of Karmarkar's queueing theoretic lot sizing approach
to production systems of the job shop type.
In what follows we shall outline the conceptual framework of the QT-
approach. Section 2 first gives the general idea of the new concept. Sec-
tion 3 then starts with some basic results of queueing theory which will
be used for the different steps of the QT-approach. These steps comprise
the determination and harmonisation of lot sizes at different bottlenecks.
Section 4 discusses non-stationarities and Section 5 shows the relationship
of the QT-Approach with short term job shop control. Finally, Section
6 provides some concluding remarks stressing again the particular type
of hierarchical coupling used in the QT-approach as compared to the
traditional PPS-concept.

2. Conceptual Outline of the QT-Approach

The QT-Approach is concerned with a restructuring only of the work

order planning module, i.e. medium term production planning and short
term job scheduling are left unchanged. At the medium term planning
level the production program is calculated and, what is most important,
the necessary capacities are determined. The work order planning module
250

then decides upon the production lot sizes and the dates of release. With
these dates the short term production scheduling can then dispatch the
individual lots according to a specific due date rule.

A multi-stage production system may be described as a system of (G/G/l)-

queueing models. Confining on serial production structures our main
concern is the time a work order i takes on its path (m = 1, ... , m(i))
through the system. This so-called mean throughput time is given by

m(i)
Ti= L Tim (1)
m=1

with
(2)
being the throughput time at stage (or work station) m . Tim is the sum
of lead time W m , manufacturing time Mim and setup time Sim ' Lead
time W m is given by the mean waiting time at work station m which
particularly depends on the lot sizes Qim(m = 1, ... , m(i)) , i = 1, . . . , N
of all products at each station:

Hence, our main concern is to determine lot sizes such that the through-
put time be minimum.
Adopting a JIT-attitude, no inventories other than those built up by the
waiting lines in front of a work station will be allowed. Hence, for a .
product i one has Qi = Qim Vm and the total throughput time is given
by
m(i)
Ti = L T im(Ql, ... ,QN)' (3)
m=1

Economically, total throughput time is a rather complex quantity. It

does not only influence work in process (WIP) but also the flexibility
of a system. Hence, it seems to be reasonable to determine lot sizes
Ql, " " QN such that
(4)
 251

In doing so, it is necessary to calculate lead times W m ( Q 1, ... , Q N ). As

is well known this cannot be done analytically. Two approxiations may
be adopted:
(1) One can start with a non-Jackson-net, i.e. a net consisting of G/G/1-
nodes and can use several kinds of approximations up to the second
moment of the involved probability distributions. Such approximati-
ons are discussed e.g. by TIJMS and WHITT.
(2) One takes, as a starting point, a Jackson-net and tries to adapt the
exponential distributions of the M/M/1-queues to more general and
realistic situations.
In what follows let us describe the second approach in some more detail.
The so-called QT-Approach consists of two planning phases (s. Fig. 3):

Medium Term Production Planning

stationary stochastic multiproduct

capacitated one-stage model

time optimal mean Phase I

lot production times

Feed-
beck Adaptation to the stationary
control multi-stage situation

product specific stationary

smoothed lot sizes

non-stationary deterministic
non-capacitated multi-staAe model Phase II

adapted lot sizes

Short Term Production Scheduling

Fig. 3 Conceptual framework of the QT-Approach

(1) Phase I describes the job shop as a stationary queueing network. It

consists of two steps.
Step A identifies the bottlenecks of the system and calculates for
each bottleneck separately lead time-optimal lot sizes. Technically
252

speaking, a stationary stochastic multi-product capacitated one-stage

inventory-production model is solved. The result of this step is a
simultaneous optimization of capacity, lead time, and lot size for one
stage.
Step B then adapts the lot sizes at the different bottlenecks to each
other. Formally speaking, step A is extended to the multi-stage situa-
tion. This results in stationary product-specific lot sizes being smoo-
thed over all stages (i.e. bottlenecks).
(2) Phase II introduces non-stationarity. This is achieved by solving a
non-stationary deterministic non-capacitated multi-stage lot sizing
model. As a result one obtains adapted lot sizes which are then
released to the job shop.

Considering the two phases one realizes that the optimal stochastic DP-
problem shown in Fig. 2 is split up mainly into two simpler problems:
Phase I copes with stochastics and capacities, whereas Phase II takes into
account the non-stationary character of a real world production system.
This hierarchical separation is not only done because of computational
convenience, it also reflects the different information status of the phases.
Thus Phase I may be considered to be still 2 months in advance of the
release date. Hence stochastics plays a dominant role. Because of the
comparatively short time range it can even be described to be stationary.
(Note, that seasonal fluctuations are coped with at the medium term
level, s. Fig. 1). Phase II, on the other hand, often is far nearer to the
release date so that demand or at least forecasts of demand are well known
and non-stationarity outweights stochastic fluctuations. Fig. 4 shows the
time scale of planning levels we typically have in mind.

Planning hori~on

~-_·"·/Vm""·-II-------+----mL>=.:::::>=.:<:~>""".:~=:::::::=:i::::=:m=::<:i." , =:w:md
Release
Medium Phase 1 Phase II
date
term
planning

Fig. 4 Time scale of planning levels

 253

3. Phase I: Stationary Analysis

Let us start with the one-stage-case (step A) which, in Step B (Sec. 3.2),
will then be extended to a complete network.

3.1 Phase I, Step A: One-Stage-Case

Starting point will be an M/M/1 queueing model for which, in Sec. 3.1.1,
well known analytic results will be derived. The next Section 3.1.2 will
then extend the results to M/G/1- and G/G/1-models and finally, Sec.
3.1.3, gives some useful statistical results for the multi-stage case discus-
sed in step B.

3.1.1 Analytic results for M/M/1

First we summarize some results given by KAMARKAR or ZIMMER-

MANN.
Let us use the following notation and correspondence
1\
1) Production batch - customer (entering the
queueing system)
1\
2) Characterisation of - mean arrival rate of
material flow
1\
customers : t
4) Capacity utilization - service degree p := ~J1.
1\
5) Production lead time - mean waiting time W

D : Capacity demand
C : Capacity supply
(both measured in processed units per planning period)
Q Lot size (units per lot)
S : Setup time (time per lot)

Derived quantities:
254

Production coefficient ._ 1
q .-(5
Arrival processing time per lot - !i
M -C
Mean setup and processing time T=S+M
Mean production rate J.L=~
,._ D
Mean rate of lots A.- Q
Service rate .- .:lp. -_ DS
P .- Q
+ DC
Degree of capacity utilisation by
setups PS .-
·-
DS
Q
Degree of capacity utilisation by
pure processing Pp .-
.-
D
C

Hence, using the elementary result of MIMl1 queueing theory W = .x':p.'

one obtains for the lead time

(5)

and, accordingly, for the minimizing lot size

D+ In
Q* C Vc S (6)
- 1 ( D)·
(5 1- C

Hence, the mean optimal processing time per lot M* = !f- can be ex-
pressed as
M* = pp + vfji; S. (7)
1- pp
Consequently, substituting Q* in (5) the optimal lead time W* is a func-
tion of M*. Furthermore, since M* depends on PP' optimal lead time is
a function of capacity utilisation by pure processing

(8)

This function is depicted in Fig. 5 showing a monotone relationship bet-

ween the optimal lead time and pp. For Pp ~ 85% lead time increases
rather steeply. On the other hand, for Pp ~ 70%, e.g., lead time is
so small that the production station may no longer be considered as a
bottleneck. Hence, let us define a bottleneck by 70% ~ pp ~ 85%. Of
 255

w·(/)p )

-
50~ 55~ 60~ 65~ 70~ 75~ 80~ 85~ 90~

Pp f f

Fig. 5: Minimal lead time as a function of pp

course, these limits are quite arbitrary and depend on the lead times
which management deems to be acceptable.
In fact, because of pp = g, Fig. 5 shows the optimal relationship between
lead time and capacity supply C. Since C is determined by medium
term production planning, Fig. 5 exhibits the reaction of the short term
operative level with respect to medium term capacity planning. Hence,
within a multi-criterion decision procedure capacity supply has to be
outweighted against the length of lead time.

3.1.2 Simulation Results for MIMI1, MIG/1 and GIG/1 Models

Having derived analytical results for the MIMI1 model we will now pre-
sent simulation results for more general models with general distributi-
ons G having variation coefficient smaller than 1. These distributions are
approximated by Gamma-distributions. Figure 6 exhibits the following
important facts:
256

Lead time W
200,----------------------------------------,

100

OL-L-~~~~_L_L~~~~~L_~~~~_L_L~

~ U U U 8 U ru ~ ~ ~
Mean processing time M

- M/M/1 ---+- M/G/1 ----- G/G/1

Fig. 6: Lead time as a function of the lot size (pp = 80%)

1. The minimum is rather robust.

2. The robustnes property is even stronger for models M / G /1 and G / G /1.
3. The optimal processing time M* is almost the same for all three types
of queueing models.
4. There is a substantial reduction in lead time in moving from M / M /1
to M/G/1. This shows that reducing the variation coefficient of the
production rate has a positive effect with respect to a shorter lead
time.
5. There is also a substantial reduction in lead time changing from
M/G/1 to G/G/1. This underlines the positive effect of reducing
the variation coefficient of the arrival process of lots.

3.1.3 Multi-Product Case

So far we considered only the single product case. The main idea, ho-
wever, is to interprete M* as the general mean processing time for N
different kinds of products. Hence, the optimal lot size Qi of product
 257

type i is taken to be

Qr = M* ,(i = 1, ... , N)
qi
(9)

with qi denoting production time per unit. Taking Qi, garantees that
each product type is produced "in its lead time minimum" , showing again
the high priority we are giving to its optimum and hence to the JIT-
principle. Of course, for different product types, lot sizes Qi and the
number of setups may differ.
Moving to the more realistic case (step B) of more than one bottleneck,
however, it is no longer possible to fix Qi by (9). In fact, it will be
necessary to harmonize the individual optimal lot sizes Qi at different
bottlenecks resulting in lot sizes Qf!. Consequently, qiQf! (i = 1, .. . , N)
fluctuate about M*. Hence it is of importance to study the influence
of this fluctuation on optimal lead time W*. Fig. 7 shows the typical
result. It is comforting to recognize that for all values of PP' as long as
VCM ::; 0.5, W* does only insignificantly depend on VCM, with VCM
being the variation coefficient of processing time M.

The results in Fig. 7 are of similar importance than those in Fig. 6. They
will subsequently be used within several test calculations.

3.2 Phase I, Step B: The Multi-Stage-Case

Let us now consider the multi-stage case of having more than one bott-
leneck. Typically, for each work station m we will now have specific lead
time-optimal lot sizes, and hence Qim =1= Qiml' Persuing a pure JIT-
attitude we shall try to harmonize lot sizes without any WIP other than
that waiting for processing in front of a machine. In Sec. 3.3 we then
present an illustrating example which summarizes the whole procedure
of Phase 1.
For a serial structure a harmonisation of lots Qim can be achieved by
258

w·

.9 .8 .7 .8 .5 .4 .3 .2 .1
VC M

rio!" ~ 90"" -+- 80 "" -+- 70"" --e- 80 ""

Fig. 7: Lead time as a function of processing time fluctuations (V eM)

solving the following optimization problem

m(i)

m=l
2:= IQim - Qf I ~ mw'Vi
Q;

s.t. 'Vi (10)

Qf :::; Qim + dim
Qf ~ Qim +d~

with dim and dim describing positive and negative limits for the "har-
monized lot size" Q~ to deviate from Qim ' Instead of the symmetric
criterion (10) one could also use the non-symmetric criterion
m(i)

2:= aim' Lim + f3im . Lim ~ min (11)

m=l

with
 259

In choosing aim and f3im appropriately one can take into account the
non-symmetric feature of the minima as shown in Fig. 6.
Optimizing (10) or (11) results in an optimal harmonized lot size Qf1*.
For a complete symmetric harmonisation, taking e.g. as a criterion
2::~i (Qim - Qf1)2, one simply obtains the arithmetic mean
m(i)
H* 1 ~ * (12)
Qi = m(i) ~ Qim'
m=l

Having calculated Q{l* one has now to check whether the new lot size
still implies minimal lead times. Of course, such a test calculation could
be incorporated in the above minimization models. Since one of the test
criteria is non-linear (s. equ. (16) below) this would imply, however, the
solution of a non-linear model.

Test calculation

In testing Qf1* we first calculate the harmonized number of setups during

the planning period
(13)

with Di being cumulated expected demand of product i during the plan-

ning period and [x] denoting - as usual - the greatest integer smaller than
x. This leads to mean harmonized lot sizes

(14)

resulting finally in processing times

-H*
Mim := qim . Qi (m = 1, . . . , m(i))

with qim being the corresponding production coefficient. Calculating the

mean value over all products

N
- 1 ~H
M m := J( ~ni . M im (15)
i=l
260

with
K
]{:= 2: n fi
i=l

we have now to make sure that M m does not deviate too much from M* ,
i.e. M ~ M* (s. Fig. 6). Similarly, we calculate the variance

(16)

and must check whether ~m :::; 0.5, at which point (s. Fig. 7) the variance
of processing time is acceptably smalL
If the test fails then at least two measures may be taken: .
(1) Adapt the parameters in the above optimization problems.
(2) Adapt capacities. This adaptation is along with the ideas of a JIT-
attitude. Only if we accept additional WIP one would turn to diffe-
rent lot sizes at different bottlenecks.

3.3 An Illustrative Example

The following example may illustrate the steps of Phase I.

Consider the following 2-machine-problem stated in Table 1, showing
particularly that product 2 is not processed on machine 1 and the same
is due for product 4 on machine 2.

products
1 2 3 4 5
Di 1000 1000 500 440 400
qil 0.05 - 0.2 0.15 0.1
qi2 0.04 0.07 0.1 - 0.2

Table 1: Demand and production coefficients qim(m - 1,2) of the

example

The planning horizon is assumed to be 8 weeks with 40 hours per week.

Hence, measuring capacity supply in hours one has C = 320h. . Mean
setup time S = 1.5h is assumed to be equal for all products and both
 261

machines. Taking into account the dates of Table lone obtains a manu-
facturing demand for machine 1 of Dl = 1000·0.05 + 500 ·0.2 + 440·
0.15 + 400·0.1 = 256h resulting in

Dl 256
PPI = Cl = 320 = 80% and M; = 12.8h.

Analogously one obtains M; = 9.6h. Hence, the product specific lot sizes
are Qim = q:m M:'n. For i = 1 and m = 1 this results in Qil = ~~o~ = 256
units. The first 2 columns of Table 2 give the results for all products.

-H*
product Qil Qi2 Q{l* n!f
t Qi
1 256 240 248 4 250
2 - 137 137 7 143
3 64 96 80 6 84
4 85 - 85 5 88
5 128 48 88 5 80

Table 2: Lot size calculation

Applying the simple arithmetic mean (equ. (12)) one obtains the harmo-
nized lot sizes in column 3 of Table 2. These lot sizes are not yet those
that can be used in practice, since it seems to be reasonable to produce
a certain (integer) number of lots during the planning horizon. Hence,
the number of lots is given by equ. (13) resulting in the figures shown in
column 4 of Table 2. As a consequence one obtains adjusted lot sizes (s.
equ. (14)) as shown in the last column of Table 2.

Finally, a test calculation has to be performed. For each product and each
machine we calculate the processing times yielding the values given in
Table 3. This results in mean values (equ. (15)) and standard deviations
(equ. (16)) as also shown in Table 3
262

-H*
product Qi n·, Mil Mi2
1 250 4 12.5 h 10.0 h
2 143 7 - 10.0 h
3 84 6 16.8 h 8.4 h
4 88 5 13.2 h -
5 89 5 8.0 h 16.0 h
number of lots 20
Mm 12.8 h 10.9 h
Sm 3.3 h 2.8 h
VC M 0.26h 0.26h

Table 3: Test calculation

Finally the variation coefficient VC M = !.t: of lot-processing time is
given in Table 3.
The test calculation shows that only for Machine 2 there is a deviation
from the mean value M; = 9.6h of lO .~~9.6 ~ 14%. Since this deviation
is positive and comparatively moderate one should accept it. Note that
VC~ = 0.26 < 0.5. Here, as we know (s. Fig. 7), deviations of 14% do
not affect the lead time.

4. Phase II: Introducing Non-Stationarities

We are now in a position to introduce non-stationarities. In doing so (in

Sec. 4.2) we first (in Sec. 4.1) have to consider the timing of work orders.

4.1 Calculating Optimal Release Dates

As we know from equs. (1) and (2) the optimal total throughput time is
given by
m(i)
T; = 2: (W~ + Mim + Sim), (17)
m=l
 263

and hence, denoting the lot-finishing date (due date) by tf, the release
date tf is given by
(18)
Though the optimal lot processing time M:nis almost the same for all
kinds of queueing models, this is not the case for the lead time (s. Fig.
6). Therefore, in calculating W:n we do not use equ. (5), but the well
known Pollaczyk-Khinchin formula

Wm = Ai(T! + V ARTm). (19)

2(1 - Pm)
Equ. (19) is valid for MjGjl models and may therefore be considered
as a better approximation than (5). We still assume, however, that the
arrival process is Poisoon. This assumption is justified for the following
reasons:
1. At the time when work orders have to be determined there is no
information to specify the arrival processes.
2. Shantikumar and Buzacott have shown (SHANTIKUMARjBUZA-
COTT) that for job shop systems the assumption of poisson arri-
val processes is not critical. They got values for variation coefficients
of 0.9 and 1.0.
3. As shown in Fig. 8 smaller ve>. lead to smaller lead times. Hence
assuming the arrival process to be Poisson is conservative. There
remains, however, the problem of overestimating the lead time.

4.2 Taking into Account N on-Stationarities

In Sec. 2 (Fig. 3) we characterized Phase II as a planning situation in

which the stationarity assumption no longer holds. In most cases, some
weeks before the actual start of production, demand will be known at least
by its forecasts. Since these forecasts in general will not be constant, a
modification of the lot sizes calculated in Phase I might result in a further
improvement.
The general idea consists in calculating the number of setups nf* (s. equ.
(13)) and setup dates within the release time tf -tf (s. equ. (18)). Using
264

Lead time
50r-------------------------------------~

40

30

20

10

O~--~---L--~----L_ _ _~_ __ L_ _~_ _ _ _L_~

1 .9 .8 .7 .6 .5 .4 .3 .2 .1
VC ).

Fig. 8 Lead time as a function of VC A

a simple demand covering lot sizing procedure (( Q, T)-policy with cycle

FR'
n: H
time T := tjn i ) one might then adapt stationary lot sizes Qi * (s. equ.
i

(14)) to non-stationary demand, leaving, however, nf unchanged.

Let us explain this idea by a simple example given in Table 4. For a
single item let us assume QH* =: Qt = 250 and n H = 4.
Table 4 shows that changing at a workstation lot sizes Qt = 250 (t =
1,3,5,7) to demand covering lot sizes Ql = 200, Q3 = 230, Qs = 320, and
Q7 = 250, one obtains a considerable reduction in inventory of 36,5%.

weeks
1 2 3 4 5 6 7 8 L:
demand 130 70 110 120 190 130 180 70 1000
Qt 250 - 250 - 250 - 250 - 1000
stock 120 50 190 70 130 0 70 0 630
Qt 200 - 230 - 320 - 250 - 1000
stock 70 0 120 0 130 0 80 0 400
Table 4: Non-stationary adaptation
Of course, an adaptation of lot sizes to non-stationarities should not in-
crease optimal lead times. Hence, a test calculation has to be performed.
 265

5. Short Term Scheduling

Having calculated lot size adaptations and release dates for each work
order we may now enter the short term scheduling module. Let us assume
that for each work order i due dates are known. As mentioned earlier,
it seems to be reasonable to use a due date criterion. Hence, for each
work station m let us use a priority rule giving priority to that work
order i which has the highest delay Dim . This delay can be calculated as
follows. Before entering workstation m the remaining throughput time
RTim is given by

m(i)

RTim := Mim + Sim + L Tij . (20)

j=m+l

Hence, the dispatching time tPrn is (s. also equ. (18))

which, with T NOW being the actual scheduling time, results in the delay

(21)

Comparing the 6im-dispatching rule with other priority rules like the
well-known FIFO-rule or the shortest processing-rule one obtains promi-
sing results (s. HAFNER). This holds particularly for criteria like the
"percentage of delayed orders" , the "variation coefficient of delays" , and
the "maximum delay" .
Often in practice a due date is not given. Instead a reordering point or
a safety stock for finished goods has to be taken into account. In that
case one would have to forecast a due date or use a 6RISK-rule as shown
in SCHNEEWEISSjSCHROEDER. With this rule, that work order is
given highest priority for which the probability of being out of stock is
greatest.
266

6. Summary and Concluding Remarks

In this paper we proposed a new design for PPS-systems. From a con-

ceptual point of view the main problem with PPS-systems is to perform
a smooth stepwise transition from a more or less non-stationary deter-
ministic medium term production planning stage to a short term sequen-
cing procedure. The key idea is, first, to introduce stochastic elements
within the framework of a stationary model. This allows the calculation
of the throughput time taking into account capacities within a multiple
item and multiple stage situation (s. Fig. 9). Optimizing throughput
times for given capacities not only improves flexibility but also redu-
ces WIP. Furthermore, it supports the JIT-principle and MRPII (s. e.g.
BERTRAND et al.). Using lot size smoothing between stages WIP is
further reduced. Finally, in the adaptation to non-stationary demand,
leaving throughput time fixed, we further reduce stocks.

I
medium term planning
determination of
stationary capacities
stochastic.
optimal lot.iua and Phase I
throughput time.
(stationary ca.e)

- relea.e date.
adapted size of lots
(non-stationary cue)
Phue II

forecasts (lot sizea

release dates
short term scheduling I
Fig. 9 The information feedback in the QT-Approach

With these lots and throughput times we were then able to use as a short
term scheduling rule a simple delay rule. More complicated rules, taking
into account, e.g., capacities turn out to be not necessary since capacity
adaptations have already been considered on a higher level (s. Fig. 9).
The QT-Approach, as it was presented in the main body of the paper,
used a markovian Jackson-Net approximation to calculate optimal lot
 267

sizes. The pertaining lead times, however, were then calculated for more
realistic distributions. All calculations were restricted to realistic serial
production structures. Non-serial structures, however, can also be taken
into account (HAFNER).
One of the nicest features of the presented approach are the interfaces of
the work order planning module (Phases I and II) with the higher and
lower modules. With respect to the medium term planning module there
is no abrupt disaggregation of the medium term quantities of finished
goods to individual items. Instead, only general lot processing times
are calculated taking into consideration the capacities determined at the
medium term production planning stage and the stationary stochastics
of the short term scheduling stage (s. lower feedback arrow in Fig. 9). On
the other hand, with respect to short term scheduling taking forecasts (s.
feeback arrow in Fig. 9), non-stationarity is smoothly introduced by the
adaptation procedure of Phase I.
In fact, Phases I and II describe two different hierarchical planning levels.
Phase I, in calculating setup frequencies, is mainly concerned with aiding
to adapt capacities on the medium term planning level (see upward arrow
in Fig. 9). Hence, the calculation of this phase should only be done e.g.
every three months. Having the setup frequencies one may then in Phase
II, adapt lot sizes (Q --* Q) much more frequently, which, in fact, has
to be done, if one applies the proposed QT-Approach within a rolling
horizon regime.
268

References

AXSATER, S., SCHNEEWEISS, CH., SILVER, E.A. (eds.): Multi-Stage

Production Planning and Inventory Control, Berlin, Heidelberg,
New York 1986
BERTRAND, J.W., WERTMANN, J.C., WIJNGAARD, J.: et al.: Pro-
duction Control - A Structural and Design Oriented Approach,
Amsterdam 1990
HAFNER, H.: Ein Warteschlangenansatz zur integrierten Produktions-
planung, Heidelberg 1992
HAFNER, H.: Lot Sizing and Throughput Times in a Job Shop, in: Int.
Journal of Production Economics 23(1991), pp. 111-116
HEINRICH, C., SCHNEEWEISS, CH.: Multi-Stage Lot-Sizing for Ge-
neral Production Systems, in: S. Axsater, Ch. Schneeweiss, E.A.
Silver (eds.): Multi-Stage Production Planning and Inventory Con-
trol, Lecture Notes, Berlin, Heidelberg, New York 1986
KARMARKAR, U.: Lot Sizes, Lead Times and In-Process Inventories,
in: Management Science (1987), pp. 409-418
SCHNEEWEISS, CH., SCHRODER, H.: Planning and Scheduling the
Repair Shops of the Deutsche Lufthansa AG - A Hierarchical Ap-
proach, in: Production and Operations Management (POM), 1
(1992)
SHANTIKUMAR, J., BUZZACOTT, J.: Open Queueing Network Mo-
dels of Dynamic Job Shops, in: International Journal of Production
Research 3(1981), pp. 255-266
SCHNEEWEISS, CH.: Planung 2, Konzepte der ProzeB- und Modellge-
staltung, Berlin, Heidelberg, New York 1992
 269

TIJMS, H.C.: Stochastic Modelling and Analysis: A Computational Ap-

proach, John Wiley & Sons, Chichester, New York, Brisbane 1990
VOLLMANN, T., BERRY, W., WHYBARK, C.: Manufacturing Plan-
ning and Control Systems, Homewood, Ill. 1988
WHITT, W.: The Queueing Network Analyzer, Bell System Tech. J.
62:2779-2815, 1983
WHITT, W.: Performance of the Queueing Network Analyzer, Bull Sy-
stem Tech. J. 62:2817-2843, 1983
ZIMMERMANN, G.: Quantifizierung der Bestimmungsfaktoren von
Durchlaufzeiten und Werkstattbestanden, in: ZfB 10(1984), pp.
1016-1032
 Issues in Specifying Planning Horizons for Production
Planning witbin elM Environments
By

S. D. Thompson, Assistant Professor; J. A. Jewell,

Graduate Student; and W. 1. Davis, Professor
Department of General Engineering
University oflllinois at Urbana-Champaign
117 Transportation Building
104 South Mathews Ave.
Urbana, IL 61801 USA
Telephone: (217) 333-2730

Abstract

The development of production plans for the CIM hierarchy directs the firm's
manufacturing activities over an extended time horizon. All of the subordinate CIM functions
(e.g. production scheduling, material and capacity requirements planning, purchasing, and
others) are impacted by this plan. The fIrst step in the development of a production plan is the
specifIcation of an appropriate planning horizon--a complex task that is plagued with numerous
uncertainties. To address this task, this paper views the production planning problem as a two-
point boundary value problem where boundary conditions must be specifIed at both the
beginning and at the end of the planning horizon. A diffIculty arises in the specifIcation of the
ending boundary conditions since they are generally unknown. Furthermore, an incorrect
specification can have a significant, perhaps detrimental, effect upon the quality of the
developed plan. An additional requirement for the planning horizon is that it be of suitable
duration to permit an effective integration of production planning with the other CIM functions.
The reported research addresses the uncertainties in specifying the boundary conditions.
An integrated solution approach that employs both Monte Carlo simulation and mathematical
programming is used to consider the inherent uncertainties associated with the production
planning problem. For the considered computational examples, it is shown that the
specifIcation of boundary conditions do effect the quality of the derived production plan.
Furthermore, the results also suggest that planning epochs exist. The observation of planning
epochs implies that a variable length planning horizon, rather than a constant length horizon,
should be employed within the context of a rolling horizon implementation of production
planning. Using this approach, the planning would be performed to the end of the current
epoch where boundary conditions can be specifIed with greater accuracy. Finally, the issue of
establishing the length of an epoch is discussed.
 271

1. Introduction

Since finished product demands vary over time and often demonstrate seasonal
variation, production planning (PP) over an extended horizon becomes essential in preventing
myopic resource allocations. The considered planning horizons are typically subdivided into
discrete time units, commonly known as production periods. PP is then performed for each of
these production periods so that production resources are efficiently allocated, given the
realized and anticipated product demands.
A hierarchical production planning (HPP) approach (see Bitran and Hax, 1977; Bitran,
Haas, and Hax, 1981 and 1982) has been predominantly advocated for modeling PP decisions.
HPP decomposes the overall PP problem into a series of decision levels and corresponding
subproblems. At the highest or aggregate production planning (APP) level, products and
resources are typically aggregated into common groups. Often, a planning horizon of a year or
more with monthly planning periods is commonly employed in planning for the aggregate
product groups. At the lowest or the detailed production planning (DPP) level, planning
horizons of a week or two with daily planning periods are considered as more frequent and
specific production decisions must be made. The eventual outcome of the HPP process is a
master production schedule (MPS) that specifies the quantities and timings for producing
various end products.
The importance of an "integrated systems approach" in providing more robust computer-
integrated manufacturing (elM) systems is widely recognized by researchers and practitioners.
To achieve higher levels of integration, more manufacturing functions must be eventually
addressed by elM systems. The incorporation of PP into the elM decision-making and
control hierarchies appears to be especially important as the developed production plans will
direct the manufacturing activities over the extended time horizons. As a result, the efficiency of
all subordinate elM functions, such as production scheduling, are ultimately impacted by these
plans. If existing HPP frameworks are to be incorporated into elM systems in an effective
manner, these hierarchical models must be significantly enhanced (see Davis, Thompson,
White, 1990; and Davis, Thompson, White, 1991). Specifically, improved methods must be
developed for considering interactions among hierarchical decision levels, integrating the
associated decision making and control, dealing with the high level of uncertainty in the
planning environment, and determining the most appropriate planning horizons for the various
PP decision levels. While all of these issues are very important, this paper focuses on the issue
of specifying the appropriate planning horizons for the HPP decision levels, which is
necessarily the first step in the development of a production plan. The specification of the
planning horizon length will directly influence the planned production quantities for the final
products during each included production period. It will also affect the computational efficiency
272

of the modeling as the number of decision variables increases with each additional production
period included in the planning horizon.
While planning horizon studies have been performed for both the PP and the lot sizing
problems (see Baker and Peterson, 1979; Bookbinder and Hng, 1986; Carlson et aI., 1982;
Chung, et al., 1988; Kunreuther and Morton, 1973, 1974; McClain and Thomas, 1977; Miller,
1979; Nagasawa, et al., 1982, 1985; Sastri and Feiring, 1989; and Sridharan, et al., 1987),
these investigations have made one or more of the following restrictive assumptions: (1) a
deterministic planning environment, (2) a single finished product, or (3) unconstrained
production capacity. The relaxation of one or more of these assumptions can have a significant
impact on the planning horizon. In practice, the most likely PP scenario would require the
relaxation of all these assumptions. As a result, the objectives of this study were to examine
planning horizon behavior within an uncertain, multiple product, and a capacitated PP
environment. The objectives further included the identification of methods that could be
employed to specify planning horizons for this type of production environment.
In pursuing these objectives. section 2 defines planning horizons and discusses their
delineation on the basis of conceptual arguments and constraints within the production
environment. In section 3, an integrated solution approach, employing both Monte Carlo
simulation and linear programming, is presented as it was used to experimentally investigate the
consequences of various planning horizon specifications for a few computational examples.
Computational results are then presented and discussed in section 4 with reference to planning
horizon specifications. The summary and conclusions are provided in section 5 and include
suggestions for specifying planning horizons and areas for future research. Finally, the
Appendix provides the data used in the computational examples of this study.

2. Planning Horizon Definition and Delineation

The PP planning horizon problem can be viewed as a two-point boundary value

problem. For example, if the initial and final boundary conditions for the selected planning
horizon (production periods 0 and T) are known with certainty, the length of the planning
horizon is of little consequence. In regard to the PP problem, these boundary conditions would
normally specify the net product inventory level and the production capacity level. Net
inventory is the difference between product inventory (lot) and product backorders (BnV, where
"n" denotes the specific product group and "t" denotes the production period. The capacity level
 273

(Pt), as defined here, would normally be a function of the available personnel and equipment,
where "t" again refers to the production period.
Therefore, the initial and final boundary conditions required for the two-point boundary
value problem can be formally stated as follows:

{Il.O, .. . ,IN,O; Bl .O,... ,BN.O; Po} E I.C.(O) [1]

{Il,T, ... ,IN.T; Bl,T, .. . ,BN,T; PT} E F.C.(T) [2]

With respect to the initial conditions [1], it is reasonable to assume that they are known or can
be specified with near certainty. On the other hand, the final boundary conditions [2] are
generally not known with certainty and may be very difficult to estimate.
Given that the final boundary conditions cannot be specified with certainty, selecting an
appropriate planning horizon is an important issue. Planning horizons which are too short lead
to inefficient production quotas and planning horizons which are too long result in
computational inefficiencies and difficulties. The minimum appropriate planning horizon length
(T*) might be defined as the length of time into the future for which planning must be
performed so that production quotas for the current production period are not significantly
different from those obtained when planning is performed using successively longer horizons.
That is, a planning horizon length of T* is indicated when it provides roughly the same
production quotas for the current production period as planning horizons of length T*+ 1
through T*+oo, where T*+ 1 and T*+oo indicate the addition of "1" through an infinite number
of production periods, respectively. The underlying observation is that once the planning
horizon becomes of sufficient length, the production quotas for current period become relatively
independent of the final boundary conditions. The existence of T*, commonly referred to as
the strong planning horizon, has been reported by Wagner and Whitin (1958) , Eppen et al.
(1968), Pekelman (1975), and Chung et al. (1988). Please note, however, that in some
instances the existence of production constraints, such as procurement lead time, may require
the consideration of a planning horizon longer than T*. If this occurs, the longer of these two
planning horizon lengths would be employed.
If we assume strong planning horizons exist for all PP problems, then the resolution of
the planning horizon issue largely becomes a process of determining methods for efficiently
finding T*. For a few extreme PP scenarios, the identification of T* appears to be trivial. For
example, T* would be a single production period if finished product demand can always be
satisfied in the period in which it occurs without using inventory. On the other hand, if product
demand is expected to continually overtax production capacity and there will never be an
opportunity to inventory finished products, T* would again be a single production period. In
more realistic production environments, where inventory use or production capacity changes are
274

expected, the identification of T* is not straightforward. In these more complicated scenarios, it

may be possible to identify regenerative points that will delineate PP planning epochs, which
would be analogous to the regenerative method employed in the statistical analysis of discrete
event simulation results. The regenerative PP points would be expected to occur in production
periods where cumulative fmished product production is nearly equal to the cumulative fmished
product demand (see Figure 1). When this occurs, the boundary condition for net inventory
becomes zero, and the planning in the subsequent periods is effectively decoupled from the
earlier periods. If this decoupling occurs, planning epochs can be delineated.

Units of
Demand/ - - Cumulative Demand
Production
- - Cumulative Production

8000
7000
~rs: Planning I ~;~;:d
I
6000 Planning
~yce .,:... •
5000
4000
3000
2000
1000 ~----------"= Regenerative Points

o
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Production
Period

Figure 1 -- Planning Horizon Delineation Based on Regenerative Points.

To further investigate planning horizon specification and the potential existence of

planning epochs, a number of computational experiments were performed. The modeling
approach and experimental procedures will now be described.
 275

3. Integrated Modeling Approach and Experimental Procedures

An integrated solution approach was employed in this planning horizon study. This
integrated solution approach combined linear programming with Monte Carlo simulation (see
Figure 2) to assess the consequences of uncertainty and capacity constraints while examining
the effects of different planning horizon lengths.
The modeled PP scenario for this study was a three-resource, three-product, fixed
workforce linear program (LP). This is the same linear programming structure that was
previously used by Thompson and Davis (1990) for modeling PP, but it is repeated here to
provide a complete presentation of the work (see equations [3] through [10] below).

Maximize:

NT NT NT MT
L L
n=l 1=1
SntSnt - L L
n=l t=l
XntXnt - L L
n=l 1=1
intInt - L L
m=l 1=1
OmtOmt [3]

subject to:
In,t-1 + Xnt - Int - r nt Bn,t-1 + Bnt = d nt [4]

[5]

L
n=l
PnmtXnt - Omt ~ emt [6]

[7]

[8]

N
L I nt ~ q ~otal [9]
n=l

[10]

Where n = 1, ...,N; m=I, ... ,M; t=l, ... , T (N=3; M=3; T=2, 3, ... , 39) and the remaining
LP notation is as follows:

N total number of products,

M total number of resources,
276

Input number of products,

resources, initial planning
horizon length, probability
distributions, and other
pertinent planning data

Construct LP Structure

Use Monte Carlo

simulation techniques to
sample distributions and
then specify a randomly
sam led LP to be solved

Use XMP to solve the

Increment the planning randomly sampled LP
horizon length by one
production period

Save the Optimal

No LPSolution

No

Determine the means and

Yes
standard deviations for the
first period production
quantities of each set of
optimal LP solutions,
associated with a given
planning horizon length

Figure 2 -- Integrated Solution Approach/or Studying Planning Horizon Length.

 277

T total number of production periods in the planning horizon,

Snt units of product n to be sold by the end of period t,
Xnt units of product n produced by the end of period t,
Int units of product n in inventory at the end of period t,
Bnt units of product n back-ordered at the end of period t,
Omt units of resource m overtime capacity used by the end of period t,
Snt selling price per unit of product n in period t,
Xnt production cost per unit of product n in period t,
int inventory holding cost per unit of product n in period t,
Omt cost per unit of resource m overtime in period t,
d nt units of demand for product n in period t,
emt units of regular time capacity for resource m available in period t,
Pnmt rate at which product n consumes resource m in period t,
r nt fraction percentage (s 1.0) of product n backorders that are retained
from period t-l as demand for period t,
Umt units of resource m overtime available for period t,
qnt maximum number of product n units allowed in inventory in
period t, and
q ~otal maximum units of total inventory for all products in period t.

Regarding the above LP notation, the decision variables and coefficients (constants) are
denoted by upper and lower case letters, respectively. The specific data, statistical distributions,
and sampling procedures used in constructing the randomly sampled LPs are described in the
Appendix. Randomly sampled LPs were solved by using XMP FORTRAN subroutines (see
Marsten, 1981).
In reference to the experimental procedures, the effects of planning horizon length were
mainly examined in terms of the first period production quantities (Xn 1's). The reason for
focusing on the Xn 1's is that within the typical rolling horizon approach, only the production
quotas (quantities) of the upcoming or current production period are actually employed. The
production quotas for the subsequent production period are planned at the end of the current
period by moving the planning horizon ahead one period and then replanning when the end of
the current period is realized. In this manner, planning is based upon the most up-to-date
forecasts, which are maintained for a constant T production periods into the future. When the
Xnl's define the production plan, as is the case in rolling horizon implementation, they become
the principal concern with respect to specifying a planning horizon.
While studying planning horizon behavior, cyclic product demand patterns were
employed. To consider the effect of planning for X n l 's at different points within the demand
cycle, four computational scenarios were investigated. In Figure 3, these scenarios are labeled
as Base 1, Base 2 , Base 3, and Base 4 Demand. For each of these four demand scenarios,
planning horizon lengths of 2 to 39 production periods were systematically examined. Note that
1000 randomly sampled LPs were solved for each planning horizon length. Overall, 152,000
278

LPs were generated and solved using the integrated solution approach. The size of these LPs
range from 30 decision variables and 32 structural constraints (the 2 period problem) to 585
decision variables and 624 structural constraints (the 39 period problem). The computational
results provided by these optimal LP solutions will now be summarized and discussed.

4. Computational Results

To examine relationships between planning horizons and the selected Xn l's, two
approaches were employed. The first approach was to graphically study the behavior of Xn l's
with regard to planning horizon length and also net inventory. The second approach attempted
to distinguish the appropriate planning horizon len~th by statistically comparing the means of
the Xnl's. More specifically, an attempt was made to identify a T* that had Xnl's that were
statistically equivalent to those of T*+ 1 through T*+oo (see Figure 4).
When the means of Xn l's and planning horizon length were graphed, the results
suggested that Xn l's do eventually stabilize or reach a steady state value. It was further
observed that the total first period production quantity (XNO stabilized with the Xn l's of the
individual products (see Figure 5). For instance, all Xnl's of the Base 1 Demand appeared to
stabilize once the planning horizon length became 8 production periods or greater. Similar
behavior was observed for the other three base demands, but the stabilization point was
dependent on the specific base demand scenario (see Figure 6). As an example, the XNl of the
Base 2 Demand appeared to become stable once the planning horizon length reached 4
production periods. The observation that planning horizon length varies with the location of the
first production period in the overall demand cycle has also been reported by Chung et al.
(1988). Another important observation with reference to Figure 6 is that the Base 3 and Base 4
Demand scenarios seem to be relatively stable from the start. These two base demands appear
to be approaching special planning circumstances in which a planning horizon of only a single
production period may be sufficient. In the case of the Base 3 Demand, the first production
period is at highest point on the total demand cycle. As a result, the production capacity would
generally be for meeting total product demand for that period and the opportunity to build
inventory would rarely occur. Similarly, Base 4 Demand also results in a tight relationship
between demand and production capacity in the first production period.
To investigate the relationship between net inventory and the stabilization point (SP) of
the XN1's, average net inventory was computed for the 1000 optimally solved LPs of each base
demand in which a planning horizon of 39 production periods was employed. The reasoning
behind these computations was to try to estimate points where the net inventory would likely be
 279

-- Demand for Product 1

3000 - Demand for Product 2
~ 2500
·2
::::> 2000
.s
"d lS00
§
e 1000
o 500
04-~~~~~~-+-+-+~~~~~~~~+-+-~~~~~~~
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
Production Period
3000

~ 2500
·2
::::> 2000
.s
"d 1500
§
e 1000
o 500

o1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
Production Period
3000

l!l 2S00
·2
::::> 2000
.s
"d 1500
§
e 1000
Q 500

2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
Production Period
3000

l!l 2500
8 2000
.s 1500
"8
ro 1000
5
Q SOD

2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
Production Period

Figure 3 -- Average Product Demands Illustrated over 26 Production Periods for the
Four Scenarios.
280

X~ t = planned production quantity for product n in

, period t using a planning horizon of h periods

T
Xu

, ,
T
X2,1 Planning Horizon of
T Production Periods
T
X3 ,1
t2 p~~
I I
II .-
1 3 4 T

XT+l
1,1

, , , , ,II ,
XT+l Planning Horizon of
2,1
T+1 Production Periods
XT+l
3,1
~

1 2 3 4 5 T+l

Hypotheses Tested:
-T -T+I
1. Xu = XI,I
-T
2. X2,l =-T+I
X 2,l

3• -T -T+I
X3,l = X3,l
3 3
-T
~
4. 4J Xn,l = 4J -T+I
~
Xn,l
n=l n=l

Note: the mean for each product quota was determined

from 1000 sampled observations

Figure 4 -- Statistical Comparisons of First Period Production Quantities.

 281

Xnl in

--..
Units
1400
--~--~---- ..... -,----...,------................-----~-."...,..---,,-- ....---..
1200 - Product 1
Base 1 Demand Product 2
1000 - -- Product 3

800
. . . . . . " . · .. _ - _ . . . . oo _ • • • • - ,, _ _ _ _ _ - " ' - - . - - _ . - . - . . . . . . . . . . . . . . . . . . . - .. - . - - - - - -. . . - - - . . . . . . . . . . . . . , , - - _ . - . " - . - . . . - • • • _ - , - - _. . . . - _ .

600

400

2 3 4 5 6 7 8 9 10111213141516171819202122232425262728293031323334 3536 373839

Planning Horizon Length in Production Periods

(a) average first period production quantities of the individual products

XNl in
Units
2800

Base 1 Demand
2600

2400

2200

2000

2 3 4 5 6 7 8 9 1011121314151617181920212223 2425 2627 2829 30 3132 3334 35 3637 3839

Planning Horizon Length in Production Periods

(b) the average of the totalfirst period production quantities

Figure 5 -- Average First Period Production Quantities versus Planning Horizon

Length, Base 1 Demand.
282

XN1 in
Units
2800

2600

2400 Base 2 Demand

2200

2000

1800
2 345678 91011121314151617181920212223 242526 27282930313233343536373839
Planning Horizon Length in Production Periods
~1in
Units
2800

2600

2400
Base 3 Demand
2200

2000

1800
2 3 4 5 6 7 8 9 10 1112 13 14 1516 17 181920 21 22 23 24 25 262728 293031323334 3536373839
Planning Horizon Length in Production Periods
XN1 in
Units
2800

2600
Base 4 Demand
2400
-
2200

2000

1800
2345678 9 10 111213 1415 1617181920212223 24 25262728 29 3031323334 3536 37 3839
Planning Horizon Length in Production Periods

Figure 6 -- Total of the Average First Period Production Quantities versus

Planning Horizon Length (Base 2, Base 3, and Base 4 Demand.)
 283

zero if an infinite planning horizon was considered. While a 39 period horizon is far less than
an infinite one. XN l's generally stabilized in planning horizons of 8 periods or less. and
therefore. it is expected that the zero net inventory points of the optimally solved 39 period
problems would approximate the corresponding points from optimally solved problems with
infinite planning horizons. Recall. the significance of these zero net inventory points is that
they should indicate the decoupling points in planning. and thus. delineate planning epochs.
In Figure 7. the net inventory is shown as a function of the production periods for the
four base demands. Note in the cases of Base 1 and Base 2 Demands. the stabilization points
for the XN1 's and the first zero point for net inventory appear to occur in the same production
period (see Figures 5(b). 6. and 7). For the Base 3 and Base 4 Demands. the net zero inventory
point appears to be different than the stabilization point for the XNl 'so However. as discussed
above. the Base 3 and Base 4 Demand scenarios represent special planning circumstances where
the first period demand often exceeds capacity and backordering occurs. Furthermore. it is
difficult to clearly distinguish the stabilization point from the data presented in Figure 6.
Regardless of planning horizon length. the relatively differences in the means of XN 1's were
small (less than 3%) for the Base 3 and Base 4 Demands.
As discussed above. the means for the first period production quantities were also
statistically compared at 0.01 and 0.05 (type I errors) levels of significance. As indicated in
Figure 4. the purpose of these comparisons was to explore the possibility of identifying
stabilization points through statistical procedures. During subsequent comparisons. the mean of
each stabilization point (X~l) was identified graphically (see Figures 5(b) and 6). These means
were then systematically compared to the means beyond the stabilization points
(X~/ t : t=1 ..... ,39-sp). In both sets of comparisons, statistical differences in the means
were evident well after the stabilization points. For example. in the case of Base 4 Demand.
there was sufficient statistical evidence to suggest that ~1 and ~1 were different at both a 0,01
and 0.05 level of significance. The same result occurred when Xf(l and X&'i were compared as
well as for many other comparisons made for the Base 4 Demand. Similar statistical results
were also observed for the Base 2. Base 3. and Base 4 Demands. These outcomes were
somewhat surprising as the relative difference in these means beyond the stabilization points
were generally small (less than 5%). More importantly. though. these results suggest that it
may be difficult (or even impossible) to identify a planning horizon of length T* that provides
the same Xnl's (statistically) as planning horizons of length T*+ 1 through T*+oo.
284

Netlnv,
1200
1000
800
600
400
200
o
-200
-400
-600
-800
1 23 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536373839
Production Period
Netlnv,
1200
1000
::::~~~~:~:ii~~p$!:---··-- ---... ------... -----... -----... -----... ------... -----.
800
600
400
200
o
-200
-400
-600
-800
23 45 6 7 8 9101112131415161718192021222324252627282930313233343536373839
Production Period
Netlnv,
1200
1000
800
600
400
200
o
-200
-400
-600
-800
123 456 7 8 9101112131415161718192021222324252627282930313233343536373839
Production Period
Net lnv,
1200
1000
-----····-::::::-::$~~~:ii~iQ~it<i==-=::::::-::::=::::==:===::::: ... -----.... ----
800
600
400
200
o
-200
-400
-600
-800 -+-r-1r-t--t--t-,_"_,_"_,_,_,_"_,_"_,_"_,_,_"_"_,_, I I I I I I I I I I
23 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536373839
Production Period

Figure 7 -- Net Inventory for the Four Base Demands Over a 39 Period Planning Horizon,
 285

5. Summary and Conclusion

In conclusion, these computational results do support the premise that a T* can be

defined as Xn l's do stabilize once the planning horizon reaches a sufficient length. It also
appears that T* can be identified on the basis of total product demand, rather than the demands
of the individual products. Another important observation is that points of zero net inventory
did show some correlation with the stabilization points for the production quantities. As a result,
there is a possibility that an algorithm can be developed to specify planning horizon length based
on planning epochs. It is also clear that T* varies with the location of the first production period
within the overall demand cycle. This observation would suggest that it is inefficient to use a
constant planning horizon length within the rolling horizon implementation. Similar to the
experimental results of Chung et al. (1988), the computational results of this study suggest that
T* is shorter than the minimum length of seasonality. For at least some PP scenarios, this
observation implies that planning horizons selected as a full demand cycle are inefficient.
In reference to determining an appropriate planning horizon, or equivalently delineating
a planning epoch, it appears that a graphical analysis or a search procedure can be performed to
identify T*. Note that Chung et al. (1988) have already proposed a cost based search
algorithm. However, a more direct search procedure might attempt to determine when the
production quantities have stabilized by comparing the differences in these quantities and
comparing them with a minimum allowable error. One problem with this type of search is that a
minimum allowable error must be specified and this minimum error will likely vary with the
specified problem being modeled. Another problem is determining the number of comparisons
that must be performed to insure stabilization. A graphical approach, as performed in this
study, can also be performed to identify T*, but this approach is rather cumbersome.
Considering that there is correlation between the zero net inventory points and the stabilization
points for production quantities, it seems likely that a concise analytical model can be developed
to specify T*, or equivalently, the regenerative points for planning epochs. While some
analytical based algorithms have been specified in the literature for delineating T*, they have
been proposed for more restrictive PP realizations (see Nagasawa et al., 1982; and Kunreuther
and Morton, 1973, 1974). Future research will focus on determining the applicability of these
earlier algorithms to the more general PP problem. This research will also continue to focus on
the relationships between the boundary conditions and the planning epochs to determine a more
generalized planning horizon algorithm.
286

6. References

Baker, G. R, and Peterson, D. W., "An Analytic Framework for Evaluating Rolling
Schedules," Management Science, Vol. 25, 1979, pp. 341-351.

Britran, G. R, Haas, E. A, and Hax, A C., "Hierarchical Production Planning: Single

Stage System," Operations Research, Vol. 29, 1981, pp. 717-743.

Bitran, G. R., Haas, E. A., and Hax, A. C., "Hierarchical Production Planning: A
Two-Stage Systems," Operations Research, Vol. 30, 1982, pp. 232-251.

Britran, G. R., and Hax, A C., "On the Design of Hierarchical Production Planning
Systems," Decision Sciences, Vol.8, 1977, pp. 38-55.

Bookbinder, J. H., and H'ng, B. T., "Rolling Horizon Production Planning for
Probabilistic Time-Varying Demands," International Journal of Production
Research, Vol. 24, 1986, pp. 1439-1458.

Carlson, R. C., Beckman S. L. , and Kropp, D. H., "The Effectiveness of Extending

the Horizon in Rolling Production Scheduling," Decision Sciences, Vol. 13, 1982,
pp. 129-146.

Chung, C., Chen, I., and Cheng G. L. Y.,"Planning Horizons for Multi-Item
Hierarchical Production Scheduling Problems: A Heuristic Search Procedure,"
European Journal of Operational Research, Vol. 37, 1988, pp. 368-377.

Davis, W. J., Thompson, S. D., and White, L. R, "The Importance of Decompositions

in CIM Control Architectures," in Proceedings of CIMCON '90, ed. Albert Jones,
National Institute of Standards and Technology, NIST Special Publication 785,
Gaithersburg, MD, May 1990, pp. 466-486.

Davis, W. J., Thompson, S. D. , and White, L. R., "Decision Making and Control
Schema for Production Planning in CIM Systems", Joint US/German Conference on
New Directions for Operations Research in Manufacturing, National Institute of
Standards and Technology, Gaithersburg, MD, 1991 (in press).

Eppen, G.D., Gould, F.J., and Pashigian, B.P., "Extensions of the Planning Horizon
Theorem in the Dynamic Lot Size Model," Management Science, Vol. 15, 1968, pp.
268-277.
Kunreuther, H. c., and Morton, T. E., "Planning Horizons for Production Smoothing
with Deterministic Demands - I. All Demand Met from Regular Production,"
Management Science, Vol. 20, 1973, pp. 110-125.

Kunreuther, H. C., and Morton, T. E., "General Planning Horizons for Production
Smoothing with Deterministic Demands - II. Extensions to Overtime, Undertime, and
Backlogging," Management Science," Vol. 20, 1974, pp. 1037-1046.

Marsten, R. E., "The Design of the XMP Linear Programming Library," A C M

Transactions on Mathematical Software, Vol. 7, 1981, pp. 481-497.

McClain, J. 0., and Thomas, J., "Horizon Effects in Aggregate Production Planning
with Seasonal Demand," Management Science, Vol. 23,1977, pp. 728-736.
 287

Miller, L. W., "Using Linear Programming to Derive Planning Horizons for a

Production Smoothing Problem," Management Science, Vol. 25, 1979, pp. 1232-
1244.

Nagasawa, H., Nishiyama, N., and Hitomi, K., "Decision Analysis for Detennining
the Optimum Planning Horizon in Aggregate Production Planning," International
Journal of Production Research, Vol. 20, 1982, pp. 243-254.
Nagasawa, H., Nishiyama, N., and Hitomi, K., "Decision Analysis for Detennining
the Optimum Planning Horizon in Aggregate Production Planning. Part 2: Difference
Between Planning Horizons in Weekly and in Monthly Schedulings," International
Journal of Production Research, Vol. 23, 1985, pp. 423-436.
Pekelman, D., "Production Smoothing with Fluctuating Price," Management
Science, Vol. 21, 1975, pp. 576-590.
Sastri, T., and Feiring, B. R., "Sequential Optimization and Revision of Production
Plans Over Time," Computers in Industrial Engineering, Vol. 17, 1989, pp.
372-377.
Sridharan, V., Berry, W. L., and Udayabhanu, V., "Freezing the Master Production
Schedule Under Rolling Planning Horizons," Management Science, Vol. 33, 1987,
pp.I137-1149.
Thompson, S. D., and Davis, W. J., "An Integrated Approach for Modeling
Uncertainty in Aggregate Production Planning," IEEE Transactions on Systems,
Man, and Cybernetics, Vol. 20, 1990, pp. 1000-1012.

Wagner, H.M., and Whitin, T.M., "Dynamic Version of the Economic Lot Size
Model," Management Science, Vol. 5, 1958, pp. 89-96.

Appendix. Data and Distributions Used in Randomly Sampling LPs

During the implementation of the integrated solution approach, the limits on inventory
(qnt and q~otal), were modeled as being deterministic. The qnt values were set at 1000 while the
value given to each q~otal was 3000. The remaining coefficients (constants) in the LP structure
(equations [3] through [10]) were treated as random variables. The data and randomly sampling
procedures used to generate values for these random variables will now be discussed.
First period (t=l) cost coefficients for the objective function (sn1, Xn1, in1, and omI)
were randomly sampled from uniform distributions. Their upper and lower unifonn
distribution limits are given in Table 1. To model trends in cost, the values of the cost
coefficients for subsequent periods (t ~ 2) were determined using a sequence of generalized
Bernoulli trials. This process involved using the value of the cost coefficient from the previous
period and a randomly sampled change coefficient (see Table 2). The process is summarized by
the recursive relationship,

(c. coeff.l nt = (c. coeff.ln t-1 + (.-1) (c. coeff.)n,t-1, n=l, ... ,3; t=2, ... ,T [11]
288

where: c. coeff. is a cost coefficient (s, x, i, or 0), and

A is the corresponding change coefficient (B:" ~.. B:.. B~.. etc.).

Note that m replaces n in the case of overtime cost coefficients.

The retention of backorders from period to period (rnt) was simply modeled with a
uniform distribution. The upper and lower limits of this distribution are also listed in Table 1.

Product 1 Product 2 Product 3

LP Coefficient Lower Upper Lower Upper Lower Upper
5.90 6.47 6.20 6.47 5.63 6.10
2.70 2.98 2.85 3.15 2.56 2.83
0.34 0.45 0.36 0.48 0.32 0.43
0.60 0.90 0.60 0.90 0.60 0.90
Resource 1 Resource 2 Resource 3
Lower Upper Lower Upper Lower Upper
°ml 2.80 2.90 2.85 2.95 2.75 2.85

Table 1. Uniform Distribution Limits for Selling Price (Snt), Production Cost (Xnt), Inventory
Holding Cost (int), Backorder Retention (rnt), and Overtime Cost (omt) .

Change
Coefficient Percent Change (Associated Probability)
~f, -2.00 (0.40) 0.00 (0.20) 2.00 (0.40)
01, 0.00 (0.20) 2.00 (0.40) 3.00 (0.40)
OJ, -2.00 (0.40) -1.00 (0.40) 0.00 (0.20)
~~, -2.00 (0.40) 0.00 (0.20) 2.00 (0.40)
~, 0.00 (0.20) 1.00 (0.40) 2.00 (0.40)
Of, -2.00 (0.40) 0.00 (0.20) 2.00 (0.40)
~i. -1.00 (0.35) 0.00 (0.30) 1.00 (0.35)
~, -2.00 (0.35) -1.00 (0.35) 0.00 (0.30)
~i, 0.00 (0.20) 2.00 (0.40) 3.00 (0.40)
~?, -1.00 (0.25) 0.00 (0.50) 1.00 (0.25)
~, 0.00 (0.50) 1.00 (0.25) 2.00 (0.25)
~, 0.00 (0.60) 1.00 (0.30) 2.00 (0.10)

Table 2. Change Coefficients and Associated Probabilities for Selling Price, Production Cost,
Inventory Cost, and Overtime Cost.

Resource capacity coefficients (cmt and Omt) were modeled as being normally
distributed. The means and standard deviations for these coefficients are given in Table 3. The
 289

consumption rates of the resources (Pnmt) by the products were also modeled with normal
distributions. The means and standard deviations for these consumption rates are provided in
Table 4.

LPCoefficient Mean Std. Dev.

Clt 400.0 16.7

C2t 746.6 31.1
C3t 480.0 20.0
Ult 100.0 4.2
U2t 186.7 7.8
U3t 120.0 5.0

Table 3. Means and Standard Deviations of Regular Time and Overtime Time
Resource Capacities.

LP
Coefficient Mean Std. Dev.
Pllt 0.2 0.010
P12t 0.4 0.020
P13t 0.0 0.000
P2lt 0.2 0.010
P22t 0.1 0.005
P23t 0.2 0.010
P31t 0.0 0.000
P32t 0.3 0.015
P33t 0.2 0.010

Table 4. Means and Standard Deviations for Product Consumption of Resource Capacities.

The demand for each product during each period (dnt) was modeled with a normal
distribution. Table 5 gives the means and standard deviations for the first 13 periods of the
base product demand. Note that when longer planning horizons were examined, the means and
standard deviations of the first 13 periods were simply repeated, starting with the first
production period.
Finally, note that the initial and final conditions for inventory and backorders (In,O,
In,T, Bn,O, and Bn,T) were set at zero for the computational experiments.
290

Product 1 Product 2 Product 3

Period Mean Std. Dey. Mean Std. Dey. Mean Std. Dey.
1 290.0 29.0 1352.0 135.2 336.0 33.6
2 435.0 43.5 1014.0 101.4 672.0 67.2
3 580.0 58.0 676.0 67.6 1008.0 100.8
4 725.0 72.5 338.0 33.8 1344.0 134.4
5 870.0 87.0 676.0 67.6 1008.0 100.8
6 1015.0 101.5 1014.0 101.4 672.0 67.2
7 1160.0 116.0 1352.0 135,2 336.0 33.6
8 1015.0 101.5 1014.0 101.4 672.0 67.2
9 870.0 87.0 676.0 67.6 1008.0 100.8
10 725.0 72.5 338.0 33.8 1344.0 134.4
11 580.0 58.0 676.0 67.6 1008.0 100.8
12 435.0 43.5 1014.0 101.4 672.0 67.2
13 290.0 29.0 1352.0 135.2 336.0 33.6

Table 5. Means and Standard Deviations for Product Demand (dnt).

IV. Scheduling
 A Closed Loop Control Heuristic for
Scheduling a Manufacturing System
Subject to Random Setup Times
by

Mitchell H. Burman and Stanley B. Gershwin

Operations Research Center
Massachusetts Institute of Technology
Cambridge, Massachusetts, USA

Abstract

A real-time lot-sizing heuristic is developed for a high volume manufacturing system

subject to significant, random setup times. The scheduling problem is modeled as a dynamic
program, generating a threshold-type control policy. Simulation experiments indicate that
closed loop feedback control is a feasible option for shop floor scheduling. A heuristic based
on characteristics of the optimal control policy is developed. Numerical tests are provided
which compare the heuristic with other controllers. Results show the heuristic performs
identically to an optimal control policy and better than other widely used approaches.

1 Introd uction

A real-time lot-sizing heuristic is developed for a high volume manufacturing system subject
to significant, random setup times. This heuristic allocates random excess available capacity
to future production while fulfilling current requirements. It is motivated by the scheduling
requirements of a medical device manufacturer. The facility consists of a multiple stage process
with a long and short term bottleneck (Goldratt, 1986) at the first stage (Stage One) where
machines have setup times which are significant and of random durations. Burman and Gershwin
(1992) suggests a two-step process for scheduling one of these machines which focuses first on lot-
sequencing and then separately on lot sizing. It is assumed that the first step of this process has
been completed and that a given part type sequence called the L list must be adhered to in the
current week of production. The amount of each part type to produce must be determined while
maintaining this established sequence. By making too little, an opportunity for reducing setup
time requirements in a future week may be lost. By making too much, present requirements
may be jeopardized.

Observations at a number of facilities suggest that variability must be incorporated in any

scheduling methodology. It is assumed that the only randomness in this system is the duration
of setup times. Gershwin (1993) suggests modeling systems subject to random events with
dynamic programming, using feedback control. The approach requires a temporal hierarchical
294

I Monthly Customer Orders I

I Monthly Orders Reduced to Weekly Demands I

I Part Types are Sequence, to MInimIZe Setup Times I

IDynamic Lot-Sizing Subject to Random Events I

Figure 1: Typical production hierarchy found within manufacturing organizations.

decomposition of the system. Each level in the hierarchy deals with decision making at different
time scales which may be modeled separately (Figure 1).

In this paper, we concentrate on the lowest level of the hierarchy (lot sizing). The proposed
heuristic avoids the complexity often encountered in more common combinatorial optimization
approaches to such discrete event problems (Baker and Scudder, 1990). Dynamic lot sizing
may also be approached as a variation of the Economic Lot Sizing Problem (Elmaghraby, 1978;
Wagner and Whitin, 1958). These approaches analyze aggregate dynamics based on average
values for production rates and setup times. Lot sizes are adjusted based on weekly demand and
capacity fluctuations. This overlooks the details required for responding to short term random
events, as we do in this paper. Others treat the problem deterministically (Trigeiro, 1989) or as
a queueing system (Shioyama and Kise, 1989). These methods are frequently restrictive, tend
to be more descriptive rather than prescriptive and are not practical in real time.

The goal of our work is to develop a policy which is based on a real problem and can be
implemented. Recent literature reviews of production research identify practical research as a
domain which has been neglected (Graves, 1981; Rodammer and White, 1988; Burman, 1992a).

Paper Summary Section 2 provides a list of key definitions for the paper. Section 3 details
the manufacturing environment. Section 4 develops the dynamic program which is used to
characterize the optimal controller and provides an example problem. Section 5 discusses the
results of numerical experiments and the heuristic development. Section 6 compares the heuristic
with other controllers and current manufacturing practices. A list of future research directions
and summary remarks are provided in the Conclusion in Section 7.
 295

2 Definitions

The concepts of machine setup, lot sequencing, lot sizing and closed loop feedback are
essential to understanding this paper.

A setup is a change in the machine configuration. The setup time is the time required to
switch machine configurations between the production of two different part types. Stage One
has setup times which are significant and of random durations.

A lot is a number of parts, all of the same part type, made contiguously within a production
schedule. Lot sequencing is the determination of the order of setup changes. Lot sizing is
the determination of the amount of each part type to produce. Burman and Gershwin (1992)
suggests a combined two-step scheduling process which focuses first on lot-sequencing and then
separately on lot sizing for our problem. Lot sizing is often complicated by random production
times, variable demand, machine failures and random setup times. In this paper, it is assumed
that the only randomness in the system is the duration of setup times.

The system state is the critical information required to make a control decision. Closed loop
feedback control specifies the action to be taken at every moment as a function of the system
state. Figure 2 illustrates a feedback cycle: the system state is determined ( e.g., how much time
is left in the week, remaining demand requirements, etc.); a control action is implemented based
on a measurement of the system and a feedback law; the system state changes in reaction to the
control and the environment; and the cycle is repeated. The system state in our problem is the
current time, the number of projected lot requirements completed and the number of completed
setups. The control action is to build more within the current configuration or to set up to
another configuration.

I
Evolution of
System State Measurement
Due to Control of the System
Action and the State
Environment
Action Based on System
t State and Feedback
Control Policy

Figure 2: A Simple Feedback Loop

296

3 System Description

This paper concentrates on the development of an algorithm for the control of one machine
with significant and random setup times. Every Monday morning, demand for a number of
different part types is provided which must be scheduled to be produced for delivery by Friday
evening of that same week. It is assumed that the number of hours available for production is
fixed. In establishing the weekly demand, management makes certain that the maximum amount
of time required for production and setups is less than the number of hours available, for a given
level of confidence. Orders required by the end of the current week are frozen (Sridharan and
Berry, 1990) and may not be altered. Projections for the following week are provided at the
same time. These projections rarely differ from the actual orders. Current demand must be
completed by the end of the week with a high level ofreliability.

System conditions are such that enough raw material upstream and enough buffer space and
capacity downstream exist so that Stage One is almost never blocked or starved. Therefore,
Stage One is both a long term and short term system bottleneck (Goldratt, 1986) and the focus
of our work. The cost of multiple tooling prohibits the production of any part type on more
than one machine. We thus suggest that the entire system be scheduled by controlling each of
the first stage machines independently by the same efficient algorithm.

The system must have excess short term capacity to account for variability. This excess
time may be used to create inventory. In this facility, inventory has a very low cost, little
obsolescence, and requires little space. Although management limits inventory to the level of
projected demand, its low cost and the excess short term capacity in the long term system
bottleneck (Stage One) suggests a limited build-ahead policy. This assumption opposes current
thinking such as Just-in- Time (Golhar and Stamm, 1991).

Combining orders of identical part types from consecutive weeks reduces the average time
required for setups. This reduction of setup time allows greater demands on the system. The
objective of our control policy is to minimize the number of setups required in the future by
efficiently allocating excess capacity in the current period. This must be accomplished without
jeopardizing the c~mpletion of the current week's requirements.

4 Model Development

The purpose of the control policy is to minimize the number of required future setups while
meeting current production requirements. It is assumed that a sequence of part types is provided
in the £ list. Burman (1992b) proves that under realistic assumptions regarding setup times, if C-
is an optimal sequence of setups (with respect to minimizing total setup time), any subsequence
 297

of this sequence is also optimal. This lot-sizing algoritlun determines the size of each lot in real
time, while maintaining the part type sequence dictated by the L list.

The system is modeled as dynamic program (Section 4.1) and an example is detailed (Section
4.2).

4.1 The Model

Setups are the only random events in our system and there is a countable number of them.
Since decisions are only required upon the completion of random events, a discrete dynamic
progranuning model indexed to setup completions is formulated. Upon the completion of each
setup, the system may produce only the current requirements and immediately set up for another
part type; or produce the next week's requirements in addition to the current requirements and
then change setups (Figure 3).

System Parameters

T = The amount of time (in hours) available each week for production
and setups.

N The maximum number of setups required within in a week.

Cim = Production time for current demand (Cn or projected demand for the
following week (C;) for Part Type i.

S = Setup time. This is a random variable.

P(S) Probability mass function for S.

P}, P2 = Penalty coefficients in the model objective function.

'l'hi. W.ak'. Proc1uction

bquir-.nt ~or "art
Type i

Next Weak.'. ion

bquir-.nt ~or Part Type i

Figure 3: Decision Points.

298

System State The system state contains all the relevant information for deciding whether
to switch machine configurations or continue production of the lame part type. Due to the
enforcement of the established sequence of setups, the index i of the most recently completed
setup and the index i of the part type currently being produced are one and the same. Therefore,
the system state, upon the completion of a setup i only consists of current time (Ti) and the
number offuture lots completed (R;):

Ti = The time of the ith setup completion. (1)

Rj = The number of extra lots completed at Tj • (2)
Zi = (Ti,R;) (3)

Control Function The control function is a mapping from the system state to a control
action. The control actions are to build inventory or to switch setups:

Build Inventory with the current machine configuration.

(4)
Switch Setups.

Objective Function The goal of the dynamic program is to develop a control policy which
results in a good system performance. System performance can be measured in terms of the
number of future lots completed and whether all current requirements are met. The objective
function is designed to reward a good system performance by increasing linearly with the number
of future lots completed (Ri) and decreasing quadratically with the number of current jobs not
completed by the end of the week (J"). There is no precise economic interpretation of the cost
function. Therefore, different coefficients (PI and P2 ) were tested to determine which provided
a policy which best met the above objectives (Section 5.1). The objective function is:

if Ti+I 2: T and Ti :5 T
(5)
otherwise

where

r _
-
{N -+
N
i
1-i
if T-Ti 2: ei
if T-Tj < ei

System Dynamics The random variable tj represents the amount of time between decision
points as a function of system state, the environment and the control action. H the control
action is to build an extra job, ti is the time to produce the present requirement (ei) plus the
time required to produce a second job (en plus the length of the random setup time (5).

Ti+I - Ti = ti(Zi,J'i(Z;)) = ct + Pi(Zi)C~ + S w.p. P(S) (6)

The reward for each job built ahead is 1. Therefore, the accumulated reward is governed by:

(7)
 299

Then:
(8)

General Program Formulation Since the objective function has a non-zero value only at
the end of the week (5), the cost-to-go function (J) at T; is written as an expectation (E;) of
the objective function value at the last decision point (iO) of the week. Then:

Cost-to-Go Function

J;"(T;,R;) = maxEdg;.(T;.,R;.)] where iO = max[k: Tic < T] and J.' = J.';, .. . ,J.';. (9)
I'

subject to (8).

Dynamic Progranuning Recursive Equations

We generate the solution to (9) by recursively solving the following equations:

J;"(T,R) = maxE [Jt+I(f;(T;,R;,J.';,t;))], for i

1'.
= 1 to 10 (10)
= maxE [Jt+I(T;
1'.
+ t;,R; + J.';)]

4.2 An Example

An example is described to get an understanding of how an optimal controller behaves (Table

1). We make no initial assumptions about the penalty coefficients PI and P2 values. A number
of different values are tested (Section 5). This example is chosen because of the similarity of the
magnitudes of parameter values to the actual facility and for simplicity in running the numerical
tests.

T = 100 Hours.
N = 10.
em = 5 Hours. m = 1,2 i = 1 to 10.
0.5 for S = 2.5

1
P(S) = 0.5 for S = 5
0 otherwise.

Table 1: Example Parameters

5 Heuristic Derivation

In this section, the characteristics of the control policy generated by the dynamic program-
ming algorithm (Bertsekas, 1987) are described. Numerous values for PI and P2 are tested
(Section 5.1). Based on these characteristics, a heuristic is recommended (Section 5.2).
300

5.1 Numerical Derivation of Control Policy

An analytical expression has not been found for the cost-to-go function in (9). Numerical
experiments have been conducted to determine the control policy associated with each system
state, for different PI and P2 • After numerous tests, it was concluded that the best results were
obtained when PI and P2 were both set to 1 and consequently these values are used in all further
experiments. In addition, it was observed that variations in R had no effect on the control policy.
Therefore, R is not accounted for when partitioning the state space to identify control actions.
Different shadings of the remaining state space in Figure 4 represent different control actions
associated with each possible system state (i.e., light grey represents a build more action, dark
grey a switch setup action and white is an indifference region between the two policies). Since
the value of the cost-to-go function is insensitive to the control action for system states in the
indifference region, we decided to chose the build more option for all such points.

Figure 4 also indicates to always build more when Ti is between 8 and 18. The model
determines that there is not enough time to complete an another job. To maximize the cost-to-
go function, the controller builds extra inventory, treating incomplete jobs as unavoidable. In
practice, one would switch setups to minimize the amount of unfulfilled current requirements.
It is assumed that the existence of this region is an end effect due to the nature of discrete finite
horizon dynamic programming modeling and that to switch setups is a better decision.

By eliminating the indifference region and the anomalous build more region, a threshold
policy represented by the threshold IJpprozimlJtion (Figure 4) is suggested. For every number of
jobs (11-i) left in the system, there is an associated time ai. If more time than ai is left in the
week, build inventory; otherwise, do not .

...o
- 100
T~e
90 80 70 ~ 50 4(1 30
Left Within The Week (T ~,)
20 10 0

Figure 4: Threshold Approximation

 301

5.2 Other Thresholds and the Heuristic

In actual practice, business conditions may vary from week to week (e.g. anticipated future ca-
pacity shortages, lack of available overtime, inventory shortages, etc.). Variability in conditions
may alter the objective function, resulting in a shifted threshold. For more aggressive strate-
gies, where inventory is maximized at the expense of the completion of current requirements,
the threshold curve moves down and to the right. For more conservative strategies, where the
completion of current requirements takes precedence, the threshold moves up and to the left.
In addition to the threshold approzimation of the optimal control policy when the penalty co-
efficients are 1, three other thresholds are represented in Figure 5. These are the aggressive
threshold, the certainty equivalence (GEG) threshold and the conservative threshold.

The risk-taker's or aggressive threshold (Figure 5) is the rightmost feasible threshold, given
our assumptions. If all future setup times are known to be the smallest value of S, this would
be the threshold boundary. Any decision to produce extra inventory when the system is in a
state to the right of this line guarantees that the week's requirements cannot be built. The
conservative threshold (Figure 5) is the leftmost feasible threshold, given our assumptions. If
all future setup times were known to be the largest value of S, this would be the threshold
boundary. A decision to produce extra inventory when the system is in a state to the left of
this line could never jeopardize current production. A third threshold exists between these two
extremes which accounts for all future events based on their expected amount of time, labeled
the GEG Threshold in Figure 5. This threshold is used commonly in certainty equivalence
controllers (Bertsekas , 1987).

The threshold approzimation starts close to, but below the conservative threshold and ap-

'rIuwbold -A"'"
IAppromaatioD
~
~ ~
1\ ~~"/

UCoaoenadn I
\
....
-A ~~ A._lye
l r
Threshold
A~ .. /
~ ~ ~ -fTlireellold

A ~. ~
~ V..
... ro/
M~EC
TIIresholdII- ~
';'
10 h I I
~ ~ ~ ~ ~ ~ ~ ~ ~ 0
OJ!.:ilM Le~t WitlUn 'rh. Week (OJ!.)
1

Figure 5: Other Thresholds

302

proaches it as the week progresses. Therefore, the conservative threshold is suggested as the
heuristic because it is very easy to calculate.

6 Numerical Comparison Among Policies

In the previous section, a heuristic is suggested which is believed to effectively schedule the
facility. In this section, the heuristic is tested. A number of policies are described to be simulated
(Section 6.1). This is followed by a description of the simulation methodology (Section 6.2) and
a summary of the results (Section 6.3).

6.1 Various Policies

Heuristic Our heuristic is a closed loop feedback back policy using the conservative threshold
as described in Section 5.2.

Certainty Equivalence Controller (CEC) Section 5.2 also describes an expected value
threshold line between the conservative and aggressive extremes. The certainty equivalence
controller (CEC) uses this threshold.

Open Loop Policies Open loop policies are policies in which the system state has no effect
on the control action. Casual observations indicate a number of open loop policies are used in
industry: the open loop aggressive policy, the open loop conservative policy, and the open loop
expected value policy.

Open Loop Aggressive Schedulers frequently realize that combining setups has advantages.
The machine needs fewer setups, which is an unpleasant task. Over time, capacity is increased
and more can get done. They establish an open loop policy in which the control decision is
to always build inventory. Unfortunately, the results of such a policy are that the current.
requirements are 'never completed. Such a policy is unacceptable in the due date framework
assumed in this paper and is therefore not examined further.

Open Loop Conservative Some machine operators are overly concerned with not completing
the required work. They only produce the current requirements until they are completely done,
If excess time exists at the end of the week, then the machine may be set up to build inventory.
This is a common practice in factories. Unfortunately, the time required to do extra setups at
the end of the week may substantially reduce capacity. By using excess capacity as it becomes
available, extra setups are avoided.
 303

Open Loop Expected Value The open loop ezpected value controller calculates the expected
value of the duration of events in the system. For our example, the expected setup time is 3.75
hours. 10 setups require, on average, 37.5 hours and 50 hours are required for for current
production. This takes a total of 87.5 hours to complete. The scheduler plans two extra jobs
for the week in the available 12.5 hours, leaving enough leeway for more than than the expected
number of large setups. This is the open loop counterpart of the CEC. Unfortunately, this policy
only allows for a maximum of two extra jobs in the week, potentially resulting in lost inventory
opportunities. In addition, tests showed this method frequently resulted in current production
requirements not being completed. Therefore, this policy is also not examined any further.

6.2 Numerical Test Design

Cost-to-go functions do not provide a clear picture of the resulting events when they are
employed to generate control policies. Simulations often provide a good picture of what actually
happens when a control policy is applied. Given the limited number of random events, it is
feasible to generate all 210 different random event sequences (10 setups, each with two possible
durations). Each controller is applied to each sequence of events to determine its performance.
By applying the probability of each sequence to the results, probability distributions of the
outcomes from the use of any controller can be obtained.

6.3 Test Results

Four control options are tested: a) the optimal controller with PI and P2 equal to 1, b) a
certainty equivalence controller, c) the heuristic controller and d) the open loop conservative
policy. For each controller, the expected value and variance of the number of jobs completed is
provided. In addition, the percentage of weeks in which all current jobs are not completed and
the percentage of weeks in which 0 through 6 extra jobs are completed is provided (Table 2).

There are three important results from our tests. First, the heuristic performed identically to
the optimal controller. Second, it performed significantly better than its open loop conservative
counterpart. Finally, the performance of the expected value controller was poor.

The expected number of reduced future setups is 2.25 for the optimal controller and the
heuristic. Also in both cases, all current requirements were always met. This is because, in
our example, even slightly aggressive action resulted in a substantial number of weeks with
incomplete current requirements. Therefore, a conservative policy is required to meet the due
date constraint.

The fact that both policies yielded the same results should not be so surprising. A careful
examination of Figure 4 indicates that the conservative threshold lies within the indifference
304

% Incomplete % of Weeks completing x future Jobs

Weeks x=

Controller 0 1 2 3 4 5 6

Optimal Controller 0.0 1.1 16.1 45.1 32.2 5.4 0.1 0.0

E = 2.25, (72 = 0.7

Expected Value Controller 31.6 0.0 0.0 4S.5 44.9 6.4 0.1 0.0

E = 2.5S, (72 = 0.4

Heuristic 0.0 1.1 16.1 45.1 32.2 5.4 0.1 0.0

E = 2.25, (72 = 0.7

Open Loop Conservative 0.0 11.3 69.3 19.2 0.1 0.0 0.0 0.0

E = LOS, (72 = 0.3

Table 2: Summary of Statistics Test for All Controllers

region. This does not prove that both should be the same but provides some intuition for the
results.

The heuristic performed significantly better than both open loop policies. Both the conser-
vative open loop controller and the heuristic managed to always complete present requirements.
However, the heuristic completed an average of 2.25 extra jobs, while the open policy completed
only LOS.

The expected. value scheduler performed poorly. This point is emphasized because many
current schedulers and software packages use expected value methods for scheduling. The re-
sults of these tests may begin to explain why these schedulers perform so poorly in practice in
environments with random events.

7 Conclusion

We began by looking at a machine with random, significant setups times. The objective
was to schedule the machine in real time in response to random events. The problem was
 305

formulated as a dynamic program with the simplifying asswnption that there was no other
randomness in the system other than setup times. An example was presented for analysis, from
which a threshold-type approximation was obtained from the partitioned state space. From this
approximation, a heuristic which is easy to calculate was suggested. Nwnerical simulations were
conducted which compared the heuristic with other policies, including the optimal policy. The
results showed that the heuristic performed very well and that feedback is a feasible way of
dealing with system randomness.

One direction for further research involves modifying the dynamic program to incorporate
machine failures and production rate variability. We hope the results of this extension will
also provide a threshold-type policy. If so, the simplicity of the solution structure would be
easily implementable. We would like to look at more examples and cost functions, as well as
incorporate phenomena such as variable-sized lots and demand levels at different percentages of
capacity. We would like to do more work on determining which future requirements to produce,
with different objective functions and setup time distributions. We are also in the process of
preparing an implementation and would like to compare its performance with previous facility
performance.

References
[1] K. Baker and G. Scudder. Sequencing with earliness and tardiness penalties: A review.
Operations Research, volwne 38{No. l):pages 22-36, (January-February) 1990.
[2] D. Bertsekas. Dynamic Programming: Deterministic and Stochastic Models. Prentice Hall,
New Jersey, USA, 1987.
[3] M. Burman. Recent trends in production research. Technical report, in preparation, MIT,
1992a.
[4] M. Burman. Scheduling a medical device manufacturing system. Master's thesis , in prepa-
ration, MIT, 1992b.
[5] M. Burman and S. Gershwin. An algorithm for the scheduling of setups and production
to meet due dates. In IEEE International Conference on Robotics and A utomation, Nice,
France, 1992.
[6] S. Elmaghraby. The economic-lot scheduling problem (ELSP): Review and extensions.
Management Science, volwne 24(No. 6):pages 587-597, 1978.
[7] S. Gershwin. Manufacturing systems engineering. Textbook in preparation, 1993.
[8] E. Goldratt and J. Cox. The Goal, A Process of Ongoing Improvement. North River Press,
Inc., New York, 1986.
[9] D. Golhar and C. Stamm. The just-in-time philosophy: A literature review. International
Journal of Production Research, volwne 29(No. 4):pages 657-676, 1991.
[10] S. Graves. A review of production scheduling. Operations Research, volwne 29{No. 4):pages
646-675, (July-August) 1981.
[11] F. Rodammer and K. White, Jr. A recent survey of production scheduling. IEEE Transac-
tions on Systems , Man and Cybernetics, 18(6):pages 841-851, (November-December) 1988.
306

[12J T. Shioyama and H. Kise. Optimization in production systems: A survey on queueing

approaches. Journal of the Operations Research Society of Japan, volume 32(No. l):pages
34-55, 1989.
[13J V. Sridharan and W. Berry. Freezing the master schedule under demand uncertainty.
Decision Sciences Journal, volume 21(No. 1):pages 97-120, 1990.
[14J W. Trigeiro. A simple heuristic for lot sizing with setup times. Decision Sciences Journal,
volume 20(No. 2):pages 294-303, Spring 1989.
[15J H. Wagner and T. Whitin. Dynamic version of the economic lot size model. Management
Science, volume 5(No. l):pages 89-96, October 1958.
 Resource- and Time Window-Constraint
Production Scheduling With Alternative Process Plans:
An Artificial Intelligence Approach

Andreas Drexl and Arno Sprecher

Institut fur Betriebswirtschaftslehre

Christian-Albrechts-Universitat zu Kiel
D-2300 Kiel 1, FRG

1. Introduction

Make-to-order production of customer-made products has to be done with respect to

product-specific time windows by efficient utilization of manufacturing resources. Most
of the traditional production scheduling approaches are based on the assumption that
for each part only one process plan is available. A manufacturing system with
alternative process plans is supposed to increase throughput rate via eliminating
bottlenecks and handling machine breakdowns [1] .

The problem considered here is closely related to project scheduling (cf. [6,7, 15]) in
the following way: Each product corresponds to a project, i.e. we are dealing with
multi-project scheduling [11]. Production uses renewable resources and consumes
nonrenewable resources [15]. The manufacturing of each product requires the assembly
of parts. Each part requires a number (and sequence) of operations to be done speCified
by a process plan. Here we consider the case where alternative process plans are
available for each part, one of which has to be chosen. The part level in production
scheduling corresponds to the job level in project scheduling and process plans are
shortly denoted as modes [8] . For the sake of shortness we will use the words mode
and job here, too . Like in project scheduling we use the general concept of temporal
constraints [2] in order to model the partial order of jobs as well as the time windows
of products.

Our focus here is not on the operations, but on the part level, i.e. we simultaneously
have to choose one mode per job and the sequence of jobs. Thus we are not dealing
with shortterm (operational), but with mediumterm scheduling. This approach seems
to be adequate in manufacturing situations (such as small batch machine tool assembly
[8]) where on the operational level a lot of work has to be performed manually.
308

We address the constraint satisfaction-problem, i.e. we want to determine a production

schedule (i.e. a mode and start time assignment to each job) which satisfies resource-
and time window--constraints.

The paper is organized as follows: In section 2 we introduce the underlying assumptions

and the basic notation. Section 3 deals with establishing the time bounds for the
production (scheduling) of the parts. A mathematical programming model with binary
variables is provided in section 4. In section 5 we present a (conceptual) integer
programming model. In section 6 we outline the implementation of the integer
programming model in CHARME, a PROLOG-based constraint propagation program-
ming language. Preliminary computational results based on several (fully reproduced)
instances are provided in section 7. Conclusions for future research can be found in
section 8.

2. Assumptions and Notation

The make-t<K>rder production scheduling problem under consideration will be

introduced based on the following assumptions:

AI. P products have to be manufactured; p (=I, ... ,P) denotes a specific product.
Product p has the release date ,z and the deadline a'; w := [,z , .{I denotes
p p p p p
the manufacturing time window of product p.

A2. Product p consists out of J partially ordered jobs. 1'. (S .) denotes the set of jobs
P J J
immediately preceeding (succeeding) job j. An arc interrelates dependent jobs h
and j E Sh . W.l.o.g. we assume that each product has one unique source (start
job) and one unique sink (end job).

A3. We assume one additional (super-) source (j=I) and one additional (super-) sink
(j=J). Furthermore jobs are labeled consecutively, i.e.
p
j = 1,2, ..., J 1+1, J 1+2, ... , E J +2;
p=l p

thus the job number (index) j identifies the related project exactly.

A4. Temporal constraints between jobs hand j E Sh are modeled via arc weights
(Th~n, Thj) , 1
where Th n denotes the minimum finish-to-start lag and ThjX
denotes the maximum start-to-finish lag between dependent jobs hand j . It is
1
assumed that Th n = - 00 (ThjX = (0) is associated with the case, where only the
 309

maximum (minimum) lag exists. Note that Th~n < 0 enables us to model schedule-
overlapping of dependent jobs.

A5. The time windows of the products are modeled as follows based on the temporal
constraints introduced in A4: We generate additional arcs from the unique source
j=l to the start job h (with 'Ph = ¢) of each product p with the arc weights
(Tf~n = ,zp , Tf~x = (0). Furthermore we introduce additional arcs from the job
j=l to the end job i (with s.
I
= ¢) of each product p with the arc weights
( ".min
, 1i
= - 00 ".max
' ' 1i
= Up)'
-/\

A6 . Job j may be processed based on one of the modes m = 1, .. .,M .. We consider the
J
nonpreemptive case, i.e. each job, once initiated in a specific mode, has to be
finished without preemption. Referring to preemptive scheduling preemptions can
only take place at points of time specified a priori. Note that nonpreemptive
scheduling implies that modes must not be changed.

A7. Scheduling job j in mode m takes d. time units (mode dependent deterministic
Jffi
duration).

A8. There exists a set of 1. renewable resources (=machines, etc.), where resource r E 1.

is available with /'i,P resource units per period; scheduling job j in mode muses
r

k~ units per period.

Jrnr

A9. Additionally we have a set of )/ nonrenewable resources (=budgets, etc.), where

resource r E )/ is available with /'i,v resource units in total; scheduling job j in mode
r

m consumes k~ units.
Jffir

Besides the temporal constraints introduced in A4 there exist six additional ones
(start-to-start, finish-to-finish ; minimum and maximum in all cases; cf. [2]) . Per arc
at most one out of the four minimum (maximum) time lags has to be specified. In order
to save additional symbols we restrict to using only (Th~n, Thr), the temporal
constraints which are supposed to be most important for modeling our make-to-order
production scheduling problem.

Assumptions Al to A9 setup a rather general time-resource-tradeoff problem. The

following definitions will enable us to give a more formal treatment of the problem
below.
310

Definition

(a) A schedule)l := {(j,m,t) I j = 1, ... ,J} is a set of triplets, which assigns to each
job j a mode m and a completion time t.

(b) A schedule )I is called time-feasible, if the starting and completion times of all
jobs fulfill the temporal constraints.

( c) A schedule )I is called resource-feasi ble, if per period- and overall-resource availa-

bilities are satisfied.

(d) A schedule )I is called time-resource-feasible, if it is both time- and resource-

feasible.

3. Determination of Feasible Time Bounds

Now we are going to consider the problem of establishing feasible time bounds. We will
use the following additional symbols:
ES. the critical path g,arliest .§.tart time of job j
J
EF . the critical path g,arliest finish time of job j
J
LF . the critical path latest finish time of job j
J
LS. the critical path latest .§.tart time of job j
J

Unlike in the single-mode case it is not possible to derive the digraph of temporal
constraints by some simple equivalent transformations [2] of all given temporal
constraints to just one type (e.g. finish-to-start time lags). This is due to the mode
dependent job durations d. rendering the "real" job duration not as a simple input
Jm
data but as the consequence of a decision.

Lower and upper bounds for the earliest and latest start and finish times of the jobs
may be derived using the minimum job durations b. . := min {d .
J Jm
Im= 1, .. . ,M .}, the
J
temporal constraints 7~~n only and omitting the maximum time lags 7~jX.

Simple modifications of well-known longest path computation methods [9] provide

these lower and upper bounds for the earliest start and finish times ES. and EF . as well
J J
as the latest start and finish times LS. and LF ..
J J
Obviously [ES. , LF.] represents a bound on the maximum time interval, within which
J J
job j can be scheduled without violating temporal constraints. In the case where the
 311

earliest finish time of the end job of one of the projects exceeds the projects deadline
definitely no time-feasible schedule exists at all.

Now we are going to describe the problem of determining a time-resource-feasible

schedule mathematically. We address the constraint satisfaction-problem, i.e we want
to find a production schedule (assignment of modes and sequence of jobs) which
satisfies resource and temporal constraints.

4. A Mathematical Programming Model With Binary Variables

In the following we are going to formulate the constraint satisfaction-problem as a

mathematical programming model with binary variables [6, 15, 16] . In section 5
another formulation based on integer variables will be provided.

Defining binary variables as follows

x. := 11, job j is processed in mode m and finished in period t

Jmt
0 , otherwise

we get the feasibility problem:

Does there exist a feasible solution to the set of equations (1) - (6) ?

M . LF.
J J
E E X jmt 1 (V j) (1)
m=l t=EF.
J

Mh LFh M. LF.
J J
E E t· x + rm~n < E E (t-d. )x. t (V j and h E 1'.) (2)
m= 1 t=EF h hmt hJ m=l t=EF . Jm Jm J
J

(V j and h E 1'.) (3)

J

J M. LF.
J v J
E E k. E x. (V r E JI) (4)
·-1 m-
J- -1 Jmr t-EF
- j Jmt
312

J M. t+d . -1
J Jm
E E k~ E XJ.mt < ~ (VrElandt) (5)
j=l m=l JIm q=t

x. t E {0,1} (Vj,m,andt) (6)

Jm

The feasibility problem requires the determination of a vector x which is feasible with
respect to the following constraints: (1) requires completion of each job j in exactly one
of the available modes within the interval [EF., LF.]. (2) and (3) state precedence
J J
relations between dependent jobs with minimum and maximum temporal constraints,
respectively. (4) and (5) allow only nonrenewable and renewable resource-feasible
schedules, respectively. To sum it all up (1) - (6) define via linear constraints with
binary variables time-resource-feasible schedules.

Solving (1)- (6) with general purpose (0-1 programming) software necessitates to store
and manipulate a large number of 0-1 variables and constraints. In order to avoid this
we are now going to present an integer programming formulation which needs less
variables.

5. A Conceptual Integer Programming Model

Based on the idea introduced in [17], i.e. a conceptual integer programming formula-
tion of the single-mode case, our scheduling problem may be reformulated as follows :

M.
J
E x.
Jm = 1 (V j) (7)
m=l

y~ + ~~n < y~J (V j and h E 'P.)

J
(8)

yg + r~jX > yfJ (V j and h E 'P .)

J
(9)

M.
J
y~ + . x.
E dJm Jm = yf
J
(V j) (10)
J m=l

J
M.
J v
E E k~ x. 5 K,
r
(V r E JI) (11)
JIm Jm
j=l m=l
 313

M.
E EJ k~ x. < ,,/ (V r E 1, and t) (12)
jEA ID=l JIDr JID r
t

y~ E lls , yf E llf (V j) (13)

J J J J

x. E {O,l} (Vjandm) (14)

JID

Here At denotes the set of jobs active (i.e. using renewable resources) in period t.
ll~
J
= {ES J. ,... , LS.}J denotes the set of possible (integer) start periods , llf
J
=
{EF . , ... , LF.} denotes the set of possible (integer) finish periods. x. is a binary
J J JID
variable denoting whether mode m is assigned to job j (x. = 1) or not (x . = 0).
JID JID

The shortcoming of this model with respect to mathematical programming approaches

is due to the absence of a mechanism for a priori identifying the elements of the sets
At·

In the following we are going to present a PROLOG-based implementation of (7) -

(14) which overcomes this shortcoming in a simple way.

6. A PROLOG-Based Implementation

In the following we will use CHARME [3], a PROLOG-based programming language,

in order to implement (7) - (14). CHARME, a UNIX-based C-reimplementation and
enhancement of CHIP (constraint handling in PROLOG), has been designed for the
following purposes : Models of combinatorial (optimization) problems are transformable
by the use of powerful constraint fourmulation commands into an equivalent declara-
tive code. CHARME-immanent constraint propagation techniques are provided to the
user by simple options in order to (efficiently) search for solutions .

Recently, efficient search via "intelligent" backtracking has attracted much attention
in the artificial intelligence literature [4, 5, 13, 14]. The main emphasis is on static
(before starting variable instantiation) and dynamic (during variable instantiation)
search tree reduction techniques. Loosely speaking, a constraint (hyper-) graph is
constructed, in which one node is associated with one variable and edges correspond to
constraints. The aim is to produce an arc- and path--consistent [10] constraint graph
via making implicit constraints explicit .
314

CHARME provides several high-level commands for formulating constraints in a

compact way. For demonstrative purposes we will reproduce some lines of code.

Based on the definition of

X[j ,m] two-dimensional array for the binary mode-selection variable x.
Jm
M[j] array for denoting the number of modes M. of job j
J
the "one-mode-per-jokonstraints" (7) may be stated via the following CHARME-
command:

for I in 1. .J do
sum (X[I,1 .. I[I]],1);

In addition using

N number of nonrenewable resources

NRS[r] supply of nonrenewable resource r

NRR[j ,m,r] amount of nonrenewable resource r requested from job j being

processed in mode m

the nonrenewable resource constraints (11) may be coded via the following CHARME-
subroutine:

define nonrenres (J,NRS,NRR,N,X)

{
array Req :: [1. .N];
for I in 1..N do
{
Req [I] in O.. NRS [I]
scalyrod (NRR[1..J,1. .5,1] .. [1. .J*5] ,X[1..J,1. .5] .. [1. .J*5],Req[I]);
}
};

Note that Req is an array defined locally in order to accumulate the consumption of
nonrenewable resources (induced by the actual instantiation of variables). In order to
use scalyrod correctly the two-dimensional arrays NRR and X have to be redimen-
sioned. For simplicitity the maximum number of modes has been set equal to 5.
 315

7. Computational Results

In the following we will provide computational results for several (small-sized)

instances. Table 1 provides the data of instance J7/1, i.e. of the first variant of an
instance with J=7 jobs . The other variants of J7 are the following:

Instance J7/2 is identical with J7/1 despite "'~ = 15 and d 21 = 10.

Instance J7/3 is identical with J7/2 despite "'i = 6 and d42 = 6.

Instance J7/4 is identical with J7/3 despite r~i~ = 12.
,
Instance J7/5 is identical with J7/4 despite r~a4
, = 29.

P=1 J=7 111=2 I,VI =2 "'i="'~=100 "'~="'~=100

j s.J M.
J
m d.
Jffi k~ffil k~Jffi 2 k~Jffi 1 k~Jffi 2
f---"
1 {2} 1 1 0 0 0 0 0
2 {3, 4} 2 1 12 1 1 6 5
2 10 1 1 8 3
3 {5} 1 1 11 4 3 2 5
4 {6} 2 1 21 6 1 7 1
2 16 4 3 2 1
5 {6} 1 1 9 2 2 2 2
6 {7} 1 1 1 3 3 3 3
7 ¢ 1 1 0 0 0 0 0

P 1

,z 4 (=rmin)
p 1,2
dp 35 (=rmax)
1,6
(r~~n = 0 and r~jX = w otherwise)

Table 1: Data of instance J7/1

Table 2 provides time-resource-feasible schedules .J{ for the instances J7; the following
remarks should be made:
 316

J7/1 Neither renewable nor nonrenewable resources are scarce. Job j=4 has
slack, i.e. four alternative schedules with (4,2,15), (4,2,16), (4,2,17) or
(4,2,18) exist.
J7/2 Due to lack of nonrenewable resource 1, mode 2 is no more assignable to
job 2. Once more job j=4 has slack and four alternative schedules with
(4,2,15), (4,2,16), (4,2,17) or (4,2,18) exist.

J7/3 Jobs 3 and 4 must not be processed in parallel due to lack of renewable
resources, i.e. job 4 has to be delayed. Three alternative schedules based on
(4,2,26), (4,2,27), (4,2,28) exist .

J7/4 Due to T~i4, = 12 the schedule provided in table 2 for J7/3 is no more
time-feasible. Thus only three time-resource-feasible schedules )( exist
based on (4,2,26), (4,2,27), (4,2,28).

J7/5 Only two time-resource-feasible schedules )( exist with (4,2,26) and

(4,2,27).

j (l,m,t) (2,m,t) (3,m,t) (4,m,t) (5,m,t) (6,m,t) (7,m,t)

J7/1 (1,1,4) (2,2,4) (3,1,14) (4,2,14) (5,1,25) (6,1,34) (7,1,35)

J7/2 (1,1,4) (2,1,4) (3,1,14) (4,2,14) (5,1,25) (6,1,34) (7,1,35)

J7/3 (1,1,4) (2,1,4) (3,1,14) (4,2,25) (5,1,25) (6,1,34) (7,1,35)

J7/4 (1,1,4) (2,1,4) (3,1,14) (4,2,26) (5,1,25) (6,1,34) (7,1,35)

J7/5 (1,1,4) (2,1,4) (3,1,14) (4,2,26) (5,1,25) (6,1,34) (7,1,35)

Table 2: Time-resource-feasible schedules )( for instances J7

A larger problem instance with P=2 and J=12 is provided in table 3. The precedence
relations between jobs 2 to 6 are identical with the ones of jobs 2 to 6 in J7.

Table 4 provides one time-resource-feasible schedule )( for instance J12. Instance J12
has 31 alternative time-resource-feasible schedules )( based on other start times of jobs
4 and 7 to 11. All the i = 1, 2, ... , 32 time-resource-feasible schedules )( for instance
J12 may be found in table 5, where only the triplets (j,m,t) for the jobs 4 and 7 to 11
are reproduced. The sequence (i = 1, 2, ..., 32) of schedules listed in table 5 corresponds
 317

to the sequence in which they have been generated by CHARME. Thus we are able to
look at the variable instantiation and backtracking process of the inference engine in
some detail. Entries of succeeding schedules which have changed have been underlined.

P=2 J=12 111=2 l.vl =2 x;i=10 x;~=100 x;~=40 x;~=100

j S. M. m d. k~Jrn 1 k~Jrn 2 k~Jrn 1 k~Jrn 2

J J Jrn

1 {2, 7} 1 1 0 0 0 0 0
2 {3, 4} 2 1 10 1 1 6 5
2 10 1 1 8 3
3 {5} 1 1 11 4 3 2 5
4 {6} 2 1 21 6 1 7 1
2 6 4 3 2 1
5 {6} 1 1 9 4 2 2 2
6 {12} 1 1 1 3 3 3 3
7 {8, 9} 1 1 5 3 4 5 4
8 {10} 1 1 6 3 2 2 2
9 {10} 1 1 8 3 3 3 1
10 ill} 1 1 2 3 6 8 3
11 {12} 1 1 13 5 3 6 3
12 ¢ 1 1 0 0 0 0 0

P 1 2

,z 4 (=rmin
1, 2
) 2 (=rmin)
1,7
p
dp 35 (=rmax ) 1,6
48 (=rmax )
I,ll

(rmin = 12 r max = 29, rmin = 12, rmin = 10 ,

2,4 ' 2,4 7,8 7,9
r~~n = 0 und r~r = ro otherwise)

Table 3: Data of instance J12

(1,1,2) (2,1,4) (3,1,14) (4,2,26) (5,1,25) (6,1,34)

(7,1,2) (8,1,19) (9,1,17) (10,1,32) (11,1,34) (12,1,48)

Table 4: One time-resource-feasible schedule Ji for instance J12

318

i (4,m,t) (7,m,t) (8,m,t) (9,m,t) (10,m,t) (l1,m,t)

1 (4,2,26) (7,1,2) (8,1,19) (9,1,17) (10,1,32) (11,1,34)

2 (4,2,26) (7,1,2) (8,1,19) (9,1,17) (10,1,32) (l1,l,M)

3 (4,2,26) (7,1,2) (8,1,19) (9,1,17) (10,lcll) (11,1,35)

4 (4,2,26) (7,1,2) (8,1,19) (9, I..!§) (10,1,w (l1,lcl!)

5 (4,2,26) (7,1,2) (8,1,19) (9,1,18) (10,1,32) (l1,l,M)

6 (4,2,26) (7,1,2) (8,1,19) (9,1,18) (10, 1cll) (11,1,35)

7 (4,2,26) (7,1,2) (8,l,lQ) (9,l,11l (10,l,w (l1,lcl!)

8 (4,2,26) (7,1,2) (8,1,20) (9, I..!§) (10,1,32) (11,1,34)

9 (4,2,26) (7,1,2) (8,1,20) (9,lJ1) (10,1,32) (l1,l,M)

10 (4,2,26) (7,1,2) (8,1,20) (9,1,1§) (10,1,32) (11,1,35)

11 (4,2,26) (7,1,2) (8,1,20) (9,lJ1) (10,lcll) (11,1,35)

12 (4,2,26) (7,1,2) (8,1,20) (9, I..!§) (10,1,33) (11,1,35)

13 (4,2,26) (7,1,~ (8,1,20) (9,1,18) (1O,l,W (l1,lcl!)

14 (4,2,26) (7,1,3) (8,1,20) (9,1,18) (10,1,32) (l1,l,M)

15 (4,2,26) (7,1,3) (8,1,20) (9,1,18) (10,lcll) (11,1,35)

16 (4,2,n) (7,ll) (8,l,w (9,lJ1) (10,l,W (11,lcll)

17 (4,2,27) (7,1,2) (8,1,19) (9,1,17) (10,1,25) (l1,lcl!)

18 (4,2,27) (7,1,2) (8,1,19) (9,1,17) (10,1,25) (ll,l,M)

19 (4,2,27) (7,1,2) (8,1,19) (9,1,17) (10,lcll) (11,1,35)

20 (4,2,27) (7,1,2) (8,1,19) (9, I..!§) (10,1,33) (11,1,35)

21 (4,2,27) (7,1,2) (8,1,19) (9,l,1J!l (10,1,33) (11,1,35)

22 (4,2,27) (7,1,2) (8,l,lQ) (9,lJ1) (10,1,33) (11,1,35)

23 (4,2,27) (7,1,2) (8,1,20) (9, I..!§) (10,1,33) (11,1,35)

24 (4,2,27) (7,1,2) (8,1,20) (9,l,1J!l (10,1,33) (11,1,35)

25 (4,2,27) (7,1,2) (8,l,ill (9,1,17) (10,1,33) (11,1,35)

26 (4,2,27) (7,1,2) (8,1,21) (9,1,1§) (10,1,33) (11,1,35)

27 (4,2,27) (7,1,2) (8,1,21) (9,l,w (10,1,33) (11,1,35)

28 (4,2,27) (7,1,1) (8,1,lQ) (9,1,1§) (10,1,33) (11,1,35)

29 (4,2,27) (7,1,3) (8,1,20) (9,l,w (10,1,33) (11,1,35)

30 (4,2,27) (7,1,3) (8,l,1!.l (9, I..!§) (10,1,33) (11,1,35)

31 (4,2,27) (7,1,3) (8,1,21) (9,l,1J!l (10,1,33) (11,1,35)

32 (4,2,27) (7,1,1) (8,1,21) (9,1,19) (10,1,33) (11,1,35)

Table 5: The time-resource-feasible schedules ){ for instance J12

 319

The computational times required on a 80486 PC with 33 Mhz are as follows : The
generation of all time-resource-feasible schedules ){ for the instances J7 and J12
required some milliseconds and one second of CPU-time, respectively. Note that the
CPU-time required depends upon some factors , among which characteristics of a
specific instance as well as details of a specific implementation seem to be most
important. Experiments have to be made in order do get more insight.

8. Future Work

This paper addresses a rather general time-resource tradeoff production scheduling

problem. Its focus is besides mathematical modeling on the usage of a PROLOG-based
(standard) software. In the future it will be interesting to deal with the following :

• The performance of the PROLOG-based approach for solving the constraint

satisfaction-problem presented here should be compared with the capabilities of
general purpose MP-based software for finding feasible solutions [12] .

• The constraint satisfaction-problem dealt with here can easily be generalized to an

optimization approach with the objective function of minimizing the cost of
consuming nonrenewable resources . Getting insight into the relative performance of
AI- and MP-based approaches for solving this optimization problem would be
interesting, too .

References
[1] Ahn, J. and A. Kusiak, "Scheduling with alternative process plans", in: Fandel,
G. and G. Ziipfel (Eds.): Modern production concepts, Berlin 1991, pp. 387-403.
[2] Bartusch, M; R.H. Mohring and F.J. Radermacher, "Scheduling project networks
with resource constraints and time windows", Annals of Operations Research,
Vol. 16 (1988), pp. 201-240.
[3] BULL, Artificial Intelligence, Charme VI, Users Guide and Reference Manual,
Bull S.A., Cedoc-Dilog 1990.
[4] Dechter, R., "Enhancement schemes for constraint processing: Backjumping,
learning, and cutset decomposition" , Artificial Intelligence, Vol. 41 (1989/90), pp.
273-312.
[5] Dechter, R. and J. Pearl, "Network-based heuristics for constraint satisfaction
problems", Artificial Intelligence, Vol. 34 (1988), pp. 1-38.
[6] Domschke, W. and A. Drexl, "Kapazitatsplanung in Netzwerken - Ein Uberblick
uber neuere Modelle und Verfahren" , OR Spektrum, Vol. 13 (1991), pp. 63-76.
[7] Drexl, A. and J . Grunewald, "Nonpreemptive multi-mode resource-constrained
project scheduling", to appear in lIE Transactions (1992) .
[8] Drexl, A. and R. Kolisch, "Production scheduling in small batch machine tool
assembly", in preparation (1992).
320

[9] Gallo, G. and S. Pallottino, "Shortest path algorithms", Annals of Operations

Research, Vol. 13 (1988) , pp. 3-79.
[10] Kolen, A. and E. Pesch, "Constraint Propagation Programmierung in Charme" ,
Discussion Paper, University of Maastricht 1991.
[11] Kurtulus, I.S. and E.W. Davis, "Multi-project scheduling: Categorization of
heuristic rules performance'., Management Science, Vol. 28 (1982), pp. 161-172.
[12] Maes, J .; J .O. McClain and L.N. Van Wassenhowe, "Multilevel capacitated lot
sizing complexity and LP-based heuristics", European J . of Operational Research,
Vol. 53 (1991), pp. 131-148.
[13] Mese~uer, P., "Constraint satisfaction problems: an overview", A/COM, Vol. 2
(1989), pp. 3-17.
[14] Mohr, R. and T.C . Henderson, "Arc and path consistency revisited", Artificial
Intelligence, Vol. 28 (1986), pp. 225-233.
[15] Patterson, J .; R. Slowinski; B. Talbot and J . W~glarz, "Computational
experience with a backtracking algorithm for solving a general class of precedence
and resource-constrained scheduling problems", European J. of Operational
Research, Vol. 49 (1990), pp. 68-79.
[16] Pritsker, A.A.B. ; W.D. Watters and P.M. Wolfe, "Multiproject scheduling with
limi ted resources: a zero-one programming approach", Management Science, Vol.
16 (1969), pp . 93-103.
[17] Talbot, F.B . and J .H. Patterson, "An efficient integer programming algorithm
with network cuts for solving resource-constrained scheduling problems",
Management Science, Vol. 24 (1978) , pp. 1163-1174.
 ROBUST SCHEDULING AND GAME-THEORETIC CONTROL FOR
SHORT-TERM SCHEDULING OF JOB-SHOPS

Victor Jorge Leon

Texas A&M University
College Station, TX 77843-3367

S. David Wu and Robert H. Storer

Industrial Engineering
Lehigh University
Bethlehem, P A 18015

1. Introduction

The successful operation of a manufacturing system depends on the economical

scheduling of resources used in the production of goods. Given the impossibility of
scheduling with explicit consideration to the entire system, often the global problem is
decomposed into a hierarchy of decision levels. This research deals with scheduling at
the operational level. Important to this level are decisions concerning the order and
timing in which products are to visit machines, and rescheduling as new information
becomes available (e.g., unforeseen disruption, or changes in the demand). The solution
to this problem is of great practical relevance because it allows the continual operation
of the system while maximizing system performance. Simply stated, the problem
consists in deciding how and when to reschedule jobs on machines as new information
becomes available.

To illustrate the system's operation, consider the following scenario: (1) Given
the production requirements for a day, a schedule is produced "off-line" in advance of
execution, (2) the plan is released to the shop-floor for its execution, (3) as unforeseen
disruptions occur or new information becomes available, "on-line" rescheduling and
control is required to adjust the off-line schedule such that the performance is
maximized. This operation mode continues until all the jobs are completed (i.e., the
end of the planning horizon). This research assumes job-shop as the production
configuration, and the completion time of all jobs (i.e., the makespan) as the objective.

In this environment, it is important to consider the role of the off-line static

schedule. An off-line schedule is typically needed as a basis for the planning of other
system activities such as, machine setup, tooling allocation, fixturing and labor
 322

planning. Therefore, when considering rescheduling, it is desirable to adhere closely to

the prescribed schedule so that low impact on other planning systems can be achieved.

Previous approaches to the problem can be classified into four groups: (1) match-
up scheduling, (2) complete rescheduling, (3) stochastic scheduling, and (4) dynamic
dispatching. In match-up scheduling, a transient schedule is computed after the
occurrence of a disruption. After a finite amount of time this transient schedule will
match-up with the off-line schedule. In complete rescheduling, a new schedule is
generated every time a disruption occurs. In stochastic scheduling, the decision
considers only the jobs to be processed at that time given all the information available
up to that moment. Finally, in dynamic dispatching, the decision on which job to
process after a disruption is usually determined by myopic rules or heuristics. Among
these approaches, only match-up scheduling considers the off-line schedule, and only
stochastic scheduling makes explicit use of (stochastic) information concerning future
disruptions.

The approach in this research is to consider two subproblems: (1) off-line robust
scheduling, and (2) on-line rescheduling and control. In the first subproblem, a schedule
is defined to be robust if it can maintain good performance in the event of unforeseen
disruptions. Further, this should be achievable using only minor adjustments in the
sequence in order to minimize the impact on the rest of the system. Research issues
include the definition of measures of robustness and the generation of robust schedules.

The on-line rescheduling and control subproblem explicitly considers both

deviation from the off-line schedule and makespan as components of the objective
function. The problem is modeled as a special game-tree that enables "look-ahead"
analysis. The decision nodes are associated with control decisions, the chance nodes are
associated with uncertain events, and solution sub-trees are associated with a control
policies. Research issues include the modeling uncertain information, the generation of
control actions, and the determination of optimal policies in the game-tree.
Computational tractability is partially achieved through the development of efficient
algorithms for the solution of a scheduling problem with vacations, and the solution of
game trees containing chance nodes.

The methodologies developed here have important practical relevance since it

allows the application of static schedules in uncertain environments. Further, although
both methodologies work well together, the concepts and algorithms developed can be
applied independently.
This paper is organized as follows. Section 2 contains the discussion about
robustness measures and robust scheduling. Section 3 describes the rescheduling and
control methodology. Finally, Section 4 presents the experimental results and analysis.
 323

2. Measures of Robustness and Robust Scheduling

Scheduling is an integral part of production systems planning for two important

reasons. First, the schedule serves as an overall plan upon which many other system
activities are based. Examples of these activities include short-term labor planning,
ordering and preparation of raw-material, and the planning of tooling and set-up
activities. Second, scheduling serves to maximize performance of the manufacturing
facility . By properly planning the timing of shop-floor activities performance criteria
such as makespan, flow time, or tardiness can be optimized.

To achieve the goal of coordinating system activities, it is necessary to compute

the schedule in advance. However, a major drawback of precomputed schedules is that
after they are released for execution, continual updating is required due to the random
nature of shop floor conditions. As a result , the performance of the schedule will
degrade. It is therefore desirable to generate schedules that are "robust" within a
reasonable range of disruptions such as machine breakdowns and processing time
variability . By robust it is meant that the performance of the schedule remains high in
the presence of disruptions. In this chapter, measures of schedule robustness are
developed and algorithms are devised to generate robust schedules.

Almost a decade ago, GRAVES (1981) suggested as a future direction in

scheduling research the need to understand schedule robustness through the explicit
consideration of uncertain information. Dynamic scheduling methods have been
developed which often achieve good performance by dispatching jobs dynamically to
account for random disruptions as they occur. Such is the case for stochastic scheduling
where information uncertainty is considered explicitly. Representative of this line of
research are GITTINS-GLAZEBROOK (1977), GITTINS (1979), PINEDO-ROSS
(1980) , GLAZEBROOK (1981, 1984, 1985 and 1987), and PINEDO (1983). These
solutions methods , however, do not specify in advance the sequence in which jobs will
be processed on machines. Rather the approach is a "dynamic policy" which "allows
the decision maker to determine his/her actions at any moment in time, while taking
into account all the information that has become available up to that moment"
(PINEDO 1983, p.561) .

A characteristic of the approaches in the published literature is that there is no

explicit treatment of the robustness characteristics of precomputed schedules. In this
research , measures of robustness are developed which enable existing local-search
techniques to solve the job shop scheduling problem where the objective is to maximize
robustness. An important property of the proposed measures are the minimal
computation requirements which allow the use of genetic algorithms in searching for
robust schedules.
324

2.1. Schedule Robustness

Let S be a schedule which specifies the sequence in which operations are
processed on machines, and Mo(S) be its deterministic makespan assuming no
disruptions. Let the random variable M(S) denote the actual makespan of S after
disruptions. Random disruptions are specified by their inter-arrival time distribution,
f(.), and duration distribution, g(.). No job can be processed during a disruption.

In order to compute M(S), the control policy used to recover the system from a
random disruption must be specified. It is assumed that the sequence specified by the
initial schedule remains unchanged. It is also assumed that job preemption is not
allowed. This control policy entails a "right-shifting" of the schedule so that the
disrupted operation must restart its processing immediately after the disruption period.
There are two advantages for using this control policy. First, it is simple and its
implementation requires minimal computation. Second, the right-shift policy is
optimum when the deviation from the initial schedule is defined as the sequence
difference from the initial schedule (i.e., minimal disruption to the planning function).

Schedule delay, 5(S), is defined as the difference between the deterministic

makespan and the actual makespan after execution. 5(S) is thus a random variable
which can be expressed as follows:

5(S) = M(S) - Mo(S) (2.1)

Since Mo(S) is deterministic we can write the expected values of M and 5 as

E[5(S)J = E[M(S) - Mo(S)J = E[M(S)J - Mo(S) (2.2a)

and

E[M(S)J = E[8(S)] + Mo(S) (2.2b)

From (2.2a), E[8(S)J characterizes one aspect of a robust schedule since it captures
makespan variability. In other words, when E[8(S)J is small, the makespan of S is less
sensitive to random disruptions. Small E[5(S)J could be achieved with large amounts of
inserted idle-time that will "absorb" any interruption without incurring delay, that is,
E[8(S)J = O. However, this measure is not of practical significance since makespan is
unbounded from above. Clearly E[5(S)] does not capture all properties of a good
schedule. From expression (2b) it can be observed that the expected makespan,
E[M(S)J , contains information about the makespan in addition to information about
expected delay. Both the expected makespan and the expected delay have practical
interpretations. For instance, minimizing the expected makespan can be related to
maximizing the expected throughput (jobs per unit time.) On the other hand,
minimizing expected delay can be related to minimizing unplanned overtime. In this
chapter schedule robustness is defined as a linear combination of both objectives as
 325

follows.

Definition 1. Let r be a real valued weight in the interval (0,1). The robustness
of a schedule S, R(S), is defined as R(S) = r*E[M(S)] + (1- r)*E[c(S)].

2.3. Measures of Robustness

One problem with Definition 1 is that the expected values are difficult to
determine analytically except for a special case in which only a single disruption is
considered in the planning horizon. If more than one disruption is considered, the
problem soon becomes intractable. This is because in job shops it is impractical to
assume independence between successive disruptions. Our approach to the problem is
to develop measures that are strongly correlated with the expected delay and makespan.
Although this approach will result in an approximation of the true robustness value, the
measures are easy to compute and are applicable to a variety of cases present in real
situations.

The slack of operation i, si' is the maximum amount of delay that can take place
during the processing of operation i without incurring an increase in makespan. From
several candidate measures evaluated, the average slack, s, showed good (linear)
correlation with expected delay for a variety of job-shop problems tested. This result
makes intuitive sense because a sequence that contains large amounts of slack-time has
the tendency to "absorb" disruptions without delay. Thus, s is selected as the
surrogate measure of expected delay. Letting Nf denote the set of operations that are to
be processed on fallible machines, this measure is computed as follows,

(2.3)

In this case the correlation is negative, i.e. the larger s the smaller the expected delay.

The computation of s is based on the graph-theoretical representation of job-

shop schedules (BALAS, 1968). Associated with a schedule is an acyclic directed graph
G(N, A). N is the set of operations including artificial source and sink nodes, and A is
the set of arcs representing precedence constraints for operations of the same job, and
sequence relations for operations on the same machine. The computation of (2.3)
entails solving 2*INI longest path problems on G(N,A). O(N) longest path algorithms
can be developed by exploiting the special nature of this graph (ADAMS et ai., 1988).

Using s, . a surrogate measure for expected makespan is computed by Mo - s

(follows from expression 2.2b and the fact that there is a negative correlation between s
326

and expected delay).

2.4. Robust Scheduling Based on Genetic Algorithms

From the study in the previous section,8 showed good correlation with expected
delay. By Definition 1, schedule robustness is a linear combination of E[6(S)] and
E[M(S)]. Since direct computation of these expected values is intractable for multiple
disruptions and since simulation is too computationally expensive, a linear combination
of the surrogate measures 8 and Mo -8 is rather used here for search purposes.

A genetic algorithm modified from (STORER et aI., 1991) is used to generate

robust schedules based on the above measure. Genetic algorithms are probabilistic local
search methods motivated by genetic evolution. The main components of a genetic
algorithm are a finite population of solutions, a chromosomal representation, a fitness
evaluation function, and genetic operators. The chromosomal representation is an
encoding that fully specifies a member in the population. The fitness evaluation
function relates each member of the population to a probability of survival and
reproduction, that is, the better fit members are assigned higher probability of survival
and reproduction. Reproduction is achieved using genetic operators. Asexual
reproduction enables the better fit members to survive from one generation to another,
i.e. members are sampled and installed directly into the new generation. In sexual
reproduction two members are sampled from the population and their chromosomes are
combined to generate offspring solutions. A crossover operator make possible the
combination of chromosome. Finally, mutation introduces random alterations to the
chromosomal representation in order to maintain variety within the population.

A genetic algorithm for robust scheduling, GAROB, is developed. The main

difference between GAROB and the algorithm of STORER et al. (1991) is in the
objective function Z(S). In GAROB, the objective function used is

Z(S) = r *8 + (1 - r) * (M o -8) (2.4)

while Storer et al. use the deterministic makespan,

Z(S) = Mo(S) . (2.5)

Experimental results suggest that schedules obtained using (2.4) significantly

outperform schedules obtained using (2.5) when both expected makespan and expected
delay are considered.

3. ON-LINE RESCHEDULING AND CONTROL

In practice, the moment the off-line schedule is released to the shop, a myriad of
unforeseen events will dislocate the scheduled jobs from the prescribed sequence of
 327

start-times and render the off-line schedule obsolete. Obsolescence also results from
new information not available when prescribing the schedule. The on-line methodology
developed here allows the updating of static off-line schedules in dynamic environments.
This is accomplished using a look-ahead analysis which explicitly incorporates new
information and information concerning future disturbances.

Past research on this problem can be broadly classified into two groups. One
line of research proposes job dispatching, where sequencing decisions take place
dynamically in order to account for disruptions as they occur. Another line of research
proposes control strategies which guide recovery of the system from disruptions giving
consideration to a pre-computed schedule. YAMAMOTO-NOF (1985) propose to
reschedule every time a machine breakdown occurs. BEAN-BIRGE (1986) and BEAN
et al. (1991) propose to "match-up" with the original schedule after a transient period in
which the disruption is compensated for. SALEH (1988) and ODREY-WILSON (1989),
use a cyclic static schedule generated off-line. An on-line controller recovers the
systrem during non-steady state periods caused by disruptions. The recovery entails
minimizing inventory and match-up to the next steady-state loading period. ROUNDY
et. al. (1989) use pricing functions obtained from solving a precomputed deterministic
problem to guide dynamic dispatching decisions. WU et al. (1991) consider a
formulation which explicitly considers the costs associated with rescheduling jobs before
and after their original start times.

The approach used in this paper falls in the class of methods that give
consideration to the off-line schedule. The control problem entails resequencing to
minimize the expected value of a bi-criterion objective function. One criterion is
makespan, and the other is deviation from the off-line, precomputed schedule. The
consideration of schedule deviation is important when the schedule serves as the basis
for other planning functions such as tooling and raw-material delivery.

In this paper, the on-line control problem is treated as a game where the
controller plays against "nature." This view of the problem originates challenging sub-
problems and a number of opportunities for the integration of techniques from
operations research, artificial intelligence, and stochastic control. Some of the
techniques incorporated here include sequencing, heuristic search, variance reduction,
and discrete event dynamic system (DEDS) models.

3.1. Problem Statement and the DEDS Model

We consider a job shop where machines are prone to failure. Each job has a
deterministic processing time, and job preemptions are not allowed. A pre-computed
makespan schedule is released to the shop at time zero and random machine disruptions
328

are destined to occur while processing jobs. As time proceeds, new information about
gradually becomes available to the controller. The objective is to utilize this
information to adjust the schedule such that its expected performance is optimized.

The above control problem can be viewed as a game played against nature. As
game moves, the controller proposes alternative ways to alter the schedule. In return,
nature places stochastic events on the schedule such as machine disruptions. Like most
games, it is intractable to evaluate the entire game tree. Alternatively, the best policies
among those available, are determined for the next N disruptions through an N-step
lookahead.

The game-theoretic formulation of the control problem can be modeled as a

DEDS. The basic components of the DEDS are the state variable xk E X k , a random
disturbance d k E Dk , the decision variable (or control) sk E Sk' and a state function 4>
which relates these components as follows

Xk+l = 4>(xk' d k , sk)' k = 1,2, .. . , N-l (3.1)

where, xk ' Ok' and Sk are the domains of xk' d k , and sk' respectively.

AN-step lookahead of the game is thus defined where k indexes the decision
epochs. State variable xk represents the schedule at k, the control sk is a recovery
schedule that the controller proposes from k to k+l, the random disturbance dk is
nature's move which represents an instance of the next disruption to occur between k
and k + 1. dk is a random variable defined on Ok with probability mass function
Pk(d k ). A disruption is a period of time in which a machine is interrupted from
operation. This may be caused, for instance, by a machine breakdown, or temporary
energy shortage.

A new event is defined every time control is applied. The controller takes action
when significant new information becomes available. For instance, when a disruption
occurs, or when the system reliability falls below a given threshold.

Figure 1 shows a game-tree representation of the DEDS. The tree consists of

alternating levels of decision nodes and chance nodes. Associated with each decision
node is a state (i.e., xk)' and associated with each chance node is a decision variable
(i.e., sk). Each arc emanating from chance nodes is associated with an outcome of d k
with probability Pk(d k ).

A control policy II = {hk ' k = 1, 2, ... , N} is a sequence of functions h(.) which

map states into decision variables sk= hk(xk) . h k (.) may be viewed as a rescheduling
heuristic (or control action) applied to schedule xk at the kth control epoch. Specific
control heuristics will be discussed later.

In Figure 1, a different h k (.) is associated with each arc emanating from a

 329

decision node. Consider a sub-tree consisting of only one arc emanating from each
decision node, and all the arcs emanating from chance nodes. This sub-tree can be
viewed as a simulation of the job-shop under a projected set of disruptions given a
policy II (i.e., a sequence of control actions). The whole tree thus provides an N-step
evaluation of alternative control policies based on the information on hand about future
disruptions. The DEDS model can be viewed as a "recursive simulation" of the job-
shop guided by the proposed controller.

CHANCE NODE

Figure 1. Game-tree representation of the DEDS.

Objective Function. As previously indicated, a major incentive provided for on-

line control is that other planning activities often use the static schedule as a basis.
Given this premise, it is important that the control objective takes into consideration
the impact of schedule changes. A bi-criterion objective function, Z, is proposed. Z is a
linear combination of makespan M, and a measure C of the schedule changes produced
by applying control. M, C, and Z are random variables depending on the outcome of
the DEDS trajectory e{(xk' d k , sk), k = 1, 2, ... , N}. e represents a path in the tree
defined by an instance of xk' d k , and sk at each step. C is defined in terms of each
operation's start-time difference relative to its start-time in the off-line schedule. Given
r E [0,1], we assume the following linear objective
330

Z(o) = r * M(o) + (l-r) * C(O) (3.2)

The control problem (P) can be formulated as follows.
(P):
Given the initial state, xo' find a control policy II to
minimize E[Z(O)]
where, 0 = {(xk' dk , hk(xk»' k = 1, 2, ... , N}
s.t.
xk+l = ¢(xk' dk , hk(xk»' k = 1, 2, ... , N -1
hk(xk) E Sk
In order to solve (P) in real-time, its size must be significantly reduced. Three
main issues are addressed:

1. treatment of uncertain events,

2. alternative schedule generation, sk' and

3. search of the optimal sub-tree.

3.2. Treatment of Information and Disruptive Events d k

Information plays an important role in the control methodology. At each control

epoch, it is assumed that information concerning the current state xk and information
concerning future disruptions is available to the controller.

Shop floor information is often characterized by some degree of uncertainty; in

fact, a combination of deterministic and stochastic information usually characterizes
future events. For instance, machine disruptions can be viewed as a union set of pre-
determined non-productive events (i.e., meal-breaks, preventative maintenance) and
stochastic events (i.e., machine breakdowns). In such situations, neither pure
deterministic models nor pure stochastic models are appropriate. Past research in
stochastic programming (Dantzig, 1989) (Birge, 1985) and more recently in scenario
analysis (Rockafellar and Wets, 1991) recognize this heterogeneous nature of
information. Unfortunately, (P) lacks of the mathematical structure to benefit from the
application of these approaches.

An important property of the proposed DEDS representation is that, (P) is

decomposed into subproblems that deal only with the "next" disruption. The overall
control problem is eventually solved due to the recursive nature of the representation.
An important benefit from this decomposition is that it allows the explicit incorporation
of combined deterministic and stochastic information about future disruptive events.
Information concerning random disruptions is assumed available in the form of the
inter-arrival time distributions for each machine fm (.), and the disruption duration
 331

gm(.), m = 1, ... , IMI. It is assumed that machines fail independently, and that the
parameters of the disruption duration distributions remain unchanged during the
planning horizon. This information is then used to compute the probability of having
the next disruption on machine m in a specific time interval.

Due to the impossibility of considering all possible outcomes of the next

disruption d k , sampling is necessary to estimate the effect of these disruptions. Given
the probability distribution for the arrival of the "next" disruption, it is possible to form
time intervals or "strata" of known probability Ps = P{"next" disruption arrives in
stratum s}. Thus, a stratified sampling of future events can be performed which leads
to a variance reduction of the above estimate. Each sample, d k and corresponding
Pk = Ps, are associated with a CHANCE node in the game tree.

3.3 Generation of Alternative Schedules, sk

Consider the sequence of sampled disruptions in a path in the game tree. At a

decision node in this path, the decision can be viewed as the determination of the
schedule that performs well "given" the sequence of disruptions. This motivates the
study of a scheduling problem with vacations (i.e., known periods of time during which
processing of jobs cannot take place).

It is known that the job shop scheduling problem can be decomposed into single
machine scheduling problems with ready-times, due-dates (Adams et al., 1988). LEON-
WU (1992) develop a branch-and-bound procedure that solves the scheduling problem
with ready-times , due-dates, and vacations to minimize maximum lateness. Tests on
1400 problems indicate that the optimal solution can be found in less than 100 nodes.
Further, a heuristic is developed that obtains the optimal solution for about 30% of the
problems tested.

Based LEON-WU's procedures, a heuristic, H, is developed for the generation of

alternative schedules. H is based in two main assumptions. First, the resulting
schedule from considering only the "next" disruption may be a good candidate schedule.
Second, the impact on the system is reduced if at most one machine is rescheduled at
each decision epoch.

Let na denote the branching degree of the decision nodes. H can be summarized
as follows: (1) sample na instances of the "next" disruption, (2) determine na new
sequences using the branch-and-bound algorithm treating the disruptions as vacations,
(3) associate each new sequence with an arc emanating from the decision node under
consideration. Procedure H addresses several important issues at the same time:

(1) the generation of control heuristics (i.e., alternative schedules) is reduced to

332

solving a one-machine rescheduling problem with one disruption,

(2) a prominent set of alternatives is defined by sampling disruptions. Since the
failure distribution is used for the sampling of disruptions, on-line information is utilized
to generate control alternatives, and

(3) since the sequences on non-disrupted machines remain unchanged, the overall
schedule deviation C (a component of the objective function Z) is expected to be low.

Furthermore, in H, the larger na is, the more alternatives will be generated and
the better performance expected. However, the size of the state space grows as an
exponential of na. As a result, the value of na must be determined empirically before
implementation such that alternatives of sufficient quality and diversity can be
generated within computational limits.

3.4. Solving the Game Tree

Procedure H(.) generates alternative schedules based on available information at

the time it is invoked. As a result, a large number of control policies II must be
evaluated. These control policies are submitted to the game tree for evaluation.
Solving the game tree entails the identification of the best control policy in the game
tree, for the next N disruptions. This policy is followed until the next decision epoch
when the controller is activated.

Recall that the objective function for the N-step lookahead is the expected value
of the linear combination Z(O) defined in equation (3.2). Based on Z, the expected cost
Vk(~) of a control sk at state xk can be computed based on the following backward
recursIOn.

(3.3a)

where,

V\+l(xk~l) = M}n { Vk+l(sk~l)} (3.3b)

i
Where Pk(d k ) is the probability associated with stratum i and V*k+l(.) is the
cost-to-go at state Xk+l. The cost-to-go at the last state xN is V N = Z(e). The
backward recursion simply states that at a decision point the controller applies the
control heuristic which has a minimal expected cost in the remaining steps.
The game tree cannot be solved using an ordinary (backward) dynamic
programming (DP) algorithm. This is due to the fact that a state xk is defined as the
 333

schedule before the kth disruption and state Xk+l remains unknown until xk' dk and sk
are generated. Applying a forward, enumerative algorithm, on the other hand, may
encounter serious dimensionality problems. Consider an N-step game tree with na
control alternatives and ns samples at each step k, the number of possible trajectories
will be (na x nb)N. More importantly, since each node in the tree is associated with a
complete job shop schedule (a conjunctive graph), much memory is needed for the
nodes actually generated. A solution is to use an implicit enumeration method which
generates new nodes only when absolutely necessary. This strategy is common in the
search of game trees and AND/OR trees, a generalized game tree (Pearl, 1984). WU-
LEON (1991) develop an implicit enumeration algorithm for AND/OR tree containing
probabilistic nodes. The algorithm is extended from the AO* algorithm of deterministic
AND lOR trees (Nilsson, 1980). Experimental evidence suggests that with a lower
bound within 50% accuracy (from the true value), the algorithm explores (on average)
30% of the tree before terminating with the optimal solution.

The efficiency of the implicit enumeration algorithm clearly relies on the quality
of lower and upper bounds. In addition, a set of "{-optimal" solutions are generated
during the search to increase the pruning power.

Bounds. Simple bounds for Z(O)=r.M+(l-r).C are proposed for the case when
r=1. The lower bound for a sequence sk associated with a chance node is set equal to
its makespan. The lower bound for a sequence xk associated with a decision node is set
equal to the preemptive makespan that results when dk_1 interrupts sk_1'

The upper bound is determined by the minimum of the best solution found so far
and a function defined as follows: a function of the maximum operation duration d max ,
the maximum disruption length Amax, and the remaining disruptions Nr within the N-
step lookahead horizon. We simply add (dmax+Amax)*Nr to the makespan of the
schedule associated with the node.

The {-optimal Solutions. Significant computing time can be saved by stopping

the search when a policy that yields an expected performance { units from the optimal
solution is found . That is, if V* denotes the optimal cost, and V{ denotes the cost of an
{-optimal solution, then V{ - V* ::; L This can be achieved by allowing pruning of the
node associated with alternative sL if there exists a sibling alternative tk' such that,
. )
LB(siJ + { > UB(sk)' Where LB(.) and UB(.) denote the lower and upper bound of a
given alteruative.

4. EXPERIMENTAL RESULTS

Different aspects of the proposed off-line and on-line methodologies are tested in
four sets of experiments. Experiment 1 studies the effect of the number of look ahead
334

steps and the monitoring frequency. Experiment 2 compares the control methodology
with "total rescheduling" and "right-shift" control policies under different levels of
information uncertainty. Experiment 3 considers the effect of providing the controller
with imperfect, or wrong information. Finally, Experiment 4 studies the effect of using
"robust" off-line schedules.

Experimental results suggest that the proposed controller obtains slightly better
expected makespan performance and significantly less impact on the system when
compared with a controller using the "total-rescheduling" policy. The latter is true
whether the off-line schedule is "robust" or not. However, better performance is
obtained (less expected makespan and less variability of the makespan) when "robust"
schedules are used.

ACKNOWLEDGEMENT

This research is part of a project supported by National Science Foundation

under Grant# DDM-9010033.

REFERENCES

ADAMS, J., Balas, E. and Zazwack. D., 1988, The Shifting Bottleneck Procedure for
Job Shop Scheduling, Management Science, 34(3), pp. 391-40l.
BALAS, E. , 1968, Project Scheduling with Resource Constraints, Proc. of the NATO
Conference in Applications of Mathematical Programming, Cambridge, England, pp. 24-
28.
BEAN, J. and J. Birge, "Match-Up Real-Time Scheduling," Proceeding of a Symposium
on Real-Time Optimization in Automated Manufacturing Facilities, NBS Publication
724, National Bureau of Standards, 1986, pp. 197-212.
BEAN, J.C., J.R. Birge, J. Mittenthal and C.E. Noon, 1991, Matchup Scheduling With
Multiple Resources, Release Dates and Disruptions, Operations Research, 39(3), pp.
470-483.
BIRGE, J.R., 1985, Decomposition and Partitioning Methods for Multistage Stochastic
Linear Programs, Operations Research, 33(5), pp. 989-1007.
DANTZIG, George B., 1989, Decomposition Techniques for Large-Scale Electric Power
Systems Planning Under Uncertainty, Impacts of Recent Computer Advances on
Operation Research, Elsevier Science Publishing Co., Inc., R. Sharda, B. L. Golden, E.
Wasil, O. Balsi, W. Stewart, editors.
GITTINS, J. C. and K. D. Glazebrook, "On Bayesian Models In Stochastic Scheduling,"
Journal of Applied Probabilities, Vol. 14, 1977, pp. 556-565.
GITTINS, J. C, "Bandit Processes and Dynamic Allocation Indices," J. R. Statist. Soc.
 335

B, Vol. 41, No.2, 1979, pp. 148-177.

GLAZEBROOK, KD., "On Non-Preemptive Strategies in Stochastic Scheduling,"
Naval Research Logistics Quarterly, Vol. 28, 1981, pp. 289-300.
GLAZEBROOK, KD., "Scheduling Stochastic Jobs in a Single Machine Subject to
Breakdowns," Naval Research Logistics Quarterly, Vol. 31, 1984.
GLAZEBROOK, KD., "Semi-Markov Models for Single-Machine Stochastic Scheduling
Problems," Int. J. Sys. Sci., Vol. 16, No.5, 1985, pp.573-587.
GLAZEBROOK, KD., "Evaluating the Effects of Machine Breakdowns in Stochastic
Scheduling Problems," Naval Research Logistics, Vol. 34, 1987.
GRAVES, C. S., "A Review of Production Scheduling," Operations Research, Vol. 29,
1981, pp. 646-675.
LEON, V.J. and S.D. Wu, 1992, On Scheduling With Ready-Times, Due-Dates, and
Vacations, Naval Research Logistics, 39, pp. 53-65.
NILSSON, N., 1980, Principles of Artificial Intelligence, Palo Alto, California, Tioga,.
PEARL, J., 1984, Heuristics: Intelligent Search Strategies for Computer Problem
Solving, Addison-Wesley, Reading, Massachusetts.
PINEDO, M. and S. Ross, "Scheduling Jobs Subject to Non-Homogeneous Poisson
Shocks," Management Science, Vol. 26, 1980, pp. 1250-1257.
PINEDO, M., "Stochastic Scheduling With Release Dates and Due Dates," Operations
Research, Vol. 31, 1983, pp. 559-572.
ROCKAFELLAR, R. T. and J.-B. Roger, 1991, Scenarios and Policy Aggregation in
Optimization Under Uncertainty, Mathematics of Operations Research, 16(1), pp. 119-
147.
ROUNDY, R., Herer, Y and S. Tayur, "Price-Directed Scheduling of Production
Operations in Real Time," Proc. 15th Conference of the NSF Production Research and
Technology Program, Berkeley, CA, 1989.
STORER, R.H., S.D. Wu and R. Vaccari, 1991, "Local Search in Problem and Heuristic
Space For Job Shop Scheduling," Management Science, to appear.
YAMAMOTO, M. and S.Y. Nof, 1985, Scheduling/Rescheduling in the Manufacturing
Operating System Environment, Int. J. Prod. Res., 23(4), pp. 705-722.
WU, S.D., R.H. storer and P.C. Chang, 1991, A Rescheduling Procedure for
Manufacturing Systems Under Random Disruptions, Proceedings of the Joint
US/GERMAN Conference on New Directions for OR in Manufacturing, NIST,
Gaithersburg, Maryland, to appear.
WU, S.D. and V.J. Leon, 1991, A Search Algorithm for Informed AND/OR Trees
Containing Chance Nodes, working paper, Department of Industrial Engineering, Lehigh
University.
 The Two - Stage Hybrid - Flowshop Scheduling Problem
with Sequence - Dependent Setup Times

Stefan Vo6

Technische Hochschule Darmstadt, Institut fUr Betriebswirtschajtslehre / FG Opera-

tions Research, Hochschulstraj:Je 1, D - 6100 Darmstadt, Germany

.Absttact
This paper coosiders the two-stage hybrid-flowshop scheduling problem with
sequence-dependent setup times where the objective is to minimize the makespan. In
comparison to the familiar two-machine flowshop problem we are given two stages
instead of two machines each containing a given number of parallel machines. This
problem combines hybrid-flowshop scheduling and the consideration of sequence-
dependent setup times. In a first part we &l1'Vey the existing literature on the respective
problems.

Since the problem is NP-hard, in a second part we aim in comparing several

heuristic approaches with respect to solution quality. In addition improvement
procedures are investigated including recent ideas of tabu search.

1. Introduction

Suppose that n jobs are to be scheduled for processing in a two-stage hybrid-flowshop. In

comparison to the familiar two-machine flowshop problem we are given two stages instead
of two machines each containing a given number of identical parallel machines. As in the
flowshop problem each job has to be processed in the same order, first at stage 1 and then
at stage 2. However, this is no restriction in general. Preemptions are not allowed but
additional setup times are taken into account. Each setup time of a job depends on its
immediately preceding job as well as on the contemplated stage, i.e., we focus on sequence-
dependent setup times. The objective is to minimize the rnakespan Cmax .

The problem at hand combines two important features of production planning in a

flowshop environment, namely:
• the hybrid-flowshop problem
• consideration of sequence-dependent setup times
 337

The hybrid-flowshop problem (without consideration of sequence-dependent setup times)

originally has been formulated in the early seventies (see, e.g., Arthanari and Ramarnurthy
(1970, 1971) as well as Salvador (1973». Since then a vast amount of literature has been
reported by several authors on this special case of our problem. Independently, the
literature on sequence-dependent setup times has produced a great number of references
also starting in the seventies (see, e.g., Uskup and Smith (1975».

The outline of this paper is as follows. First we present a review of the existing literature
for both problems as described above with special emphasis on the restriction of two stages.
Then some necessary definitions are given resulting in a more precise formulation of the
two-stage hybrid-flows hop problem with sequence-dependent setup times. Subsequently we
restrict ourselves to the problem with identical parallel machines at the first and exactly
one machine at the second stage. Some heuristic algorithms and improvement procedures
are considered, respectively. In addition, we investigate tabu search as a recent technique
for guiding improvement procedures to overcome local optimality.

Computational experiments are undertaken to analyze the algorithms with respect to solu-
tion quality. For the two-machine flowshop problem with sequence-dependent setup times
one of our proposed heuristics clearly outperforms the best known heuristics from the
literature with respect to solution quality. Finally ideas for further research are given.

2. Literature Review

In a flowshop problem each of n jobs must be processed through m stages in the same
order, and each job is processed exactly once at each stage. Usually, there is exactly one
machine at each stage. The most common objective of flowshop scheduling (and the one
used throughout this paper) is that of minimizing the makespan, i.e., the maximum com-
pletion time C max . Equivalently, Cmax is the flow time of that job which is completed last
and therefore the time required to complete all n jobs. For an interesting thoughtprovoking
survey of" the lessons of flows hop scheduling research" see Dudek et al. (1992).
The most pioneering work on flowshop problems is that of Johnson (1954), considering the
case of exactly two machines. He proposed an exact algorithm (with running time
O(n log n» and showed that for this special case an optimal schedule exists such that the
processing order on all the machines is identical (furtheron a schedule with an identical
processing order on all machines is referred to as a permutation schedule). In general,
however, most versions of flowshop scheduling are NP-hard.

Most prevalent future directions in flows hop research might be in the fields of flexible
manufacturing systems as well as in applications in the chemical process industry. This
leads to more general problems involving hybrid- or flexible flowshops as well as
consideration of sequence-dependent setup times.
338

2.1 Hybrid-Flowshop

Instead of restricting the flowshop to exactly one machine at each stage in a

hybrid-flowshop at least one stage i consists of mj ~ 2 identical parallel machines. Initiated
by Arthanari and Ramamurthy (1970, 1971) for the case of exactly two stages several
authors have treated this problem from different points of view.

In Gupta (1988) it is shown that the two-stage hybrid-flowshop problem is NP-hard even if .
one of the stages contains only one machine. This complexity result is quite obvious since
scheduling two identical parallel machines to minimize the makespan is NP-hard by itself.
This result gives solid motivation for dealing with the problem when there is only one
machine at either of the two stages. Early references in this respect are Rao (1970), Mittal
and Bagga (1973) and Murthy (1974). Gupta (1988) gives a heuristic for the problem with
only one machine at stage 2 as well as an algorithm which minimizes the number of ma-
chines at the first stage used by a given schedule. Similar algorithms for the case of exactly
one machine at the first stage are presented in Gupta and Tunc (1991). Narasimhan and
Panwalker (1984) propose a heuristic algorithm for the case of one machine at stage 1 and
two machines at stage 2 with a modified objective considering idle times of the machines.

A branch and bound procedure for the general hybrid-flowshop problem is given by Brah
and Hunsucker (1991), who generalize the machine and job based lower bounds known for
the entire flowshop problem. A no-wait version of the general problem, i.e., a variation
where each job, once started, has to be processed without interruption until it is completed,
has been treated by Salvador (1973). For this problem, where no in-process inventory is
allowed between the stages, the author proposes a branch and bound algorithm and reports
a real world application for the production of nylon in a polymerization process with seven
stages.

It should be noted, that up to now two streams of research on hybrid-flowshops indepen-

dently, at least partly, invented similar results. This is due to the fact that the
hybrid-flowshop as defined above is also treated in the literature as a so-called flexible
flowshop. Papers dealing with flexible flowshops are usually devoted to the development of
heuristics and their behaviour. Heuristic algorithms based on a decomposition approach are
proposed by Wittrock (1985, 1988) as well as Kocbhar and Morris (1987). Worst-case
analysis results for some heuristics can be found in Langston (1987) as well as Sriskanda-
rajah and Sethi (1989). Additional worst-case analysis results for the no-wait case are given
in Sriskandarajah and Ladet (1986) (note that the latter authors define what we denote a
hybrid-flows hop as a flexible manufacturing system).

While the overall objective in hybrid-flowshops usually means minimization of the make-
 339

span, additional secondary objectives may be followed. Examples are the minimization of
the number of machines used at any stage (see, e.g., Gupta (1988» or the minimization of
queueing (see, e.g., Wittrock (1988». Furthermore, Langston (1987) considers additional
interstage transportation planning in a two-stage system.

Applications of hybrid-flowshop problems are discussed by Arthanari and Ramamurthy

(1971), Salvador (1973), Paul (1979), Wittrock (1985, 1988), and Schmenner (1990, pp.
127).

Extensions of the described two-stage hybrid-flowshop with uniform or unrelated instead of

identical parallel machines are possible (see, e.g., Narasimhan and Mangiameli (1987) and
Sherali et al. (1990». Additional features in flexibility similar to the hybrid- or flexible
flows hop extension can be found, e.g., in Sethi and Sethi (1990) and Egbelu (1991).

2.2 Sequence-Dependent Setup Times

Gupta and Darrow (1986) show the NP-hardness of the two-machine flowshop scheduling
problem with sequence-dependent setup times. Again, this complexity result is a quite
obvious one, since scheduling a single machine with sequence-dependent setup times to
minimize the makespan is NP-hard by its relationship to the well-known traveling
salesman problem. The same idea applies to permutation schedules as well as the cases
where sequence-dependent setup times appear only at one of the two machines. Corwin and
Esogbue (1974) as well as Bellman et al. (1982) describe dynamic programming algorithms
for the latter cases.

Note that, opposite to Johnson's (1954) two-machine flowshop problem, in the case of
sequence-dependent setup times there need not be an optimal solution which is a
permutation schedule, i.e., passing may give some decrease in the objective function value.
Heuristic procedures for determining permutation schedules as well as lower bounds based
on the solution of some traveling salesman problems are developed by Gupta and Darrow
(1985, 1986). A neighbourhood search technique is proposed by the same authors, too.
Independently, a similar approach has been followed by Simons (1992) who gives heuristics
for the m-machine flowshop problem with sequence-dependent setup times. A mixed
integer linear programming formulation is presented by Srikar and Ghosh (1986) and
corrected by Stafford and Tseng (1990).

3. Definitions and Properties

Let J = {l,2, ... ,n} be a set of n jobs to be processed at two stages, first at stage 1 and
then at stage 2. Let t .(i) be the processing time of job i at stage j. Because of the
J
340

existence of identical parallel machines at stage 1 there is no distinction in processing times

at this stage. Define c.(i,k) as the setup time of job k when its processing at stage j is
J
immediately preceded by job i. c.(i,i) =0 for all jobs L
J
For ease of notation an artificial job 0 will be included in J giving both an initial and a
final state of the machines. Then c.(O,k) is the setup time of a machine at stage j when the
J
sequence starts with job k. Correspondingly, c.(i,O) states job i to be the last job in a
J
schedule, Le., c.(i,O) is the time to lead back a machine at stage j to its original state. This
J
time need not necessarily to be taken into account. Therefore, it may be zero for all jobs.
The following assumptions are made for the sake of clarity:
• all jobs and all machines are available at time zero (there are no release dates)
• all processing times and all sequence-dependent setup times are known and finite
• only one job can be processed at a time on a specific machine
• once a job starts on a machine, it must be completed before another job may be
scheduled on the same machine (no preemption is allowed)
• there is unlimited storage available between the stages
• passing is allowed

Given m 1 (ml ~l) identical parallel machines at stage 1 and ~ (here we choose ~=l)
identical parallel machines at stage 2 the objective is to minimize the makespan C .
max
With respect to the above mentioned complexity results in the sequel we restrict ourselves
to the problem with exactly one machine at the second stage. Given a schedule Sits
makespan may be defined as T(S). With slight modifications to the well-known
symbolism of machine scheduling the problem may be stated as follows:
F2(Pm1,1) I cj(i,k) ~O, m(i)=2 I Cmax
where m(i) gives the number of operations of each job which will be processed following
the assumptions of a (hybrid-) flowshop. In short, we consider the two-stage hybrid-flow-
shop problem with sequence-dependent setup times while minimizing the makespan or the
maximum completion time, respectively.

Following the problem's complexity, devising efficient approximation procedures has a

solid motivation. Therefore, in the subsequent sections heuristics as well as improvement
procedures are considered.

4. Heuristic Algorithms

In the following several heuristic algorithms will be investigated. There are numerous
possibilities to evaluate heuristics, e.g. by using a decomposition or a simultaneous
approach. In this study our proceeding will be to give five algorithms which may assume
 341

exactly one machine at both stages (Le., m 1 = m2 = 1). Then resulting schedules will be
modified according to the availability of m1 identical parallel machines at stage 1.

In more detail the following decomposition approach will be used:

• 1st phase: Calculate an ordered list of all jobs (e.g. by using a heuristic for the
two-machine flows hop problem with sequence-dependent setup times)

• 2nd phase: Determine a machine assignment for all jobs according to the ordered list
of the first phase (e.g. by using a heuristic for scheduling identical parallel machines)

4.1 Finding Initial Schedules

Referring to Johnson's (1954) famous two-machine flowshop algorithm Gupta and Darrow
(1985) developed two algorithms for determining permutation schedules for the two-
machine flowshop problem with sequence-dependent setup times. These algorithms may be
adapted to find initial feasible schedules for our problem. The general outline of the first
algorithm is as follows.

Gupta - Darrow 1

Step 1 Let S be a schedule with n empty places. Define 1) and kr to be the leftmost and the
rightmost index for an empty place in S, Le., kJ : = 1 and kr : = n. 1 and r denote the
leftmost and the rightmost job in S (initially I : = r : = 0).
Step 2 Repeat
Define for all unscheduled jobs i:
A(i) : = c/l,i) + t1 (i) - c2(1,i)
B(i) : = t 2(i)
C1(i) : = c1(l,i) + t1 (i)
C2(i) : = tli) + cli,r)
Find unscheduled jobs i* and j* according to minimize A(i) and B(j) for all
unscheduled jobs i and j, respectively.
ifA(i*) <B(j*) then placei*onpositionkl'l:=i*,I):=kJ+1 endif
if A(i*) = B(j*) then
if i* t j* then
if min {Cl(i*), C2(j*)} ~ min {Cl(j*), C2(i*)}
then place i* on position kl' I : = i*, kJ : = kJ+ 1
else place j* on position k,r r:= j*, kr := kr -1 endif
else
342

if C1(i*) ~ C2(i*)
then place i* on position~, 1: = i*, ~ := ~+1
else place i* on position k,r r:= i*, kr := kr -1 endif
endif
endif
if A(i*) > B(j*) then place j* on position k,r r := j*, kr := kr -1 endif
until ~ = kr .
Step 3 Insert the last unscheduled job at position k) in S.

A and B as well as C1 and C2 give dominance conditions similar to that of Johnson's

algorithm, with the inclusion of sequence-dependent setup times. These conditions are used
to augment a partial schedule from the outside to the inside. The second algorithm,
Gupta-Darrow 2, is similar to Gupta-Darrow 1 with the following changes in Step 1:
Replace the definitions of A(i) by that of C1(i) and of B(i) by that of C2(i) and vice versa.
The complexity of both algorithms is O(n2).

Based on some relationships with the traveling salesman problem we may develop a third
heuristic for our problem. In principle this so-called Insertion-Method proceeds as follows:
Randomly choose a yet unscheduled job and place this job such that the increase of the
new schedule length is minimum.

The temporal distance between processing two jobs in succession is fixed by the largest
setup time between them either at stage 1 or at stage 2, i.e., the distance d(i,j) between
two jobs i and j is described as d(i,j) := max {c1(i,j), cli,j)}.

Insertion-Method
Step 1 Select jobs i* and j* (i* *j*) such that d(i*, j*) minimizes d(i,j) for all jobs
i, j e{1, ... ,n}, i*j. Define an initial schedule as S = {i*, j*}, k:= 2.
Step 2 Repea.t
Insert an arbitrary yet unscheduled job j from J in S such that this insertion
causes the smallest increase of the makespan T(S). (The placement of j in Swill
be checked for all k+ 1 possible places).
k:= k+1
until k = n.

In the following a new so-called Setup-Method will be described The main orientation for
choosing the jobs is their setup time with respect to the last scheduled job. The first
scheduled job (called job i*) will be chosen in a way which minimizes the idle time at
stage 2. This idle time is determined while job i* is processed at stage 1 and only the setup
 343

is carried out at stage 2, i.e., job i* will be chosen to minimize

A(i) : = c} (O,i) + t} (i) - clO,i)

for all jobs i = 1, ...,n. The job which succeeds job i* will be selected from the remaining
jobs such that it has the shortest setup time as an immediate successor of job i* at stage 2.
All other jobs will be selected as described for the second one, i.e., they have the shortest
setup times with respect to the previously scheduled jobs.

Setup-Method

Step 1 Select job i* to minimize A(i) for all i = 1, ... ,n.

1 := i * (last scheduled job), k] : = 2 (next free place in the schedule).
Step 2 Repeat
Select an unscheduled job i * such that
c2(1, i *) = min {c2(1, i) I i is an unscheduled job}.
Place i* on position kl' I : = i*, k] : = k] +1
until k] = n+l.

Both algorithms, the Insertion-Method as well as the Setup-Method, may be implemented

in 0(n2).

By observing the mode of working and the results of this algorithm one will recognize that
in general the more jobs are scheduled the worse are the setup times which have to be
accepted because the number of unscheduled jobs decreases. Therefore we use an algorithm
which tries to overcome this disadvantage by scheduling a certain number of jobs by
complete enumeration. Given this number as nenum, then Step 2 will be repeated until
k] =n - nenum + 1 is reached and a third step will be added:
Step 3 Schedule the remaining nenum jobs by complete enumeration on the places
n - nenum + 1, ... , n (while considering the last scheduled job 1).
In our implementation below nenum equals 7 (Setup-Method 7), considering a reasonable
tradeoff with computation times.

4.2 A Second Phase Algorithm

All algorithms described above will be used to find initial schedules without considering
more than one machine at stage 1. The makespan of these schedules will be computed by
an algorithm of Gupta (1988) which tries to minimize the number of machines in use at
stage 1, too. The main aspect of this algorithm, however, is to assign the jobs to specific
machines at stage 1.
344

Assignment

Step 1 Place the jobs at stage 2 in order of a given schedule S.

Step 2 For all jobs according to S do
Place the job on that machine at stage 1 on which the processing of this job will
be finished as late as possible without getting idle time at stage 2. If there is no
such machine then schedule the job on that machine where the placement causes
the smallest idle time at stage 2. Ties are broken arbitrarily.

In order to minimize the makespan all heuristics will be combined with this algorithm
Assignment.

5. Improvement Procedures

All algorithms described in the previous section may be used to compute feasible solutions.
In the sequel improvement procedures are proposed which start with a feasible schedule
and seek to improve this schedule via a sequence of interchanges. The following definition
of a neighbourhood will be used. Let S be a schedule, e.g. as determined according to the
initial phase of any algorithm of section 4.1. Its makespan T(S) is the objective function
value after having applied Assignment. When exchanging two jobs this is done with respect
to S resulting in a schedule S I. The corresponding change in the objective function is
computed by re-applying Assignment for S I and comparing the resulting makespan T(S I)
with T(S). Three improvement procedures using this neighbourhood definition will be
described.
The first algorithm is a simplified neighbourhood search technique adapted from Gupta and
Darrow (1985). It may be characterized as a 2-optimal exchange procedure.

2-opt Method

Step 1 Choose an initial schedule S with a makespan T(S).

Let i : = 1 and j : = 1.
Step 2 Repeat
j := j+1
if j =i then j : = j +1 endif
Get a new schedule S I with makespan T(S I) by exchanging the jobs on positions
i and j.
if T(S I) <T(S) then accept S I as the current best schedule and redefine S : = S I
endif
until j = n.
Step 3 i := i +1, j : = 0
If i < n then goto Step 2 else stop endif.
 345

Starting with an initial feasible schedule given by one of the algorithms above at the end S
hopefully contains an improved schedule. The crucial assumption of the 2-opt Method is
that whenever an improvement is possible it will be realized, but no restart will be made.
However, this is compensated by a running time known in advance.

The following algorithm tries to overcome this disadvantage by computing in each iteration
the best possible exchange of any two jobs and restarting whenever an improvement is
realized. That is, for this algorithm called Steepest Descent the number of iterations is not
known in advance.

Steepest Descent
Step 1 Choose an initial schedule S with a makespan T(S).
Step 2 Repeat
S' : = S
*
For all i,j = 1, ... ,n (i j) do
get a new schedule S" with makespan T(S") from S by exchanging the jobs on
positions i and j
if T(S") < T(S') then S' : = S"
endfor
if T(S') < T(S) then S: = S' endif
until no improved schedule S' is found.

Steepest Descent proceeds in a greedy fashion and stops whenever in one iteration no
improvement is possible by exchanging any two jobs. If no improvement is possible the
current best solution, however, may be a local optimum. This leads to a "Steepest Descent
Mildest Ascent" procedure which allows for some deteriorations to leave local optima. To
avoid returning into a local optimum already encountered we use the ideas of a tabu list
(this idea is based on the static tabu search algorithm which was adapted by Widmer and
Hertz (1989) for the entire flowshop problem, cf. Glover (1989».

The tabu list contains a number of jobs and their places which led to previous solutions.
The time the jobs rest in the tabu list depends on its size. That is, with every new entry of
jobs in the list, it will be updated by probably throwing away the oldest existing entry.

With the size of the tabu list, too, the possibility of leaving a local minimum is fixed.
Besides the computational effort of the algorithm increases with this value. However, a
tabu list which is too small may lead to circles. Referring to Widmer and Hertz (1989) we
fixed the size to seven entries, i.e., the tabu list contains at most seven prohibited
combinations of jobs and places. (It could also be reasonable to vary the size of the tabu list
346

according to the problem size, i.e., the number of jobs). The computation time of the
algorithm as well as the depth of a local optimum which could be left depends on a variable
called nbmax. With this variable the maximum number of iterations without improving the
current best solution is fixed. Here nbmax will be set to n (the number of jobs). This results
in the following algorithm.

Tabu Search

Step 1 Choose an initial schedule S with a makespan T(S). Fix the length of the tabu list.
nbmax := n, nbiter:= 0, BI := 0, SI := S (T(S') = T(S) )
Step 2 Repea.t
nbiter : = nbiter+ 1
Search for the two jobs (for all jobs i,j = 1, ... ,n, ij:j) whose interchange (with
respect to their places in S I) leads to a schedule S" giving the largest improve-
ment of T(S") against T(S I) (or the smallest deterioration, respectively) while
not considering those changes which lead to combinations of jobs and places
included in the tabu list.
S I : = S", update the tabu list
if T(S I) < T(S) then S: = S I, BI : = nbiter endif
until nbiter - BI = nbmax .

It may be recognized that with every improvement BI will be set to the value of nbiter,
i.e., the difference of nbiter and BI will be zero and the algorithm continues. Otherwise,
e.g. caused by successive deteriorations, the difference between BI and nbiter will increase
until it is equal to nbmax and the algorithm stops.

6. Computational Results

The proposed algorithms are tested while varying the number of machines at stage 1 as
well as the number of jobs. We will report results for the cases when there are one, two or
five machines at stage 1 and the number of jobs is 10, 20, 30, 40, and 50, respectively. Each
possible combination is tested over a sample of 50 randomly generated problem instances.
The processing times are drawn from a uniform distribution in the interval [1,50] and the
setup times in the interval [1,5]. To have a more transparent presentation a second set of
test problems is considered with the same assumptions but a modified rate of processing
over setup times, i.e., the setup times are drawn from the interval [1,25]. All algorithms are
programmed in Turbo Pascal Version 5.5 and run on an IBM AT Pc.

For computing the relative deviation of an algorithm for each problem the best result
(lowest makespan obtained while running all twenty combinations, i.e., all five heuristics
 347

itself and combined with either of the three improvement procedures) is taken as the refer-
ence. For problem x the relative deviation of the solution of algorithm y is computed as:

makespan of algorithm y - best solution

relative deviation xy =
best solution
We present the average of the relative deviation over the sample of 50 problem instances.
Tables 1 and 2 sum up these values with respect to the average deviation (in %) where
setup times are chosen from [1,5] in Table 1 and from [1,25] in Table 2. The results are
given for all combinations of the five heuristics with the improvement procedures. For each
algorithm the first column shows the average deviation of the results gained with that
heuristic together with the procedure Assignment. The next columns in Tables 1 and 2 are
devoted to the improved solutions by the 2-opt Method (2-0), by Steepest Descent (SD),
and by Tabu Search (TS).
The number of parallel machines at stage 1 is assumed to be either 2 or 5 while for reason
of comparison the usual flowshop with sequence dependent setup times is included, i.e., the
case m l = 1. In the following we summarize some of our conclusions.
The cases m! = 2 and ml = 5 are only slightly varying, whereas for m l = 1 we get quite
different results. For m l = 1 the Insertion-Method may be recommended as it produces the
lowest average deviation values for all numbers of jobs giving significantly improved results
in comparison to the algorithms presented in Gupta and Darrow (1985, 1986). For m l =2
and m l = 5 our Setup-Method and Setup-Method 7 behave best. As m l increases the
Setup-Method may be recommended as its average deviation in comparison to the other
heuristics decreases. (Note, that there is no inconsistency when comparing the results of
Table 2 for these two methods, since feasible solutions are gained by the decomposition of
first applying one of these methods and then applying Assignment. These results together
with that of the improvement procedures confirm that in some cases the Setup-Method 7
need not be considered since its computation time is larger than that of the Setup-Method.)
Obviously, in all cases best improvements are given by Tabu Search. The 2-opt Method
reaches smallest improvements followed by Steepest Descent. The time complexity of these
three algorithms is inverted to their ability in improving given solutions. Most time is
needed by Tabu Search, i.e., 2 seconds for a problem size of 10 jobs and about 10 minutes
for a problem size of 50 jobs. Steepest Descent varies between 0.2 and 60 seconds depending
on problem size, too. Lowest time is needed by the 2-opt Method since this algorithm needs
0.1 - 8 seconds. The computation times of all initial heuristics together with the procedure
Assignment are neglectable against the computation times of the improvement procedures.
While the number of jobs increases the ability of the improvement procedures with respect
to average deviation increases, too. The improvement procedures behave usually the same
 *
Table 1: Average deviation, processing times e[I,50] , setup times e[I,5]

Gupta Darrow 1 Gupta Darrow 2 Insertion-lethod Setup-lethod Setup-lethod 7

m1 jobs 2-0 SD TS 2-0 SD TS 2-0 SD TS 2-0 SD TS 2-0 SD TS
1 10 4.8 1.6 1.0 0.4 4.4 1.5 1.0 0. 4 1.6 1.1 0.9 0.3 11.3 3. 8 1.2 0. 5 3.5 1.4 1.1 0. 3
20 6.1 1.8 1.2 0.3 5.5 1.8 1.0 0. 3 1.6 1.2 1.0 0. 2 9. 7 2.6 1.1 0. 4 7. 1 2. 0 1.0 0.3
30 6.9 1.8 0.9 0. 2 5. 7 1.7 0.9 0.2 1.4 0. 9 0.7 0.2 9.2 2.1 0.9 0. 2 8.0 1.9 0. 9 0.2
40 7.3 1.9 0.9 0.1 5.9 1.7 1.0 0. 1 1.3 0.8 0.7 0. 2 8. 6 1.9 0.9 0.2 7.5 1.9 0. 8 0. 2
50 6. 6 2.0 1.1 0. 1 5. 4 1.7 1.1 0.2 2. 0 1.2 1.0 0.2 10.2 2. 1 0.9 0. 2 9.4 1.9 1.0 0.1
2 10 5. 4 1.9 1.0 0.4 5.3 1.8 1.1 0.4 5. 4 1.9 1.3 0. 4 3.6 1.4 0. 8 0. 3 4.6 1.6 0.9 0.4
20 6.8 1.9 1.0 0. 3 6. 0 1.6 1.0 0.3 4.1 1.5 1.0 0.4 2.4 1.0 0.6 0. 3 2. 6 1.1 0.6 0.2
30 7.0 1.6 0.9 0. 1 6.0 1.5 0. 9 0.1 3. 6 1.2 0.8 0. 2 1.8 0.8 0.4 0.2 1.7 0. 7 0.4 0.1
40 7. 3 1.4 0.8 0. 1 6.2 1.5 0.8 0.1 3.4 1.0 0. 7 0. 2 1.3 0. 5 0.3 0. 1 1.3 0. 5 0.3 0. 1
50 7.5 1.4 0 . 8 0. 1 5.7 1.3 0.7 0.1 4.9 1.3 0.8 0. 1 1.7 0.7 0.4 0. 1 1.7 0. 7 0.3 0.1
5 10 5. 4 1.9 1.0 0.4 5. 4 1.7 1.1 0. 3 5.4 1.6 1.2 0. 4 3.1 1.3 0. 7 0. 3 4. 3 1.4 0.8 0. 4
20 6.8 1.8 1.0 0.3 6. 0 1.5 1. 0 0. 3 4.1 1.4 0. 9 0. 3 1.9 0.8 0.5 0. 2 2. 1 0. 9 0.5 0.2
30 7. 0 1.6 0.9 0.2 6.0 1.5 0. 9 0.2 3.7 1.2 0. 8 0. 2 1.5 0. 7 0.4 0.2 1.4 0.6 0. 4 0.1
40 7.0 1.6 0. 9 0.2 6. 0 1.5 0.9 0. 2 3.7 1.2 0.8 0. 2 1.5 0.7 0. 4 0.2 1.4 0.6 0.4 0.1
50 7.5 1.5 0. 5 0. 1 5.7 1.2 0.7 0. 1 4. 8 1.2 0.8 0. 1 1.6 0. 6 0.3 0. 1 1.6 0. 6 0. 3 0.1
Table 2: Average deviation, processing times E[1,50j , setup times E[1,25]

Gupta Darrow 1 Gupta Darrow 2 Insertion-lethod Setup-lethod Setup-lethod 7

m1 jobs 2-{) SD TS 2-{) SD TS 2-{) SD TS 2-{) SD TS 2-0 SD TS
1 10 20.6 7.2 4.1 1.9 16.2 7.3 3.6 1.3 4.6 3. 5 3.1 1.5 21.4 8.2 4.3 1.4 6.9 4.3 3.3 1.7
20 31.9 9.2 5.1 1.4 18.6 8.5 4.1 0.9 5.6 4.6 4.1 1.4 23.6 8.8 4.7 1.3 17 . 7 7.6 4.7 1.0
30 36.9 9.6 5.3 1.1 19.1 7.8 4.7 1.0 5.6 4.6 4.0 0.9 25.3 8.7 4.3 1.0 20.9 7.8 4.6 1.1
40 38.8 9.3 5.3 1.2 20.2 8.7 5.4 0.9 4.9 4.2 4.0 0.6 26.5 8.6 5. 0 1.0 23.0 8.1 5.0 0.9
50 37.9 9.5 5.7 0.8 17.6 8.3 5.8 0.7 5.6 5.0 4.5 1.0 31.4 9.0 5.2 0.8 28.3 9.0 5.6 0.8
2 10 25.2 10.0 4.8 1.8 19.3 7.4 4.9 1.8 14.7 7.2 4.8 1.8 7.8 5.3 3.8 1.7 10.7 5.9 4.4 1.8
20 31.4 9.8 4.9 1.3 23.7 8.6 4 . 3 1.6 14.2 7.0 5.0 1.4 5.0 3.1 2. 4 0.8 6.4 3.8 2.8 1.1
30 36.3 10.3 4.9 1.7 24.4 8.1 4.7 1.4 14.3 6.9 4.8 1.5 3.8 2.3 1.5 0.8 4.6 2.7 1.7 0.6
40 37 . 5 10.0 5.0 1.4 26.9 8.8 5.5 1.4 13.7 7.0 5.1 1.4 2.4 1.6 0.9 0.3 2.6 1.7 1.2 0.5
50 39.0 10.0 5.0 1.7 26.3 8.2 4.7 1.5 17.3 7.5 4.9 1.5 2.1 1.2 0. 8 0.3 3 . 4 1.7 1.1 0.4
5 10 25.4 9.8 4.7 1.7 19.5 7.3 4.8 1.8 14.9 6.9 4.9 2.1 6.5 4.6 3.5 1.5 10.5 5.6 4.4 1.7
20 31.5 9. 6 4.9 1.2 23.8 8.5 4.2 1.4 14.3 7.1 4.7 1.4 4.8 3.1 2.3 0.9 6.2 3.8 2.6 1.1
30 36.4 10.3 4. 9 1.5 24.5 8.1 4.8 1.3 14.4 6.7 4.7 1.4 3.0 2.0 1.4 0.6 3.8 2.4 1.6 0.6
40 37.6 9.8 5.1 1.4 27.0 9.0 5.6 1.4 13.9 7.0 5.0 1.5 2.2 1.4 0.9 0.3 2.4 1.6 1.3 0.5
50 39 . 2 9.3 5.0 1.8 26.4 8.2 4.8 1.5 17.5 7.5 5.1 1.7 1.8 1.2 0.8 0.3 3.0 1.7 1.0 0.5

U)

~
350

for all initial feasible solutions. With respect to solution quality in most cases the
Setup-Method (or Setup-Method 7) combined with Tabu Search should be applied,
however, by observing the tradeoff with the computation time. For m l = 1 a combination
of Tabu Search with the Insertion-Method could be recommended.

Additional computational testing gives nearly the same results (e.g. if we have the same
proportion of processing times over setup times: processing times out of [1,100] and setup
times out of [1,10]).

7. Conclusions

This paper discusses the two-stage hybrid-flowshop problem with sequence-dependent

setup times. This problem is NP-hard. Five algorithms to compute feasible schedules and
three algorithms to improve given schedules have been tested. It is shown that the
Setup-Method with tabu technique gives best solutions. This technique is a powerful tool,
however, with probably high computational effort. Taking more elaborate studies of tabu
search into account (see e.g. Domschke et al. (1992» a dynamic instead of the static
concept followed throughout this paper should give further improvements.

An open question remains with respect to lower bounds. These apparently need to be deve-
loped and a branch and bound algorithm for the exact solution of the problem should be
investigated. In addition, an obvious problem generalization motivated by real-world
applications where buffer restrictions have to be observed between the stages should be
considered.

References
T.S. Arthanari and K.G. Ramamurthy (1970). "A branch and bound algorithm for
sequencing n jobs on m parallel processors", Opsearch 7, 147-156.
T.S. Arthanari and K.G. Ramamurthy (1971). "An extension of two machine sequencing
problem", Opsearch 8, 10-22.
R. Bellman, A.O. Esogbue and I. Nabeshima (1982). Mathematical Aspects of Scheduling
& Applications, Pergamon Press, Oxford et al.
S.A. Brah and J.L. Hunsucker (1991). "Branch and bound algorithm for the flow shop with
multiple processors", Eur. J. Oper. Res. 51, 88-99.
B.D. Corwin and A.O. Esogbue (1974). "Two machine flow shop scheduling problems with
sequence dependent setup times: a dynamic programming approach", Nav. Res. Log.
Q. 21, 515-524.
W. Domschke, P. Forst and S. VoB (1992). "Tabu search techniques for the quadratic
semi-assignment problem", in: G. Fandel, T. Gulledge and A. Jones (Eds.) New
Directions for Operations Research in Manufacturing, Springer, Berlin et al.
R.A. Dudek, S.S. Panwalkar and M.L. Smith (1992). "The lessons of flowshop scheduling
research", Operations Research 40, 7-13.
 351

P.J. Egbelu (1991). "Batch production time in a multi-stage system with material-handling
consideration", Int. J. of Production Research 29,695-712.
F. Glover (1989). "Tabu search - part I", ORSA J. Computing 1, 190-204.
J .N.D. Gupta (1988). "Two-stage, hybrid flow-shop scheduling problem", J. Oper. Res.
Soc. 39, 359-364.
J.N.D. Gupta and W.P. Darrow (1985). "Approximate schedules for the two-machine
flowshop with sequence dependent setup tImes", Indian J. Management and Systems 1,
6-11.
J.N.D. Gupta and W.P. Darrow (1986). "The two-machine sequence dependent flowshop",
Eul. J. Oper. Res. 24, 439-446.
J.N.D. Gupta and E.A. Tunc (1991). "Schedules for a two-stage hybrid flowshop with
parallel machines at the second stage", Int. J. of Production Research 29, 1489-1502.
S.M. Johnson (1954). "Optimal two- and three-stage production schedules with setup times
included", Nav. Res. Log. Q. 1,61-68.
S. Kochhar and RJ.T. Morris (1987). "Heuristic methods for flexible flow line scheduling",
J. of. Manufacturing 6, 299-314.
M.A. Langston (1987). "Interstage transportation planning in the deterministic flow-shop
environment", Operations Research 35, 556-564.
B.S. Mittal and P.c. Bagga (1973). "Two machine sequencing problem with parallel
machines", Opsearch 10, 50-61.
RN. Murthy (1974). "On two machine sequencing with parallel processors", Opsearch 11,
42-44.
S.L. Narasimhan and P.M. Mangiameli (1987). "A comparison of sequencing rules for a
two-stage hybrid flowshop" , Decision Sciences 18, 250-265.
S.L. Narasimhan and S.S. Panwalker (1984). "Scheduling in a two-stage manufacturing
process", Int. J. of Production Research 22, 555-564.
RJ. Paul (1979). "A production scheduling problem in the glass-container industry",
Operations Research 27, 290-302.
T.B.K. Rao (1970). "Sequencing in the order A, B with multiplicity of machines for a
single operation", Opsearch 7, 135-144.
M.S. Salvador (1973). "A solution to a special case of flow shop scheduling problems",
Lecture Notes in Econ. and Math. Systems 86, 83-91
RW. Schmenner (1990). Production / Operations Management: Concepts and Situations,
4th ed., Macmillan, New York.
A.K. Sethi and S.P. Sethi (1990). "Flexibility in manufacturing: a survey", The Int. J. of
Flexible Manufacturing Systems 2, 289-328.
H.D. Sherali, S.C. Sarin and M.S. Kodialam (1990). "Models and algorithms for a
two-stage production process" , Prod. Planning and Control 1, 27-39.
J.Y. Simons (1992). "Heuristics in flow shop scheduling with sequence dependent setup
times", Omega 20, 215-225.
B.N. Srikar and S. Ghosh (1986). "A MILP model for the n-job, m-stage flowshop with
sequence dependent setup times" , Int. J. of Production Research 24, 1459-1474.
C. Sriskandarajah and P. Ladet (1986), "Some no-wait shops scheduling problems:
complexity aspects", Eur. J. Opel. Res. 24, 424-438.
C. Sriskandarajah and S.P. Sethi (1989). "Scheduling algorithms for flexible flowshops:
worst case and average case performance", Eur. J. Oper. Res. 43, 143-160.
352

E.F. Stafford and F.T. Tseng (1990). "On the Srikar-Ghosh MILP model for the n xm
SDST floshop problem", Int. J. of Production Research 28, 1817-1830.
E. Uskup and S.B. Smith (1975). " A branch-and-bound algorithm for two-stage
production-sequencing problems", Operations Research 23, 118-136.
M. Widmer and A. Hertz (1989). "A new heuristic method for the flow shop sequencing
problem", Eur. J. Oper. Res. 41, 186-193.
R.J. Wittrock (1985). "Scheduling algorithms for flexible flow lines", IBM J. Res. Develop.
29, 401-412.
R.J. Wittrock (1988). "An adaptable scheduling algorithm for flexible flow lines",
Operations Research 36, 445-453.
 SHOP·FLOOR SCHEDULING AND CONTROL:

A SYSTEMS APPROACH

Richard A. Wysk, Richard J. Mayer and Hyuenbo Cho

Texas A&M University

Sanjay Joshi, Jeffrey Smith and Walter Hoberecht

The Pennsylvania State University

C. Dennis Pegden
Systems Modeling Corporation

and

Alben L. Jones
National Institute for Standards and Technology

Abstract

Shop-floor control of complex manufacturing systems continues to be a major obstacle in CIM

implementation. The complexity of the problem is seen in the detailed scheduling requirements of
product through the manufacturing resources, the allocation of tooling resources within the system, the
communication requirements of the system, and the data base maintenance and access requirements of
the system. In order to effectively control these activities, it is necessary to define an architecture and
functional perspective of how a shop-floor control system operates.

In this paper, we defme the architectural requirements and an overall control schema that we feel
fits the complexity of the problem. The architectural elements include defmitions of the following
architectures: functional, control, data, computer, communication and factory. The definitions and
interactions of these architectures along with a view for automated shop-floor control is discussed. An
architecture for shop-floor control along with the application environment to implement such a system
is also presented.

BACKGROUND
Integration is one of the most imponant problems facing U.S. manufacturers today. While it is still
354

believed that Computer Integrated Manufacturing (CIM) can have a major positive impact on U.S.
productivity, it is far from commonplace in our factories. A major problem is the difficulty involved
in integrating products (software and hardware) together with information (internal and external) into a
smooth running system. The reason is that these products are typically supplied by a variety of vendors
who do not provide the necessary "integration hooks". That is, the' products from one vendor are not
meant to be (and in some cases are impossible to be) integrated with products from a different vendor.

The hooks referred to here fall into two categories: functional and control. Consider the example of
computer aided process planning (CAPP) and scheduling. Clearly, integration of these is essential in
every factory. There are, in fact, numerous products available today which claim to perform these two
functions. However, no two CAPP systems perform exactly the same functions, and the situation is
worse for scheduling systems. Furthermore, the input/output and control requirements can vary
dramatically from one product to another. This includes software (content and format), hardware
(communication protocols and capabilities), and database content and representation. Consequently, it
is almost impossible to integrate an arbitrary CAPP system with an arbitrary scheduling system. The
same situation exists for almost all manufacturing functions. To make matters worse, each user
invariably requires some special feature or option. This increases development and maintenance costs
and exacerbates the integration problem.

The underlying integration problem is both architectural and technological in nature. Without a detailed
architecture, vendors do not know what functionality and interfaces to provide in their products and what
technologies to use to build those products. Consequently, they provide a range of both. In terms of
the earlier example, this means that vendors provide several different products all claiming to do process
planning. In reality, these products can have vastly different capabilities using a Witle variety of
technologies at varying costs.

Without an architecture, users do not know what to expect from a given product They are left to tind
the one that meets their own definitions, objectives, and costs. As for integration, they typically have
four choices: go where they went before; pay handsomely for specific integration tools; do it themselves;
or ignore it Invariably, they do not have the resources for option number three. Consequently, closed-
systems and islands-of-automation continue to exist with neither vendors' incentive nor users' where-
with-all to change our current dilemma.

METHODOLOGY
The control of CIM systems continues to be a difficult applications problem for both industrial and
computer engineers. The complexity of a manufacturing system continues to perplex researchers
 355

involved with the scheduling and control issues. The data requirements also strain the communication
and computing networks used to control these systems. Our basic approach to control a complex eIM
system is similar to the approach taken by many people involved with nondeterministic systems -- that
of closed loop simulation. This approach was fIrst described by DarEI and Wysk [DarEI & Wysk 1980]
and then later in Wu and Wysk [WU & Wysk 1989]. The approach is illustrated in Figure 1. As the
fIgure shows, a model of the system is used to simulate the system's performance before a specifIc
control is implemented. This approach looks to have tremendous potential, but by itself provides no
specifIc answers. A schedule or set of schedules needs to fIrst be created for the model of the system.
These schedules then need to be evaluated for a variety of characteristics (performance, deadlocking,
local phenomena, etc.). The model of the system usually represents a parallel manufacturing system
environment that can be used to test strategies rather than to solve problems. The strategies for control
need to be developed so that they can be applied to the model. Furthermore the model needs to be
effIcient enough so that the strategies can be evaluated promptly without slowing the system operation
while decisions are being made.

ACTUAL

Figure 1. Basic Control Schema.

356

Key to developing this type of control system is to distribute the control as well as to generate robust
schedules. Robust means that the schedule remains close to optimal (relative to some perfonnance
measure(s» over a realistic range of uncertainties. This means that the schedule is rigorous enough to
absorb minor perturbations.

Control is distributed using a hierarchical shop floor control system (SFCS) containing three levels: cell,
workstation and equipment. The equipment to be controlled includes: NC machining centers, robot
systems, AGYs, or any other computer controlled manufacturing equipment and related tooling resources.
The cell controller periodically receives a list of jobs, with associated due dates from a Master
Production Scheduler. For each job, a process plan is retrieved from a database. The SFCS then
determines the sequence of activities necessary to execute that process plan. This involves the selection
of specific process routing activities, scheduling of all activities at both the cell and workstation levels,
coordinating those activities across the equipment, monitoring activities, job tracking, and some error
recovery.

Error recovery in this context means dealing with the impact of delays. This is a two step process:
determine the impact on the current schedule and update the schedule. The flfst step requires a robust
data structure for the schedule which allows one to determine the "ripple effect" of a delay on the delay
on the current start and finish times. Based on this analysis, the recovery can be a quick fix (in which
a few times get revised), a major rescheduling, anellor a major replanning.

When most people refer to a manufacturing architecture, they really refer to a number of architectures.
For our discussion when we use "architecture" without a qualifler, it is our global manUfacturing
architecture. We see manufacturing systems consisting of the following architectures:
1) Factory Architecture
2) Functional Architecture
3) Control Architecture
4) Communication Architecture
5) Information System Architecture

A factory architecture would consist of all the information concerning the physical factory. The
architecture is most closely aligned with what we associate to a "building architecture". Factory
architectures are normally represented using a standard layout drawing scheme, popUlated with
manufacturing equipment. The factory architecture forms the basis of much of the definition for the
control system, and needs to be qualified with modifiers such as: reach, capacities (spatial and load),
intersections, etc.; and other attributes that form constraints for the control system.
 357

A functional architecture is a partitioning of activities that occur in a system as they relate to the
production requirements (process. move. etc.). Although different manufacturing facilities perform vastly
different operations. the generic underlining functions are the same.

The control architecture considers all other architectures and embodies a set of procedures and messages
that are required to plan for. schedule and present (execute) the tasks defined in the functional
architecture.

The communication architecture defines how entities in the system can "listen" and "talk" with other
entities. The information system architecture consists of the communication architecture as well as the
computational and data storage resources within the system.

The three level hierarchical shop floor functional architecture. which contains cells. workstations. and
equipment as proposed in [Jones. Barlaneyer. & Davis. 1989. Joshi. Wysk & Jones. 1990. and Jones &
Saleh] is used as our functional architecture. One crucial property of the functional architecture is that
each module performs the same tasks: planning. scheduling. and execution. See Figure 2. This
functional architecture significantly can simplify software development. Key to reusable software
development is developing strategies to address the distributed elM control and the scheduling problem
[Davis & Jones. 1989. Davis & Jones. 1988 and Chang. Wu & Storer]. We believe that these strategies
are highly compatible and that the resulting approaches can be expanded to address the distributed
planning problem.

PLANNING
SCHEDUUNG SHOP FLOOR
EXECUTION

PLANNING
SCHEDUUNG WORKSTATION
EXECUTION

PLANNING PLANNING PLANNING

SCHEDUUNG SCHEDUUNG SCHEDUUNG EQUIPMENT
EXECUTION EXECUTION EXECUTION

Figure 2. Functional Architecture for Shop Floor Control System.

358

THE SYSTEM
Our approach will be illustrated using a hierarchical shop floor control system (SFCS) containing three
functional levels shown in Figure 2: cell, worlcstation and equipment

Architecture
We believe that a "systems approach" to this problem is essential to achieve the level of integration that
is required for CIM. A primary component of any system theory is the definition of the state of the
system. For any complex system like manufacturing, the deflnition of the system state could involve
thousands of variables, some discrete, some continuous, many stochastic. As the system evolves in time,
those variables take on many different values. The goal is to define a set of decision rules which govern
the way in which that evolution takes place. These rules can be generated "off-line" or "on-line", and
are invoked to "control" how and when the system changes.

The mathematical/computational problems involved in developing such rules for large systems are well
known. Decomposition, in one form or another, is a recognized technique for addressing these problems.
We believe that the two primary decomposition theories, temporal and spatial, are fundamental to this
problem. Temporal decomposition attempts to deal with the fact that the events which cause changes
in one or more state variable occur at "lots of different frequencies". This also implies that the rules
used to control the magnitude of those changes must also be determined at different frequencies. Spatial
decomposition attempts to deal with the large-scale optimization problems of planning and scheduling
whose solution yields the rules we mentioned above. Goal-directed and price directed methods are
typically used to decompose each problem (decision space) into a series of well-defined, solvable
subproblems. The hierarchical Slructure shown in Figure 2, is based on the simultaneous application of
both of these decomposition theories and forms the functional architectural basis for the remainder of
this paper. As shown in the figure, each controller performs three main functions planning, scheduling,
and execution.

Planning is responsible for selecting/updating the process plan and lower level configuration to be used
in producing assigned jobs and generating revised completion times for those jobs. Cell level process
plans contain routing summaries for each job. Each workstation plan contains the operations to be done
and the information necessary to generate part programs or their equivalent We view equipment level
process plans as the control program required to actually generate a machine. This information is
typically encoded in RS494 format. At each level, scheduling is responsible for evaluating candidate
process plans and generating/updating expected start and finish times (the schedule) for the activities in
the selected plan. Execution is responsible for interfacing with all subordinates and other applications.
It initiates start-up and shutdown, issues commands to perform assigned activities, uses feedback to
 359

monitor the execution of those activities, and oversees error recovery.

We emphasize that these functions are executed at different frequencies within the same level and across
separate levels. Developing analysis tools to determine those frequencies will be an important part of
the proposed research on the architecture. Those tools will attempt to balance the coordination required
by the cell controller against the independent scheduling of the workstation controllers that arises from
the temporaVspatial decomposition.

Planning
This function determines the best set of processing instruCtions among several candidates that appear
in the process plan file and configures control of the "virtual cells" and "virtual workstations" to be used
to produce a product. The selection is based upon an analysis to predict the impact that each candidate
will have on capacity, potential bottlenecks, and other system performance measures. Process plans are
typically viewed as graphs. Each node in the graph can be expanded at lower levels as the planning
detail increases. Planning is illustrated in Table 1.

Scheduling
Current approaches to scheduling are inadequate for the type of distributed scheduling that we feel is
required for flexible systems. They typically do not: include the complex constraints arising from
material handling and tool management; allow multi-criteria optimization; provide for the coordination
of distributed schedulers across hierarchical levels; and do not condition the decision upon the current
state of the system.

Our schema for scheduling is illustrated in Figure 1. The key feature of the schema is that these rules
and procedures are fIrst previewed using a simulation of the system in order to determine how well they
perform. This generic framework for scheduling is applicable at both the cell and workstation levels
[Davis & Jones, 1989]. The concurrent simulation is initialized to the true system state and run to
forecast the behavior of the system for a selected set of scheduling alternatives and performance criteria.
The cell controller generates start and fInish times which become constraints for the workstation level
schedulers who, in tum, do their own independent scheduling. This work extends the concepts
developed in the Multi-Pass Expert Control System (MPECS) [WU & Wysk, 1989]. MPECS uses an
expert system to detennine the set of potential scheduling rules or procedures at the cell level and
evaluates those rules using discrete event simulation. A neural network is being tested to determine if
it can provide the rule base for scheduling.

The key to developing this type of distributed system lies in the generation of "robust schedules".
Table 1. Taxonomy or the Planning Function in Shop Floor Control. (.0)

~
Level Plan A~:1Cd Route Methodology Goals Characteristics Information needed
Decompose a
shop level plan I) A task has alternative I) Order (ordec
into machine, tooling, and number, pan milt,
workstation filtturing resources quantity, due date)
PI P2 workslation2 level plans such 2) Task precedence 2) Process plan file
S.... ~ , _................. (AND/OR ",ph) 3) Shop
~ ,... ~""""'~h 3)T""".""",,,,_ ~r.""""'.k
1'3 ~ ~ serialization workstation is pans are grouped 4) Workstation
~ ~ (modified bIIanced and 4) Tasks with common staws
T T : line material tooling/fIX turing 5) Machining data
ks . 1 balancing handling time resources are grouped handbook
wor lallon problem) number of 5) Physical machining
workstation constraints
level plans) is 6) Current workstation
minimized workload is considered

Decompose a I) A task has alternative I) Order from shop

Partitioning workstation machine, tooling, and 2) Process plan file
~~ machine2 machine3 lIld level plan into filtturing resources 3) workstation
Wodt- ,,~ I I serialization equipment level 2) Task precedence configuration file
station .17.::ll plans such that (AND/OR graph) 4) Machine database
OllTQjI
K.Tl.x~ ~.
\~ 04
?
1
~ ~ (modified
d 'VI~ line worIdoad 3) Physical machining 6)Tooldatabase
~ """" 0 I I balancing assigned to each consttaints 7) Fixwre database
~ nlachinel Inachinel problem) machine is 4) Tooling and filtturing 8) Machining data
o balanced and the consttaints handbook
production rate 5) Current machine
is maximized workload is considered

Choose a I) A task has alternative I) Orlb' from

10012 10014 feasible tooling resources workstation
Equip- I I sequence to 2) Task precedence 2) Process plan file
ment I T4 Tl T4 I Serialization accomplish a (AND/OR graph) 3) Task
+ + ~
I I
equipment plan.
that is.
3) The number of tool
changes are reduced
specifications file
4) Tool database
10011 10011 serialization the 4) Tool traveling time is 5) Machining data
task precedence reduced handbook
graph 5) Task optimization
 361

Robust means that the schedule remains close to optimal (relative to some performance measure(s» over
a realistic range of uncertainties. This means that the schedule is rigorous enough to absorb minor
perturbations. This will limit the number of "reschedulings" that must be done when something goes
wrong.

Execution
Execution provides the mechanisms for carrying out the actions and manages the actual flow of
information/parts between the computer/machines. This requires a system capable of generating,
maintaining, and integrating error-free, run-time software to control the system.

Concepts from formal language and automata theory have been used already to demonstrate this
capability at the workstation level [Hopcroft & Ullman, 1979; Mettala, Joshi, & Wysk, 1989; and Naylor
& Volz, 1987]. We view each module as a layered parser. The highest layer parser is an automatically
constructed push down automata which interprets command/feedback inputs, and activates the appropriate
actions. Examples include the movement of parts in the system, consultation of schedulers, and
preparation of reports. The next layer is the Synchronization layer which is activated upon recognition
of grammatical constructs reserved for synchronization of machine interactions. A good example is the
synchronization required when a robot holding a part in a machine fixture requests that the machine
grasp the part. Error actions are executed upon recognition of error conditions and implemented via the
use of the error parser. This same concept can be expanded to the cell level where parsers and
associated context-free grammars will be generated automatically from the description of the cell.

Further Architectural Considerations

We mentioned that there are several architectures that comprise our controller. Figure 2 portrays a
functional view of our control system. Figure 3 illustrates the control view that we have taken. Again
each controller performs planning, scheduling and execution activities. The equipment level controller
receives process planning, quantity and batch information from the workstation controller. Low level
planning is performed by the equipment controller, e.g., precedence requirements are established for unit
operations, tools are allocated., etc. The scheduling at this level consists of sequencing the unit
operations so that processing time can be minimized. Finally, these instructions are fed directly to an
NC controller (RS494) and switches and instrumentation is checked for status, and a start signal is given.
All of these activities are represented using Manufacturing Message Specification (MMS). It should be
noted that the control architecture interacts with a number of data bases and in doing so interacts with
two other architectures: the information system architecture and the communication architecture. We
anticipate to use a data base query language like SQL to obtain information from the information system.
The individual communication links can be customized to each piece of equipment and computing entity,
362

i.e.• RS232C. Ethernet LAN. etc.

URP CUPS)
Facility Model
Custom interface
Part 1 _ QtJ-- St.at.ue }
~ Dul dlle _ report.lna to/from MRP
[No standards exist]
~ SHOP
Part 1_ St.at.ue
reporUna
Due dale __

WORKSTATION

Part I feature LisL St.at.us

Qly___ Due Dale reporUna
..--~_....L.. _ _,

EQUIPMENT

MMS ,..-_--,

Figure 3. Control Architecture for SFCS.

SYSTEM COMPONENTS
Part Representation
Automatic generation of shop floor control software requires considerable reasoning about the
relationships between the desired finished part requirements. the capabilities of the shop facilities and
equipment, as well as the initial and intermediate part forms. The development of a product description
scheme and the reasoning mechanisms that can perform such evaluations is critical to the success of any
control system. Understanding the finished part requirements. and those of the intermediate stages that
the part must go through is a critical capability for a flexible cell control system. For example. if the
system is to have the capability to modify the process plan (operations and/or routing summary). it must
be able to (on-demand) reinterpret the part requirements without human intervention. Even for less
drastic decisions. such as those associated with allocation of material handling resources. dynamic job
priority reassignment. or determination of the impact of a device failure requires the capability to reason
about the part in the context of the manufacturing process. In addition. adding the capability to the
 363

control software to diagnose part quality problems and to take corrective actions requires the ability to
reason about the changes to a part that result from complex interactions between the part and the many
different processes which have been applied to the part.

Previous attempts to hard code an understanding of part requirements and manufacturing capabilities by
means of a group technology scheme has resulted in impressive if limited success. These attempts
substantiate the value of imbuing the system with such an understanding. Their failures however, point
to the need for the system to be able to generate such an understanding on-demand.

A part description representation scheme that is sufficiently robust as to allow the construction of the
associated reasoning mechanism itself is required. In the development of the description representation
scheme one of the key factors is the ability to seamlessly integrate the many descriptions of a part with
the many models of that part. This amounts to the integration of facts that must be true of the resulting
part with models prescribing constraints that must be satisfied by such a part. Such a representation
scheme must be able to support both the extraction of descriptions from models (e.g., recognition of a
feature) as well as generation of models from descriptions and existing models (e.g., to allow the
generation of a stress model for the evaluation of a derived fixturing concept).

Process Planning
Process plans play an important role in the definition of any control structure, as well as in defining
alternative sequences of operations that must be known by the planning and scheduling modules. The
need for an alternative manifests itself in two important ways:

(i) it provides dynamic schedules with various alternatives available at any instant, and
(ii) the process of creating control software to implement the scheduling actions requires that alternatives
be known priori, so that control software can provide for the necessary links in the control architecture.

A process plan representation is required that is capable of representing all multi-level interactions and
possible precedence that occur among the planning and processing decisions. An AND/OR precedence
graph based representation is proposed as a compact representation of the process plan [Mettala and
Joshi, 1989]. The hierarchical nature of eIM control makes the graph representation even more
attractive. Table 1 shows the use of hierarchical AND/OR graphs to represent different levels of process
planning required. The process plan decomposition parallels that of the control architecture, and thus
ensures that the routing alternatives are provided for decision making at an appropriate level of detail
suitable for each level in the architecture. Details on what information needs to be represented at each
node of the graph needs to be determined for successful implementation. At the equipment level, each
364

task will be structured to represent a unit processing element. These elements are assembled to produce
as efficient of process plan as possible. The plan is then poned to the equipment using MMS
specifications.

A "formal model structure" for process planning has been proposed [Hoberecht, Joshi & Wysk 1991].
This structure is of a generic nature and captures the essence of process planning. It will be used to
store process planning knowledge and generate new and/or alternative process plans. The formal model
uses a directed graph (formal automata model) to describe process planning decision making. Industry
experts can create the initial graph to be used in the planning.

In our view, on-line planning means selecting from several candidates each with its own set of resources
and precedence relations among those resources. This of course implies the existence of process plans
that contain these candidates. Furthermore, there must be alternatives at each level of the hierarchy.

System Deadlock: Detection and Avoidance

In order to achieve true flexibility in an FMS, the pans must be capable of being routed through the cell
in a manner that does not inhibit pan flow. System deadlock is a situation where machines in a
manufacturing system have been allocated parts in such a way that pan flow is impossible. Deadlocking
can arise even in the simplest of automated manufacturing systems. Once a deadlock occurs, automated
control for the manufacturing system may have to evoke manual recovery procedures, resulting in
degrading performance of automated manufacturing systems. Hence the detection and avoidance of
system deadlocks should be an imponant integral pan in any automated manufacturing system.

System deadlocks have been ignored in most scheduling studies, which typically assume infinite
capacities and few limits on physical constraints imposed by material handling equipment Conventional
implementations of FMS systems allocate large amounts of storage space to avoid deadlocking, and rely
on creating batches of pans using linear flow characteristics which results in inefficient use of resources
and major loss in flexibility of systems. The deadlock system to be used in each controller is described
in [Wysk, Yang & Joshi, 1991].

Automatic Generation of Integrated Scheduling and Control Software

As mentioned earlier, one of the major obstacles in software for efficient FMS application is
development of generic control procedures for pan and information flow. Operational flexibility,
ponability, and speed of creating consistently error-free software, can be achieved through research in
the software development problem. This portion of the research addresses the following problem: Given
a manufacturing facility under hierarchical control and a set of process plans for parts to be produced,
 365

how do we develop a system capable of generating run time software to control the system?

The development of software for CIM systems is characterized by the need to carefully model machines
on the factory floor, interacting of parts on those machines as specified by a process plan, specific data
updates that occur within computers constituting the control sys~em, and, finally, to specify the
interaction between the control commands and models that characterize the computer control system.
To focus on automatic software generation, equivalence of fonnal models of the manufacturing system
and of software must be achieved. Loosely speaking, analogous techniques have been used in compiler
construction, but in order to apply those techniques to manufacturing control systems, strict equivalence
of the properties of manufacturing systems, and the corresponding messages and formal models used to
control the system must be shown.

Several authors [Naylor and Volz 1987, Biemans and Blonk, 1986, Naylor and Maletz, 1986] have
developed formal descriptive models of the factory floor with the intent of supporting software
generation, but have not shown methods for systematically extending their fonnal models into running
software. Other authors have developed rapid software prototyping systems [Maimon and Fisher, 1988]
that provide only for the structural specification of cell level control software, still demanding enormous
development of hand crafted code to realize effective systems. In order to overcome the current
impediments to the adoption of CIM technologies by small to medium firms, we believe that reliable,
affordable factory floor level control software generators deserve consideration.

Early work in this area has focused on the development of automatic generation of control software
(called CIMGEN [Joshi, Mettala and Wysk, 1991]) for a two level hierarchical system. The overall
approach taken here was based on the idea that each controller can be modelled as a context free
grammar, and control is exercised as a by product of recognition of manufacturing system commands.
The key to the system is the semantic actions associated with the productions in the grammar. The
semantic actions are executed as the result of the parsing process and provide all the control activities
such as book keeping, message composition and transmission, etc. CIMGEN currently does not interface
with a scheduling system, but hooks are provided for future interface.

Multipass Scheduling System

All decisions on part movement will be made by the scheduler, and the control system merely serves
to execute the commands from the scheduler. The scheduler forms the "brain" of the proposed system.
A multi-pass scheduling is proposed and integrates with the system as shown in Figure 4.

Scheduling in our proposed architecture will be performed hierarchically. Based on the input to the cell
366

CIO CofllTOlk,

.
!'
----------------Wo~~Co~ --~

r Plannlnl
Fundioa

r-------------------~-----SdMdutinIFu~ioa------­

p.nrouainp
processin, times
I'C.!OW'Ce requirements

,_.... - - - - - - - - - - - - - - - - - - .. Prep••--rocessc>rl
~~---~
... ----------- -,I
I
I

---
I
orr-line
Neural I
•_
•
Traininl
Propwnminl Network I
I
I
revise
U'ainin, (rulel.ruJe2) I
Jet
I
I I
• I
Data quety: move(O,LI,u)? Deadl....... I
Mulli-pass uu. I
analyzer Simulator Detection
, answer Module I
I t t
II
I
___'_'__-_-__- ~~!l
I L ______ ~
~

----------------------j .= r----~-J

I1
Eqllipmelll CofllTOlk,

Figure 4. The Multipass Scheduling Schema.

 367

level scheduler, the cell will generate start and finish times which become constraints for workstation
level schedulers, who in turn do their own scheduling. A generic framework for such scheduling has
been developed which is applicable to both cell and work station levels [Davis and Jones, 1988]. This
approach will be extended and linked to the multipass scheduling concept.

Multi-pass scheduling is an approach to resolving factory scheduling problems and has shown promise
recently. DarEI and Wysk [1980] and Nof and Gurecki [1981] have shown significant throughput
improvements can be made using a simulation model to determine the future course for a manufacturing
system. Essentially the procedure works as follows. A simulation model of the system is resident on
the system control computer. At each decision point, a deterministic simulation is run to see what
control policy (from a series of rule based policies) impacts the current system most favorably. This
control policy is then chosen and the appropriate control response is signaled for execution on the
system.

One unfortunate drawback of this procedure is that FMS's are not static (the types of parts, demand,
tooling, etc. change over time). Each time the specifics change, the simulation model must be rewritten
or updated which has to be done manually. Although the benefits can be significant, the implementation
costs can be quite expensive.

The key to the development of this type of distributed scheduling lies in the generation of "robust"
schedules. Robust in this context means that the schedules remain close to optimal (relative to some
performance measures) over a realistic range of uncertainties. This means that the schedule is rigorous
enough to absorb minor perturbations.

Simulation Model
The major function of the simulation model is to evaluate alternative control policies. The simulation
model queries the current databases and determines the future system status by making a pass of
deterministic simulation according to a rule(s) defined in the rule module. The system performance
predicted by each pass of simulation is a measure of closeness to an objective function from the higher
level factory control system. Thus, at the end of all passes of the simulation, the best schedule is then
applied to the physical manufacturing system. The simulation model will operate on parts only currently
available to the system (exogenous activities will not be permitted). This creates a deterministic system
that can be easily analyzed.

An important issue that must be considered is the length of the simulation window, that determines how
long a simulation should look ahead. Intuitively, this value is critical to overall perfonnance of the
368

multipass scheduling mechanism. Detennination of the most desirable simulation window based on its
effect on perfonnance of the scheduling mechanism. computational burden on the system and other
factors that will arise is critical.

Automatic Error Recovery

Ensuring long intervals of continuous operation must be one of the aims of development of any
automated system. The manufac~g environment is a complex environment with a large number of
interacting components. and potential for errors is high. Most current generation of FMS cannot
automatically deal with error situations and rely on human intervention and complete stoppage of the
system to correct any unforeseen situation. Errors in FMS can be broadly defined as an event that
causes some part of the resources to assume an undesired state. and the resource ceases to deliver the
expected services.

The error recovery module interfaces with the cell and workstation controller. and is invoked when errors
are detected. The "hooks" in the cell control and workstation control software. to which the error
module will connect. are created by the automatic control software generation process during the stage
of software generation. The interactions are controlled via error semantic actions which interface to the
error recovery module. Other interfaces include the deadlock detection module to ensure that the error
recovery process does not result in creation of system deadlocks, and MPECS to detennine scheduling
or error recovery actions and to simulate the actions of the various alternative error recovery actions
possible.

Several error recovery and handling procedures have been developed by researchers in areas of computer
science, robotics. electrical engineering in the domain of fault tolerant computing. electronic circuit
design. etc., but very little work has focused on FMS. The various approaches suggested include
emergency stop approach. use of redundant systems. software contingency control. fixed recovery
strategies. among others. These provide some ideas for error handling, but need more research to adapt
or develop new economic strategies for FMS.

Error recovery techniques will also have to be developed in an hierarchical manner for each module of
planning scheduling execution at various levels.

Error Recovery in Scheduling

Error recovery in the context of scheduling means dealing with the impact of delays. We envision this
as a two step process: detennine the impact on the current schedule and update the schedule. The first
step requires a robust data structure for the schedule that will allow one to detennine the "ripple effect"
 369

of a delay on the delay of all other relevant start and finish times. Based on this analysis, the recovery
can be a quick fix (in which few times are revised), a major rescheduling, and/or a major replanning.
New methods such as penurbation analysis or match up scheduling will be investigated.

Conclusions
Integration continues to be one of the most important problems facing U.S. manufacturers. While it is
still believed that Computer Integrated Manufacturing (CIM) can have a major positive impact on U.S.
productivity. it is far from commonplace in our factories. A major problem is the difficulty involved
in integrating products (software and hardware) together with information (internal and external) into a
smooth running system. The reason is that these products are typically supplied by a variety of vendors
who do not provide the necessary "integration hooks". In this paper, we have illustrated that the
integration hooks referred to here fall into two categories: functional and control and that several
architectures work concurrently in a CIM environment. Clearly, integration of these is essential in every
factory.

The underlying engineering integration problem is both architectural and technological in nature.
Without a detailed architecture, system supplies do not know what functionality and interfaces to provide
in their products and what technologies to use to build those products. Consequently, they provide a
range of both without an architecture, users do not know what to expect from a given product They
are left to find the one that meets their own defmitions, objectives, and costs. Closed-systems and
islands-of-automation continue to exist with neither vendors' incentive nor users' where-with-all to
change our current dilemma.

This paper overviews these topics, and presents a general structure to pursue CIM shop-floor control.
The key here is the development of an integrated view of computer INTEGRA TED manufacturing.

Also inherent in the paper is that several technological issues still require resolution. The closed loop
simulation control approach provides a tractable approach to scheduling of very flexible systems. The
frequency used to evoke this procedure, the application of intelligent scheduling rules, the analysis of
the simulation response and the application of deadlock resolution are all issues that need to be resolved
before this approach will bear fruit.
370

REFERENCES

Chang, P., Wu, S., and Storer, R., • A One-Machine Rescheduling Heuristic with Efficiency and Stability
as Criteria", Lehigh University working paper.

Chang, T. C., Wysk, R. A. and Wang, H. P., Computer-Aided Manufacturing, Prentice-Hall, Englewood
Cliffs, NJ, 1991.

DarEI, E. and Wysk, R.A., "Job Shop Scheduling-A Systematic Approach", Journal of Manufacturing
Systems, Vol. I, No. I, 1980.

Davis, W. and Jones, A., "Issues in Real-Time Simulation for Hexible Manufacturing Systems",
Proceedings of the European Simulation Multiconference, Rome, Italy, June 7-9, 1989.

Davis, W. and Jones, A., "A Functional Approach to Designing Architectures for Computer Integrated
Manufacturing", Systems, Man, and Cybernetics special issue on Manufacturing, Vol. 19, No.2,
164-174, March, 1989.

Davis, W. and Jones, A., "A Real-Time Production Scheduler for a Stochastic Manufacturing
Environment", International Journal of Computer Integrated Manufacturing, Vol. I, No.2, 101-112,
1988.

Gurecki, R. and Nof, S.Y., "Decision Support for a Computerized Manufacturing System", The Optimum
Planning of Computerized Manufacturing Systems, NSF Report No. APR74-15256, No. 14, 1979.

Hoberecht, W., Joshi, S., and Wysk, R.A., "A Fonnal Model for Process Planning", Proceedings llE
Research Conference, Chicago, IL, May 1992.

Hopcroft, J. and Ullman, J., Introduction to Automata Theory, Languages, and Computation, Reading,
MA,1979.

Jones, A. and Saleh, A., "A Multi-levellMulti-layer Architecture for Intelligent Shop Hoor Control",
International Journal of Computer Integrated Manufacturing special issue on Intelligent Control,
to appear.

Jones, A., Barlaneyer, E., and Davis, W., "Issues in the Design and Implementation of a System
Architecture for Computer Integrated Manufacturing", International Journal of Computer Integrated
Manufacturing special issue on CIM architecture, Vol 2, No.3, 1989.

Joshi, S., Wysk, R.A., and Jones, A., "A Scaleable Factory Architecture", Proceedings US/Germany
Conference on Operations Research in Manufacturing, 1990.

Mettala, E., Joshi, S., and Wysk, R., "CIMGEN - A Case Tool for CIM Development", Proceedings of
the FMS Conference, Boston, MA, August, 1989.

Naylor, A. and Volz, R., "Design of Integrated Manufacturing System Control Software", IEEE
Transactions on Systems, Man, and Cybernetics, Vol. 17, No.6, 1987.

Wu, S. and Leon, V., "A Search Algorithm for AND/OR Graphs Containing Chance Arcs", Artificial
Intelligence (to appear).

Wu, S. and Wysk, R., "An Application of Discrete Event Simulation to On-line Control and Scheduling
in Hexible Manufacturing", International Journal of Production Research, Vol. 27, No.9, 1989.

Wysk, R. A., Yang, N. and Joshi, S., "A Procedure for Deadlock Detection in FMSs", IEEE Transactions
on Robotics and Automation. Vol. 7. No. 6.00 853-859, Dec. 1991.
 Spatially-oriented Production Planning and Scheduling

Uwe Petersen
Siemens AG, Corporate Production and Logistics Dept.

Integrated Production Methods, Information and Knowledge Processing, ZPL I IP 31

Otto-Hahn-Ring 6

D-8000 Munich 83, FRG

1. Introduction

The present paper deals with the temporal location of manufacturing orders in limited pro-
duction areas. Such nesting problems are chiefly encountered in large-scale construction pro-
jects, e.g. in the shipbuilding industry. If the available assembly area is a potential bottle-
neck, the dimensions of large-scale orders have to be considered. In view of the increasing
automation of production, the proper utilization of assembly halls involves both the schedul-
ing of production orders and their spatial positioning.

The combination of scheduling and cutting stock gives rise to complexities, no simultaneous
computer algorithms having been developed to date. The purpose of the present study is to
describe and verify a new computer-aided procedure that closes this gap in production man-
agement. A large number of different models will be examined and the results evaluated.
Decision-making support for spatially-oriented capacity loading and scheduling can only be
presented by using heuristic simulation. Linear or mixed-integer programming models are not
available for these facility layout problems.

Order-sequencing strategies have therefore to take account of spatial as well as temporal as-
pects. Furthermore, the demands and aims of an industrial user in production planning have to
be considered by integrating knowledge-based capacity loading rules. The decision support
system proposed includes an efficient assignment algorithm, a graphical user interface for the
presentation of planning results, and a relational database for the storage of information. The
procedure is implemented on a Pc.

The dynamic layout heuristic is based on a discretization of space and time, freely combin-
able priority rules, and a flexible strategy of assignment. The particular quality of this ap-
proach consists in a substantial flexibility, so that items of any shape can be processed. Fi-
372

nally, the results of the simulation studies carried out will be summarized and, in addition, the
practical applicability of the procedure will be shown.

Searching for suitable locations is based on elementary register operations of a microcom-

puter CPU. These operations are composed of Boolean algebra rules and can therefore be
executed very fast.

2. Characteristics of spatially-oriented production planning

The conventional capacity planning problem consists in finding a schedule for n orders,
which have to be produced on m machines, so that at least one target criterion is optimized
(HOITSCH, 1985). Traditional planning approaches just take account of the time and quan-
tity aspects, while disregarding the spatial dimensions that determine capacity. Available pro-
duction area represents in many cases a bottle-neck for customer-oriented manufacturers en-
gaged in job production. Spatial dimensions should not be disregarded for this reason.

The existence of production programs is of fundamental importance for all job-shop schedul-
ing problems. Typically, manufacturers of large-scale products have a fixed amount of un-
filled orders for the coming planning period. Short-term modifications of the production
program only occur in isolated instances. To implement the program it is necessary to arrange
production orders according to logical and temporal requirements.

Currently available MRP systems are generally based on successive planning concepts and
frequently reveal an inadequate harmonization of program/quantity planning and scheduling
(ZAPFELI MISSBAUER, 1988 and KURBEU MEYNERT, 1988). It is mainly quantities
and times that are planned, to the exclusion of the workshop area. In capacity scheduling,
without exception, it is only the limitations on resources that are taken into account. The spa-
tial dimensions of manufacturing orders and means of production are disregarded, so that the
planning of parts-nesting is not possible using existing software systems.

In principle, computer-aided production planning and control is only possible with the aid of
optimizing or heuristic procedures. Heuristics are always chosen if an exact solution of the
optimization problems involved using appropriate self-contained mathematical models is
either impossible or too costly. To achieve a feasible solution to a planning problem at all,
people make do with a less than optimum result.

The implementation of parts-nesting is often influenced by sequencing, assignment and se-

lection problems. These problems are peculiarly complexity and are not amenable to an op-
timizing approach, because
 373

no mathematical layout models with temporal aspects exist as yet;

familiar LP-Iayout models are too costly and not applicable on account of their static na-
ture;
special optimizing algorithms are not available either; and
the task in question can be solved for practical requirements by using a newly developed
heuristic.

Considering scheduling heuristics, priority rules are the most widespread. These rules make
possible an assignment of items to productive units according to predetermined criteria. Pri-
ority rules are heuristic decision-making rules, derived largely from the objectives of produc-
tion process planning on the basis of plausibility considerations. They assign a priority to the
orders to be considered in the planning period, thus permitting the orders to be sorted in order
of importance, and allow preferential nesting according to available capacities. A motivation
is the need for a fast heuristic that could achieve "good" layouts (ISRANII SANDERS, 1985).

This means that, in the field of production, an instrument for the control of parts-nesting in
production areas is available. Individual investigations, however, have to find an answer to
the question, which criteria are particularly suitable. An adaptive learning heuristic can help
manufacturing planners to define combined priority rules.

Scientific research has not yet come up with efficient procedures for solving spatially-ori-
ented dynamic parts-nesting problems (SCHLAUCH! LEY, 1988 and KUHNLE/
SCHLAUCH, 1989). Innovations in this field are overdue, however, as there are various user
groups that could be helped. Practical applications exist principally for engineering compa-
nies whose production operations are conducted at internal building sites. The shipbuilding
industry is a good example of this.

In dynamic parts-nesting, n orders have to be assigned locations in m planning areas. Each

item has a specific shape and each area has fixed (usually rectangular) dimensions. In practice
applications have to take account of certain additional restrictions which limit the permissible
solution space. Basically, item- and area-oriented restrictions have to be distinguished.
Manufacturing of any part order has to take account of due dates (with time buffers if neces-
sary). Sometimes positioning can be carried out alternatively in several orientations. Further-
more, heterogeneity of production areas is frequently important for industrial engineers. Re-
strictions of this kind are evident in the shipbuilding industry, for example, where you have
big tools installed in halls. Among other things, those tools are used to vault the outer hull
sections of ships. That part of an area cannot be used for other items.

This planning problem boils down to finding spatially and temporally optimized assignments
of a set of objects to a set of locations. At the start of a nesting process the initial spatial as-
signments of specific parts derived from previous periods has to be considered.
 ~ c.>
ciQ ' ~
....... Operation Layout
Scheduling Planning
~
""
""
~
2
5'
;:s
~
~
$::>
....
t;j Spatial Intema! Conl:>anY
;!s capacity
~ Scheduling Assignment Location Location
"" Loadng
~. Planning Planning Planning
s·
So
~

8;::
~
~

~
""g. Quadratic
lhrougll>ut Capacity Machine Job
'"~ Scheduling Scheduling Loading Sequencing Parts-nesting Spatial C&P Problems IIAssignment
in production dimensions only Other Problem
~ dmensions
§ areas
I:l..
Spatial + terrporal
dimension
;::
f...
~

[
s·
()<)
Cutting Bin-Packing
Problem Problem
 375

Parts-nesting differs from job-shop scheduling through its combination of temporal and spa-
tial considerations. Dynamical locations in areas combines temporal dimensions of traditional
capacity planning with spatial dimension of cutting and packing problems (see fig. I).

The superficial notion that parts-nesting can be understood as a cutting problem and thus be
solved by using available software does not bear closer examination. Cutting software is lim-
ited in most cases to manipulating two-dimensional shapes. Furthermore, the size of the
problem is narrowed down, target criteria are not transferable and many restrictions cannot be
modeled. Not only the spatial dimensions, but above all the time taken by items-nesting has to
be regarded.

In contrast to cutting and packing problems (C & P problems) on the one hand and conven-
tional sequencing problems on the other, spatial and dynamic parts-nesting basically consti-
tutes a four-dimensional problem, since the time dimension has to be considered in addition
to the three space coordinates.

At first, however, the shape of the part is represented only two-dimensionally. When objects
are projected onto their ground space, only three dimensions (including time) remain relevant.
Discrete measurements of distance are made possible by the use of exactly scanned objects
and planning areas, which are particularly easy to handle on a computer. Finally, it should be
mentioned that parts-nesting is considered to be deterministic having fixed data constella-
tions.

The actual nesting procedure thus contains a dynamic C & P problem. The outlines of the
object have to be located in different areas while minimizing trim loss. A planning strategy
arises out of priority and scheduling rules. Apart from minimizing trim loss, target criteria
should include shortening throughput times and delivery delays and maximizing capacity
utilization. Relevant restrictions for nested objects include the various temporal parameters
(i.e. throughput time , the earliest possible and the latest permissible starting time), production
value, height, weight, and outline dimensions. Examples of area restrictions are height,
ground bearing or crane lifting capacity, surface area and a specific location value (e.g. the
book value of an assembly shop).

The modeling of spatial objects using relational databases is no simple task. Shapes have to
be easy to store and quickly retrievable. The computer-internal representation of the spatial
form of an item generally involves breaking down complex shapes into easy-to-handle geo-
metrical elements such as volume, area, border and point (SPUR! IMAM! ARMBRUST/
HAIPTER! LOSKE, 1989).

Optimized nestings of manufacturing orders have to contain verification and analysis of po-
tential collisions. Object administration must be realistic and take account of form and quan-
376

tity if it is to be adequate to practical requirements. For this reason, not only simple and
regular outlines such as rectangles and triangles, but any and all shapes have to be considered.
In one planning section about 50 to 500 different objects have to be administered. This task
can only be mastered effectively by scanning the geometrical structures (with the aid of dis-
crete locus and time coordinates). The sharpness of scanning in relation to unit of length can
be adjusted on a case-to-case basis and may - theoretically at least - vary on a scale ranging
from micrometers to kilometers. Whether standardized dimensions (mm, m or km) or seldom
used units (such as 0,5 dm or 250 m) are selected has no algorithmic significance.

Heuristic plans of parts-nesting are based on given target-dates. Throughput and buffer times
are still frequently derived (e.g. in shipbuilding) from the general technical drawing, which is
prepared immediately upon receipt of the order and only gradually assumes detailed form,
depending on urgency. This entails uncertainty with regard to data collection. Finally, the ca-
pacity scheduling of the previous planning period has to be taken into account (rolling plans).
These object-nestings represent an initial surface utilization selection for the following plan-
ning period.

3. Heuristic procedure for simultaneous scheduling and layout planning

Spatial-dynamic assignment using CALPLAN (computer-aided scheduling and layout plan-

ning) consists essentially of an area- and time-covering location search for those items which
have to be produced in a given period. Free space for objects is only available if a collision
check embracing the whole geometrical expansion and manufacturing time is executed.

Before a search for locations can start, items and planning areas have to be sorted according
to specified sequencing rules. A queue can only be sorted when the planning strategy and
system parameters are fixed and the relevant queries to the database answered. The search for
a location is largely dependent on the scheduling and assigning strategy adopted. System pa-
rameters that can be varied are

the starting time for locating an item;

the planning area in which a part is nested; and
the orientation of an item in the assembly area.

It makes a difference, therefore, whether an unsuccessful search leads to modification of

starting time, location area or the orientation of a part. If sufficient space is not available in
the current area, day and orientation, the search has to be continued systematically using
modified parameters. The algorithm's termination criterion is met when a location possibility
for the whole production time is confirmed or when the search is ultimately unsuccessful. If a
 377

place for an object is found , this location must be reserved. Then it is the tum of the next ob-
ject.

For a successful location search covering area but not confirmable in throughput time, the
planning process must be continued with a shifted object (while retaining all other parameter
settings). At first shifting follows the width (y-coordinate), then the length of a location area
(x-coordinate). At the end of the planning procedure the human planner can either call up the
graphical nesting results or restart the algorithm with modified parameters.

The area-covering loading logarithm is the most important part of the procedure. Depending
on the degree of surface area utilization, this module will be called several hundred or even a
thousand times for a single item. The program flow is illustrated in fig . 2.

A collision of several items will be recognized by bitwise logical AND operations for the
item and area columns involved. These AND operations can be executed in the computer
registers and are thus extremely fast. It is worth mentioning that INTELs 80386 requires only
two processor states for an AND operation (MORSEl ISAACSONI ALBERT, 1987). A
16 MHz 80386 can perform about 8 million comparisons per second provided, of course, that
the instructions have been loaded and decoded and the operands located in registers.

If all columns of an item and a corresponding area have been compared without a collision
being detected, this item could be placed on the location found for the current day. But the
search requires several million AND operations if a large number of attempts have failed be-
fore nesting is successful. For an item measuring 50 x 10 units of length and three planning
areas measuring 180 x 30, for example, a figure of 50 · 20 · 130 · 20·3 =7.8 million opera-
tions have to be executed for a single orientation on a production time of 20 days. This un-
derlines the significance of a hardware-oriented implementation.

The functioning of a bitwise logical AND operation is shown in fig. 3. The result of this op-
eration is one only if both the first operand and the second operand are one (that is to say
TRUE). In the other three possible cases the result is zero (or FALSE). In consequence, if
neither or only one operand shows the state TRUE, the result of the AND operation is
FALSE. An operand with TRUE content now represents an occupied element of an area and
FALSE an unoccupied one. Hence, if the result of an AND operation is TRUE, both operands
collide. The result FALSE shows on the other hand that the elements involved do not collide.
 378

[ Start
J
1
1 Initialization of variables 1
,.....
1
no
,.... yes
2
no
yes
I Set counter of item column to zero I

no y~<
1 Increment column counter and pointer of item and area columns I
.....

3f
yes
~I A place is found I V

1 Set left-shift counter to one 1

:<Y
no

I Increment left-shift counter I

I

1
no W 6

Update origin of y-coordinate

yes >1
Move item as calculated
!
1
I

,,~

\.r

~
yes
7 End
'L
t
no
no
yes 8

Reload copy of item,

increment origin of x-coordinate,
set origin of y-coordinate to zero
I
1 = Can item be moved along x-axis?
2 =Can item be moved along y-axis?
3 = Do further item columns exist and is area column & item column =0 ?
4 = Are all item columns checked?
5 = Can item be moved along y-axis and is area column & item column <> 0 ?
6 = Can this item column be moved along y-axis?
7 =Is a place be found?
8 =Can item be placed in this area after moving along x-axis?
where: & = bitwise logical AND Operator

Fig. 2: Flow diagram of area-covering location search

 379

Rules for combining bits using bitwise logical AND operator

Bits in the first operand

Bits in the second operand
Result of the operation

0 0 0
0 1 0
1 0 0
1 1 1

Examples of bitwise AND operations using 32-bit variables

1010101010101010101010101010101010101010101010101010101010101010 I
The first example produces zero, i.e. the operands do not collide

101010101010101011 10101010101010101010101010101010101010101010101
The second example does not produce zero, i.e. the operands collide

Fig. 3: Representation of a bitwise logical AND operation

A 32-bit CPU register can execute 32 AND operations simultaneously with a single assembler
command. For this reason, it is possible to recognize a potential collision between a whole
item column and an area column in a single 32-bit register comparison.

The algorithm for an area-covering location search executes successively register operations
in the direction of item length (with area origin constant), area width and, finally, area length.
The length of the part is taken into consideration by a column-counter, the area width by a y-
coordinate-counter, and area length by an x-coordinate-counter.
380

In the course of the procedure an object must therefore be shifted along the y-axis at first, if
the location search was unsuccessful. These leftward shiftings of an item can be implemented
very fast using assembler statements (left-shift or register rotating commands, see fig. 4).
INTELs 80386 requires three processor states (SHL command) for left-shifting (MORSEl
ISAACSONI ALBERT, 1987). The entire location search and collision check is thus re-
stricted to the assembler commands of a computer CPU. Shifting operations along the y-axis
can be optimized by shifting not the entire object, but only the item-column causing the col-
lision. A new collision check, embracing the total length of the part, does not restart until the
moment when a free place is found for the shifted column (or the edge of the area is reached
and shifting has to take place along the x-axis).

Cyclical left-shift operation of a (8-bit) register

Fig. 4: Left-shift operation

After incrementing the x-coordinate-counter, the part has to be shifted back to the right or a
previously saved copy has to be loaded again. Repeated right-shifting in the author's experi-
ence takes up more time on average than reloading a copy of the original object. For this rea-
son, the second alternative has been chosen.

The quality of the algorithm used has been verified by a large number of different simulation
studies. The parameters of the nesting procedure consist of planning period, production pro-
gram and operating area (including initial nestings). In addition, planning strategy is deter-
mined by priority rules for the objects and areas as well as by a scheduling strategy (rule for
the resolution of conflicts). In the course of various simulation experiments, several simple
and combined priority rules were examined. The output of the planning results is presented
by way of example in fig. 5. Here the quality of the procedure can be clearly seen.

Computer-aided parts-nesting with CALPLAN requires so little CPU time, that even a re-
peated, iterative planning process can be completed within an acceptable period (for concrete
specifications, see PETERSEN, 1992).
 381

Example of parts-nesting in a specific production area

on different shop calendar days
o 30 60 90 120 150 I 0

12~_ _~~~~~~~

oL-________~2===~L_____________~~

Day: 265

Day: no

Day· "27.

))ay: :! 0

Day: 2 5

Day: 2 9

Fig. 5: Time chart of parts-nesting

382

4. Results of computer-aided planning

A comprehensive analysis of the results of simulations for variants of the production program
shows that

some priority rules always generate unsatisfactory results;

other rules are suitable only for a specific production program; and
some object priority rules, carefully combined, produce equally good results for different
ranges of parts. The best combinations offer, even for different quality criteria, the abso-
lutely highest values.

On the whole, elementary priority rules should be used if preference is to be given to individ-
ual quality components. In any case elementary rules bring about a wide fluctuations in use-
fulness with regard to the objectives set. Although acceptable results may be attained for
some target criteria, markedly worse solutions have to be registered concerning other quality
yardsticks. For example, the rule "shortest-operation-time-first" makes possible a high degree
of area utilization, but cannot at the same time minimize delays in delivery (or a maximiza-
tion of just-in-time delivery). This result has also been documented by HOSS (1965).

Priority rules combined additively or multiplicatively should be chosen, if several aims have
to be fulfilled to an equal extent concurrently. The additively and multiplicatively combined
rules examined showed very homogeneous results, neither excellent nor very bad solutions
being registered. The choice of a given rule should be governed by the quality of solution re-
quired for the elementary components. This means that only the most successful components
should be combined. Additive and/or multiplicative components are especially suited for use
in alternative object priority rules.

A priority rule should always be combined alternatively if not only several quality criteria
have to be considered, but a different part handling according to specific object characteristics
is also necessary. Thus distinguishing between cases helps in the classification of planning
parts. The simulation studies carried out showed that, in addition to a comparatively stable
solution quality, the best results are attainable with alternative rules. The suitability of alter-
natively combined priority rules has to be verified on a case-to-case basis before they can be
applied to real problems. Fine calibration is not feasible without an evaluation of the existing
production program. Hence the requirement that company-specific conditions be taken into
account, as they have a decisive influence on the effectiveness of priority rules in achieving
target criteria (HAHN, 1989).

The specification of individual cases should be oriented toward the premises of the decision-
making party. For example, if forward scheduling is desired, the earliest possible starting time
 383

has to be considered and not the latest feasible time. The relevance of the planned starting
date is also confirmed in the technical literature. Sometimes it is regarded as the most impor-
tant assignment priority criterion (HESS-KINZER, 1976). Alternatively combined object
priority rules in turn should use those components that gained the best results in previous
simulations.

In general, it should be remembered that elementary item priority rules cannot optimize
scheduling (VOl, 1974). On the contrary, a rule should be combined alternatively within a
sequence and should not contain too many (e.g. not more than three) components.

Variations of the area priority rule generally yield no significant improvement in results.
Their importance can, however, be increased if the rule is not to have global, but object-
specific application. On the other hand, the suitability of a scheduling strategy depends on the
constellation of data in each case. If due dates have to be considered first, a strategy which
modifies nesting time last should be chosen.

5. Summary and prospects

The greatest influence of computer-aided parts-nesting arises from object priority rules. The
effectiveness of CALPLAN in meeting targets was up to 25% higher than in the case of less
successful order selection strategies (refering to the examined simulation experiments). By
means of a learning adaptive approach it is possible to increase the quality of results using
automation. In successively improved rule combinations, only those components should be
chosen which have already reached good results when not combined. Alternatively combined
priority rules generally yield the best results.

Real planning process have to offer the user additional options. At first it is desirable that a
decentralized database can be filled easily with the relevant production program. For this rea-
son, it is necessary not only to improve the user interface, but to establish a connection to the
company-wide database.

A second point of view concerns the possibility of executing not only complete new plans,
but also incremental modified (manual) plans. Net-change planning has the advantage that
satisfactory intermediate planning states need not be called in question completely. Instead,
individual relocations can be started at specific times. Furthermore, important items can be
nested manually in advance. Other situations no doubt will require execution of a new overall
plan in the future . A variant involving weekly new planning and a daily net-change, for ex-
ample, would seem to be appropriate.
384

A final aspect of the further development of CALPLAN is related to the way of scheduling
orders. With a view to nesting flexibility, backward scheduling should be offered in addition
to conventional forward scheduling. Retrograde scheduling attempts to achieve a schedule ac-
cording to estimated due dates on the basis of priority numbers (ADAM, 1988).

To recapitulate, it should be pointed out that the importance of short-term control components
will undoubtedly increase as new developments in MRP systems occur. Besides creating
autonomous decentralized decision-making fields, graphical visualized simulation becomes
more essential to support production control (ZELLI SCHEER, 1990). Spatial production
planning and scheduling in particular can be improved substantially using graphical process
visualization.

The procedure presented here fills a gap in the field of production control. Efficient solutions
to this problem yield considerable advantages for a company. Only such an approach can en-
sure that bottle-necks are recognized, that available surface area is used economically and
that, consequently, shorter lead times are achieved. The risks of delayed delivery are gener-
ally directly measurable for individual production companies through the fixing of contract
penalties. Shorter throughput times mean reductions in the amount of tied capital and in the
use of intermediate storage areas on reallocations for purposes other than originally intended.
Finally, they ensure more economical manufacturing.

Literature

ADAM, D.: Retrograde Terminierung: Ein Verfahren zur Fertigungssteuerung bei diskonti-
nuierlichem Materialfluss oder vernetzter Fertigung, in: Schriften zur Unternehmens-
fUhrung, Bd. 39, Wiesbaden 1988, pp. 90 + 92.
HAHN, D.: Produktionsprozessp1anung, -steuerung und -kontrolle - Grundkonzept und Be-
sonderheiten bei spezifischen Produktionstypen, in: Hahn, D'/ Lassmann, G. (Hrsg.):
Produktionswirtschaft - Controlling industrieller Produktion, Bd. 2: Produktionspro-
zesse: Grundlegung zur Produktionsplanung, -steuerung und -kontrolle und Beispiele
aus der Wirtschaftspraxis, Teil VI: Prozesswirtschaft - Grundlegung, Heidelberg 1989,
p.81.
HESS-KINZER, D.: Produktionsplanung und -steuerung mit EDV, Stuttgart-Wiesbaden
1976, p. 190.
HOITSCH, H.-J.: Produktionswirtschaft, Mlinchen 1985, p. 239.
HOSS, K.: Fertigungsablaufplanung mittels operationsanalytischer Methoden, Wlirzburg-
Wien 1965, p. 168.
ISRANI, S.S./ SANDERS, J.L.: Performance testing of rectangular parts-nesting heuristics,
in: International Journal of Production Research, Vol. 23, 1985, Nr. 3, p. 438.
KUHNLE, H.I SCHLAUCH, R.: Termin- und Kapazitiitsplanung bei Anlagenmontage, in:
Die Maschine, Internationale Zeitschrift fUr Fertigungstechnik und Konstruktion, 1989,
Nr. 4, pp. 76 - 80.
 385

KURBEL, K.! MEYNERT, J.: Flexibilitat in der Fertigungssteuerung durch Einsatz eines
elektronischen Leitstandes, in: Zeitschrift fUr wirtschaftliche Fertigung und Automa-
tisierung, 83. Jg., 1988, Nr. 12, p. 581.
MORSE, S.P.! ISAACSON, EJ.! ALBERT, DJ.: The 80386/387 Architecture, New York-
Chichester u.a., 1987, pp. 276 + 283.
PETERSEN, U.: Produktions- und Ablaufplanung unter Beriicksichtigung der Montage-
flachenbelegung, to appear in: Betriebswirtschaftliche Forschung zur Untemehmens-
fUhrung, Bd. 25, Wiesbaden, 1992.
SCHLAUCH, R.! LEY, W.: Flachenorientierte Termin- und Kapazitatsplanung in der Anla-
genmontage, in: Zeitschrift fUr wirtschaftliche Fertigung und Automatisierung, 83. Jg.,
1988, Nr. 5, pp. 223 - 227.
SPUR, G./ IMAM, M'/ ARMBRUST, P./ HAIPTER, J./ LOSKE, B.: Baugruppenmodelle als
Basis der Montage- und Layoutplanung, in: Zeitschrift fUr wirtschaftliche Fertigung
und Automatisierung, 84. Jg., 1989, Nr. 5, p. 234.
VOl (Hrsg.): EDV bei der PPS, Bd. 2: Fertigungsterminplanung und -steuerung, VOI-
Taschenbuch T23, 2. Aufl., Diisseldorf 1974, p. 100.
ZAPFEL, G./ MISSBAUER, H.: Traditionelle Systeme der Produktionsplanung und
-steuerung in der Fertigungsindustrie, in: Wirtschaftswissenschaftliches Studium, 17.
Jg., 1988, Nr. 2, p. 77.
ZELL, M'/ SCHEER, A.-W.: Datenstruktur einer graphikunterstiitzten Simulations umgebung
fUr die dezentrale Fertigungssteuerung, in: Reuter, A. (Hrsg.): GI - 20. Jahrestagung II,
Informatik-Fachberichte Bd. 258, Berlin-Heidelberg 1990, p. 27.
v. Lot-Sizing and Inventory Control
MODELS FOR INTEGRATED INVENTORY CONTROL BY EDI
FOR A SINGLE SUPPLIER AND MULTIPLE BUYERS

Avijit Banerjee and Snehamay Banerjee

Department of Management, Drexel University
Philadelphia, PA 19104, U.S.A.

1. Introduction

Electronic Data Interchange (EDI) is a relatively new communi-

cation technology that is expected to affect many operational
aspects of organizations. Essentially, EDI involves the ex-
change of information between trading partners (or geographi-
cally dispersed units of a firm) in a structured, machine pro-
cessable format (HANSEN and HILL, 1989). This techchnological
development permits the electronic transmission of information
from one computer application to another in a different loca-
tion, without rekeying or human intervention. EDI does not in-
clude free-form message communication, such as electronic mail
(E-Mail), or FAX transmissions, which are not machine read-
able. The basic processes of EDI are (1) directing data trans-
missions to and data collection from different application
programs, (2) Converting data between proprietary (as used by
application programs) and standard (as transmitted by the com-
munication network) formats and ( 3 ) actual transmission of
data between different locations through a communication net-
work (see SOKOL, 1989, and EMMELHAINZ, 1990, for detailed dis-
cussions on various aspects of EDI).

Although the earliest developments of EDI in the U.S.A. occurred

in the transportation industry in the 1960s, it is still in pilot
implementation stages in many other industries (TDCC, 1989).
Nevertheless, it is strongly felt by many that this relatively un-
tapped technology will play an important role in revolutionizing
many facets of managerial decision making (BANERJEE and BANERJEE,
390

1991). Particularly, in the area of inventory control, the major

benefits of EOI, i.e. reduced communication and transaction times,
greater information accuracy, reduced paperwork and improved cus-
tomer relations are likely to have profound effects on existing
procedures and policies. Some have suggested that EOI may be in-
strumental in reducing inventory levels (MALLEY and RAY, 1988) and
may offer significant advantages to firms operating under Just-In-
Time (JIT) based policies (ANVARI, 1990). We contend that in order
to derive inventory related benefits resulting from the implemen-
tation of this technology, traditional inventory management proce-
dures need SUbstantive modification. Towards this end, we focus on
the case of multiple industrial buyers purchasing a product from a
supplier who manufactures it in a batch production system.

The characteristics of EOI that can influence inventory control

are: (1) SUbstantial reduction or elimination of buyers' ordering
costs and suppliers' order processing costs, (2) faster informa-
tion transmission and processing, leading to shorter supply lead
times and (3) increase in the volume and frequency of information
exchange between buyers and suppl iers, resulting in higher levels
of cooperation on purchase related decision processes. Recent li-
terature on EOI and inventory control have discussed some of these
issues in general (SOKOL, 1989) and have suggested orderless in-
ventory systems, where the responsibility of inventory sizing and
timing decisions rests solely on the supplier (BANERJEE and BANER-
JEE, 1991). In such integrated (or coordinated) systems, the nece-
ssity and the associated costs of buyer order actions are elimina-
ted and the supplier or manufacturer is permitted to make inven-
tory replenishment decisions with the objective of minimizing
total system costs (i.e. costs incurred by all parties involved).
Needless to say that EOI may place additional financial burdens on
small suppliers to large companies (TOCC, 1989). So considering
only the buyer's advantage may not be conducive to developing
long term cooperative buyer-supplier relationships, hence the em-
phasis on total relevant costs incurred by all parties.

In a traditional industrial purchasing situation, each buyer inde-

pendently determines its own optimal ordering pattern, leaving the
 391

supplier to formulate its own production/inventory policy in res-

ponse. The existing literature on inventory management suggest a
variety of models primarily from the perspective of a buyer
(see, for example, SILVER and PETERSON, 1985). Recently, JOG-
LEKAR and THARTHARE (1990), among others, have examined the
case of multiple buyers and a single vendor with the objective
of total system cost minimization. such analyses, however, are
based on traditional replenishment concepts. With the EDI link-
age between buyers and supplier, the product consumption infor-
mation from each buyer can be automatically and instantly
transmi tted to the vendor's computer, so that the latter can
determine the replenishment decisions for each buyer. Under
such a system the buyers do not generate purchase orders, but
specify other system parameters, e. g. carrying cost, demand
rate, acceptable level of stockouts, etc. Thus, as a departure
from current practice, the decision making process for
inventory management is transferred from the buyers to the
supplier. such systems are not discussed widely in the extant
literature, but the proponents of EDI are already advocating
such ideas for the future. The potential accounting control and
audit problems that may stem from our suggested orderless, co-
ordinated systems have been investigated by NORRIS and ELAINE
(1989) and POWERS (1989).

In this study, we present analytical models for two proposed

orderless production/inventory systems for a vendor and N in-
dustrial buyers. Each of these models is based on the concept
of a common cycle time for all buyers (determined by the supp-
lier, as a means for coordination). In the first case, which is
based on an earlier model developed by BANERJEE and BANERJEE
(1991), the supplier's production policy is essentially lot-
for-lot, Le. its production cycle time is the same as the
common replenishment cycle. Furthermore, in each cycle, the
product is shipped individually to each buyer through separate
shipments. In our second model, the vendor may take advantage
of economies of scale and is allowed to replenish all buyers
through a single shipment in each cycle, if such economies
exist. Under this model, the vendor's production lot size is an
392

integer multiple of the total demand from all customers over a

replenishment cycle, thus permitting the former some flexibili-
ty in terms of prorating its setup cost among a larger number
of items, if this parameter value is relatively high. For simpli-
city, we limit our focus to a deterministic scenario and assume
a constant demand rate at each buyer's end.

2. Notation

We adopt the following notation in this paper:

Di demand or usage rate for buyer i, i 1, 2, ...• , N:
T.
1
replenishment cycle time for buyer i, i = 1, 2, ...• , N:
hi unit inventory carrying cost per time unit for buyer i,
i = 1, 2, .••• , N:
Ai ordering cost/order for buyer i (without EDI), i=1,2, .. N:
ci vendor's shipping cost/lot to buyer i, i = 1, 2, ..•• , N:
C vendor's shipment cost/cycle if all buyers are supplied
through a single shipment:
S vendor's production setup cost/setup:
A vendor's order processing cost/order, without EDI:
P vendor's production rate per time unit:
H vendor's unit inventory carrying cost per time unit:
T = common replenishment cycle time under integrated policies:
K integer multiplier for vendor's cycle time under model 2:
N
D E D. total demand rate on vendor:
i=l 1
TRC = total relevant cost per time unit:
JTRC = joint total relevant (i.e. system) cost per time unit.
Note that, of necessity, P > D and, due to the value added
concept, H < hi' for all i.

3. Individual Optimal Policies without EDl

We examine first the single supplier, multiple buyers case when

they are not linked by EDI, such that each party acts indepen-
dently and formulates its own optimal production and/or inven-
 393

tory policy. without the integration implied by EDI, the

classical economic order quantity (EOQ) based optimal order
cycle time for the ith purchaser is

T1. * = J[2A./D.h.]
111
(1 )

and the resulting minimum total relevant cost is given by

J[2D.A
111
. h. ] ( 2)

When each buyer adopts its individually derived optimal policy

and the vendor's production cycle length is Tv' the latter's
total relevant cost per time unit can be expressed as

N
TRCv(Tv ) = S/Tv + (DTH/2) (l-D/P) + E (A + c.)/T.*. (3)
i=l 1 1

The first two terms in (3), representing the supplier's setup

and carrying costs per time unit, respectively, are based on
the assertion by JOGLEKAR and THARTHARE (1990) that in the pre-
sence of a large number of buyers with unequal order cycles,
the timings of their orders may not be known deterministically.
Therefore, the vendor is likely to seek to minimize its own
costs by resorting to the classical production lot size for-
mula (see SILVER and PETERSON, 1985, for instance). The last
term in (3) denotes the supplier's order processing and ship-
ment costs per time unit, which are separated from the setup
cost, although these three costs are often combined into a
single fixed cost. It is to be noted that (3) is an approxi-
mation of the vendor's true total relevant cost, which is diff-
cult to formulate algebraically, due to the different discrete
inventory depletions occurring at different time intervals for
the various buyers. Nevertheless, the approximation gets better
as the number of buyers becomes larger and their individual or-
der sizes get smaller. In any event, this is not a critical is-
sue here, since (3) is used only for making rough cost compari-
sons between traditional practice and our EDI based approaches.

Setting the first derivative of (3) with respect to T at Tv =

Tv * yields the vendor's optimal production cycle length, i.e.
Tv * = J[2S/DH(1 - DIP)]. (4)
394

Substi tuting ( 4 ) and ( 1 ) into ( 3 ), we obtain the supplier's

minimum total relevant cost per time unit, i.e.

N
TRCv(Tv *) = j[2DSH(1-D/P)] + E (A+c.)j(D.h./2A.). (5)
i=l 1 1 1 1

Thus, adding (2) and (5), we obtain the joint total relevant
cost per time unit resulting from individual optimization

N
j[2DSH(1-D/P)]+ E(A+c.+2A.)j(D.h./2A.).
i=l 1 1 1 1 t6)

4. Model 1 For Common Replenishment Cycle Policy With EDl

As mentioned earlier, each buyer's ordering cost and the sup-

plier's order processing cost are likely to become nergligibly
small with a fully functional EDl system installed. In such a
situation, we suggest the first of our two inventory replenish-
ment systems based on the notion of a common cycle time for all
buyers. Figure 1 depicts the inventory time plots for three
purchasers and a vendor operating under such a policy, demon-
strating the nature of coordination involved. without loss of
generality, it is assumed in this figure that the supply lead
time is zero. The vendor's production batch equals the sum of
the replenishment lot sizes for all the buyers, each of these
lots representing the demand (or usage) over the common cycle
time for the buyer in question. The supplier coordinates the
system in such a way that all the buyers' replenishment lots
are delivered sequentially through the production of a single
batch. The decisions involving inventory replenishment for all
the purchasers are made by the vendor, Hence this suggested
approach is an integrated, orderless control system.

For purchaser i, in the absense of an ordering cost, the total

relevant cost per time unit consists of only holding cost, i.e.
TRCi(T) = (D i T/2)h i . (7)
Similarly, without a significant order processing cost, the
vendor's total relevant cost per time unit is now
 395

T
:........

BUYER 2

T
I...:

~~:.
O~··
ffi~
>..J
:.
'.
·~I
~ D, :
.'
•
'.
Z· . ' .

: .: .1. . : .: T : .1 : ::
. . . . . . ' , I

' .

: .: . '
.,: :. ::
SUPPLIER

T
TIME~

FIGURE I: Inventory Time Plots for a Single Supplier and Three Buyers Under
Common Cycle Replenishment Policy (Modell)
396

N N 2
(S +.E Ci)/T + (HT/2P).E Di (8 )
1=1 1=1
From (7) and (8) the total system cost per time unit is
N N
JTRC(T) = (S +.E Ci)/T + (T/2)[(H/P).E Di 2 + E Dihi' (9)
1=1 1=1
It can be easily shown that (9) is strictly convex in T and the
jointly optimal common cycle time, T*, obtained from the first
order optimality condition, is given by

N N N
T* = J[2(S + E c.)/{(H/P) E 0. 2 + E D.h.}]. (10)
i=l 1 i=l 1 i=l 1 1
substituting (10) into (9) yields the minimum joint total rele-
vant cost resulting from the adoption of a common cycle time by
all the parties, i.e.

N N 2 N
JTRC(T*) J[2(S + E Ci){(H/P) E Di + E D.h.}]. (11 )
i=l i=l i=l 1 1

If the buyers wish to operate under a JIT based purchasing

policy (since, in the absense of an ordering cost it would be
in the interest of each buyer to minimize its delivery lot
size, i.e. cycle time) it may be desirable for the common cycle
time to be relatively small, such as a day, or even a fraction
of a day. From the vendor's standpoint, however, such a short
production cycle length may not be economically attractive,
particularly if its setup cost is relatively high. Thus, the
latter must reduce this cost, so that a predetermined, rela-
tively short cycle time is justifiable. In this context, we re-
write (10), replacing S with So and T* with To' as

(12)

where So represents the reduced setup cost that will result in

an optimal common cyle time, To' consistent with JIT concepts.
 397

Proposition 4.1: For any buyer i, the necessary and sufficient

condition for the cost effectiveness of adopting a common cycle
time, T*, is T* < 2T.*.
1

Proof: If adopting a common cycle time, T*, is to be cost eff-

ective, TRCi(T*) must be less than TRCi(T i *), i.e. from (7) and
(2), (D.T*/2)h. < J(2D . A. h . ), which simplifies to the above
1 1 111
condition.

The implication of this proposition is that a jointly optimal

common cycle time policy yielded by model 1, would be desirable
for any buyer from a cost perspective, as long as its delivery
lot size under such a policy is less than double its EOQ.

5. Model 2 For Common Replenishment Cycle Policy With ED!

As alluded to above, our second suggested integrated inventory

policy is also based on a common cycle approach, albeit with
two important distinctions. First, the supplier's inventory
cycle length under this policy can be an integer multiple of
the replenishment cycle time common to all buyers, i. e. the
former's inventory cycle length is KT, where K is a positive
integer and T is the replenishment cycle time. Secondly, all
the buyers' lots are delivered simultaneously at the beginning
of each replenishment cycle, either through separate individual
shipments or a single joint shipment. Obviously, the latter
delivery procedure is chosen by the vendor, if significant eco-
nomies of scale in transportation exist. Either of these trans-
portation methods is in marked contrast with the sequential,
staggered shipment pattern dictated by the first policy. Figure
2 illustrates the inventory time plots for a two-buyer, single
supplier case with K = 3 under our second integrated policy.

As before, a fully operational EDI makes each buyer's ordering

cost and the vendor's order processing cost negligibly small
and the latter's total relevant cost per time unit can be
expressed as
(1/T) (S/K+C) + (DTH/2)[{D/P(2-K)+K-1}]. (13)
398

BUYER 1

~
. .

BUYER 2:

T
01( »:

SUPPLIER:

KT
TIME~

AGURE 2: Inventory Time Plots for a Single Supplier and Two Buyers Under
Common Cycle Replenishment (K = 3) with Joint Shipment (Model 2)
 399

The second term in (13) represents the inventory holding cost

per time unit for the vendor, which has been derived earlier by
BANERJEE (1986). Furthermore, ROSENBLATT and LEE (1985) have
proved in an earlier paper that K must be a positive integer
in any solution that is optimal for the vendor, or for the
total system of buyers and supplier. Note that in formulating
the cost function (13) we have implicitly assumed that a bulk
joint shipment to all the buyers is more economical than sepa-
rate deliveries. If this is not true, the cost parameter C in
(13) and subsequent expressions is replaced by the term EC i "
i

From (7) and (13), the joint total relevant (i.e. system) cost
per time unit can be expressed as

N
JTRC(K,T) = (S/K+C)/T + (T/2)[DH{D(2-K)/P+K-1} + E D.h.]. (14)
i=l ~ ~

Once again, it can be shown easily that (14) is strictly convex

in both K and T. Thus, the first order optimality condition
for a given K, leads to the following expression for the opti-
mal common replenishment cycle time, T**:

N
T** = J[2(S/K+C)/{DH(D(2-K)/P+K-1) + E D.h.}]. (15)
i=l ~ ~

Substitution of (15) into (14) results in

N
JTRC(K,T**) = J[2(S/K+C) {DH(D(2-K)/P+K-1) + E D.h.}]. (16)
i=l ~ ~

Note that in (16) K is a positive integer. Thus, if the total

system cost per unit time is minimized at K = K*, the following
conditions must hold:
JTRC(K*,T**) ~ JTRC(K*-l,T**) and JTRC(K*,T**) ~ JTRC(K*+l,T**)
Using (16) in the above conditions, we obtain

N
K*(K*-l) ~ S[(l/DH) . E Di h i +2D/P-1]/[C(1-D/P)] ~ K*(K*+l). (17)
~=1
400

In order to determine the jointly optimal integrated EDI based

policy under model 2, K* is first calculated from conditions
(17), then K is replaced with this K* in (15) to determine T**.

Proposition 5.1: For any buyer i, the necessary and sufficient

condition for the cost effectiveness of adopting a common re-
plenishment cycle time of T** is T** < T.*.
~

Proof: Similar to that for Proposition 4.1.

Proposition 5.2: The necessary and sufficient condition for the

equality of the vendor's cycle time and the buyers' common re-
plenishment cycle length is

N
S 5 [2C(1-D/P]/[(1/DH).E Dihi + 2D/P - 1]
~=1

Proof: Substitution of K* 2 into ( 17) leads to the above

condition.

Finally, as before, it can be readily shown that if the buyers

wish to operate under a JIT replenishment policy with a prede-
termined, relatively short cycle time, To', the supplier's set-
up cost must be reduced to So ' (so that the latter's lot-for-
lot JIT production and replenishment policy under model 2 is
optimal from a system cost standpoint), where

N
S ' = (T ,2/2) [HD 2/P + E D.h.] - C. (18)
o 0 i=l ~ ~

From (12) and (18), it is clear that under identical delivery

costs per cycle and shipment frequency requirements under JIT,
the setup cost reduction necessary on the supplier's part is
always higher for model 2.

6. Numerical Illustration

Suppose that the following data is available for a system con-

 401

sisting of three purchasers and a supplier:

01=1000 units/year, 02=5000 units/year, 03=800 units/year,
A1 = $20/order, A2 = $30/order, A3 = $25/order,
h1 = $5/unit/yr., h2 = $4/unit/yr., h3 = $4.50/unit/yr.,
S = $200/setup, A = $15/order, H = $3/unit/yr.
P=25000 units/yr., c 1 =c 2 =c 3 =$90/shipment, C = $50/cycle.

6.1 Individually Optimal Policies Without EDI

Optimizing individually, without the benefit of EOI, the cycle

times for the three buyers are 0.0894, 0.0548 and 0.1179 year,
respectively. The resulting minimum TRC values are, respective-
ly, $447.21, $1095.45 and $424.26. Using (4), the supplier's
cycle time is 0.1641 year, yielding an annual TRC of $4143.51.
Thus, the joint total relevant (i.e. system) cost amounts to
$6110.43 per year. In this case the vendor's cycle time is re-
latively long, which may be reasonable in view of its substan-
tial setup cost, but results in relatively high order process-
ing and shipment costs.

6.2 Integrated Policy 1 with EDI

Using (10), the optimal common cycle time for model 1, T*, is
found to be 0.1351 year. As a result, the three buyers' and the
supplier's TRC values are, respectively, $337.65, $1350.59,
$243.11 and $2363.09 per year, yielding a total system cost of
$4294.44 per year (representing approximately a 30% savings
over independently derived individually optimal policies). Note
that the adoption of this integrated policy substantially redu-
ces the costs incurred by all the parties except for buyer 2,
for which the TRC value, in fact, increases. This occurs
because the condition stated in Proposition 4.1 is violated for
the buyer in question Le. 0.1351 > 2(0.0548). Nevertheless,
through a side payment or some similar scheme, the 30% savings
in total system costs may be shared by all the parties involved
402

in a fair and equitable manner ( see BANERJEE, 1986 or ROSEN-

BLATT and LEE, 1985, for instance), such that buyer 2 may have
adequate inducement for accepting this EDI based inventory con-
trol system.

In addition, if the supplier is able to reduce the common re-

plenishment cycle time to less than 2(0.0548) = 0.1196 year by
an appropriate reduction in its setup cost as per (12), the
second buyer's TRC value will decrease under this T* policy and
the total system cost reduction will be even higher. It can,
therefore, be said that for an EDI based common cycle approach
to be economically attractive, setup cost reduction (which is
also a basic thrust of JIT) plays an important role, indicating
the compatibility of orderless, coordinated policies with JIT
principles.

6.3 Integrated Policy 2 With EDI

Under our second suggested coordinated policy developed in

section 5, K* is determined to be 3 from (17) and, using (15),
the jointly optimal replenishment cycle time of 0.06045 year
(implying a cycle length of 0.18135 year for the vendor). As a
result, the TRC values of the three purchasers and the vendor
are $151.31, $604.51, $108.32 and $2995.82 per year, respec-
tively. The total system cost now is $3859.87 (representing a
savings of about 37% over individual optimization without EDI).
Under this policy, the supplier's TRC is larger than that re-
sulting from model 1 (primarily due to higher inventory hold-
ing costs). Nevertheless, total system cost yielded by model 2
is about 10% less than that stemming from model 1. Once again,
some equitable cost savings sharing scheme may make the former
policy more attractive from the supplier's viewpoint; and with
suitable setup cost reduction, total system cost savings can be
even higher. Overall, it appears that the common cycle policy
based on model 2 is superior to that derived from model 1, due
to the inherent flexibility of the former, as mentioned
earlier.
 403

7. Conclusions

The numerical example above amply demonstrates a major poten-

tial advantage of integrated, orderless inventory control sys-
tems over individually optimal systems, namely, significant re-
ductions in buyers' order cycles (and lot sizes), as well as
the vendor's production cycle (and batch size). These reduc-
tions not only yield sUbstantial savings in total system cost,
but also point out the vi tal importance of EDI towards the
successful adoption and implementation of JIT inventory con-
trol techniques.

A significant limitation of this study involves the assumption

of deterministic conditions. The concepts developerd here are,
however, not without merit. Under stochastic conditions, the
notion of a common cycle time can be an important part of an
orderless, periodic review policy for the buyers. Based on the
existing variabilities in demand and lead time for each buyer,
its appropriate order-upto-level can be determined. In such a
case, the monitoring of inventory levels, and the supplier's
production and shipping decisions can be greatly facilitated by
EDL

Regardless of existing operating conditions (deterministic or

stochastic), an important accomplishment of EDI in any common
cycle based policy is the elimination of order actions and the
associated costs on the part of the buyers. Our suggested
models, although deterministic, indicate that this can contri-
bute towards total system cost reduction, which can be shared
by all parties concerned. In both our decision models, the res-
ponsibility of making replenishment decisions rests entirely on
the vendor, indicating the need for tight coordination and con-
trol. Such close coordination can be achieved easily and
economically through the implementation of EDI. Even under de-
terministic conditions, the potential benefits of this techno-
logy have been made clear by our analysis.
404

References

ANVARI, M. (1990), "Electronic Data Interchange and Invento-

ries", Working Paper, Concordia University, Canada.
BANERJEE, A. (1986), "An Integrated Lot Sizing Model for a Pur-
chaser and a Vendor", Working Paper, Louisiana State Univer-
sity, Baton Rouge, LA, U.S.A.
BANERJEE, A. and BANERJEE, S. (1991), "Coordinated, Orderless
Inventory Replenishment for a Single Supplier and Multiple
Buyers Through Electronic Data Interchange, Working Paper,
Drexel University, Philadelphia, U.S.A.
EMMELHAINZ, M. (1990), Electronic Data Interchange: A Total
Management Guide, New York: Van Nostrand Reinhold.
HANSEN, J.V. and HILL N.C. (1989), "Control and Audit of Elec-
tronic Data Interchange", MIS Quarterly, Vol. 13, 403-413.
JOGLEKAR, P. and THARTHARE, S., "The Individually Responsible
and Rational Decision Approach to Economic Lot Sizes for One
Vendor and Many Purchasers", Decision Sciences, Vol. 21,
492-506.
MALLEY, J.C. and RAY, R. (1988), "Informational and Organiza-
tional Impacts of Implementing a JIT System", Production &
Inventory Management, Vol. 29, 66-70.
NORRIS D. and ELAINE, M. (1989),"Control of Electronic Data In-
terchange Systems", Journal of Systems Management, Vol . 40,
21-25.
POWERS, W.J . (1989), EDI Control and Audit Issues: EDIA Mono-
graph, Alexandria, VA, U. S. A.
ROSENBLATT, M.J. and LEE, H. (1985), "Improving profitability
with Quantity Discounts Under Fixed Demand", lIE Trans-
actions, Vol. 17, 388-395.
SILVER, E.A. and PETERSON, R. (1985), Decision Systems for In-
ventory Management and Production Planning, (2nd. ed.), New
York: John Wiley & Sons.
SOKOL, P.K . (1989), EDI: The Competitive Edge, New York: Inter-
text Publication (McGraw-Hill).
TDCC: The Electronic Data Interchange Association(1989), EDI in
North America: status of Usage and Technology, EDIA Mono-
graph, Alexandria, VA, U.S.A.
Optimal Inventory Policy in Assembly Systems with
Component Commonality
by

Srinivas Bollapragada and Ram Akella

Graduate School of Industrial Administration
and the Robotics Institute
Pittsburgh, PA 15213, USA

1 Introduction

The problem of multiplant coordination involves coordinating a chain of facilities in which

the upstream faciliti es of the chain supply components and subassemblies to the ones
downstream and the final assembly facility meets the demands for the finished goods. In
order to meet the externa.l demand which is stochastic, with a certain high probability,
it is necessary to maintain inventori es at va.rious levels. The presence of commonality of
components among various products can be exploited to reduce the safety stock levels in
such systems.

Two different production environments occur in manufacturing, the first is an assembly

in which the assembly operation is long. As a consequence, it is necessary to assemble
the finished goods to forecast (ATF). III contrast , in the other environment the assembly
operation is short and as a result it is possible to observe the realization of the actual
demand and assemble to order (ATO). In both environments it is necessary to order
(procure) raw materials (components) well ahead of time because of the long procurement
leadtimes of these components.

The problem of component commonality in assemble to order environments has been

dealt with by several authors. Baker et al (198.5), Gerchak et al (1988), and Gerchak and
Henig (1989) addresed the issue of component commonality in simple assembly systems
and demonstrated the characteristics of the solution. In this paper, we address the assem-
ble to forecast environment where a single component is assembled into several products
for which stochastic demands exist. The component is procured from an external supplier
and arrives after a fixed lead time. Once a component shipment is received it is allocated
among the various products. The product assembly takes a significant lead time and
hence production is done to meet forecasted demands. Linear penalty and holding costs
406

are incurred at the end of every period on backorders and excess inventory respectively.
The procurement and allocation decisions are made to minimize the expected total costs.

The rest of the paper is organized as follows. Section 2 deals with the notation and
problem formulation. The optimal allocation policy is described in Section 3. The ordering
problem is addressed in Section 4 and the conclusions are in Section 5.

2 Notation and Problem Formulation

The notation used in the paper is as follows:

N = number of products,

i = index of products, .

U; = usage rate of the component in product i,

ho = holding cost per period per unit of the common component,

h; = holding cost per period per unit of product i,

p; = penalty cost per period per unit of product i,
L = procurement lead time of the common component,

I; = assembly lead time for product i,

d; = one-period demand for product i, a random variable,
/;(.) = probability density function of d;,

F;(.) = cumulative distribution function of d;,

dl = sum of I i.i.d. random variables each identical to di ,

/!(.) = probability density function of dL

F!(.) = cumulative distribution function of dL
v = 'Lf;ol 'L~l U;d;,h the total demand for the N products over L periods,

Pi = mean one-period demand for i,

0'; = standard deviation of one-period demand for product i,

 407

fLv = L L:!'::1 UifLi, the mean of V.

The state and decision variables of the system are as follows .

Xt = inventory of the common component available for allocation at the beginning of

period t,

Yi,t = work in process and finished product inventory of product i,

Vi ,t = allocation to product i in period t,

Wt = quantity of the common component ordered in period t,

Zt = the total system inventory on hand and in transit in terms of the component.

The state variables are governed by the following equations.

N
Xt+1 = Xt - L UiVi,t - Wt-L+l (1)
i=1

Yi ,t+1 = Yi,t + Vi,t - di ,t (2)

N
LU iVi,t::; Xi,t (3)
i=1
N
Zt+1 = Zt + Wt - L uidi,t (4)
i=1

Every period the following events occur in that order. The state of the system is
observed and an order tv for the component is placed on the external supplier.The order
placed L periods earlier is received. The allocation of the component to the products is
then made. Dem ands for that period for the products occur at the retailers. Holding
costs for excess inventory and penalty costs for backlogged demands for the products are
then assessed.

We first consider a single cycle problem involving a procurement decision in period 0

and an allocation decision L periods later when the ordered quantity is received. Holding
costs on the component due to the procurement decision are incurred in period L. The
allocation to the product i in period L comes out of the assenbly line and is available to
meet the external demand in the period L + Ii. Therefore, the holding and penalty costs
on product i due the ordering decision in period 0 are incurred in period L + Ii. The
ordering and the allocation decisions are made to minimize the expected value of the sum
of these costs. Clearly, the quantity to be ordered depends on the allocation policy used.
We first solve the allocation problem and use the obtained optimal allocation policy in
solving the ordering problem .
408

3 The allocation problem

The problem can be stated as follows . Given the work in process and finished product
inventory of each product and the common component stock available, decide the quan-
tities to be released to each product. Using the notation described earlier, with the time
indices supressed, the problem can be formulated as follows.

subject to the constraints

N
"u·v,<x
L...J ' , - (6)
i=1

Vi ~ 0, i = 1, .. . ,N (7)

Problems of this form have been discussed earlier in literature (Luss and Gupta, 1975
and Zipkin, 1980). The problem described above is convex since the constraints are linear
and the Hessian, H of the objective function is a diagnol matrix with positive diagnol
elements (Hii = (pi + hdfJ·+l(Yi + Vi)) and hence is positive definite. The allocation
quantities, Vi can be obtained by solving the first order Kuhn-Tucker conditions given
below.

(pi + hi )pf·+l(Yi + Vi) - Pi - ho = A + Ai, i = 1,2, ... ,N (8)

N
A(X ...:. L UiVi) = 0 (9)
i=1

AiVi = 0, i = 1,2, ... ,N (10)

Ai ~ 0, i=1,2, ... ,N (11)

where A is the lagrange multiplier associated with the constraint in equation (6) and Ai, i =
1,2, ... , N are the lagrange multipliers associated with the nonnegativity constraints on
the release quantities.

From the first order conditions it can can seen that Ai = 0 when Zi > O. Let I =
{i : Ai = O}, i.e., I is the set of retailer indices to which positive quantities are shipped.
Define
(12)

The solution of the first order Kuhn-Tucker conditions yeilds the following allocation
policy.
 409

Vi = Si - Yi, i = 1, ... ,N
• Case 2 ;r < L;:'l Ui(Si - yd
V=(F 1i +1 t 1 (Pi+).)_y iEI
I I Pi + hi "
where the lagrange multiplier). is chosen to satisfy LiEf UiVi = x.
Thus the allocation policy is characterized by the critical numbers Si associated with
the products. When there is sufficient stock of the common component the product
inventories are raised to their respective critical numbers. This is shown in case 1. When
there is not enough stock of the component to release up to the critical numbers, the
available stock is allocated among the products and this is done as described in case 2.

4 The Ordering Problem

The ordering problem concerns the determination of the quantity U of the common com-
ponent ordered under the optimal allocation policy discussed in the previous section. The
quantity to be ordered in period 0 is determined to minimize the expected value of the
sum of the holding cost of the common component in the Vh period and the holding and
penalty costs due to product i in the period L + li for i = 1,2, ... , N. Let s be the the
total system inventory after the order has been placed, i.e, s = w + z. The cost func-
tion consists of two terms, ea.ch representing a different case which occurs with a certain
probability. The first is the cost incurred when the release upto levels for the products
are achieved in period L. This happens when V S; s - L~l UiSi which occurs with a
probability of Fv{s - L;:'1 U;Si). The cost incurred in this case is
N N
ho(s - /lv - L: UiS;) + E{L: hi(Si - d~i+1)+ + Pi(d~i+1 - Si)+}
i=l i=l

The second term is for the case when the relese upto levels cannot be met and therefore
the available stock of the component is divided among the products. The cost incurred
in this situation is given by

(13)

where
N
L:1Ii(Vi + Yi) = s - V (14)
;=1
410

Vi 2: 0, i = 1, . .. ,N (15)

Note that C is a function of 8 - V and the ViS, the product inventories. The ViS are
random variables with distributions hard to compute. So we make an approximation to
eliminate the dependence of C on the ViS. This is done by relaxing the nonnegativity
constraints on the release quantities Vi. A similar approximation has been addressed in
Fredergruen and Zipkin (1984). The approximation is also analogous to the equal fractile
assumption in Eppen and Schrage (1981). Under this approximation C is a function of
8 - Valone and is given by

where
N
L Ui(Vi + Vi) = 8 - V
i=1

Let .\(8 - V) be the value of the lagrange multiplier for the equality constraint at the
optimal solution. Therefore,
dCjd8 = .\(8 - V) (16)
The lagrange multiplier .\(8- V) equals ho when 8 - V = 'L~1 Ui5i and .\(8- V) = -Pmax
when V 2: 8 where Pmax is the maximum of PhPZ,'" ,PN. For values of V in between the
two, .\ is an decreasing function of V.

The cost function, g(8) can be written by combining the two terms as follows .
N N N
g(8) = FV(8 - L ui5i )[ho(8 -I.tv - L ui5i ) + E{L hi(5i - d~·+I)+ + Pi(d~·+I - 5i )+}J
i=1 i=1 i=1
+ fv:s-'L:'..t u.s. C(8 - V)dFv(V) (17)

The single cycle cost function g( 8) is thus a function of one variable s, which is the
total inventory in the system after the external order is placed. The quantity to be ordered
is obtained by minimizing g(8) with respect to w. We now show that (Lemma 1) g(8) is
convex in 8 which implies that the optimal policy has an order up to structure. The value
of 8 which makes the first derivative of g( 8) equal to 0 is the order upto level 50.

Lemma 4.1 The cost function, g(8) i8 convex in 8 and hence in w.

Proof: The first derivative of the cost function w.r.t 8 is

N roo
hoFV(8 - ~ tli5i ) + }v=s-'L:'..t u.s• .\dFv(V)
 411

The second derivative of 9(S) w.r.t. sis

fv:S-z:.':,,1 u,s, d>../dsdFv(V)

which is nonnegative since>.. is a non decreasing function of s.
The ordering problem can be reduced to the standard Newsboy problem with the
single period cost function g( z + tv) (z is the stock on hand and tv is the quantity to
be ordered) and a zero ordering lead time. The results available from the literature can
directly be applied here.

5 Conclusions

We address the problem where a single component is assembled into various products with
lead times for product assembly and component procurement. Optimal policies for the
allocation of the component to the products and the procurement of the component are
described. The optimal allocation policy is characterized by critical numbers associated
with each product. The policy is to release up to these critical numbers if there is sufficient
stock of the common component available; otherwise the available stock is allocated among
the products as described in the paper. The procurement problem can be reduced to a
Newsboy problem with zero lead time under certain approximations. The order up to
level for the component procurement is obtained by solving an implicit equation.

References

K. J. Arrow, S. Karlin and H. Scarf, Studies in the Mathematical Theory of Inventory and
P1'Oduction, Standford University Pl'ess, 1958.

K. R. Baker, M. J. Magazine and H. L. W. Nuttle, "The Effect of Commonality on

Safety Stock in a Simple Inventory Model," Management Science, Vol. 32, No.8, pp.
982-988, August 1985.

D. A. Collier, "The Measurement and Operating Benefits of Component Part Common-

ality," Decision Sciences, Vol. 12, pp . 85-96, 1981.

G. Eppen and L. Schrage, "Centralized Ordering Policies in a Multiwarehouse System

with Lead Times and Random Demand," Multi-level P1'Oductionlnventory Control
Systems: Theol'Y and Practice, L.Schwarz (ed.), North Holland, Amsterdam, pp.
51-67, 1981.
412

A. Federgruen and P. Zipkin, "Approximations of Dynamic Multilocation Production and

Inventory Problems," Management Science, Vol. 30, No. 1, pp. 69-84, 1984.

Y. Gerchak and M. Henig, "Component Commonality in Assemble-To-Order Systems:

Models and Properties," Naval Research Logistics, Vol. 36, pp. 61-68, 1989.

Y. Gerchak, M. J . Magazine and A. B. Gamble, "Component Commonality with Service

Level Requirements," Maganement Science, Vol. 34, No.6, pp. 753-760, June 1988.

H. Luss and S.K. Gupta, "Allocation of Effort Resources among Competing Activities,"
Operations Research, Vol. 23, pp. 360-366, 1975.

P. Zipkin, "Simple Ra.nking Methods for Allocation of One Resource," Management Sci-
ence, Vo1.26, No.1, pp. 34-43, 1980.
 Valuation of Leadtime Reduction in Multi-Stage
Production Systems

by

Karl Inderfurth
Faculty of Economics
University of Bielefeld
D-4800 Bielefeld 1
Germany

Abstract

Projects for leadtime reduction in production system have to be evaluated economically

with respect to their impact on logistic costs and customer service. In this paper a pro-
cedure for valuation of these effects is presented taking into consideration uncertainty in
customer demands as well as in manufacturing leadtimes. This procedure can be applied
to general multi-stage production systems with convergent or divergent structure and can
be used as a helpful tool for assessing the cost/service performance of investments III
reducing length and variability of leadtime in complex production systems.

1. Introduction

In many companies efforts to reinforce competitiveness by improving cost position and

customer service are focussed on measures to reduce leadtimes for operations in the ma-
nufacturing system. This can be done by reorganizing the production system following
a Just-in-Time concept, by promoting the use of flexible manufacturing systems, by sim-
ply installing extra capacity, and other actions more. Whatever the specific project for
leadtime reduction may be, for checking its value to the fir1p besides a possible strategic
judgement it has to be evaluated economically by comparing its costs and benefits as it
is done for investment decisions by using capital budgeting techniques (see Wouters 1991).

There are yet some approaches that tackle this problem for specific investment situa-
tions as, for instance, presented by Moerman (1988)"or Fine and Freund (1990) . In this
paper we will focus on the problem of how the benefits of any leadtime reduction project,
that may affect leadtimes anywhere in the manufacturing systems, can be valuated in
terms of its corresponding effects on customer service and operating costs. Corbey (1990)
has broadly desrcibed these effects in a more qualitative way. Here we will analyze conse-
quences of leadtime reductions quantitatively using an Operations Research approach to
414

determine its implications on safety stocks and work-in-process in the whole prmluction
system. Lot size effects of leadtime reduction are not taken into account following the
arguments by Corbey (1990) that a direct lot size impact in the econonomic valuation of
lead time investment is quite negligable.

Measuring the influence of leadtime reduction on customer service is quite difficult be-
cause there are at least two dimensions of service performance that have to be taken into
consideration, service time and service level. Now, while changes in leadtimes directly
will change the delivery time to customers in a make-to-order situation, in make-to-stock
systems there just will be an impact on the degree of customer service levels. Thus, as
shown in figure 1, the relevant type of service effect depends on the location of the custo-
mer order decoupling point (codp) in the total production system (see Wouters 1991).

make - to - stock make - to - order

Figure 1: Production system and customer order decoupling point (codp)

Whereas the service time effect goes hand in hand with the lead time variation and the-
refore is easy to calculate, the service level effect in an multi-stage setting is very much
harder to evaluate.

In order to tackle this difficult problem we will concentrate on make-to-stock systems

and present a procedure for the numerical evaluation of the optimal reaction on leadtime
reduction possibilities with respect to service level improvement and holding cost savings.
This is done for a situation where not only customer demand but also lead time itself is
assumed to be stochastic. Thereby two kinds of leadtime reduction can be investigated
seperately, a decrease in the leadtime expectation (=LT-length) and a reduction of the
leadtime variance (=LT-variability). We will discuss leadtime effects in simple one-stage
as well as in complex multi-stage manufacturing systems with convergent and divergent
production structure.

2. Leadtime reduction in single-stage systems

At first the effect of leadtime reduction on service and cost is investigated for the or-
dinary one-stage production/inventory system. This is done because this system is the
basic element of the more complicated multi-stage systems and can be evaluated in an
easy manner using traditional inventory theory.
 415

Throughout this paper it is assumed that we work under periodic inventory control with
d
a lot-for-Iot ordering philosophy. Both, demand per period and leadtime periods ~ are
allowed to be stochastic. They are mutually independent and i.i.d. random variables - at
least approximately - described by normal distributions with mean and variance {ti, O"~}
d
for and {X, O"U for ~. The ordering policy has to guarantee a predetermined customer
service level 0: which is defined as the probability of not going out of stock during re-
plenishment time. Cost is charged to inventory holding with h being the cost per unit
of stock. In determining expected stocks and costs for a single-stage system we have to
differ between the case of deterministic and stochastic leadtimes.

In the deterministic LT-case (0",\ = 0) we know from basic inventory theory that
safety stocks (sst) have to be

(1)

with qa = cIl- 1 (0:) and cIl(.) as standardized normal d.£.

The expected pipeline stock (pst) will be equal to the expected lead time demand

pst = d·'\ (2)

meaning that both inventory categories and hereby total stocks and inventory holding
cost depend on the LT length i It is evident that leadtime reduction leads to decreasing
stocks and cost, in linear terms for work-in-process and progressive for safety stocks.

In the stochastic LT-case (0",\ > 0) two different planning situations must be consi-
dered. For a single production stage which is part of an internal production and supply
chain the stochastic leadtime ~ is converted into a deterministic leadtime \ for reasons of
integrated multi-level production planning as we know from MRP systems. Usually this
fixed planned lead time ~ is set in such a way that it will be sufficient in a reasonable
percentage 7r of all cases, thus being

(3)

It is assumed that in case of extraordinary long leadtimes, not covered by \ a delay in

delivery can be avoided by using emergency actions supported by a sufficient level of
operating flexibilitiy in the total production system. Thus, using fixed planned leadtimes
means that each (internal) demand will be satisfied just after \ periods producing expec-
ted stocks as follows
416

(4)

and

(5)

Here stocks and cost depend on both LT-length (X) and LT-variability (u.x). It can
be seen that a leadtime reduction in terms of decreased LT-variability is connected with
reduced stocks in a linear way for work-in-process and in a slightly progressive way for
buffer-inventory.

If a single production/inventory stage is located at the final level of a multi-stage produc-

tion system, directly facing external demand, the service process is usually arranged in
such a way that customer demand will be fulfilled as soon as the fluctuating final-stage
leadtime will allow it. So in this planning situation we are dealing with flexible leadt-
imes. As a consequence the desired customer service has to be guaranteed with respect
to a stochastic demand with in a stochastic leadtime resulting in a safety stock that ap-
proximately - supposing lead time demand to be normally distributed - can be described
by (see Silver, Peterson 1985)

sst = qa . JA• u~ + iF . u~ (6)

The pipeline stock size will not differ from the deterministic LT case in (2)

pst = d· A
We see that with flexible leadtimes a reduction of LT-variability will not decrease work-
in-process. It just will result in degressively decreasing safety stocks.

Whatever the planning situation may be, it is evident from (1) to (6) that with sto-
chastic leadtimes total holding costs H = h . (sst + pst)" in a simple way depend on
LT-parameters X and U.x on the one side and on service level 0: on the other

(7)

From this equation we can evaluate the impact of leadtime and its possible reductions on
the two performance criteria cost and service.
 417

3. Leadtime, cost, and customer service in multi-stage systems

For multi-stage production systems valuation of leadtime reduction is connected with

additional problems. While - furtheron regarding stationary (external) demand - the pi-
peline stock still is simply computable from expected demands and single leadtimes of
production processes, the system safety stock no longer is directly determined by custo-
mer service levels. In multi-level systems desired service levels can be reached by various
ways of allocating safety stocks in the system (upstream vs. downstream) generating the
problem of optimal safety stock determination with respect to connected holding costs.
The solution of this problem, being a necessary information for the valuation of lead time-
reduction projects, depends on the leadtime properties of all processes in the system.
According to the comments in section 2 in case of stochastic lead times in a multi-stage
situation we suppose the application of fixed planned leadtimes for all production levels
except the final one where flexible leadtimes are used.

In this planning situation safety stock optimization depends on the way of integrating
safety stocks in the inventory control and reordering process. In multi-stage systems for
material coordination under uncertainty a base-stock control policy is recommended (see
Silver, Peterson 1985). From inventory theory we know that - assuming a lot-for-lot or-
dering situation - this kind of policy is optimal for serial production structures (Clark,
Scarf 1960) as well as for convergent systems (Rosling 1989) and divergent ones (Langen-
hoff, Zijm 1990), represented by an echelon order-up-to-S policy. According to this policy
local safety stocks at each production level are determined.

The size of these safety stocks depends on external customers service levels on the one
and on internal supply conditions on the other hand. For fulfilling internal production
orders within the planned leadtime facing uncertainty in requirements each production
stage mainly will operate with safety stocks according to a prespecificed internal service
level which has to guarantee availability of products even for unexpected demand up to a
reasonable amount. In case of extraordinary requirements it will be assumed - reflecting
to arguments introduced by Simpson (1958) and broadly discussed by Graves (1988) - that
operating flexibility in the production system is high enough to enable a complete supply
of succeeding production stages by ot.her means than by safety stocks. Based on this kind
of total reliability of the internal supply process the problem of safety stock allocation
concentrates on the question how the total replenishment times of all final-stage products
in a multi-stage system have to be divided into local replenishment times at different
stock points determining adjoined safety stocks which will guarantee desired internal or
external service levels. At first we will describe the problem formally for the elementary
multi-stage case of a serial production system.

3.1. Serial production systems

As depicted in figure 2 a serial production system is characterized by a sequence of ope-

rations beginning with outside supply of raw material up to the final manufacturing steps
for of the finished product. Each intermediate product has just a single predecessor and
418

a single successor.

Figure 2: A 3-stage serial production system

Stock holding is assumed to be possible for each product in the chain connected with
specific unit holding costs hi depending on the value of product i (i = 1, 2, 3). End-item
customer demand per period d3 is normally distributed with mean d3 and variance 0'33'
The properties of the stochastic lead times also are described by their means and varian-
ces characterizing X; as (expected) LT-Iength and O'~; as LT-variability parameters. In
periodic review situations final-stage leadtime mean X3 additionly contains the length of
the review period. Assuming - without loss of generality - a 1:1 input-output-relation at
each processing stage, under the base-stock control policy as discussed the internal requi-
rements per period for intermediate products also are normally distributed with means
d; = d3 and variances O'~; = O'J3 (i = 1, 2). The required customer service is described
by a service level 0'3 , whereas for internal supply - possibly different - service levels 0';
(i = 1, 2) are postulated. Safety stock optimization now means to determine product
specific replenishment times Ti for building up protective stocks in such a way that all
given service levels are satisfied and total inventory holding costs are minimized. For
the probelm formulation we have to differ between the case of strictly deterministic and
stochastic leadtimes.

In the deterministic LT-case (O'A; = 0), with X; being product specific leadtimes that
are known with certainty, safety stocks ssti for all products i depend on the respective
replenishment period T; analogously to formulation (1) in the single-stage case

sst; = q; . O'd; vn
• with q; = cf>-l (O'j) (8)
Thus expected safety stock holding costs are

H••t; = Ci' vn with c;:= hi' q; . O'd; (9)

Using T; as decision variables in this w.ay the safet)t stock optimization problem in serial
system with deterministic leadtimes (sstopjsd) can be fomulated according to the ap-
proach of Simpson (1958) :
 419

L Ci·.;r:
3
H"t = i=l =} min!

s.t. <1> Tl < '\1

<2> Tl +T2 < Xl + X2

<3> Tl + T2 + TJ = Xl + X2 + XJ
and <4> Ti > 0 (i = 1,2,3)

Restrictions < 1 > and < 2 > mean that replenishment times for safety stock protection
at intermediate levels may not exceed cumulative leadtimes while at the end-item level,
as shown in < 3 >, the sum of replenishment times must be equal to the total leadtime
for the whole process to guarantee immediate satisfaction of customer according the pre-
specified customer service level. The non-negativity conditions are self-evident.

It is easy to show that sstop/sd is a concave minimization problem forcing the opti-
mal solution to be restricted to the extreme points of the linear polyhedron < 1 > to
< 4 >, thereby meaning
i
T;" € {O , L Aj - TLd (10)
j=l
This solution property reduces the set of relevant Ti-combinations by far and can be
used to develop a very efficient dynamic programming solution procedure (see Inderfurth,
1991). It is evident that as result of the optimization the total stock holding costs depend
on ailleadtimes Xi (i = 1, 2, 3) .

In the stochastic LT-case for intermediate stages Xi (i = 1, 2) are replaced by fixed

planned leadtimes ~i - yet known from (3) -

~i = Xi + ki . (J)i (11)
only having impact on the restrictions of the optimization problem. For the final custo-
merside stage a flexible leadtime ~J has to be considered lea,9.ing to a TJ-dependent safety
stock according to the stochastic leadtime condition in (6)

sst3 = qJ . J(J'JJ . T3 + 4 . (J'~J (12)

SO safety stock holding costs at the finaJ stage can he expressed as

(13)
420

Using replenishment time T3 in that way means that it has to be identified as expected
covered leadtime which accordingly has to restricted by the expected LT-Iength in the
restriction set.
Thus the appropriate reformulation of the safety stock optimization problem in serial
structures with stochastic leadtimes (sstop/ss) leads to
2
H .. t = ;=1
L: c;' JT; + C3 . '/"'(3 . T3 + 63 => min!

s.t. <1> Tl < ~1

<2> Tl +T2 < ~1 + ~2

<3> Tl+T2+T3 = ~1 + ~2 + ~3

and <4> T; > 0 (i=I,2,3)

Now, sstop/ss also turns out to be a concave minimization problem with the same be-
nefits from the extreme point solution property as problem sstop/sd.

It can be seen that in the general stochastic LT-case safety stock costs depend on both
the expectations and the variances of all single lead times. The same is not completely
true for the expected pipeline stock costs Hp• t in serial systems which in extension of the
single stage considerations in (5) and (1) are given by
2
H p• t = Lh;.J;.(~;+k;.o);) + h3·J3·~3 (14)
;=1
showing dependency on all lead time parameters except the variance of the final-stage
leadtime.

3.2 General multi-stage production systems

In practice of multi-stage production simple serial structure are seldom found. Most
manufacturing processes are of an assembly type whereas many chemical processes and
a lot of distribution processes which often follow production stages have an abo res cent
structure. From that point of view it is important that the safety stock optimization
approach demonstrated for the serial multi-stage case can be extended to divergent pro-
duction systems (see Inderfurth 1991a) as well as to convergent ones (see Inderfurth 1992).
Also mixed multi-stage systems can be dealt with as will be demonstrated for the 3-stage
system depicted in figure 3.
 421

d5

Figure 3: A mixed convergent/divergent production system

In this system at the first stage two kinds of material are delivered from outside suppliers.
In a second stage they are assembled to a component which can be further processed in
a final stage to two specific end-items both facing separate customer demand.

In the deterministic LT-case the safety stock optimization problem for this mixed
multi-stage deterministic case (sstop/md) can be formulated as a certain extension of
the serial model:
5
Hsst L c; . ..;r:
= ;=1 ~ min!

s.t. <1> T1 < .\1

<2> T2 < .\2

<3> T1 + T3 < .\1 + .\3

< 4.a > T1 + T3 +T4 = .\1 + .\3 + .\4
< 5.a > T1 + T3 + T5 .\1 + .\3 + .\5 for .\1 ~ .\2

< 4.b > T2 + T3 + T4 > ).2 + ).3 + ).4

< 5.b > T2 + T3 + T5 > .\2 + .\3 + .\5

and <6> Ti > 0 (i = 1,2, .. ,5)

In the assembly part of the structure (at intermediate stages) the (longest cumulated)
lead times for each product must not be exceeded by (cumulative) replenishment times
as prescribed in < 1 > to < 3 > . Restrictions < 4.a > and < 5.a > say that for both
final-stage products (in the divergent substructure) the maximum of cumulative leadtimes
must equalize the respective total replenishment times while for pathes in the production
structure with lower cumulated leadtimes the inaquality will hold as formulated in < 4.b >
and < 5.b >. It is easy to see that in sstop/md we again find a concave minimization
422

problem with all the nice solution properties mentioned above.

In the stochastic LT-case a straightforward extension of the optimization model is

possible just as in the serial case. The equivalent mixed multi-stage stochastic problem
sstop/ms then turns out to be

H. ot = ~ min!

with c;. "'t;. and OJ( for j = 4,5) as defined in (13)

s.t. < 1 > to < 6 > from sstop/md

with ~i defined in (11) instead of Xi (for i=I,2,3)

It can be stated that even for the most general problem sstop/ms (notice that sstop/md
is a special case of sstop/ms with (7).i = OVi ) the concave minimization property still
holds. This fact allows an easy computation of safety stocks and their respective holding
costs as a function of the distribution parameters of uncertain lead times even for compli-
cated convergent/divergent production structures. Holding costs for work-in-process are
simply computable analogously to (14) by
3 s
= L hi' di · (Xi + k i · (7).i) + "h
L.J)·· d·X
) )· (15)
i=1 j=4

Noticing that customer service levels OJ (j = 4, 5) are incorporated in the cr coefficient of

the H ..t objective function the sstop/ms program gives us a powerful tool - in addition
to the analytic relation in (15) - for evaluating the dependence of cost and service on de-
terministic and stochastic leadtimes and their different reduction opportunities in general
production systems. The investigation of lead time reduction effects now is demonstrated
for a 3-stage mixed system.

4. Leadtime reduction in a mixed convergent/divergent system

As an illustrative example for studying reduction effects we will use the 3-stage conjoined
system shown in figure 3. The specific data are as follows

- for demand: d4 = 60 , ds = 80 , (7di = d;/2 (i = 4,5) , no correlation,

- for holding costs: hi = h2 = 3 , h3 = 3 , h4 = hs = 4

- forleadtime: X2 =1, XI='X3=X4=X~=2, (7).i=X;/2(i=I,2, .. . ,5)

- for safety factors/service levels: ki = 1 ~ 1r = 84% (i = 1,2,3)

qj = 2 ~ OJ = 97.7% (j = 1,2, .. . , 5)
 423

With this information the optimal policy, computed from the sstop/ms model;' turns
out to recommend safety stocks only for the two final stage products (4 and 5) and for
the first- stage product with the larger leadtime (1). This policy is connected with safety
stock holding costs Hut = 1,937 while pipeline stock costs according to (15) sum up to
H p• i = 3,010.
Now we will consider what happens if LT-variability (7). or LT- length X is reduced at
the raw material (products 1/2), component (product 3) and end-item level (products
4/5). At each stage leadtime is reduced by the same percentage for each product respec-
tively. LT -length reduction is studied for the case of deterministic leadtime (== 100%
(7).-reduction).

..'.Iy alock coala

• p P.nn. alock coala

lolal coal. 10lal coala

8000 ,ooo ~~~~----------------,

5000
3000
'000

3000 2000

2000
1000
1000

0 o
0 10 20 30 40 50 80 70 &0 110 100 o 10 20 30 40 50 80 70 &0 110 100
It reduction In .. LT reduction In ..

Fig. 4: LT-variability reduction Fig. 5: LT-length reduction

at material level at material level

From figures 4 and 5 for the material level we see that a reduction of LT-variability
generates a slight linear drop of both safety and pipeline stock costs whereas a LT-length
reduction gives chance for a larger decrease of costs. In principle, the same holds for
leadtime reduction at the component level. Leadtime reduction effects at the final stage
have a bit different form as figures 6 and 7 show.
424

IOt~ c~la
eooo ~~~I~ ~~I~a__________________~
~~C~

3000

2000

1000

o 10 20 30 40 50 110 70 80 110 100 '" 20 30 40 50 eo 70 eo 110 "'0

It reduction In " LT reduction In "

Fig. 6: LT-variability reduction Fig. 7: LT-length reduction

at end-item level at end-item level

Reducing LT-variability is connected with constant pipeline stocks and degressively de-
clining safety stock costs while a LT-length reduction leads to a considerable fall of both
costs, for safety stocks even in a clear progressive manner.
In investigating lead time effects at different production levels we more or less found a
confirmation of the analytical results for the respective single stage cases. However, this
cannot generally be expected since the total safety stock distribution in the system -
according to the optimization procedure - can change when leadtimes are reduced at
single stages or for single products. A further interesting question is, how total costs will
react on a simultaneous leadtime reduction all over the system. Figures 8 and 9 give the
answer.

IOIIll eoa,a (In 1000) totlll coata lin 1000)

5 ~--~~~~------ ______-,
8000 ~~~~~~------------,

5000 4 t---------.--------------~

4000
3

3000

2000

1000

o
10 20 30 40 50 eo 70 80 110 100 o 10 20 30 40 50 eo 70 eo 110 »0
It reduction In " LT reduction In "

Fig. 8: Total LT-variability reduction Fig. 9: Total LT-length reduction

 425

Figure 8 makes evident that for total LT- variability reduction the final-stage degtessive
cost character dominates. In figure 9 we observe a linear decrease of work-in-process
combined with a serious progressive reduction of safety stock costs if general actions
for diminishing leadtime are accomplished. In the extreme case of totally avoiding any
leadtime inventory holding costs, of course, can be reduced to zero.
Finally the simultaneous effects of leadtime and LT -reduction on cost and service level is
visualized in figure 10. The curves in figure 10 represent the trade-off between cost and
service level for given leadtime and stem from a parametric solution (with respect to cu-
stomer service levels 0'4 and as) of the safety stock optimization problem. We see that, in
general, an improvement of customer service only can be performed at progressively rising
costs as we yet know from single-stage inventory theory. A general leadtime reduction
now leads to a more favorable trade-off curve, again being more considerable for LT-
length than for LT- variability reduction. These trade-off considerations, of course, can
be worked out for any leadtime reduction activity at a specific process in the multi- stages
production system.

coat (In 1000)

5.5r------------,

4.5
- - atanderel LT-data
- - 30.. -LT -variab.red.
- - 30"-LT-langth red.

2.5 L-_---''--_-"-_ _--'-_ _-'

80 85 90 95 100
service level in %

Fig. 10: Cost and service for different kinds of LT-reduction

5. General conclusions

The analytical and numerical results presented in this paper give insight into structural
effects of leadtime reduction efforts, especially demonstrating the important cost savings
that can be gained from LT-length reduction.

Additionally a powerful numerical approach is offered that can be used for evaluating
LT-reduction benefits with respect to cost and service for concrete multi-stage production
systems taking into consideration a lead time depending optimal safety stock policy. In
this context safety stock optimization is an important module because inappropriate sa-
fety stock distributions may cause severe drawbacks in costs. Using a buffer policy with
426

this context safety stock optimization is an important module because inappropriate sa-
fety stock distributions may cause severe drawbacks in costs. Using a buffer policy with
holding safety stocks for each single product in the standard example of section 4, for
instance, leads to a 13% increase in safety stock costs.

For a general valuation procedure of leadtime reduction projects the cost/service trade-off
function plotted in figure 10 can be a very useful tool. After specification of the desired
customer service that shall be achieved with leadtime reduction activities the respective
cost can be read and incorporated in a financial evalutation procedure. Furtheron, critical
service levels connected with stock holding cost improvements which just outweigh the
cost of capital invested in a reduction project could be determined. By these means a
high degree of transparency can be given to the total valuation process.

The valuation tool for leadtime reduction described above can - according to the app-
licability of the safety stock optimization procedure - be used for a wide class of produc-
tion systems. It can even be extended to problems incorporating deterministic lead time
optimization decisions along with the safety stock problem (see Inderfurth, 1991b). On
the other hand the procedure presented here is restricted in its accuracy by some as-
sumptions for the safety stock determination module which may be critical in specific
actual situations. These may be suppositions of lot-for-Iot ordering, of total reliability of
internal supply processes, and of normal distributed lead time demands. Nevertheless, it
has to be noticed that the stock optimization results are used as an information within
a superior investment valuation model where a limited precision of data usually is met.
Thus, despite of these limitations the approach presented may be a useful tool in many
situations.

References
Clark, A.J . and H. Scarf (1960), "Optimal Policies for a Multi-Echelon Inventory Pro-
blem", Management Science 6, 475-490.
Corbey, M. (1991), "Measurable Economic Consequences of Investments in Flexible Ca-
pacity" , in Production Economics : Issues and Challangers for the 90 's, Grubbstrom,
R.W., Hinterhuber, H.H. and J. Lundquist (eds.), Elsevier, Amsterdam, 47-57.
Fine, C.H. and R.M. Freund (1990), "Optimal Investment. in Product-flexible Manufac-
turing Capacity", Management Science 36, 449-466.

Graves, S.C. (1988), "Safety Stocks in Manufacturing Systems", Journal of Manufactu-

ring and Operations Management, 1,67-101.

Inderfurth, K. (1991a), "Safety Stock Optimization in Multi-Stage Inventory Systems",

International Journal of Production Economics, 24, 103-113.

Inderfurth, K. (1991b), "Combined Optimization of Safety Stocks and Processing Lead

Times in Multi-Stage Production Systems", in Modern Production Concepts: Theory
and Applications, Fandel, G. und G. Ziipfel (eds.), Springer, Berlin, 291-301.
 427

Inderfurth, K. (1992), Safety Stock Allocation in General Multi-Stage Production 54j6tems

with Uncertainty in Demands and Leadtimes, Discussion Paper No. 250, Fakultat
fUr Wirtschaftswissenschaften, Universitat Bielefeld.

Langenhoff, L.J.G. and W.H.M. Zijm (1990), "An Analytical Theory of Multi-Echelon
Production/Distribution Systems", Statistica Neerlandica 44, 149-174.

Moerman, P.A. (1988), "Economic Evaluation of Investments in New Production Tech-

nologies", Engineering Costs and Production Economics, 13, 241-262.

Rosling, K. (1989), "Optimal Inventory Policies for Assembly Systems under Random
Demands", Operations Research, 37, 565-579.

Silver, E.A. and R. Peterson (1985), Decision Systems for Inventory Management and
Production Planning, 2nd Ed., Wiley and Sons, New York.

Simpson, K.F. (1958), "In-Process Inventories", Operations Research, 6, 863-873.

Wouters, M.J.F. (1991), "Economic Evaluation of Leadtime Reduction", International

Journal of Production Economics, 22, 111-120.
 Nervousness and Reorder Policies
in Rolling Horizon Environments
by

Thomas Jensen
Faculty of Economics, University of Bielefeld
D-4800 Bielefeld 1, Germany

Abstract

In this paper measures for characterizing planning stability in material-coor-

dination systems are described and applied to analyze the stability performance of
(s , nQ)- and (s,S)-policies for a single-stage system simulated in a rolling horizon
environment. Introducing a stabilization parameter into reorder-point lot-sizing
policies a method to improve planning stability is proposed and compared to the
efficiency of the stabilization method when freezing a part of each planning cycle.

1 Introduction

Frequent replanning and rescheduling of orders when updating plans rolling forward the
forecast horizon in material coordination systems has been well characterized in literature
because of the negative impacts on system performance (see for example STEELE [16] who
comments on the effects of this so-called "system-nervousness" in MRP-systems).

Besides practical problems to handle and implement extensive replanning and reschedul-
ing messages due to variations of external or internal stochastic variables of the production
system (e.g. demand, machine reliability) the planning process gets more complicated and
mainly capacity planning decisions become more difficult because of permanently altered
information. In multistage production systems even changed information near the end of
the forecast horizon at the end-item level can result in precipitate plan revisions for up-
stream stages in the first periods of each planning cycle and, passing stochastic impacts
in a distorted and probably intensified way to upstream stages, planning revisions by
themselves become stochastic inputs for preceding stages and hierarchical subordinated
planning procedures. As a result throughput times and inventories may increase. At
 429

the same time service-levels may decrease due to wrong priorities given by the planning
system.

In order to accept not all replanned data on lower planning levels (see for example
PENLESKY ET AL. [9]) or to restrict replanning activities, different methods have been
established (see MINIFIEjDAVls [8] for a classification of dampening procedures). Be-
cause of the important role of lot-sizing decisions in aggregating stochastic impacts of
several planning periods and because of their sensitivity to data changes, early attempts
to deal with the problem of nervousness focus on the performance of various deterministic
dynamic lot-sizing models in rolling horizon environments (e.g. CARLSON, JUCKER and
KROPP [3, 6, 7] who extended the dynamic deterministic lot-sizing problem by incorpo-
rating change costs for undesired replanning of setups) .

In addition to this lot-sizing oriented approaches various simulation studies have been
performed which are primarily concerned with the analysis of potential impacts on plan-
ning stability and the performance of various stabilization strategies and their interde-
pendence. E.g. , SRIDHARAN with several co-authors [14, 11, 12, 13, 15] analyzed different
methods of freezing the master production schedule in combination with using safety
stocks whereas BLACKBURN ET AL. [1, 2] studied the performance of different stabi-
lization strategies (e.g., the implementation of frozen zones, extension of the forecast
horizon beyond the planning horizon, holding safety stocks on the end item level) for
elementary multistage environments. YANO and CARLSON [4, 18, 17] analyzed the in-
teractions between safety stocks and combinations of alternative replanning policies for a
single-product, two-stage assembly system.

In this paper key factors for the stability performance of (s, nQ)- and (s, S)-lot-sizing
policies in an elementary single stage rolling horizon environment are analyzed using two
suggested measures to characterize different types of planning stability.

Introducing a tolerance factor for the studied reorder-point lot-sizing policies a sta-
bilization strategy is proposed and the trade-off between service-level performance, total
costs and the achieved planning stability is shown and compared to the stabilization
method of freezing a part of each planning cycle. The paper concludes with suggestions
for further research.

2 Measures for planning stability

In previous works different concepts to define planning stability have been suggested (see
for example Ho [5]). In this paper two standardized measures are proposed to define setup-
430

oriented- (only taking into account the number of new planned or canceled setups when
considering succeeding planning cycles) as well as quantity-oriented planning stability
(looking for changed lot size-quantities between succeeding planning cycles).

To define the proposed measures, several design parameters of rolling horizon environ-
ments as depicted in figure 1, showing two succeeding planning cycles (k -1) and k (with
a length of N periods) are used.

N-P
k-l

Figure 1: succeeding plan generations

At the beginning of each planning cycle (define M/ to be the starting period of plan-
ning cycle 1), the planning horizon is rolled forward one period, updating period two up
to period N of the preceding plan. Taking into account the stabilization parameter P
(1 ~ P ~ N) which determines the length of a frozen zone during which plan revisions
are not allowed, there are (P - 1) periods frozen from the preceding planning cycle and,
as can be seen in figure 1, (N - P) periods have to be considered when adding up plan
changes referring to the preceding planning cycle. For further use define L to be the
leadtime and Q~ denotes the lot-size quantity for period t, planned in the kth planning
cycle.

Measuring setup oriented planning stability

The proposed setup-oriented planning stability measure 'Irs is given by equation 1.

~s
'Ir. := 1- - - - (1)
~SMa",

Defining ~S as given by equation 2 to be the average number of changed setups per

planning cycle (for a total of ]( planning cycles following an initial planning cycle with
 431

k = 0) and using the maximal number of changed setups per planning cycle D.SMax =
N - L - 1 as upper bound for D.s (incorporating the leadtime L takes into account that
demand has to be covered by lot-sizes planned L periods ahead by the latest), 11'. is
normalized between values of 1 and 0 referring to situations of extreme setup planning
stability and extreme setup planning instability and can be interpreted as the average
percentage of planning periods without setup changes per planning cycle.

(2)

11'. can be applied directly to characterize the planning stability of multi-stage pro-
duction systems by aggregating 11'. stage-by-stage thereby using an appropriate weighting
of values for different stages (e.g. referring to different capacities at various stages). An
extension in order to differentiate between new and canceled setups as well as weighting
plan changes relating to their distance to the forecast horizon can easily be incorporated.

Measuring quantity oriented planning stability

To formulate a comparable normalized measure for quantity oriented planning stability

is a more complicated task because the maximum quantity D.qMax which can be changed
between two succeeding planning cycles depends on characteristics of the demand distri-
bution. Given D Max to be the maximum reasonable demand per period, D.qMax can be
calculated by D.qMax = (2(N - L) -l)D Max , based on the assumption that the maximum
quantity changed per planning cycle is given when the maximum demand per planning
cycle is aggregated to one lot-size which is then replanned to another period in the next
planning cycle. The proposed quantity based stability measure 11'q is defined by:

(3)

where D.q is given by

(4)

z>O
16 (z) is defined as the Kronecker-symbol 6 (z) = { 0
else
432

The quantity stability measure 11'q can be interpreted as the average percentage of
quantity per cycle which is taken to the new plan without changes related to the maximum
possible amount of changes per planning cycle. Instead of D Max = const., time-variant
values for DMax can be used in order to consider instationary demand situations.

As for the setup based measure, weighting of plan deviations referring to their distance
to the planning date is possible. Again -applying an appropriate weighting of the values
of 11'q for different stages- 11'q can be used to characterize planning stability in multi-
stage production systems. For component stages of multi-stage production systems D Max
depends on the maximum reasonable demand per period for all end-item level products
the component is needed for.

This quantity-based planning stability measure can be compared with a measure for
planning instability proposed by SRIDHARAN ET AL . (see [11]). Dividing the average
quantity changed per setup by an average lot-size given by the economic order quantity
EOQ, their measure can be interpreted as the percentage of an average lot-size which
is changed by replanning activities. On the one hand providing an illustrating measure,
their concept is not standardized in the sense of being restricted to a certain range of
values, bounded by a maximal and a minimal value of planning instability. Instead, their
measure for planning instability depends on the simplifying assumptions of the EOQ (e.g.
no capacity restrictions) as well as on cost parameters.

3 Planning stability of (8, nQ)- and (8, B)-lot-sizing

policies in rolling horizon environments

In this section results are presented, applying the suggested measures to characterize
the planning stability of (s, nQ)- and (s, S)-lot-sizing policies operating in an elementary
single-product, single-stage environment.

To analyze the impact of demand characteristics and combinations of the reorder-point

s with reorder-levels S and lot-sizes Q for the different policies, planning lot-sizes in order
to satisfy forecasted demand is simulated for various scenarios in a rolling horizon envi-
ronment. In the basic scenario the design parameters of the rolling horizon environment
are given with N = 25 and P = 1 (no frozen zone allowed). The demand is assumed to
be stationary distributed according to a beta-distribution with different parameters. The
leadtime L is set to a length of 1 period. The inventory cost parameters are set to holding
costs of h = 0.01 and setup costs of R = 0.1.
 433

At the beginning of each planning cycle the inventory position is actualized taking
into account incoming orders and the realized demand for the first planning period and
-assuming a constant demand forecast equal to the expected demand per period- lot-sizes
for the whole planning cycle are determined. Demand that can not be satisfied in the first
period is back-ordered. In addition to the stability measures 7r. and tr q the achieved j3-
service-level as well as total costs C (consisting of setup and holding costs) are calculated
during simulation.

In a first series of simulations the impact of different order quantities and several
characteristics of the demand distribution has been analyzed. Setting the reorder-point s
to the expected demand during leadtime plus review period (s = (L + l)JL) the parameter
Q is varied from 1 up to 40 units (for the (s, S)-policy the order-up-to-point S is calculated
by adding an amount of Q to the reorder-point s).

In figure 2 the effects on the suggested stability measures 7r. and tr q are shown for three
different variances of the demand distribution for the (s, nQ)-policy. Fixing the expected
demand per period to 5 units this results in a coefficient of variation varying from 0.21
up to 0.46 with the demand distribution covering a range of 1 up to 9 units.

Considering the setup-oriented planning stability measure 7r. it can be seen that start-
ing with a high setup stability given small values of Q so that in each period an order
has to be planned, the setup stability dramatically decreases for Q being slightly greater
than the expected demand per period. As shown in the first part of figure 3, further
increasing of Q results in increasing 7r s . This can be explained by the reduced number of
setups per planning cycle when increasing Q. In contrast to the development of 7r. for
values of Q being greater than the expected demand per period (JL), the quantity-oriented
planning stability is independent of the lot-sizing policy parameter Q. The high stability
achieved for Q being equal to one unit and equal to the expected demand per period
can be explained by the fact, that for this parameter-combination the lot-sizing policy is
completely coordinated with the constant demand forecast.

Considering the impact of the variance of the demand it can be seen, that (for Q > JL)
- as expected - with increasing u 2 (increasing coefficient of variation) setup- as well as
quantity-oriented planning stability decreases.
434

1)-". ,,=5
----
(s. nO) = «L+ nO)

"",12

0.
--
"...2J
"...5,3

.t 0.'

Ii 0.1

0.1

0.1
10 \I 20 :III :10 311 ..,
Q

(8, nQ}-policy
pl&lllliug-atability measures: 11' .. 11',

(s. nO) = ({l+I)·". nO) ,,=5

o.o- t - - - - - - - - - - - - - - - - - l

0'1~~~~~,O~~II~~2O~~:m~~:IO~~~3II~~..,
Q

Figure 2: 'Ir. and 'Irq for (s,nQ)-policy, varying Q for different variances of the demand
distribution

The given results are confirmed when varying the expected demand per period for
a given variance of demand (see figure 3). The typical patterns in the development
of both stability measures are preserved but shifted according to the different demand
expectations. For values of Q being slightly greater than the constant demand forecast
1', the quantity stability measure is unaffected by variations of Q whereas 'Ir. increases
'Irq

with decreasing coefficient of variation (by increasing the expected demand per period) as
described above. Considering 'Irq it is shown that for Q being a multiple of the expected
demand per period a very high quantity-oriented planning stability is achieved.
 435

Concerning the impact of the reorder-point s on stability performance further simula-

tion results show that both stability measures do not react to an increase of the reorder-
point which means that available inventory is not used to prevent replanning.

=«l+ I )' .,.

--
(s. nO) nO) .,.'=2 .3

.....
". 5
" • 10

r:Il
0,'
o· 15

- --
U
10 I. ac
Q
. XI . ..,

(&, nQ)·policy
planning-stability meuures: '1'"11 'l'"t

(s. nO) = «l+ I ). ". nO) ""=2.3

, ~

0,

~
•• 10
0·15

fu
f 0,7

0.&1

0._
.0
"
ac
Q
.. XI . ..
Figure 3: 'Irs and 'Irq for (s, nQ)-policy, varying Q for different demand expectations

Comparing the planning stability performance of the (s, S)- and the (s, nQ)-policy,
figure 4 shows distinct cyclic patterns in the development of both stability measures for
the (s, S)-policy dependent on the value of the parameter Q (again the reorder-point s
is fixed to the expected demand during leadtime plus review period) whereas the general
tendency for both stability measures is similar to the development of 'Irs and 'Irq described
above for the (s, nQ)-policy.
436

The oscillating patterns in the development of 1r. and 1rq for the (s, S)-policy can be
explained as follows: for Q being an integer multiple of the expected demand per period
the order-quantities are adjusted to the stationary demand forecast. In this case due to the
regenerative characteristic of the (s, S)-policy slight positive deviations between realized
and forecasted demand cause an updating of planned setups for the whole planning cycle
whereas for other combinations of sand S (with S = s + Q) deviations of the realized
demand are partially absorbed by the first planned setup so that replanning is probably
not necessary. Further experiments show that this cyclic patterns diminish with increasing
the variance of the demand distribution.

(s. nO) ... (s.S) V=5 ~·=2 . 3

0.'

fU
t ~, ~-------

0.'

UI~~~~,o~~,a~~a~~a~~~~~z~~~
Q

($, S) vs. (8, nQ )-po~cy

planniDS,otability _lI8ur..: ......,

.s. (s. 5) 1'=5 ~·=2.3

~.

Figure 4: Comparing 1r. and 1rq for the (s, nQ)- and (s, S)-policy, varying Q
 437

In figure 5 in addition to 'Irs and 'Irq the ,B-service-Ievel and total cost,s C are depicted,
varying again Q. It can be seen that in order to obtain a high planning stability by
using low order quantities at the same time high setup-costs and a low service-level (or
additional holding costs to guarantee a higher service-level) have to be accepted. Using
large order quantities to get high setup-stability as well as a high service-level on the
other hand causes substantially increasing holding costs. Due to the cyclic patterns in
the development of both types of planning stability for the (s, S)-policy, adjusting the
policy-parameters for low values of Q enables to achieve substantially increased planning
stability levels without reducing service-level or increasing total-costs at the same time.

(s. nO)=«l+ I )' ". nO)

~·~~~~'O~~,~.~~~~~a~~~~~~~~~
a

(8,S) vs. (s.nQ)-policy

planning·mbility, fJ aervice-I~vel. cos...

(s. S)=«l+1 )' " •• +0) JJ"S ... ·=2.3

u'.............:-r-rr-r;';IO:""""'''''7.',• .,...,..,.'''':~:!:''''''~a!:'T''''.,...,..,~~~
.....,..,...,...,J~
Q

Figure 5: Comparing stability measures, ,B-service-Ievel and total costs for the (s, nQ)-
and (s, S)-policy, varying Q for a given demand distribution
438

4 Improving the planning stability of reorder-point

lot-sizing policies

In this section an extension of the analyzed reorder-point lot-sizing policies is suggested

by introducing a stabilization parameter a (0 :$ a). This parameter specifies the amount
of available inventory which is used to prevent new planned setups and which determines
the extent the inventory position is increased in order to prevent cancelling planned setups
from preceding planning cycles.

case inventory-position Y~ Q~-l under-/overshoot of s Q~

(s,nQ) (s, S)
(1) Y~ < s =0 y~~(I-a)s =0 =0
(2) y~ ~ s >0 y~«I+a)s Q (S-y;)

Table 1: Modified reorder-point lot-sizing policies

Using this modified policies, planned orders for each period t do not only depend
on the actual inventory position y~ for this period when determining the kth plan, but
also on the lot-size determined in the last planning cycle (k - 1) for this period. If the
reorder-point s exceeds the actual inventory position though according to the unmodified
lot-sizing policy an order has to be planned, this order is suppressed in order to create a
stabilization effect if the undershoot of the reorder-point s does not exceed a percent of
the given value for s (case (1) in table 1). In the same wayan order is planned according
to the considered reorder-point policy if the actual inventory position exceeds the reorder-
point but the overshoot of s is within the range of as (case (2) in table 1). For a = 0
this policy leads to the unmodified lot-sizing policy. By variation of the parameter a the
trade-off between planning stability, costs and the achieved service-level can be studied.

It can also be seen that using these extended policies, planning stability is not only
affected by the lot-sizing policy parameter Q but depends on the given value for the
parameter s too. For fixed a (a> 0) setup as well as quantity-oriented planning stability
increase with increasing s.

Possible extensions of this stabilization strategy are to use different stabilization pa-
rameters for the two cases Yt < sand Yt ~ s and to define time-dependent stabilization
parameter( s) for each planning period according to the distance to the end of the planning
cycle.
 439

(s. nO: 1 0)
0,.,, . . - - - - - ' - - - - - - -- - - -------"]

0,1

<) 0.-

f 1" '"-.. . . . . .<! ~:.:-::::'~~-=..=

0 . . -~.~-~.~=-=.
.• . . . . . =·-:~·\-.. 1 c t;IWl,

f ~I
~

stabilization strategy a
plaDDiDg·.t.ability, P service-level. costs

(s. S:,+10) 1':5 ... ·:2.3

--.
0,'
..-9.
o.l
~

............ -<-
n Q
0.'

i
<) ~

C t;IWll

f ~I
.J
~ ....
....
~.
0 0.1 0.2 0.3 0.' 0.5 0.. 0.7 0.. 0..

Figure 6: Comparing stability measures, f1-service-level and total costs for the (s, nQ)-
and (s, S)-policy, varying the stabilization-parameter a

The results when comparing both extended lot-sizing policies are shown in figure 6.
In this figure relative deviations of total costs, f1-service-level and the suggested stability
measures 'Irs and 'Irq related to the values received for a = 0 (which results in the unmod-
ified lot-sizing policy) are depicted for a fixed lot-size of 10 units and a given demand
distribution.

For small values of a both stability measures substantially increase whereas the f1-
service-level slightly decreases. Further increasing of a however does not affect planning
stability while the f1-service-level continues to decrease. Comparing total costs and f1
440

service-level for the (s, S)- and the (s, nQ)-policy it can be seen that the {3- service-level
for the (s, S)-policy is reduced even stronger than for the (s, nQ)-policy when increasing a.
At the same time total costs for the (s, S)-policy decrease compared to slightly increasing
total costs for the (s, nQ)-policy. This fact is due to the size of the planned lot-sizes for
the case (2) depicted in table 1, which are restricted for the (s, S)-policy to the range of
(S - y;) whereas for the (s, nQ)-policy each planned lot-size has a magnitude of at least
Q units.

(s. S) : ,:0.99 .,.':2 .3

------------------,
~:5
o.tI$-

",
C toIWll

.... .-.. ~- ...

o..2!{1.2!!1 Q.W' .35 D.. ., •

o/m

stabilization strategy a VB. P

planniD,-stabili~y. ,B-rervice-leveJ, costs

(s. s) : ,=0.99 ~=5 .,.'=2.3

Otl$-

0-9.-

IJ

'"
I
~
"'' '
;.

--.--- .-.-'-

00.""
11'311
Ptm
''''' . 11118
2QlU,

Figure 7: Comparing stability measures, {3-service-Ievel and total costs for an approximate
optimal (s, S)-policy using stability strategies a and P

Compared to the stabilization method to freeze a part of each planning cycle the pro-
 441

posed lot-size oriented stabilization strategy is more flexible and can create an stabilization-
effect in every period, whereas the stabilization method to freeze a part of each planning
cycle has an stabilization effect only in the first P periods and results in an "all-or-
nothing" stabilization without considering the actual inventory position. In order to verify
this judgment both stabilization methods have been studied in further simulation experi-
ments. For this simulations approximate optimal policy-parameters s· and S· = s· + Q.
have been calculated using approximations developed by SCHNEIDER/RINGUEST (see
[1OJ). In order to maintain the achieved ,B-service-level of 99% when increasing the sta-
bilization parameters a or P, a modified reorder-point s = ms· is determined using the
optimal value s* multiplied by a parameter m (m ~ 1). Based on this reorder-point the
reorder-level S is given by S = s + Q*.

The results for the (s, S)-policy varying the stabilization parameters a (0 :::; a :::; 0.5)
and P (1 :::; P :::; 20) are depicted in figure 7 showing the relative deviations relating to
the unmodified reorder-policy which is given for a = 0 and m = 1. In order to achieve
the stabilization effect given by the lot size-oriented stabilization method for a = 0.2 the
length of the frozen zone has to be greater than 20 periods (for a planning cycle length of
25 periods). At the same time total costs increase substantially so that the stabilization
method a clearly dominates the stabilization method of using a frozen zone.

5 Further research

For further research the analyzed environment has to be extended for several aspects.
Assuming dynamic demand-forecasts as well as instationary demand situations, further
stochastic impacts can be introduced into the model. In addition to this the effect of
forecasting models on planning stability should be analyzed.

In a next step the single-stage analysis has to be extended to multi-stage systems

allowing for a more comprehensive analysis of the trade off between reducing planning
instability and service-level and cost performance of the system under consideration.

Finally, modeling the interactions between control parameters of lot-sizing policies, lot-
size oriented stabilization strategies and the proposed planning stability measures might
lead to extensions of lot-sizing models which follow cost optimization objectives and take
into account restrictions of prespecified service-levels as well as given planning stability
requirements.
442

References

[I] BLACKBURN, J.D.; KROPP, D.H .; MILLEN, R .A.; A Comparison of Strategies to

Dampen Nervousness in MRP Systems, Management Science 32, 1986,413-429.

[2] BLACKBURN, J.D .; KROPP, D.H.; MILLEN , R.A.; Alternative Approaches to

Schedule Instability: A Comparative Analysis, Int. Journal of Production Research
25,1987, 1739-1749.

[3] CARLSON, R.C .; JUCKER, J .V .; KROPP, H.; Less Nervous MRP Systems: A
Dynamic Economic Lot-Sizing Approach, Management Science 25, 1979, 754-761.

[4] CARLSON, R.C.; YANO, C .A.; Safety Stocks in MRP-Systems with Emergency
Setups for Components, Management Science 32, 1986,403-412.

[5] Ho, C.; A Examination of Dampening Procedures Considering Rescheduling Costs,

OMEGA 20, 1992, S. 257-269.

[6] KROPP, D.H.; CARLSON, R.C.; A Lot-sizing Algorithm for Reducing Nervousness
in MRP Systems, Management Science 30, 1984,240-244.

[7] KROPP, D .H.; CARLSON, R.C .; JUCKER, J .V.; Heuristic Lot-sizing Approaches
for Dealing with MRP System Nervousness, Decision Sciences 14, 1983, 156-169.

[8] MINIFIE, J.R.; DAVIS, R.A .; Interaction Effects on MRP Nervousness, Int. Journal
of Production Research 28, 1990, 173-183.

[9] PENLESKY, R.J.; BERRY, W.L .; WEMMERLOV, U.; Open Order Due Date Main-
tenance in MRP Systems, Management Science 35, 1989,571-584.

[10] SCHNEIDER, H.; RINGUEST, J .L.; Power Approximation for Computing (s,S) Poli-
cies Using Service Level, Management Science36 , 1990, S. 822-834.

[11] SRIDHARAN, S.V.; BERRY, W .L.; UDAYABHANU, V.; Measuring Master Produc-
tion Schedule Stability under Rolling Planning Horizons, Decision Sciences 19, 1988,
147-166.

[12] SRIDHARAN, V .; BERRY, W.L.; Freezing the Master Production Schedule under
Demand Uncertainty, Decision Sciences 21, 1990, 97-120.

[13] SRIDHARAN, V.; BERRY, W.L.; Master Production Scheduling Make-to-stock Prod-
ucts: A Framework Analysis, Int. Journal of Production Research 28, 1990,541-558.
 443

[14] SRIDHARAN, V.; BERRY, W .L.; UDAYABHANU, V.; Freezing the Master Produc-
tion Schedule under Rolling Planning Horizons, Management Science 33, 1987, 1137-
1149.

[15] SRIDHARAN, V.; LAFoRGE, R.L.; An Analysis of Alternative Policies to Achieve

Schedule Stability, Journal of Manufacturing and Op. Management 3, 1990, 53-73.

[16] STEELE, D.C.; The Nervous MRP System: How to do Battle? Production and
Inventory Management 16, 1976,83-89.

[17] YANO, C.A.; CARLSON, R .C.; Interaction between Frequency of Rescheduling and
the Role of Safety Stock in Material Requirements Planning Systems, Int. Journal
of Production Research 25, 1987,221-232.

[18] YANO, C.A.; CARLSON, R.C.; Safety Stocks for Assembly Systems with Fixed
Production Intervals, Journal of Manufacturing and Op. Management 1, 1988, 182-
201.
 SOME VALID CONSTRAINTS FOR THE CAPACITATED

ASSEMBLY LINE WTSIZING PROBLEM

by

Heinz D. Mathes
Lehrstuhl ffir Produktionswirtschaft
Rechts- und Wirtschaftswissenschaftliche Fakultat
Universitat Bayreuth
Postfach 10 12 51, D-8580 Bayreuth

I. Introduction

In modelling general multifacility, capacitated lotsizing problems structures of the single

end-item (single product) assembly line type quite frequently emerge. On higher aggregated
levels of production planning this may be the case, because assembly line structures in many
instances provide sufficiently close approximations of actually more general production
environments. On more detailed levels of production planning (material requirements
planning) single product assembly line structures can be found mainly for two reasons: First,
because the lotsizing problem under consideration simply happens to exhibit exactly this
structure. Second, because the general structure of numerous MRP-problems contains
substructures, which can be formulated as single product assembly line problems. It would
seem therefore rather promising to examine this type of problem in greater detail since many
actual applications could be expected to benefit from improved solution procedures.
Though single product assembly line structures appear to represent comparatively simple
production environments finding an optimal sequence of lots and inventories can result in a
formidable task even when the product structure (number of subassemblies, components and
materials) is only of medium size and capacities are unrestricted (see for instance
AFENTAKIS et al., 1984, CROWSTON et al., 1973, 1972). Solution procedures in this
 445

field in general rely heavily on certain properties of optimal solutions, which are used to
construct algorithms, wich are more efficient than the direct approaches such as dynamic
programming or mixed integer programming. Unfortunately the main properties hold only
in the unrestricted case. Therefore due to the mathematical difficulties imposed by capacity
constraints solution procedures for capacitated problems are dominated by heuristic methods.
These methods are partly based on results of optimisation theory (see for instance
ZAHORIK, 1984, THIZY et al., 1985) or apply something like common sense rules to
accelerate the solution procedure (BLACKBURN et al., 1984).

The approach put forward in this paper follows a different rationale. First the lotsizing
problem is presented in the well known mixed integer programming formulation (e.g.
STEINBERG et al., 1980) tractable by any commercial integer programming code. Given
this formulation a preprocessing phase (see NEMHAUSER et al., 1988) joins before the
numerical solution is performed. The objective of preprocessing can be thought of as
accelerating the speed of the algorithm applied by improving or simplifying the original
formulation of the problem. This can be achieved by measures such as tightening bounds on
variables, fixing variables or adding inequalities on logical grounds thus reducing the set of
feasibel solutions to be searched and possibly truncating the optimization procedure. The
results of preprocessing here are based on conditions holding for feasible or optimal solutions
of the lotsizing problem, which can be transformed into sets of valid constraints on the mixed
integer formulation. These additional constraints reformulating the problem bear some
resemblance to what is called valid inequalities. In integer optimization valid inequalities are
added to approximate the convex hull of solutions (see NEMHAUSER et al., 1988). Since
the inequalities employed to reformulate the present problem do not suffice to describe the
convex hull completely, the the following the term valid constraints is preferred (for an
example of a complete approximation for a problem of WAGNERlWHmN type see
BARANY et al., 1984).

In view of the variety of optimization procedures for the uncapacitated case one could argue
that there may exist more sophisticated algorithms to solve the capacitated assembly line
problem, i.e. algorithms, which are superior in efficiency, because they are better adjusted
to the special structure of the problem under consideration. Yet the method put forward has
been chosen, because it seems to offer some advantages over more specialized procedures:
First, the reformulated problem can be solved employing generally available software for
mixed integer programming. Second, it can be easily adapted to changes in the production
446

environment (e.g. extension to additional types of constraints). Third, the reformulated

problem is based on linear programming and can therefore without major modifications form
part of more complex models of production and inventory control (such as MRP systems).
Fourth, recently published studies in multi-item, capacitated assembly line problems (MAES,
1991), based on results of complexity theory, seem to indicate that mathematical program-
ming may prove the ·common denominator· for the development of solution procedures in
the future.

ll. Dermition of the problem

The product structure in an assembly line production process can be represented by a directed
network, G(V,E), where the set of nodes, V, describes the set of items, and E, the set of
directed arcs, defines the technologically feasible assembly operations: An arc (ij) leads from
node i to node j iff item i is a component of item j. The total number of parts is IV I = N.
It is assumed that the nodes have been numbered in topological order by the integers
{1,2, .. . ,N} where the node k corresponding to the end-item has been labeled v(k) = 1. This
labelling scheme is sometimes referred to as technological ordering (THOMPSON, 1965).
A number av is associated with every arc (ij), designating the production coefficient of the
processing operation. Without loss of generality one may define av = 1 for all ij E E (see
BLACKBURN et at., 1984). The unique immediate successor of i is denoted by s(i) and the
set of immediate predecessors by B(i). For single product assembly line structures one
obtaines s(1) = { } and I s(i) I = 1 for all other nodes. In such a directed graph there
exists a unique path, P(i), from every node i E V - {l} to node 1 with P(i) = (i = i.. iz, ... ,
it = 1), ~ E V, satisfying (~, ij + l) E E for all j = 1, ... , k - 1.

In the remainder of this paper the following conditions are assumed to hold:
Demand for the end-item is given for the periods t E {I, ... ,T} with ~ representing the
quantity of demand in period t. No external or independent demand exists for the other items.
Demand must be satisfied from on-hand inventory and no backlogging is permitted.
Item i is manufactured at facility i (node, production stage), which has limited capacity Cj

(e.g. machine hours) remaining constant over all periods. This type of resource is not being
used by any other item. Resources, which are employed in the manufacturing process and
have to cover the demand of more than one item, are considered as non-limiting.
 447

An item i can only be manufactured when all required components k, k E B(i), are available.
To facilitate the notation lead times at all production stages are assumed to be constant,
independent of lotsizes and shorter than the chosen length of period (t, t+ 1]. The necessary
modifications to cover lead times exeeding the length of period are obvious. In the manufac-
turing process set-up cost are incurred at each facility i when a lot of item i is produced. This
set-up cost is independent of the lotsize and equals zero whenever production amounts to
zero. For every produced unit which is not employed to satisfy the demand of the succeeding
facility (or the external demand), in its period of production a constant holding cost is
charged at the end of the period. Production costs per unit at each facility are supposed to
remain constant over all periods and can therefore be omitted. The objective of the lotsizing
problem in assembly systems is to choose a sequence of lots at each facility such as to cover
the external demand and to minimize the sum of set-up and holding costs over the periods
{1,2, ... ,T}. To formulate the capacitated lotsizing problem as a mixed integer program the
following notation will be applied:

<lit: lotsize (quantity of production, production) of item i in period t

Yit: inventory of item i at the end of period t (Yio = 0 for all i)
K;: set-up cost for item i
Ii: holding cost per unit for item i
ci : capacity of facility i
~: external demand for item 1 in period t

lifClit>O
~I = { 0 otherwise

The following formulation can be considered as a capacitated version of the model presented
by AFENTAKIS et aI. (1984):
Problem MIP
N T
min E E (Ki ~I + Ii Yil) (1)
i-I I-I

subject to
t = 1, ... , T (2)

YI,I-l + <ltl - <1.(1),1 - Yit =0 i = 2, ... , N; t = 1, ... , T (3)

448

i = 1, "" N; t = 1, .. , , T (4)

~t € {O,I} i = 1, .. " N; t = 1, .. " T (5)

i = 1, .. " N; t = 1, .. ,' T (6)

(1) is the objective function of the lotsizing problem. Constraint (2) represents the flow
conservation constraint for the end-item. Constraint (3) assures the flow conservation for the
remaining items 2, ... ,N taking into account that the demand for item i is determined by the
lotsize of item k with k = s(i) , Constraint (4) serves a twofold purpose: First, it limits the
lotsize to the amount of available capacity C;. Second, it assures that whenever a positive
amount of capacity is consumed a set-up cost will be incurred.
In this paper, an additional assumption will be used:

I, ~ E Ik for i = 1, .. ,' N (7)

k€B(I)

This assumption can be motivated by considering that inventory holding cost is directly
related to the value of the item produced at facility i. (7) appears therefore quite reasonable
an assumption, since it requires only that the value added by each processing operation is
nonnegative (for a similar argument see for instance LOVE, 1972).
To facilitate the mathematical reasoning, the following shorthand notations will be applied
in the remainder:
t
D:: = E d~ and
~·k
t
C,~: =E c, = (t - k + 1) c, for k s; t
~·k

(i,t) will be used when reference is made to facility i in period t.

m. Valid Constraints for the Problem MIP

1. Preparatory Remarks and Results

It has been outlined in the first section that one of the features of the solution procedure put
forward in this paper is the use of conventional mixed integer programming software for the
numerical solution of the lotsizing problem. One objective of this approach is to facilitate its
 449

integration with other planning procedures as for instance such of MRP-systems. Practically
all commercially available software in this field applies for (mixed) integer problems a
branch and bound routine solving a sequence of LP relaxations of the original problem in
order to obtain the optimal solution. From practical applications it is well-known that LP
based branch and bound algorithms on average spend only a minor fraction of the required
computing time to find a candidate for the optimal solution. The major part of computing
time (frequently more than ninety percent) is required to confirm that no better solution
exists. This effect is due to the number of branches which have to be generated and pursued
in course of this search process for given bounds. To accelerate the branch and bound
routine, one will therefore have to focus on methods to reduce the number of required
branching operations by tightening the LP formulation. This objective can be achieved by
exploiting available information on feasible and optimal solutions of the lotsizing problems
to exclude a-priori certain combinations the set-up variables may take on. This approach rests
on the anticipation that the more combinations of values can be excluded at early stages of
the branch and bound routine, the less nodes have to be generated and investigated thus
probably curtailing the search for an optimal solution.

In the remainder of this section it will be demonstrated how available information on problem
parameters such as demand and cost structures or capacity profiles can be employed to
formulate conditions on feasible and optimal solutions. These conditions are then transformed
into valid constraints on the MIP-problem, which basically either fix certain set-up variables
to predetermined values or establish relations between set-up variables of different production
stages and/or periods.
Before the first constraints can be stated some auxiliary results are needed. The following
represents a generalization of a theorem by BAKER et al. (1978) for the single stage lotsizing
problem. Though there exists no straightforward manner to transform it into a linear
inequality, it provides some mathematical insight into the structure of optimal solutions and
will be used in later proofs.

Proposition 1: For optimal solutions holds Yj,t-I (Cj - 'l;J xjt = 0 for allj,t.

Proof. In contradiction to the assumption let Yj,t-I > 0 and 0 < 'IJt < Cj with period t as first
period for which this condition holds. Period T denotes the last period with positive
production at facility j preceeding period t. To obtain an alternative feasible solution the
following steps must be taken:
450

(i) increase production at (j,t) by h(j,t) = min(Yj,t-h Cj - CW and decrease production at

(j,T) by the same amount.
(ii) decrease inventory at facility j in periods T to t by h(j,t).
(iii) decreasing production at (j,T) by h(j,t) requires a corresponding increase of
inventory for periods T to t at each facility k E B(j).
With all other production and inventory variables remaining unchanged steps (i) - (iii) toget-
her with assumption (7) lead to a cost reduction amounting to at least

(1j - E It> h(j,t) ~ 0 .•

t€B(j)

2. Constraints Derived from Demand and Capacity Prordes

When external demand exceeds the available capacity at facility 1 in some periods as for
instance could be the case with a seasonal demand pattern, it may occur that the accumulated
capacity fails to cover the sum of demands of a number of periods, say periods lt to lz, lz ~

tl • Consequently, to obtain a feasible schedule a positive inventory is required at the

beginning of period th e.g. YI,tI-1 > O. Motivated by this observation a lower bound (yo IJ
for inventory on-hand at facility 1 can be easily established:

yuo = max [0, Dt+1

t+l -
C I,t+l'
t +1 Dt+l
t+l -
C l,t+l'
t +1
---,
DT
t+l -
Cl,t+l'
T ]
(8)
t = 1, _-0' T - 1

Taking (2), (4) and (6) the proof is straightforward and can therefore be omitted. It follows
from (8) that the demand, which has to be satisfied by facility 1 in each period is given by
~ + Yllo for a given optimal solution. If this demand exeeds the capacity of facility 1 in any
period t proposition 2 holds for optimal solutions of problem MIP:

Proposition 2: If ~ + yO 11 ~ CI then qll = CI for all t = 1, ... , T.

Proof qll ~ (0, cl) follows immediately from Proposition 1, because otherwise for all feasible
solutions YI,t-1 ~ ~ + yO It - qll > O. For a given optimal solution let qll equal zero and T

< t denote the first period preceeding t with positive production at facility 1. Because of ~
+ yO 11 Ct Proposition 1 assures
~ qb = Ct. To obtain an alternative feasible solution the
following steps have to be taken:
 451

(i) decrease production at (l,T) by CI and increase production at (l,T) by the same
amount
(ii) decrease inventory at (l,T) , (l,T+ l), ... ,(l,t-l) by CI
(iii) increase inventory at (k,T) , ... ,(k,t-l) by CI for all k f B(I)
With all other production and inventory variables remaining unchanged the alternative
solution considering (7) incurs a cost reduction amounting to (II - I1.. 8(1) IJ ci ~ 0

Proposition 2 together with (4) provides a first valid constraint:

Corollary 1: If ~ + yO 11 ~ ci then XII = 1 for t = 1, . .. , T.
For an actual demand greater than zero but less than cl, e.g. 0 < ~ + yO II < ch the
produced lotsize cannot be predetermined in general. Yet it is possible to formulate a lower
bound on the lotsize if period t production is positive.

Proposition 3: Let 0 < ~ + yO II < CI . For optimal solution holds: If qll > 0 ,
then ~ + y011 :s; qll :s; CI for t = 1, .. . ,T

Proof Considering (4) qll :s; CI is obvious. Let 0 < qll < ~ + yO 11. To obtain a feasible
solution this requires YI,I.I ~ ~ + yOIl - qll > 0 in contradiction to Proposition 1. •

The comparison of accumulated demand and capacity profiles at each (j,t) can be used to
derive further valid constraints.

Proposition 4: If at (j,t) Cjll < DII+I + yOI,I+1 is fullfilled for a j = 1, ... , N,

t = 1, ... , T - 1, then the following statements (a) and (b) form valid
constraints for all feasible solutions:
1+1
(a) E ~t = (t + 1) for all k € P(j) with ck S cj
• -I

(b) Xt,t+ 1 = 1 for all k € P(j) with Ck > cj

Proof To beginn with, statement (a) will be shown for facility j . Combining and rearranging
(2), (3), (4) and (8) leads to Cj E,I= \1 Xj, ~ DII+I + yOI,I+1 for all feasible solutions. Since
Cjll = cj t, one obtains the inequality
452

(9)

which feasible solutions must satisfy. Considering Cil < 0lt-I + Y\t+1 and (5) the right
hand side of (9) can only hold for E~t_ \1 xjt- = (t + 1).

Since c tl t ~ c jlt for Ct ~ cj the statement obviously holds for all k E PO) with Ct ~ Cj. To
prove (b) suppose contrary to the assumption that Xt,t+ I = 0 for at least one k E PO) with Ct
> Cj. Hence for feasible solutions E/. I CIt~ ~ Olt+1 + yO I, t+1'

(k,t) is linked to at least one predecessor (i,t) (not necessarily an immediate predecessor) the
accumulated capacity cnt of which fails to cover the actual demand Ott+1 + yOI,t+I'
Consequently the flow conservation constraints at facility r for periods 1 to t with r E PO) and
r immediate predecessor of k cannot be satisfied simultaneously. -

For optimal solutions a relation between set-up variables of directly adjoining facilities can
be established based on the succeeding straightforward consideration. Suppose facility k is
immediate predecessor to facility i with Ct ~ Cj. To any schedule with 'lit = 0 and 'Itt > 0
one can expect to fmd another schedule with lower cost when transferring the production
quantity 'Itt to later periods. This is always possible since in periods beyond t facility k must
exhibit a capacity surplus as compared to facility i.

Proposition 5: For optimal solutions holds: If Ct ~ Cj with i = s(k),

then Xtt - ~ ~ 0 for all t = 1, ... ,T.
Proof. For a given optimal solution {y 0, x 0, q O} suppose on the contrary q \, > 0 and
q °it = 0 with t representing the last period for which this condition holds. Let s denote the
first period succeeding t with positive production at facility i. If no such period exists the
proof is obvious. If s exists, to obtain an alternative feasible solution {y, x, q} the following
steps have to be taken:
(i) set 'Itt = 0 and Yb = y\~ - q\~ for t ~ T ~ s-1.
The inventory in these periods remains nonnegative since it amounted to at least q Ott
per period. Increase inventories for all r E B(k) in
periods t ~ T ~ s-l: y,.. = yO,.. + qO tt.
 453

(ii) in periods T, SST S T let Ct. denote the maximal unused capacity at facility k. Set
production at facility k equal to 'It. = q\. + C(k,T) for periods T = S, S + 1, ... ,T
starting with period s. The production increase c(k, T) is recursively defined by

c(k, t) = min ['1:T ' max ( 0, chT - Uj<T

E C(k,j»)] (10)

(iii) starting with period s determine recursively the resulting inventories at facility
k and at facilities r E B(k) by:
Yt. = Yt,,..1 + q\. + C(k,T) - q\. ,
y.. = y",..1 + qO .. _ C(k,T) - qOb

Leaving all other production and inventory variables of the given solution unchanged together
with (7) set-up and inventory costs are effected as follows: Set-up cost are reduced by Kt in
period t. Transferring qO kt to periods SST S T may lead to at most one additional set-up
since Ct ~ qO kt thus the sum of set-up cost is decreased by Kt or remains constant. From step
(i) follows a reduction of inventory cost of Ot - E,.II(It) 1,) qO kt ~ 0 for all periods T, t S T

S S - 1. From (ii) and (iii) one obtains for periods SST S T that inventory at facility k
decreases by

Yko, - Yh = Qr..~ - E c(k,j) ~ 0 ,

Uj<T

whereas at facilities r E B(k) inventories increase by

Yr: - Yn = -~ - E c(k,j) ~ 0 ,
SSj<T

The comparison of these expressions reveals that per period decreases at (k, T) are
accompanied by increases at all (r,T), r E B(k) of identical size. Consequently applying (7)
the sum of inventory cost for SST S T never exceeds the cost of the given solution for the
corresponding periods. Summing up these results the alternative solution in toto leads to cost
savings greater or equal to zero -
454

3. Constraints Derived from Cost Parameters

The valid constraints developed in the previous section were based on the available in-
formation about demand and capacity profiles. So far neither inventory nor set-up costs nor
cost relations have been used to motivate further constraints on problem MIP. This section
will take a closer look on the given cost parameters of the lotsizing model. The objective is
to identify conditions on cost parameters which will guarantee that all facilities k E P(i)
following identical schedules may be a-priori determined and to transform these conditions
into valid constraints. Similar conditions have been identified for the uncapacitated case (see
for instance AXSATER et al., 1986). Unfortunately they rest on results which cannot be
obtained unless facilities draw from unlimited resources (see AFENTAKIS et al., 1984). As
will be observed, the derived conditions bear some resemblance to those of the uncapacitated
case. But unlike their counterparts in the uncapacitated case they are not generally applicable
but restricted to facilities forming paths whose capacities are non-decreasing as one moves
down through the product structure. Such paths shall be denoted PO(i) with i representing the
facility with the largest number for which this condition holds. Before valid constraints
referring to cost parameters can be stated some auxiliary results are needed.

Proposition 6: Let Ct ~ Cj with j = s(k). For optimal solutions holds: If <Itt > 0 and 'lit > 0
then <Itt ~ 'lit for all t = 1, ... , T.
A formal proof can be omitted, since supposing the contrary shows immediately that savings
in inventory cost could be obtained.

Proposition 7: For optimal solutions holds: If there exists a path p0(j) with 'lit > 0 and ~ +
yOu> 0 then

Proof. Proposition 5 assures CLt > 0 for all r E p0(j) if 'lit > O. Let ~ + yO u ~ cl • From
Proposition 2 one has immediately qu = CI ' Proposition 6 together with the transitivity
property of the relation "greater or equal than" provides 'lit ~ cl • For 0 < ~ + yO It < cl
considering Proposition 3 an analogous argument may be advanced. -

For facilities k with only minor set-up operations and accordingly low set-up cost, there may
exist criticallotsizes q(k,t) with the property
 455

Kk - (lk - E Ir) q(k,t) ~ 0 (11)

reBOO

In (11) the expression in brackets corresponds to the so-called echelon holding cost at facility
k first introduced by CLARK et al. (1960). (11) can be interpreted as follows: If the required
production at (k,t) exceeds q(k,t), producing this quantity in periods prior to t and carrying
inventory cannot be preferable to a set-up at (k,t). This consideration motivates

Proposition 8: Let Ct ~ Cj, j = s(k). For optimal solutions

holds:
If Q,;t > 0 and Kt - (II< - E" 11(1<) I,) <!,;t S 0,
then Xtt = 1.

Proof. For a given optimal solution suppose on the contrary ~ = O. To obtain an alternative
feasible solution the following steps can be taken (leaving all other variables unchanged):
(i) replace ~ = 0 by ~ = <!,;t.
(ii) decrease production successively at all (k, T) with 'h:~ > 0, 1 S T < t,
moving from period t downwards until the sum of reductions equals <!,;t. In
each period, in which a reduction is made inventory is decreased by an equal
amount.
(iii) in each period with a production decrease at facility k inventory at all facilities
r, r E B(k), is increased by the same amount.

Step (i) adds set up cost of Kt. A comparison of steps (ii) and (iii) reveals that in terms of
quantity, inventory decreases at facility k correspond exactly to increases at facilities
r E B(k). Since the sum of all inventory changes at each facility equals <!,;t, expression (11)
with q(k,t) = <!,;t can be used to establish a lower bound on the change of set-up and in-
ventory costs effected by (i) - (iii). Considering the assumption this result does not accord
with the assumed optimality of the given solution. -

Combining Propositions 1,3,5,7 and 8, one arrives at valid constraints on problem MIP
which are derived from set-up and inventory costs.

Proposition 9: Let ~ + yO It > O. For optimal solutions holds:

456

If Kj - (lj - E I) 11. ~ 0 for all iEP*(k) - {l}

j £B(i)

then "tt - x.(l),t = 0 for all i E P *(i) - (l).

Proof. For optimal solutions must hold either qll = 0 or qll > O. Let qll = 0 and Proposition
5 assures the statement without having to make use of the cost assumption. In case of qu > 0
Propositions 2 and 3 provide qll ~ 1\. Applying Propositions 8 and 7 successively to all
facilities i E P*(k) - {I} leads to the required result. -

Proposition 9 states conditions on set-up and inventory costs and on the structure of resources
which will guarantee that sets of facilities follow identical schedules. Besides facilitating the
numerical solution of lotsizing problems these conditions may as well be employed in
designing production environments which shall be operated according to the just-in-time
principle since they provide some insight into the required structural characteristics of such
systems.

IV. Concluding Remarks

Currently research efforts are directed towards testing the performance of the derived valid
constraints in greater detail and possibly to incorporate them into a branch and bound code.
To date, their efficiency has only been studied in the context of what may be called
preliminary computational experiments. The best problems were run under LINDO LP code
on a conventional personal computer with a 80386 processor. In order to allow for a
comparison of computing times tests had to be restricted to rather smalliotsizing problems
comprising only eight nodes and ten periods. These first results therefore can only be
considered as indicative of the possible improvements to be achieved by adding valid
constraints.
The following experimental design was used: Inventory cost remained fixed for all tests
whereas set-up cost were assigned randomly within upper and lower bounds. The capacity
at each node was determined in a random fashion considering the independent demand to
avoid infeasibilities. Two demand patterns were examined. Seasonal demand, i.e. demand
exceeding in more than three periods the capacity of the first facility and normal demand,
 457

which in all periods remained within the bound set by the capacity constraint of facility 1.
Demand patterns were further distingished by their capacity tightness. Capacity tightness
refers to the absorption (in %) of the accumulated bottleneck capacity by the accumulated
independent demand. Two cases were investigated: 60 - 80% and more than 80% tightness.
In all cases the established sets of valid constraints were checked for redundancies. It showed
that the most substantial improvements occurred when seasonal demand was given: With
capacity tightness amounting to 60 - 80% (> 80%) computing times were on average reduced
by approx. 60% (75%). For normal demand the corresponding figures dropped to 52%
(63%). It was further observed that on average the first integer solution of the reformulated
problem deviated less that 5 % from the optimum. This observation motivated the use of a
simple heuristic principle: The search process was stopped if the next integer solution did not
improve the previous one by at least b%. Averaging over all test problems, this heuristic
showed its best performance (time reduction of 95 % with an optimality gap of less than 2 %)
when was fixed to 0,5 %.

Since only relatively small problems were investigated on a preliminary basis, further
research certainly remains to be done. Yet the results seem rather encouraging thus providing
some justification for the proposed approach. By the use of structural information on problem
parameters a set of valid constraints could be developed which accomplished a reasonably
tight LP formulation of the lotsizing problem. It is conceivable that the efficiency of this
approach could be further improved by exploiting informations or patterns emerging in the
LP solutions of the investigated nodes. Moreover the last results point to the possibility that
valid constraints may be employed to refine heuristic procedures or to develop new
heuristics.
458

References

Aftentakis, P., Gavish, B., Karmarkar, U., Computational efficient optimal solutions to the
lotsizing problem in multistage assembly systems, Management Science 30, 1984,222 - 239.
Axsater, S., Nuttle, H.L.W., Aggregating items in multi-level lot sizing, in: Axsatter, S.,
SchneeweiB, Ch., Silver, E. (eds), Multi-stage production planning and inventory control,
Berlin 1986, 109 - 118.
Baker, K.R., Dixon, P., Magazine, M.J., Silver, E.A., An algorithm for the dynamic lotsize
problem with time-varying production capacity constraints, Management Science 24, 1978,
1710 - 1720.
Barany, I., Roy, T.J., Wolsey, L.A., Strong formulations for multi item capacity lotsizing,
Management Science 10, 1984, 1255 - 1261.
Blackburn, J.D., Millen, R.A., Simultaneous lotsizing and capacity planning in multistage
assembly processes, European Journal of OR 16, 1984, 84 - 93.
Clark, A., Scarf, H., Optimal politics for multi-echelon inventory problems, Management
Science 23, 1960, 475 - 490.
Crowston, W.B., Wagner, M.H., Dynamic lot size models for multistage assembly systems,
Management Science 20, 1973, 14 - 21.
Crowston, W.B., Wagner, M.H., Henshaw, A., A comparison of exact and heuristic
routines for lot size determination in multistage assembly systems, AllE Transactions 4,
1972, 313 - 317.
Nemhauser, G.L., Wolsey, L.A., Integer and combinatorial optimization, New York 1988.
Love, S.P., A facilites in series inventory model with nested schedules, Management Science
18, 1972, 327 - 338.
Steinberg, E., Napier, H.A., Optimal multilevellotsizing for requirement planning systems,
Management Science 26, 1980, 1258 - 1271.
Thizy, J.M., Van Wassenhove, L.N., Lagrangean relaxation for the multi item capacitated
lotsizing problem: A heuristic implementation, AllE Transactions 17, 1985,308 - 313.
Thompson, G.L., On the parts requirement problem, Operations Research 13, 1965,
453 - 461.
Zahorik, A., Thomas, L.J., Trigeiro, W.W., Network programming models for production
scheduling in multistage, multi-item capacitated systems, Management Science 30, 1984, 308
- 325.
 DLSP for Multi-Item Batch Production
by
Wolfgang Briiggemann and Hennann Jahnke
Institut fUr Logistik: und Transport
Universitiit Hamburg
2000 Hamburg 13
Gennany

Abstract

The discrete lot-sizing and scheduling problem (DLSP) has been suggested for the simulta-
neous choice of lot sizes and production schedules. However, the common assumption of
instantaneous availability of the manufactured units is not satisfied in practice if the units arrive
in inventory in one batch no earlier than completion of the whole lot Therefore, we present
additional constraints for the case of batch production. The resulting extended DLSP is
formulated as a mixed-integer linear program. The feasibility problem of this modification to
the standard DLSP is again NP-complete. A two-phase simulated-annealing approach is
suggested for solving the modified DLSP. Since DLSP is a finite-time-horiwn model, sensible
target inventories have to be determined. Numerical results are presented for different problem
classes.

1. Introduction

One of the most challenging problems in production planning has been the simultaneous
choice of lot sizes and production schedules in order to minimize cost. In the early 196Os,
mixed-integer linear models were proposed in the economic literature to cope with this
situation, e.g., Adam [1963] and Dinkelbach [1964]. Their principal idea is to divide the finite
time horizon into (small) time intervals in which the machines can be used either for
production of at most a single product, or can be setup for such a production. Recent interest
in these models, now popular as the single- or multi-machine discrete lot-sizing and
scheduling problem (DLSP), stems from the development of new algorithms to solve these
problems. Fleischmann [1990], for example, suggests an exact algorithm for the single-
machine multi-item case with zero setup times. A comprehensive reference to these algorithms
460

is given by Salomon [1991]. After showing that generating feasible solutions for the single-
machine multi-item DLSP with non-zero setup times is NP-complete, he suggests a heuristic
based on Lagrangean relaxation.

All these modelling approaches pertain to the case of instantaneous availability of the
manufactured units prior to the completion of the lot. However, this assumption is not satisfied
in practice if the units arrive in inventory in one batch no earlier than completion of the whole
lot. Therefore, we introduce additional constraints in order to model the case of batch
production. The resulting extended DLSP is formulated as a mixed-integer linear program. For
solving this problem a two-phase simulated-annealing (SA) approach is suggested because of
the general applicability of SA (Kuik and Salomon [1990] apply SA to the related multi-level
lot-sizing problem).

In many cases, there is no natural time limitation for production processes and demand will
be stochastic. Hence, finite-time-horizon models are applied repeatedly (rolling application)
in order to solve approximately the underlying infinite planning problem and to incorporate
new estimates based on more available data in each planning instance. In this context, a
special emphasis is put on the final inventories. Therefore, we determine sensible inventories
by a parametric application of the modified DLSP-model to the expected demand. These
inventories can be used as target (final) inventories for production planning in rolling ap-
plication, while actual (final and therefore new initial) inventories result from the deviation
of realized from estimated demand. Positive initial inventories yield more feasible production
schedules with respect to the demand restrictions with additional cost saving potential.
Therefore, optimal target inventories will also not be zero for all products simultaneously.

2. The DLSP in the Case of Batch Production

Ordinarily, DLSP is formulated for N items and a finite-time horizon of T time units. Here,
production scheduling and demand should be considered on different time scales: Typically,
demand can be estimated for example as demand per day, while the DLSP requires a less
coarse discretization of the time axis due to the underlying idea that during one time unit (e.g.,
hours or 30-minute intervals), the machine can be used at most for setup or production of only
 461

one item (see constraints (4) below). Hence, the time units for DLSP will be chosen to be the
greatest common divisor of setup times and minimal production times for all items. The
different time scales of demand and scheduling yield a division of the planning period T into
M (demand) subintervals of lengths T ..-T... \ (m=l, ... ,M; To=O; TM=T) where demand di, for

product i is positive only at times T.. (diT.~O and di,=O t=l, .. .,T; t~T.. ; m=l, ...,M;

i=l, ...,N).

A production schedule is a matrix (v,y)=«vi,), (Yil» where

Vi' are zero-one variables indicating setup in period t for production of product
i and
Yi' are zero-one variables indicating production of product i in period t with
YiT+J=O for formal reasons.

Then the objective is to minimize the sum of setup and inventory holding cost

N T
Min C«v,y» = LL
i.l ,.1
r i Vi' + hJi' (1)

s.t.
i=l, ... ,N; t=l, ... ,T (2)

i=l, ... ,N; t=ai+l, ... ,T; t=O, ...,ai-l (3)

Y· = 0 i=l, ... ,N; t=l, ... ,ai (3a)

"
N

L
i-I
(Yi'+V) ~ 1 t=l, ... ,T (4)

i=l, ... ,N; t=l, ... ,T (5)

i=l, ... ,N; t=l, ... ,T (6)
ei , ~ (l-yi ,+\) i=l, ... ,N; t=l, ... ,T (7)
T.

Ei.. =Le
,-\
i, i=l, ... ,N; m=l, ... ,M (8)
T.

Vi .. =LV
,.\ i, i=l, ... ,N; m=l, ... ,M (9)

i=l, ... ,N; m=l, ... ,M (10)

462

i=I, ... ,N; m=I, ... ,M; t=I, ... ,T.. (11)

i=I, ... ,N; m=I, ... ,M; t=I, ... ,T.. (12)

T.

i;". =E (OiYi~'" - di~) + 1;0

~el
i=I, ... ,N; m=I, ...,M (13)

1;" I iO' i;". ~ i=I, ... ,N; t=I, ...,T; m=I, ... ,M (14)

YiT+1 = 0, eil' Vi" Yil, Yi/m E {O,ll i=I, ... ,N; t=I, ... ,T; m=I, ... ,M (15)

The objective (1) and constraints (2), (3) and (4) are taken from the standard DLSP [Salomon
1991, pp. 34 and 43]. Again, ri in (1) are sequence independent setup costs per setup period
and the inventory holding costs are given by the product of parameter hi (costs per unit and
period) with nonnegative inventories I ir The demand-satisfaction constraints (2) of standard
DLSP describe the dependence of current inventories on the inventories of the preceding
period, the quantity of the corresponding item produced (where 0i is the production speed) and
the demand. The correct sequence of setup and production periods is modelled in equation (3)
where ai>O is the corresponding number of setup periods. Constraints (3a) are additional to
the DLSP formulation due to Salomon [1991] and prohibit production of item i in periods
1, ... ,ai with no preceding setup. Constraints (4) are used to prevent simultaneous action (setup
or production) on the machine.

These constraints alone do not pertain to the case of batch production. While the inventory
including work in process IiI (as described above) is required for computing holding costs in
(1), the last batch begun before time t is not available for satisfying demand if the production

process of this batch is not fmished by t. Therefore, we introduce inventory i;", which is

already available in demand instance T",. In order to model this available inventory, some
auxiliary variables and constraints (5) - (12) are needed: Here, zero-one end-of-batch variables
eil indicate the production of the last unit of the current batch in t by eil=l. End-of-batch
constraints (5) - (7) are similar to setup constraints (3). Ei"" Vi'" and Di", are set in equations
(8) - (10) to the cumulated number of completed batches, number of setup time units and a
correction term, respectively. This correction term is required to determine zero-one variables

Yi/m which are then used in (13) to define the auxiliary inventory ~". of available units. It

should be noted that Yi/m vanishes even if the decision variable Yil is equal to one, if period
 463

t is used for production of item j in a batch which was started before but has not been
completed by Till' This behaviour of the auxiliary variables Yillrl is guaranteed by constraints

(11) and (12), where the latter consists of two parts both of which are zero in the case of
incomplete production at time Till: In this case, the difference between cumulated number of
setup periods before Till and before t S Till is greater than zero for all batches that are
completed before T"" while the second bracket is positive only in the case of no production
extending in time over T",. Moreover, Yillrl may also vanish despite completion of the batch by
T", if not all of the units produced are needed to satisfy demand in Till' Note, that the auxiliary

inventory ~'" will be smaller than the actually available inventory only in this case.

Constraints (14) and (15) are the standard non-negativity and binary constraints, respectively,
for the decision and auxiliary variables. Initial inventories can be thought of as given constants
while final (target) inventories are implicitly included in the demand of period T.

Salomon [1991] shows NP-completeness of the feasibility problem of the DLSP with non-zero
setup times. The modified DLSP is also in NP, because feasibility of any given structure
(production schedule) for any problem instance can be checked in polynomial time. Since any
instance of the standard DLSP can be transformed to an instance of the modified DLSP (again
in polynomial time) and a feasible production schedule for the modified DLSP is also feasible
for its standard version given essentially by constraints (2) - (4), the standard DLSP reduces
to the modified DLSP which is then also NP-complete [Florian et al. 1980].

3. The Two-Phase Simulated Annealing

Simulated annealing is a multi-purpose heuristic for the solution of combinatorial optimization

problems that was first suggested by Kirkpatrick et al. [1983] and Cerny [1985]. A neighbour-
hood structure is superimposed on the (usually finite but large) space of feasible solutions
(configurations or in this context production schedules). Given such a feasible configuration
(vew,yew) a candidate solution (ve"" ,ye...) is drawn randomly from the corresponding
neighbourhood. This new configuration will be accepted subject to either improvement of the

objective function or another random experiment with acceptance probability given bye -l!.C/y

where .1C=C(ve"",Ye",,) -C(vew,Ye..,) is the difference of the cost function values of the
candidate and the current configuration. y is a control parameter corresponding to temperature
in the original physical analogue in thermodynamics [Metropolis et al. 1953]. Infinite
464

repetition of this procedure with a fixed value of control parameter 'Y can be viewed as one
realization of a homogeneous Markov chain where the current state of the Markov chain is
the last accepted configuration. Iterative reduction of the temperature (i.e., 'Y) yields a sequence
of such Markov chains and it can be shown [Mitra et al. 1986] that, roughly spoken, the
sequence of configurations converges asymptotically to a globally optimal solution, almost
surely. Besides this convergence behaviour (although efficiency compared to tailored
algorithms is usually poor), the main advantage of SA relative to other (tailored) solution
methods suggested is its general applicability. Solving a specific problem with SA requires
only determination of a neighbourhood structure and an appropriate cooling schedule (i.e.,
choice of the sequence of control parameters 'Y and number and length of the finite
approximations of the homogeneous Markov chains). Van Laarhoven [1988] discusses
different cooling schedules and application of SA to selected combinatorial optimization
problems. However, it is important to point out that the neighbourhood choice is usually
performed on the set of feasible configurations only. Feasibility in the context of the modified
DLSP is mainly given by the demand-satisfaction under batch production (13) and setup
constraints (3) - (4). For a general problem instance, it is therefore necessary to construct an
initial feasible solution disregarding costs and to optimize with respect to the objective
function in a second phase. The overall structure of the two-phase algorithm outlined in the
following two sections is similar to the two-phase simplex method for the solution of linear
programs.

The neighbourhood structure for a given production schedule (v,y) is defmed by reducing (v,y)
to a "pure" production schedule y"'" by physically elimipating setup periods. An element (vc"",
Yc,.,.) out of the neighbourhood of (v,y) is then obtained by exchanging the activities of two
arbitrary periods in y,..,. and expanding (i.e., inserting setup periods in front of each
production batch and shifting later production by the corresponding number of time units)

y::; to (vc"",yc,.,.)' During expansion of y::.' to (vc... ,yc,.,.) a necessary condition for feasibility
(the last batch must be finished in or before period n is checked and if this condition is
violated another configuration is drawn out of the neighbourhood of (v,y).
 465

3.1 Phase 1: Search for a Feasible Production Plan

The procedure described above does not ensure feasibility in the sense of (2) to (15) since the
demand-satisfaction constraints might not be satisfied. Therefore, a first production schedule
is chosen to consist of a single batch for each item where the batch size is determined to
satisfy the cumulated demand in T and production is carried out immediately. For this
situation, initial inventories are computed in order to fulfil the demand satisfaction constraints
(13) in each demand instance. Afterwards the sum of the positive deviations of these
hypothetically needed inventories from the actual inventories form the objective function and
are minimized in phase 1. As in the simplex method, a feasible production schedule is found
if this sum vanishes. SA will yield more feasible (in the sense that production is finished by
n production schedules by restricting the exchange to active production periods only contrary
to the exchange of activities which also include the final idle periods.

3.2 Phase 2: Search for a Cost-Optimal Production Plan

The actual optimization with respect to the cost function is carried out in phase 2. Here, the
same neighbourhood structure is used, but exchanges are now carried out between arbitrary
(active and idle) periods of y p"r<. Production schedules that are not feasible in the sense of
(13) are not considered as candidates. Thtis, feasibility is preserved before each acceptance
in phase 2 subject to the acceptance criterion corresponding to the original cost function (1).

4. Parametric Optimization of Target Inventories

In chapter 2 the modified DLSP was introduced for the situation of deterministic (possibly
estimated) demand and given inventories. In most practical situations, the scheduling problem
arises in the context of rolling application with stochastic demand. Especially here, it is
important to find a sensible set of inventories which can be used as target inventories at each
planning instance. Actual final inventories are the initial inventories of the next interval. It
should be noted that these inventories will differ from the target inventories in practice due
to the deviation of realized from estimated demand. In this situation, demand is to be modelled
466

by random demand variables. In order to find a set of sensible target inventories, the stochastic
demand variables are replaced by their expected values. For this demand structure, final and
initial inventories are assumed to be equal and varied parametrically in a modified DLSP
application in order to find a cost optimal set of target inventories. Additionally. sensible
target inventories for the stochastic application will usually contain a safety stock component
in order to compensate for the deviations of the realized from the estimated demand.

5. Numerical Results

Demand schedules are generated randomly for the numerical evaluation of the modified DLSP
solution procedure suggested in the preceding paragraphs. We consider planning horizons of
maximal length of 60 periods and up to six products which are distinguished into up to three
different categories depending on their demand expectations and variances. These consist of
products with high expected values and low variances, intermediate moments and finally of
products with low expected demand and high variances. Every tenth period is a demand
instance for all products. Demand for each product is measured in how many units can be
produced during one production period. Thus. all production speeds are equal to one.
Similarly. inventory costs for all products are assumed to be equal and unity. Setup of one
period is required for each product and the corresponding costs are set to 60.

The solution procedure is programmed in FORlRAN 77 and implemented on an IBM PC with

an 80 486 / 25 MHz processor for the parametric application of DLSP, and on an IBM
compatible NEC PC with an 80 486/20 MHz processor for the case of random demand. The
computation times required for solving the parametric DLSP on the first computer turn out
to be approximately 80 % of the computation times on the slower processor. We use a
geometric cooling schedule. the number of repetitions is given by "acceptances max constant"
and no acceptances at one temperature stage is used as the stopping criterion (for notation see
Collins et al. [1988]; the parameter values are given in table 1). Note that SA parameters used
for the numerical experiments below are constant while they should be problem-size dependent
for application to general problems. In most practical applications. finding a feasible
 467

Parameter Phase 1 Phase 2

Initial Temperature 10 100

Reduction 0.9 0.9

Max Number of Acceptances 120 70

Max Number of Repetitions 10000 1000

Number of Reductions 50 50

Table 1: Parameters of Simulated Annealing Cooling Schedule

production plan will be of primary interest when cost parameters are not easily estimated.
Therefore, the cooling schedule for phase 2 (optimization) is chosen to be coarser than for
phase I (feasibility). Suboptimal solutions obtained by this rough procedure are improved by
shifting batches to the right in order to fill unnecessary gaps.

To determine a set of sensible target inventories, we use the rounded expected values for the
demand in the demand instances. Due to computational restrictions (e.g., there are 106
different inventory combinations for six products if inventories vary from 0 to 9), only
preselected initial and final inventories are tested in the parametric application of modified
DLSP. For each inventory configuration, this problem is solved 25 times with different
initializations of the random number generators due to the stochastic nature of the solution
procedure. The results of the best five of all investigated inventory configurations are given
in table 2.

Inventory Average CV') of Number Average CPU CV') of

Cost Cost of Best Time (sec) CPU Time

202000 1026.84 3.58 6 301 9.9

201000 1027.52 3.15 5 298 9.9

201100 1041.96 3.35 3 295 11.0

302000 1045.60 2.23 I 309 7.9

402000 1048.80 3.15 1 301 11.7

Table 2: Parametric application of the modified DLSP to selected inventory configurations

.) (Coefficient of Variation) 100
468

Typically, asymmetric initial inventories yield the lowest costs despite the underlying
symmetric demand structure where each product category consists of two identical elements.
This result can be explained by the trade-off between the size of the set of feasible solutions
(which is larger for higher initial inventories) and the contribution to the final objective-

function value from the holding costs for the initial inventory.

The inventory vector (2,0,2,0,0,0) yields the lowest average cost value and outperforms the

other invetigated inventory configurations in 6 out of 25 repetitions. Safety stocks are set to
one unit for every product which yields target inventories for the DLSP of (3,1,3,1,1 ,1) for
products 1 to 6. This inventory vector is used as initial and target inventories for the modified
DLSP.

During the numerical experiments for the modified DLSP with randomly generated demand
structures it turns out to be necessary to differentiate between hard and easy problems. As an
elementary indicator of difficulty, we use the ratio of the total demand plus one setup period
per product and the length of the planning period. Problems with an 85 % or higher degree
of difficulty are considered to be hard. Note that two problems with the same elementary
degree of difficulty can be quite different with respect to actual difficulty due to different
demand distributions over time.

One 3-product / 30-period (3/30) problem with a 63.3 % elementary degree of difficulty is
solved exactly by LINDO-386 on the IBM 80486/25 MHz after more than 48 hours of
computation. The same optimal solution is found much more rapidly (with an average CPU
time on the 20 MHz machine of 95.2 seconds) by the two-phase simulated annealing algorithm

in five out of ten repetitions while the other five solutions are near optimal with a coefficient
of variation of 2.63 %. Note that the CPU times for our solution procedure can be improved
by adapting the parameters of the cooling schedule (Table 1) to the actual problem size. Due

to the extremely high computation time required for the exact solution, we use the coefficient
of variation of the final objective function value for ten different realizations as a criterion of
stability and quality of the solutions found instead of the true optimum.
 469

Degree Number Average Average CV') of Unsolved

3/30 of of CV') CPU average Problems
Difficulty Problems of Cost Time (sec) CPU Time

Easy .70 - .84 15 2.27 103 9.4 -

Hard .85 - .92 15 2.07 103 16.7 -

Table 3: Results for the 3-product I 30-period DLSP

.) (Coefficient of Variation) 100

The results obtained for the two different classes of difficulty are given in table 3 for the 3/30
problem, and in table 4 for the 6/60 problem.

Degree Number of Average Average CV') of Unsolved

6/00 of Problems CV') CPU average Problems
Difficulty of Cost Time (sec) CPU Time

Easy .65 - .84 15 3.25 409 10.9 -

Hard .85 - .92 15 2.92 747 41 13.3 %

Subclass 1 .85 - .92 10 3.48 540 15 -

Subclass 2 .85 - .92 5 1.81 1185 8.8 40 %

Table 4: Results for the 6-product / 6O-period DLSP

.) (Coefficient of Variation) 100

For each of the 15 demand situations we compute ten repetitions with new seeds for the
random number generator. In both cases, the coefficients of variation for the objective are at
most 5 %, indicating stable results. Computation times tend to be higher for problems with
a high elementary degree of difficulty. In particular, phase 1 can require a significant amount
of computation time for the 6/60 problems. The two-phase solution procedure seems to be
sensitive to higher demands at the beginning of the planning period. This observation might
be an explanation for the substantial differences in computing times of the hard problem
instances. A further natural subdivision of this class is found into problems with an average
computation time similar to the class of easy problems and on the other hand problems which
470

exceed 1000 seconds of required CPU time. In this latter class only. it happens that phase 1
of the algorithm is unable to find a feasible solution on several occasions. This may again be
due to the fact that the set of feasible solutions is in general smaller for a high demand density
towards the beginning of the planning period. This hypothesis is supported by the small
coefficients of variation of the fmal objective function value. A specific problem instance is
given in table 5 as an example for high demand distribution during the frrst two subintervals.

Demand 10 20 30 40 50 60
I 1 4 I 2 3 I
2 3 2 I 2 2 2
3 4 2 0 0 1 0
4 I 2 2 I I I
5 0 0 1 2 0 1
6 0 0 3 1 0 0
Table 5: Demand structure for a hard 6160 problem

The elementary degree of difficulty is 88.3 % and a feasible solution is given in figure 1
(R stands for setup periods while production is indicated by 1).

*.*.*.**** ••• ****.********************** •• **** •••

Produktionsplan
••••• ***.*****.**** ••• ****** •• ***********.*******

1 2 3 4 5 6
123456789012345678901234567890123456789012345678901234567890
1 1
-1-1 Rlll Rlllll Rllll 1
2 1 Rllll RIll Rlllll 1
3 IRllll Rllli
4 1 Rlllll Rl Rll 1
5 1 Rllll 1
6 1 Rllll 1
1 1
Auslastungsgrad: 100.0 %
Zielfunktionswert: 1441.0

Figure 1: Production Schedule for the demand structure of table 5

This is the only feasible solution that is found by phase 1 in ten repetitions after 1204 seconds
of CPU time.
 471

6. Summary and Conclusions

The problem presented here differs from the standard DLSP given for example by
Salomon [1991] in considering the case of batch production, not necessarily vanishing
inventories, and in using different time scales for production scheduling and demand. The
feasibility problem for this modified DLSP is shown to be again NP-complete. The suggested
two-phase simulated-annealing solution method is based on intuitive ideas. The optimization
procedure is separated into phase 1 searching for a feasible solution and optimizing cost in
phase 2. Production schedules are generated by dividing, combining and shifting batches. The
numerical experiments presented indicate that our heuristic solution method yields stable
results in short computation times compared to the exact solution obtained by LINDO-386.
This is not surprising because of the large number of integer variables required in the
proposed mixed-integer linear model. The solution procedure can be readily applied to larger
problems where stable solutions will be obtained with reasonable computation times by a
problem-size dependent choice of the cooling schedule. In contrast to the prohibitive
computation times required for the exact solution, our SA approach allows the numerical
sensitivity analysis of cost parameters which are often not easily estimated in practice.

References

Adam, D.: Simultane Ablauf- und Programmplanung bei Sortenfertigung mit ganzahliger linearer
Programmierung, Zeitschrift fiir Betriebswirtschaftslehre 33 (1963),233 - 245.

Cerny, V.: Thermodynamical Approach to the Traveling Salesman Problem: An Efficient

Simulation Algorithm, Journal of Optimization Theory and Application 45 (1985) I,
41 - 51.

Collins, N.E., R.W. Eglese, B.L. Golden: Simulated Annealing - An Annotated Bibliography,
American Journal of Mathmatical and Management Sciences 8 (1988) 3 & 4, 209 -
307.

Dinkelbach, W.: Produktionsplanung, Wiirzburg Wien 1964.

Fleischmann, B.: The discrete lot-sizing and scheduling problem, European Journal of Operations
Research 44 (1990), 337 - 348.

Florian, M., J.K. Lenstra, A.H.G. Rinnooy Khan: Deterministic Production Planning: Algorithms
and Complexity, Management Science 26 (1988) 7, 669 - 679.

Kirkpatrick, S., C.O. GelatI, M.P. Vecchi: Optimization by Simulated Annealing, Science 220 (1983),
671 - 680.
472

Kuik, R., M. Salomon: Multi-level lot-sizing problem: Evaluation of a simulated-annealing

heuristic, European Iournal of Operations Research 45 (1990), 25 - 37.

Metropolis, N., A.W. Rosenbluth, M.N. Rosenbluth, A.H. Teller, E. Teller: Equation of State
Calculations by Fast Computing Machines, The Iournal of Chemical Physics 21
(1953) 6, 1087 - 1092.

Mitra, D., F. Romeo, A. Sangiovanni-Vincentel1i: Convergence and Finite-Time Behavior of

Simulated Annealing, Advances in Applied Probability 18 (1986), 747 - 771.

Salomon, M.: Deterministic Lotsizing Models for Production Planning, Berlin 1991.

Van Laarhoven, PJ.M.: Theoretical and Computational Aspects of Simulated Annealing, Ph.D. thesis, Erasmus
Universiteit, Rotterdam 1988.
Information Production Functions in Dynamic Lot-Sizing

by

Gunter Fandel and Peter Fran~ois

Institute for Production Management
Faculty of Economics
Fernuniversitat
D-5800 Hagen 1, Germany

1. Preliminary Remarks

The complex problem of describing procedures of information

processing in PPC-systems in the manner of activity analysis
was discussed in detail by FANDEL (1991 and 1992). In this con-
text notions like information processing activity and informa-
tion technology had to be defined and several definitions of
eff iciency and dominance had to be stated. This was done in
order to define information about goods, which are involved in
the production of a company, as input and output goods of pro-
duction planning and control and to describe formally proce-
dures of efficient information processing in PPc-systems by
information production functions. After all, these considera-
tions serve to evaluate individual PPC-methods I -modules or
total PPc-systems on the basis of efficient and dominant proce-
dures of information processing. This is so as the planning
procedure for solving a PPC-task or the whole complex of tasks
is certainly superior to another procedure, if it achieves
better results in the sense of a certain dominance definition
when using and generating relevant information for the maste-
ring of such planning tasks. Some smaller examples of dynamic
lot size planning and production program planning in joint pro-
duction have served to explain these considerations with
respect to their practical applications. However, information
production functions have not been derived up to now, neither
implicitly nor explicitly.
474

Thus the considerations in this paper aim at a more detailed

understanding about the derivation of information production
functions especially for the PPC-module of dynamic lot-sizing.
For this purpose three solution procedures, which are normally
used for this planning situation, will be evaluated and com-
pared with respect to the information outputs they provide for
the solution of this planning task, if the information inputs
are systematically varied in the course of a sensitivity analy-
sis. The efficient information processing activities, which can
be obtained from this procedure, represent a part of the infor-
mation production function for this case of planning.

2. Description of the Information processing by the concept of

Activity Analysis for a specific Planning situation of
Dynamic Lot-Sizing

2.1. Information processing Activities and Information

Technology

In the planning situation considered here the cost minimal lot

sizes qk-10 = = 11, ... ,18, for eight planning periods and
vk' k
the related costs C = v 19 are to be determined for the demands
xk = vk of eight periods k = 1, ... ,8 with the storage cost rate
1 = v9 and the set-up costs a = v 10 given. In order to generate
the data outputs vk' k = 11, ... ,19, from the data inputs v k ' k
= 1, ... ,10, mainly physical data processing inputs v 20 and v 21
in form of programming and computer operating-times are neces-
sary. As admissable procedures of information processing for
getting the desired data outputs from the data inputs with the
help of the data processing inputs only

- the Silver-Meal-algorithm (SILVER/MEAL 1969 and 1973),

- the Part-Period-algorithm (DEMATTEIS 1968, MENDOZA 1968) and
- the Wagner-Whitin-algorithm (WAGNER/WITHIN 1958)

are employed here for the specific planning situation as

adequate solution procedures, because they promise "good
 475

results" as simulation studies suggest (GLASER 1975, KNOLMAYER

1985, WEMMERLOV 1982, ZOLLER/ROBRADE 1987).

On the basis of the costs of set-up and inventory

TV+1
a + 1 ~ (t-T v -1)X t ,
t=TV+ 1

which have to be minimized for the respective lot size cycle

[T v+1,T V+1 ] of the periods TV+1,T V+2, ..• ,T V+ 1 ' with Tv ,T V+1
€{0,1, ... ,8} and TV < Tv+1 , the sequence [T*v+1, T*V+1 ] of the
optimal lot size cycles with given demands XT, T = 1, ... ,8, and
known quanti ties 1 and a is determined (cf . ZOLLER/ROBRADE
1987) according to

- the Silver-Meal-algorithm by

V 0,1, ... ,n-1,

o TO* < T1* < ... < Tn* ~ 8;

- the Part-Period-algorithm by

V = 0,1, ... , n-1,

476

- and the Wagner-Whitin-algorithm by

* j
Cr,j = Cr - 1 + a + I . ~ (t-1)X t
t=1
j = 1, ••• ,8, 1 = 1, .•• ,8,

with

o.

In all three cases for the resulting optimal dynamic lot sizes
of the eight periods one obtains:

and q1 = 0 *
for 1 4= 1V+1,

v = 0,1, ... ,n-1,

o 10* < 11* < ••• < 1n* ~ 8.

If the three procedures of dynamic lot-sizing mentioned above

are implemented on the computer, only the physical information
processing inputs for generating the data outputs as the
desired information about the problem solution from the data
inputs are of further interest. Because then it is obvious how
an information processing activity

v = (v 1 , •.. ,v 21 ) I

(X1, ... ,x8,I,a,q1, ... ,q8,c,v20,V21) I

of the planning situation in question can be accomplished.

These activities are to be differentiated by means of an addl-
tional upper index in v~ with ~€{SMA,PPA,WWA} depending on the
application either of the Silver-Meal-algorithm (SMA), the
Part-Period-algorithm (PPA) or the Wagner-Whitin-algorithm
(WWA).
 477

For the ranges Dk of the components k of these information pro-

cessing activities

R U {9}, k 1, ... ,19,

R_u {O}, k 20,21,

applies.

In this context Rand R is the set of real or negative real

numbers. For the values, which can be adopted by the data
inputs or outputs kE: {1, ... ,19} as information components of
the information processing activities, it applies that they are
either real numbers, if the information occurs in the activity,
or they are 9, if the corresponding information does not occur.

As the information processing input amounts v k ' k = 20,21, are

ordered in the usual way with regard to production theoretical
aspects, so that according to the preference ~ or > in the
p p
sense of "better or equal" and "better"

or

applies, this is not the case for the data elements vk'
k€{1, .. . ,19}, as information about the quantities of produc-
tion, because it is quite possible that for example the infor-
mation v9 = 1 1 or Vg 1 = 2 about the storage cost rate can
be of different quality and it is not clear from the start that
v9 < Vg implies v9 >
P
Vg.
From the point of view of information processing only the
following orders can be defined on the ranges Dk of the d~ta
inputs (k€{1, ... ,10} = K) and data outputs (k€{11, ... ,19} = K).
478

vk = VJc

or vk = 8 and vJccR,

or vk > v'k
p

kcK:

vk ~ v'k
p

++

that is, with regard to information processing it is better if

ceteris paribus one has less knowledge and less information
processing inputs to be used or if one receives more informa-
tion from such a process. If

21
D X Dk ' then vcD.
k=l

As information technology of the dynamic lot-sizing in question

the set

TI = {v~cD I v~ generated by the process ~c{SMA,PPA,WWA}}

can be defined. In this way TI fulfils the axioms formulated

for information technologies as well as the postulates of rele-
vance of data inputs and outputs and the realizability of
information processing activities (cf. FANDEL 1991).
 479

2.2. Efficiency Definitions and Information Production

Function

Efficiency definitions serve to exclude obviously bad activi-

ties of information processing and to reduce further considera-
tions to the remaining activities. In the following, three
definitions of efficiency are introduced.

Definition 1 : Original Efficiency

An information processing vector v€T I is said to be

originally efficient,
only if if and there is no other
information processing vector w€T I with

1) wk ~ vk for all k€{l, . .. ,21} and

p

2) wk > vk for at least one k€{l, ... , 21}.

p

Data processing vectors w€T I which satisfy the above condi-

tions should be called originally dominant. This can be
expressed formally by w > v. The set of all originally
p
efficient information processing vectors is to be denoted by
TI TI C TI.
0' 0

In analogy to the activity analysis of production theory, the

formal description of the set of all originally efficient
information processing activities renders itself to be
considered as the information production function of this plan-
ning situation for dynamic lot-sizing, i . e . let

fI: D .... R,

then fI is an information production function for the case

analysed here, if

<= >
480

Certainly it would be desirable if the information production

function in the planning case at hand was generated by only one
of the three solution algorithms employed for dynamic lot-
sizing. This solution algorithm would then from an information
processing per spec tive at least never be inferior to the other
two approaches, possibly even be superior. For such a robust
selec tion the notion of original efficiency may, in some
instances, prove insufficient, the reason being that each
activity v' is already originally efficient in comparison to
another activity v as far as it features a different informa-
tion content in just one data component, i.e.

vk + vk + e for one kE{1, ... ,19}.

For that case the assessment-efficiency provides a stronger

selection criterion.

Definition 2: Assessment-Efficiency

Let z: D -+ RL be a real and vector-valued assessment

function, Z = (z1"",ZL) " which assigns to each
information processing activity vED the assessment
vector z(v)ER L in accordance with criteria l, ... ,L.
Then the information processing vector VET I is called
assessment-efficient with respect to z, if and only if
there exists no wET I with

1) Zl(w) ~ zl(v) for all IE{1, ..• ,L} and

2) Zl(w) > zl(v) for at least one IE{1, ..• ,L}.

Information processing vectors WET I that fulfil the

above conditions are said to be assessment-dominant.

The assessment-efficiency permits the evaluation of information

processing activities under diverse criteria. possible criteria
are for example computing times, costs as well as profits or,
in general, decision objectives to be taken as a basis for, or
 481

typical for the problem of production planning at hand. The

notion of assessment-efficiency has the advantage that the
efficient picture of information processing activities under
the function z can be studied in the real number-space RL,
this being the actual objective of analyses in many instances
of production planning.

In the case at hand the following modified approach may be

recommended, in order to analyse and to evaluate in a robust
manner the information processing activities of the dynamic
lot-sizing on the basis of original efficiency. Let be

p(v) v = (v 1 ,··· ,v10,c,v20,V21)'

(x 1 , ... ,x 8 ,l,a,c,v
-
20 ,v 21 )'

10 21
with p: D-+/5 xD uRux D
k=l k k=20 k

C = !(J(v)f.R,!(J: D-+R,

a new information processing activity, that is created from the

original information processing activity by substituting the
information C on the costs by the actual costs of the policies
of production run and storage and neglecting the determined lot
sizes. (Applied correctly, the respective problem solving
method and proper calculation will yield C = !(J(v) = v 19
= C. C ~
p
C' <=> C $ C' or C > C' <=> C <
p
C'. The lot sizes have
to be available anyhow according to the postulate of the
relevance of data-output and they are indeed available as
information by construction of the respective solution method).
In addition then, the following definition of efficiency in a
broader sense is to be introduced.
482

Definition 3: Efficiency in a Broader Sense

An information processing activity veT I is termed efficient

in a broader sense, if and only if there is no weT I with

pew) w> v p (v) ,

p

i.e. w does not dominate originally the activity v.

This type of approach focuses the evaluation of information

processing activities involved in dynamic lot-sizing on the
inputs of information, the costs related to the solution and on
the programming- and computing-times spent in the process. To
include the data-input in the evaluation makes sense as diffe-
rent data-inputs characterize different decision cases that
cannot be compared with each other. The consideration of the
other components is evident. By contrast, such a restriction of
evaluation may be considered problematic for the dynamic lot
sizes q1, ... ,q8 as data-outputs of the information processing
vector v. The reason is, that the current lot sizes are in this
situation of decision desired data outputs, the knowledge of
which is absolutely required for the solution of the production
planning task of lot-sizing. On the other hand the bare infor-
mation about a sequence of dynamic lot sizes is of no benefit
wi thout knowledge of the costs that the sequence involves,
because ultimately, the information about such a sequence of
(optimal) dynamic lot sizes that transgresses into a decision
makes only sense when the resulting costs for the company are
known.

Instead of the information production function fI, sections

of the modified information production function fI are now in-
vestigated for the current case of dynamic lot-sizing with
 483

TIo = {vEDI there is no wED with w > v,

p

v P(v), w = p(w), v,W,ET 1 },

3. Example

3.1. The Planning Cases and the Data Inputs

In the modified information processing activities

v (x1 8
, ... ,x -
,I,a,C,v 20 ,v 21 ) I

that will be discussed the various decision cases of the

dynamic lot-sizing are differentiated by the values of data
inputs

Xl, ••• , x 8 , 1, and a.

For the solution and information processing the three

algorithms M = SMA, PPA, WWA mentioned above are applied . In
each case this is connected with the data processing input
quantities v 20 and v 21 and results in the costs C determined by
the lot size policy.

Bases for the selection of data inputs are six alternative

structures of demand (constant, increasing,
decreasing, u-shaped, inverted u-shaped and evenly fluctuating
demands) wi th the respective storage cost rate 1m = 0,2 and
the set-up costs am = 100 as basic decision cases m = 1, ... ,6
(cf. table 1).
484

Table 1: Basic decision cases

X1m x2m x 3m x4m x 5m x 6m x 7m x 8m 1m am

m=l 300 300 300 300 300 300 300 300 0,2 100

m=2 66 133 200 267 333 400 467 534 0,2 100

m=3 534 467 400 333 267 200 133 66 0,2 100

m=4 550 350 200 100 100 200 350 550 0,2 100

m=5 50 250 400 500 500 400 250 50 0,2 100

m=6 500 100 500 100 500 100 500 100 0,2 100

In order to get a condensed idea of the course of the

information production function fI for such discrete descision
cases, the data inputs x4m, 1 m and am, that are respectively
the demand of the fourth period, the storage cost rate and the
set-up costs, are systematically varied in a single-stage sen-
sitivity analysis within the following ranges and steps star-
ting from the basic decision cases

x 4m ,sl x4m - 100 + sl sl 1,2, ... ,200,

1 m,s2 1m - 0,2 + 0,01S 2 s2 1,2, •.• ,200,
a m,s3 am - 100 + s3 s3 1,2, ... ,200.

This leads to 18 sensitivity runs (m, sll)' m = 1, ... ,6, II =

1,2,3. In the course of the 18 sensitivity runs altogether
6'200'3 = 3600 different decision cases are examined, which are
solved alternatively with SMA, PPA, and WWA. This then are
altogether 10800 information processing activities of dynamic
lot-sizing which have to be ordered by the original efficiency
to illustrate the information production function fI by
sections. The information processing vectors were produced on a
IBM PS2/70 (processor: Intel 80486, 25 MHz) with the software
of the three solution algorithms programmed on our own.
 485

3.2. Results

The analysis outlined above provides the following findings and

results.

(1) The data processing input quantities v 20 and v 21 in the

information processing vectors v~ hardly provide any indi-
cations for a qualified evaluation of these activities, for
they are almost equal in all vectors. The reason for this
is that the equally sized times for the computer and its
user are - due to the rigid combination of computer and
personnel - almost exclusively caused by the times for put-
ting in data, which are normally in the range of minutes.
To the contrary, the computing times for the solution pro-
cedures within these computer operation times comprise only
milliseconds, so that the computing times in the planning
situation mentioned above can hardly provide a sensible and
interesting tool of differentiation.

(2) The dependencies of the actual set-up and stock keeping

costs C from the announced variations of parameters are
illustrated in the figures 1-8 for the sensitivity runs
(m,s2)' m = 1, ... ,6, as well as (3,sl) and (3,s3)' Here
already considerable changes can be seen, if the data
inputs are varied. At the same time one can realize that
among the three solution procedures for every comparable
planning case the Wagner-Whitin-algorithm provides the
cost-optimal solution of the dynamic lot-sizing, as is well
known from formal considerations in the literature (OLIVIER
1977, KISTNER/STEVEN 1990, SCHENK 1991, SCHNEEWEISS 1981,
TEMPELMEIER 1988) and is once again confirmed by the
results of the sensitivity runs. That means that in the
given planning cases the Wagner-Whitin-algorithm forms the
information production function fI, which is presented in
figure 1-8 in the respective sections.
 ....
(X)
a>

C
-
1000,1--------------------------------------------------~

800 f-

600 I-

400 I-

Silver-Meal Aig.
200
Part-Period Aig.
Wagner-Whitin Aig .
O~I----~------~-- __~____~______~____~______ L __ _ _ __ L_ _ _ _~L__ _~
I S2
o 20 40 60 80 100 120 140 160 180 200

Fig. 1: Sensitivity Run (m-1,s2)

 C
1000 " ---------------------------------------------------,

800

600

400

Silver-Meal Aig.
200
Part-Period Aig.
Wagner-Whitin Aig.
,~ ~ ~ ~ ~
o ____ ______ ____ ______ ____ _ L_ _ _ _ _ _L __ _ _ __ L_ _ _ _ _ _~_ _ _ __ L_ _ _ _~
S2
o 20 40 60 80 100 120 140 160 180 200

Fig. 2: Sensitivity Run (m=2,s2)

.j>.
ex>
-..j
 i
C
1000 i~--------------------------------------------~

800 ,'" ",-\

600

400

Silver-Meal Aig.
200
Part-Period Aig.
Wagner-Whitin Aig.

o I I 52
o 20 40 60 80 100 120 140 160 180 200

Fig. 3: Sensitivity Run (m-3,s2)

 C
1000 f,-------------------------------------------------~

800

,.
f :

600

400

Silver-Meal Aig.
200
Part-Period Aig.
Wagner-Whitin Aig.

o f f S2
o 20 40 60 80 100 120 140 160 180 200

Fig. 4: Sensitivity Run (m,.4,s2)

&;
<0
 .,..
:g
C
1000 ~i------------------------------------------------~

800

600 "

400

Silver-Meal Aig.
200
Part-Period Aig.
Wagner-Whitin Aig.
--~ I 52

20 40 60 80 100 120 140 160 180 200

Fig. 5: Sensitivity Run (m=5,s2)

 C
1000 1~--------------------------------------------------'

800

600

400

Silver-Meal Aig.
200
Part-Period Aig.
Wagner-Whitin Aig.
o ~I _ _ _ _- J_ _ _ _ _ _~_ _ _ _- L_ _ _ _ _ _~_ _ _ _~_ _ _ _~_ _ _ _ _ _L __ _ _ __ L_ _ _ _~_ _ _ _~
52
o 20 40 60 80 100 120 140 160 180 200

Fig. 6: Sensitivity Run (m=6,s2)

."..
~
 ~
C
700 ~1----------------------------------------------~

650 I-

600 I-

550

500

Silver-Meal Aig.
450 Part-Period Aig.
Wagner-Whitin Aig.
I i I I I I I -'---_---11 S 1
400
o 20 40 60 80 100 120 140 160 180 200

Fig. 7: Sensitivity Run (m:03,s1)

 C
1200 ~i----------------------------------------------~

1000

800

600

400

Silver-Meal Aig.
200 Part-Period Aig.
Wagner-Whitin Aig.

o " , S3
o 20 40 60 80 100 120 140 160 180 200

Fig. 8: Sensitivity Run (m a 3,S3)

.j>.

~
494

(3) If in all sensitivity runs the isolatedly varied input data

are ordered according to their values in the range of the
real numbers it can be seen that the sections of the
information production function graph i I show courses in
analogy to concavely increasing functions. This may be less
surprising for the sensitivity runs (m,s2)' m = 1, • .• , 6
(figure 1-6), in which for given basic patterns of demand
the storage cost rates are systematically increased, as in
the beginning, one tends to supply for the total demand for
eight periods all at once in the first period whereas later
on with higher storage cost rates the demands are supplied
period by period . Exactly the reverse phenomenon can be
observed if the set-up costs a are systematically increased
(cL sensitivity run (3,s3) in figure 8). However, it is
surprising that the costs develop in the same way if, with
given storage cost rates and set-up costs within the basic
patterns of demand, the demand of the fourth period is
raised steadily (sensitivity run (3,sl) in figure 7) . All
sensitivity runs not graphically presented here principally
show the same courses of the information production
function fI .
 495

References:

DeMatteis, J.J.: An Economic Lot-Sizing Technique, I, The Part-

Period Algorithm, in: IBM Systems Journal, 1/1968, pp.
30-39.

Fandel, G.: A Theoretical Basis for the Rational Formation of

Production Planning and Control (PPC) Systems, in:
Fandel, G. , Zapfel, G. (eds.) : Modern Production
Concepts - Theory and Applications, Berlin, Heidelberg,
New York et al. 1991, pp. 3-17.

Fandel, G.: Analysis of Production Planning and Control (PPC)

Systems as an Efficient Combination of Information Acti-
vities, in: Fandel, G., Gulledge, Th . , and Jones, A.
(eds.): New Directions for Operations Research in
Manufacturing, Berlin, Heidelberg, New York et al. 1992,
pp. 40-59.

Glaser, H.: Verfahren zur Bestimmung wirtschaftlicher Bestell-

mengen bei schwankendem Materialbedarf, in: Angewandte
Informatik , 12/1975, pp. 534-542.

Kistner, K.-P., steven, M.: Produktionsplanung, Heidelberg

1990.

Knolmayer, G.: Zur Bedeutung des Kostenausgleichsprinzips fUr

die Bedarfsplanung in PPS-Systemen, in: Zeitschrift fUr
betriebswirtschaftliche Forschung, 5/1985, pp. 411-427.

Mendoza, A.G.: An Economic Lot-Sizing Technique, II,

Mathematical Analysis of the Part-Period Algorithm, in:
IBM Systems Journal, 2/1968, pp. 39-46.

Olivier, G.: Material- und Teiledisposition - die mathemati-

schen Methoden fUr ein Programmsystem, Bonn 1977.
496

Schenk, H.Y.: Entscheidungshorizonte im deterministischen dyna-

mischen Lagerhaltungsmodell, Heidelberg 1991.

SchneeweiB, Ch.: Modellierung industrieller Lagerhaltungssy-

steme, Einftihrung und Fallstudien, Berlin, Heidelberg,
New York 1981.

Silver, E.A., Meal, H.C.: A simple modification of the EOQ for

the case of a varying demand rate, in: Production and
Inventory Management, 10/1969, pp. 51-55.

Silver, E.A., Meal, H.C.: A Heuristic for Selecting Lot Size

Quantities for the Case of a Deterministic Time-Varying
Demand Rate and Discrete Opportunities for Replenish-
ment, in: Production and Inventory Management, 2/1973,
pp. 64-74.

Tempelmeier, H.: Material-Logistik: Quantitative Grundlagen der

Materialbedarfs- und LosgroBenplanung, Berlin, Heidel-
berg, New York et ale 1988.

Wagner, H.M., whitin, T.M.: A Dynamic Version of the Economic

Lot Size Model, in: Management Science, 10/1958, pp. 89-
96.

wemmerlov, U.: A Comparison of Discrete, Single Stage Lot-

Sizing Heuristics with Special Emphasis on Rules Based
on the Marginal Cost Principle, in: Engineering Costs
and Production Economics, 1/1982, pp. 45-53.

Zoller, R., Robrade, A.: Dynamische Bestellmengen- und

LosgroBenplanung Verfahrenstibersicht und Vergleich,
in: OR Spektrum, 4/1987, pp. 219-233.
VI. Quality
 HOW MUCH QUALITY IS ENOUGH?

JUles R. DORROH
Louisiana State university
Baton Rouqe, Louisiana 70803 USA

Thomas R. GULLBDGE
Georqe Mason University
Fairfax, virqinia 22030 USA
Iforaan It. WOKER
university of Kississippi
University, Kississippi 38677 USA

1. preliminary Remarks

Managers are often faced with a decision about the amount of

resources to devote to enhancing the production process. The
decision involves removing resources from immediate production,
and enhancing these resources so that they are more productive
for later production. Job training falls into this category. A
decision is made by management to divert resources from the
production process to knowledge creation, and ultimately to
enhancing the quality of the production process. This scenario is
distinctly different from the learning-by-doing scenario that has
received much attention in the economics and industrial
engineering literature. Learning-by-doing is perceived as a cost
reducing by-product of production processes. In many studies this
by-product is modeled as a simple function of time or of
cumulative production experience. We present an alternative
explanation where managers decide what resources to devote to
knowledge acquisition. Our model explains resource and output
behavior for a specialized production situation; production to
contractual order.

There are several interesting questions that should be addressed

by management. How much investment in quality enhancement should
the firm incur? Should the firm invest throughout the production
program? Does an investment in quality lead to lower total cost?
We use a simple production model to gain insight into the answers
to these and other questions. We show that the cost minimizing
producer will choose to invest heavily in process improvement
500

early in the production program, but the rate of investment will

decline over time. other influences on the optimal level of
quality are explored by examining the optimal time paths of the
control and state variables in a comparative dynamics analysis.

2. Introduction

This paper examines a dynamic production situation in which a

decision is made to invest resources in the production of
knowledge. The knowledge is treated as an intermediate product
that is used as an input in the production of output. The overall
effect is an enhancement of the production process. This idea is
similar to MODY's (1989) "investment in engineering" and
KILLINGSWORTH's (1982) "investment in worker training." This
model is consistent with an explicit management commitment to
system quality. As such, it stands in dramatic contrast to the
traditional learning-by-doing hypothesis where process
enhancement, and hence cost progress, is an endogenous by-product
of production experience.

We illustrate these ideas by modeling a firm that minimizes the

cost of producing a small number of specialized units. The
product is defined by a contractual order. The prototypal example
is aircraft production, but the model could apply as well to
other projects; e.g., ships, jet engines, construction projects,
software projects, etc. Since the contract defines the production
time interval, the design, and the number of units to be
produced, we call this made-to-order production.

We also point out that the discussion of quality in this paper

does not include modifying the contracted product. We are purely
interested in enhancements to the production process. These
enhancements flow from explicit management decisions to enhance
the process in any resource consuming way.

We hypothesize that the firm will choose to invest heavily in

production process improvement early in the production program,
but the rate of investment will decline as the program matures.
The idea is that a stock of knowledge that is available early in
 501

the program is more valuable than the same stock later in the
program. The remainder of the paper is organized as follows. The
next section presents some history of the made-to-order
production problem and motivates the hypotheses to be examined.
Then, we present a control theoretic model of the production
situation and use the optimal trajectories to examine the main
hypothesis.

3. Learning and Knowledge Discovery

When producing to contractual order, learning is often an

important determinant of cost. The background literature is
extensively discussed in GULLEDGE/WOMER (1986), hence we move
directly to the main ideas of the paper. This paper is concerned
wi th knowledge creation as opposed to serendipitous learning.
While this distinction is often unclear in the literature, it is
central to the understanding of our model. ROSEN (1972),
KILLINGSWORTH (1982), and MODY (1989) clearly model resources
that are explicitly diverted from output production to knowledge
discovery. This could be investment in training as noted by
Killingsworth, or it could be a more aggregated policy such as
Mody's investment in engineering capabilities. The distinction
between these papers and many others, especially those in the
industrial engineering learning curve literature, is that the
amount of resources to divert to knowledge creation is a decision
variable in an optimization framework.

The distinction may be further elucidated by considering the work

of MUTH (1986) . Muth shows that power function relations
(learning curves) can result from random search over production
possibilities. Our results do not emphasize these details of the
knowledge creation process. Instead we show the relation between
management decisions to use resources to produce productivity
enhancing knowledge (perhaps by searching) and the resulting
impact on production cost. Other examples of this type of
resource diversion are the creation of on-the-job training units,
the enrollment of employees in educational classes, or the
creation of quality circles. The type of knowledge creation that
we consider requires a managerial decision to divert resources
502

from the production of output to the production of knowledge;

that is, an investment in organization capital and quality.

4. A Model of Made-to-Order Production

Let q(t) denote the firm's rate of output at time t, and let Q(t)
denote the cumulative output at time t. We are interested in the
situation where the firm produces both knowledge and output.
Hence, let t(t) be the rate of knowledge production at time t,
and let L(t) denote the cumulative stock of knowledge at time t.

Assume a variable composite resource with use rate x(t) at time

t. For our analysis we assume that the contractual agreement
defines a short planning horizon, and economic conditions are
such that constant relative prices prevail for the class of
resources denoted by x(t). Furthermore, we assume that through
the managerial decision process, the firm decides which part of
x(t) to devote to knowledge creation, xl(t), and which part to
devote to the production of output, xq(t). Hence, x(t) = xl(t) +
Xq(t). Given the above assumptions, cost may be measured in units
of the variable composite resource; i.e.,

where p is the resource discount rate.

Let output rate be determined by a concave production function

that is homogeneous of degree one in Xq(t) and L(t), so that

q(t)/L(t) = f[xq(t)/L(t)] (1)

This specification is similar to the learning curve; however,

equation (1) is driven by the cumulative stock of knowledge as
opposed to cumulative output. Also, let the rate of knowledge
production be determined by a concave production function that is
homogeneous of degree one in xl(t) and L(t), so that
 503

Ht)/L(t) = g[x,(t)/L(t») (2)

The following optimization problem is used to examine the

knowledge creation problem:

Subject To:
q(t) =L(t) f[xq(t)/L(t»),

~(t) = L(t) g[x,(t)/L(t»)'

with boundary conditions; Q(O) = 0, Q(T) = V, L(O) = H, and L(T)

= free. In this formulation, T is the terminal time, PL is the
salvage price of knowledge, and V is the contracted number of
units to be produced.

For this paper a form of the Constant Elasticity of Substitution

(CES) production function is used. This function was first
proposed by ARROW/et al. (1961). Our versions are

q(t) =aL(t)[cx(Xq(t)-y + (1-cx»)~, (4)

1 L (t)

and

(5)

If equations (4) and (5) are substituted into the optimization

problem defined by (3), a numerical solution is possible. The
solution to this .problem, and related problems are discussed by
504

DORROH/et al. (1992). For this paper, we concentrate only on the

quality investment implications; but we do note that the results
of MANGASARIAN (1966) assure that the second-order conditions are
satisfied.

5. Model Analysis

The time paths for knowledge creation and resource usage depend
in critical ways on the model's parameters. These time paths are
examined relative to those generated from the base-case parameter
set that is presented in Table 1. Hence, the analysis that
follows is an exercise in comparative dynamics.

Table 1
Base-Case Parameter Values

1.00 & = -.40 p .05

.10 cr = .50 .50

y = .40 p= .50 L(O) = 2.00

The sensitivity of the time paths to parametric changes is

examined by generating new trajectories after varying selected
model parameters.

The production of output is motivated by the contractual

requirement to have produced V output units by time T. The
instantaneous production of output is characterized by
diminishing returns to resource use; that is, an increase in
output rate requires a more than proportional increase in input
rate.

Discounting requires the marginal costs of producing output to

increase over the life of the program. Given a positive discount
rate, the contractor is motivated to delay incurring costs, delay
production, and to increase output rate late in the program. This
is true in learning-by-doing models [DORROH/et al. (1986»), and
the result is independent of knowledge creation. However,
 505

knowledge creation, as presented in this model, provides even

more incentive to delay output production.

The possible sensitivity analyses and parameter combinations are

limitless, but the major results can be summarized in two
figures. The trajectory of resources devoted to the production of
output is presented in Figure 1, and those devoted to knowledge
creation are presented in Figure 2. These two isoquant maps
indicate that the optimal strategy is to invest early in process
enhancing knowledge creation, and later into output production.
This result is robust for reasonable admissible parameter values.

The relationship between the expansion path and output rate is

also illustrated with the isoquant map presented in Figure 1.
This figure also provides a linkage to the industrial engineering
learning curve literature. These isoquants can be regarded as a
family of learning curves indexed by output rate. However, unlike
most learning curves, these curves relate resource use to a stock
of knowledge that is not proportional to cumulative output. That
is, we provide an analogous learning curve that relates resource
requirements to the stock of knowledge that results from a
manager's decision to invest in knowledge creation.

The position and the shape of the expansion path depends on the
quantity of knowledge available at the beginning of the program.
High initial knowledge permits a reduction in x q , both directly
and indirectly, by easing the production of additional knowledge.

Knowledge is valuable in this model because increased knowledge

1. lowers production cost for output,

2. raises residual values for knowledge, and
3. lowers production costs for future knowledge.

At each point in time the benefits from additional knowledge must

be balanced against the cost of producing more knowledge. If the
residual value of knowledge is high relative to the cost of
producing knowledge, then the production plan may call for
knowledge to be produced up to the end of the program. At lower
506

--
30

cr
a, -/.00
X
oz- .10
z
-I0-
0
y -.40

:;)
8--.40
0
0 a -.50
0::
0- 20 - 13· .50
I-
:;)
0- P - .50
I-
:;) p •. 50
0
0 L(O)· 2.00
I-
0
w
t-
O
> 10 -
W
0
V>
W
U
0::
::>
0 /' q=5
V>
w ~
..
--- _-----_.
. •.. q=4
0:: qa 3
Iq=2 I
0 '
2 3 4 5 6 7 8

KNOWLEDGE (L)

Figure 1. Knt:7«ledge iscx.Juant map with expansion path in:licat~

the resources devoted to outp.Jt production.
 507

- 3

- a l-1.00
~
X

Z a2-· 10
-
0
~ y-.40
u
:::> 8· -.40
0
0
0:: a =.50
0..
W 2
(!) (3 - .50
0
w P -.05
...J
~ p- .50
0
z L(O)- 2.00
~

0
~
0
W
~
0
>
W
0
CJ)
W
u
0::
::>
0
CJ)
W
0::
0
2 3 4 5 6 7 8

KNOWLEDGE (l)

Figure 2. Knowledge isoquant map with expansion path in:licatin;J

the resources devoted to kncMledge production.
508

residual values, knowledge must be sufficiently valuable in the

production of output to justify spending resources on its
production. Thus, knowledge mayor may not be produced at the end
of the program, and in extreme cases may not be produced at all.

In any case the value of producing more knowledge decreases

toward the end of the production program, and the resources
required to produce more knowledge also decline. The implication
is that fewer resources are devoted to producing knowledge during
the late stages of production. The analysis of the time path of
resources devoted to knowledge creation forms a major portion of
this sensitivity analysis.

Figure 2. relates the resources devoted to knowledge production

to the level of knowledge in the base case. Since L increases
with time, the level of resources used to produce knowledge is
read from the curve which increases and then decreases. The
isoquants in the figure depict the knowledge acquisition rate
achieved for alternative resource-knowledge combinations. For the
base case, resources are devoted to the production of knowledge
early in the program. As increased knowledge permits easier
knowledge production, the resources devoted to knowledge creation
increase, increasing the rate of knowledge production. As the
value of future knowledge declines toward the end of the program,
the resources devoted to knowledge creation and the rate of
knowledge production decline.

6. Conclusions

We have presented an alternative model of process improvement in

production. The model differs from the usual model of process
improvement, the learning curve, in the following way. In most
learning curve models, required resources are assumed to be
proportional to time or cumulative production experience. They
imply that learning is just an endogenous by-product of the
production process. For our model, knowledge creation requires a
managerial decision. That is, the rate of . knowledge production is
a control variable.
 509

We have used the model to demonstrate that, for a reasonable

range of parameter values, the optimal strategy is to invest
early in quality enhancing knowledge creation, and at some later
time, invest in the production of output . This result is
consistent with the human capital literature where it is noted
that education and training are dominant early in an individual's
life cycle, and at later life cycle stages, production is the
dominant activity.

Our model is sufficiently robust to describe the types of time

paths that are typically observed on made-to-order production
programs. This is a major disadvantage of learning curve based
models where the trajectories are all exponential functions .
Furthermore, the model relationships were explained as results of
rational managerial decision making, as opposed to assuming that
resource reductions due to learning are merely by-products of the
production process.

One final feature of this model is that it demonstrates the

importance of the initial endowment of knowledge. If our firm
begins production with a large endowment of knowledge, fewer
resources must be devoted to knowledge creation in order to
produce a given quantity of output at a given cost. The quality
implications are obvious.

7. References

Arrow, K.J., H.B. Chenery, B.S. Minhas, and R.M . Solow, 1961,
Capital-labor sUbstitution and economic efficiency, Review of
economics and statistics 43, 225-250.
Dorroh, J.R., T.R. Gulledge, and N.K. Womer, 1986, A
generalization of the learning curve, The european journal of
operational research 26, 205-216.
Dorroh, J.R., T.R. Gulledge, and N. K. Womer, 1992, Knowledge
creation: an alternative to learning-by-doing, Unpublished
manuscript, 1992.
Gulledge, T.R. and N.K. Womer, 1986 . The economics of made-to-
order production (Springer-Verlag, Berlin).
Gulledge, T.R., N.K. Womer, and J.R. Dorroh, 1984, Learning and
costs in airframe production: a multiple-output production
510

function approach, Naval research logistics quarterly 31, 67-85.

Killingsworth, M.R., 1982, Learning by doing and investment in
training: a synthesis of two rival models of the life cycle,
Review of economic studies 49, 263-271.
Mangasarian, O.L., 1966, Sufficient· conditions for the optimal
control of nonlinear systems, SIAM journal on control 4, 139-
152.
Mody, A., 1989, Firm strategies for costly engineering learning,
Management science, 35, 496-512.
Muth, J.F., 1986, Search theory and the manufacturing progress
function, Management science 32, 948-62.
Rosen, S., 1972, Learning by experience as joint production,
Quarterly journal of economics 86, 366-382.
 DECISION METHODOLOGY FOR PRODUCT QUALITY
IN FLOW LINES

Layek Abdel-Malek

and

Xiuli Chao

Division of Industrial and Management Engineering

New Jersey Institute of Technology

Newark, NJ 07102

USA

1. Introduction

The need to compete in a global market, where only products of low cost and high

quality can capture their share of sales, has stimulated renewed interest in quality control.

Since the birth, in the 1940's, of what is known today as statistical quality control, re-

search in this area has focused mainly on the development of standards, charts for process
control, acceptance sampling plans, and inspection strategies, see Abdel-Malek and Benes

(1991), Banks (1989), and Kennedy, et .al. (1987) . As computer control of manufacturing

permeated industry in the decade 1970-1980, the possibilities and advantages of sequential
on-line quality control have become evident. Taguchi (1986) is credited with pioneering this

philosophy which has led to improved performance for several manufacturers. Motivated

by the current emphasis we develop in this study quality control policies and inspection

strategies for one of the most frequently encountered means of production; that is flow

lines .
512

Typically a flow line consists of a series of work stations in tandem. All arriving
jobs follow the same sequence of processing. Nevertheless, despite similarities, inspection
strategies can vary from line to line depending on the nature of the product. In this paper,

we will introduce inspection and control models for flow lines where

(1) The output of a work station is not affected by the incoming product quality; i.e.

as long as the incoming product is conforming with the quality standards, there is a known

fixed probability that the outcome of its processing is successful (process capability).

(2) The conformance to quality standards of a particular station will be dependent

on its predecessor's acceptance range (that is, even the product successfully passes the

inspection, its particular dimension still could effect the outcome of the successor station).

The objective of the aforementioned models is to develop quality control strategies to

optimize the efficiency of the production process for these types of flow lines. The decision

variables are: the location of inspection stations, the portion of the output to be sampled
after each station, and decision rules to determine whether to process further a product or

discard it.

The taxonomy of this work is as follows: We begin by reviewing the pertinent litera-
ture. Then we analyze different types of flow lines and develop their various quality control
strategies. This will be succeeded by the concluding remarks.

2. Literature Survey

Interest in quality control has necessarily produced a vast literature, so we limit this
review only to some of the papers which has taken up issues of interest to this investiga-

tion. Lindsay and Bishop (1964) considered the location of inspections at particular stages

of production. Seeking to answer the question"When, where, and how much should we

inspect?" . Under the assumption of inspection cost linearity, they developed a dynamic
programming algorithm for allocating screening efforts to achieve minimal inspection cost.
 513

White (1966), (1969) arrived at similar inspection plans by using a "shortest route" model.

Peters and Williams (1987) also analyzed monitoring efforts in multi-stage production by

using dynamic programming heuristics. However, their model was concerned with cost

minimization rather than White's profit maximization. Along similar lines, Shtub and

Chakravarty (1987) applied a shortest route algorithm for the strategic allocation of in-

spection effort in a serial multi-product system.

In a series of papers , Bhat and Ram (1990), Bhat and Sanders (1990), introduced the

use of Markov chains to help in developing sampling and acceptance plan characteristics,

and to determine the average outgoing quality. They used a two-state process describing

runs of successes and failures and augmented it to cover the sample size in use. But they

did not consider the question of cost.

Bulgak and Sanders (1990) examined the performance of an assembly system with

"automatic" inspection and repair loops. Their paper developed an approximation for a

general queue using renewal theory, and handled the "splitting of paths" due to detected

defects in assembly, and to the subsequent recycling of units to earlier nodes.

The arrival of the Taguchi (1986) philosophy of statistical process control and on-line

inspection have also led to critical studies of these ideas for their validity and impact on

particular situations. Their promotion of statistical methods for on-line control clearly
points in the right direction. Nevertheless, Vining and Myers (1990) have claimed that the

method followed by the Taguchi school (viz. optimizing a mean response while minimizing

the variance) is not always statistically sound.

As can be seen, not enough consideration has been given to the nonlinearity of in-

spection costs, the effect of prior quality information on the inspection strategy or the

interdependencies among work stations on product quality. This is our intent in this work.
514

3. The Models

In this section we introduce three types of quality control strategies for flow lines.

Based on the particulars of each line, these strategies should define where to inspect, the

portion to be inspected, process or discard an item and the acceptable standard limits of
work station output. These strategies will be derived from the line work station's specifics,

product standards and its processing dependencies, and the various machining and raw

material costs. In the following section a more detailed description for some of these

parameters as well as other necessary definitions is provided.

3.1 Nomenclature

Consider a flow line consisting of N work stations indexed by 1,···,N. The

following notations will be used throughout:

'i= the probability that the product of station i is a success (known in the quality

control literature as process capability).

d = the cost of supplying a defective at output,

V =the revenue of an acceptable item in the yield,

R=raw material cost per item,

mi=machining cost of station i.

3.2. Inspection in flow lines with independent process capability

Consider a flow line consisting of a series of stations with known process capabilities

'i, i = 1,···, N. At each work station a fraction of its output is to be inspected, and

the defective parts are discarded. Our objective in this model is to find where to locate

inspection stations along the line, and the proportion of parts to be inspected so as to

minimize the cost.

 515

Assume that a fraction Ui of the output of station i is inspected for overall acceptability

of the procedures carried out by all preceding stations, where detected unacceptable items

are discarded. The remainder of the output, a fraction 1 - Ui, proceeds to the following

station. Assume that no attempt is made to change 5i on the bases of inspection, and the

u;'s are viewed as control variables to be chosen in the unit interval. Obviously Ui =0
implies that output of stations i is not inspected.

Define Pi as the fraction of acceptable in the output of station i. Further define qi as

the fraction of defectives in the output of station i, and ri the fraction of parts "absent"

due to inspection and removal from stations 1 through i. Let ro = 0, and assume Po and

qo are known. Note that po is the fraction of acceptable of the raw material input. For

convenience let us assume that Po = 1, i.e., the raw material is perfect. Then one can show
that the fraction acceptable in the output of station i + 1 can be expressed as

i = 1"" N - 1,

and the fraction absent is

i=I,···,N-l.

Note that Pi + qi + ri = 1, for i = 0"", N.

Assume that tPi(Ui) is the cost for inspecting a fraction Ui at station i, which is an
increasing convex function, and ri is the salvage value of an item discarded after processing
at station i.

From these assumptions, one can verify that Pi =nj=l 5j, i =1"", N, and
•
qi+1 = qi(1 - ui+d + II 5j(1 - sj+d(1 - Uj+d,
j=l
516

Consequently,
i-I i i-I
qi = L II 'j(1- 'HI) II (1 - U1)
i=Oj=1

i j-l
ri = L Uj(q;-1 + Po II ,.l(1 - Ij)).
j=1 i=1

The total average net profit may be computed as follows:

N N N
C(Ul! ... ,UN) =R- VPN + Lmi(Pi + qi) - L
1i(ri - ri-I) + I/Ii(Ui) + L dqN
i=1 i=1 i=1

N N i N i N i-I i i-I
= [R-V II 'j+ Lmi
j=1 i=1
II Ij+ Li=1 1i(I-'i) II liJ+ Lmi i=Oj=1
j=1
L II 'j(1-'j+d II (l-uI)
i=1 i=1

N i-2 i i-I N i-I

+ L 1i Ui L II Ij(1 - li+d II (1 - U1) + L 1i Ui(l- 'i) II (1 - 'i)
i=1 i=Oj=1 i=1 i=1

N N-l i i-I
+L I/Ii(Ui) + d L II Ij(l- 'HI) II (1- U1)
i=1 i=O j=1

The terms that in the bracket [.J is independent of the the decision variables u;'s, hence
it suffices to minimize the rest of the cost function. For notational convenience define
Xi = 1 - Ui, then it is easy to see that the cost function could be conveniently expressed as
N N N
C( X l,···,XN) = Lai II Xi + LbiXi + Ll/li(l- x i), (1)
i=1 i=1 i=1 i=1

with a;'s and b;'s being the appropriate coefficients of the 'i'S, m;'s, 1;'S and d. Therefore
the quality control problem for this serial production model is equivalent to the following

mathematical programming problem:

man C(Xl,X2,···,XN)

s.t.

o::; Xi ::; 1, a = 1, ... ,N

where C(Xl,··· ,XN) is defined in (1).
 517

This nonlinear programming problem is known to have no explicit form solution.

Iterative algorithms, however, can be developed to solve it. In the case of linear inspection

cost, it can be shown that the optimal solution is reached at boundary point. That is, at

each station, either inspect all the product, or do not inspect at all. This phenomenon is
called Bang-bang control.

Before we present the algorithm to compute the optimal solution, we provide intuition

for the methodology. Note that if u = (U1, U2,' .. , UN) is an optimal solution for the

problem, then its partial derivatives vanish at this point if it is an interior one, otherwise

this point should be on boundary. In other words , it should be either satisfying

N 1
L • 0
ai·
0'-1 Xi + bi
X ,
= 01,1 (
'+'i 1- )
Xi , (2)
.IE='

or

xi=Oor1, i = 1,· . . ,N. (3)

Therefore, we can construct a sequence using equation (3) and iteratively compute x~n), i =
1,···, N. This should converge to some value that satisfies either (2) or (3). Thus they
are the optimal solutions.

Algorithm:

Let us define I i (.) as the inverse function of derivative t/JH.).

Step O. Assign initial values to x~o), for i =1,· .. , N, and k =O. Set initial € > O.
Step 1. For i = 1,···,N, compute
N 01 (1)
-(k+1) _
x, -
1 _ [o('C"'
, ~ a,
0 i=1 Xi
(1)
+ bo, ).
l=i Xi

If lk+1)
I
> 1, the x(k+1)
I
= l' if X(k+1) < 0
1 t ,
then X(k+1)
t
= O· and otherwise X(k+1) = x(k+1)
' t t '

Step 2. If max{lx~k+1 - x~1)1} < €, stop. Otherwise, set k = k + 1, and go to Step 1.

If we consider an isolated station, say station one, then it can be thought of as a special

model with N = 1 with a non-zero qo. In this case substitute N = 1 in the formulas before,
518

we get P1 = 'lPO, q1 = (1 - ut)[qO + po(1 - ,t)J, and

where Xl =1 - U1 . Hence if we denote k = d(1 - po) + po(1 - ,t)(d - 'Yt) - 'Y1qO, then
if I(k) > 1, then Xl = 0 so U1 = 1, that is, inspect everything; if I(k) < 0, then Xl = 1"
U1 = 0, so the optimal inspection policy is to inspect nothing; if 0 ~ I(k) ~ 1, then

Xl = 1 - I(k) and U1 = I(k) is the optimal proportion to be inspected at the station.

3.3. Flow lines with dependent process capability and full interstage inspection

In this type of flow lines we assume that there is hundred percent product inspection
between the successive work stations. The semi finished product is categorized after work
station j as either acceptable or not. This is characterized by an attribute X, for which
there is an acceptable tolerance limit X E A. After the attribute is observed a value X is
known to us. An item conformance to the quality standard of the subsequent work station,
i + 1, is dependent on the incoming value of x as well as w which is a random variable

describing j + 1 work station process capability. To determine the acceptability of a job

after processing by station j+ 1, the attribute Z =Z(x,w) is introduced. For an item to
continue its processing on station j + 2 the value of Z E B. This can be generalized for
j = 1,···, N. (See the following figure.)

1st attribute

x---
accept •
--==----'-

~ reject
*'
lzl--.
process for z

W
2nd atuibute

• !~
L0~-~
data
acceptance policy
 519

To clarify we consider a product consisting of a peg-and-a hole. The tolerance of the

clearance between the peg should be within specified domain for the product functionality

(similar to a piston-cylinder arrangement of an internal combustion engine),

Let us assume that process capability of the peg radius is x and that of the cylinder is

z. Then the question is one of finding the value of x beyond which the peg is discarded in

order to minimize the cost. We focus on one of the station in the serial production system,

say station i. Further we assume that assume that the quality of the incoming material

is known. Obviously only when both x and z falls into the tolerance limit the product is

then acceptable. Our objective is to provide the quality control decision criteria.

Now consider an infallible test, which is performed on an incoming unit (a peg) which

tells us whether it meets the specification X E A. In other words our information or

observation is the indicator function I{X E A}, which takes value one if X E A and

zero otherwise. The options are to process the unit in the hope of meeting Z E B (the

clearance tolerance range between the peg and the hole). In most practical situations

the two attributes X and Z are dependent, hence we assume that X and Z take joint

distribution function F(x,y) and density function /(x,y).

The cost structure is similar to the previous section. For example, R is the cost of

raw material. Since we are focusing on station i, this R is considered as the input cost to

station i. For simplicity we omit the subscript. Again d is the defective cost, i.e., X ¢ A

or Z ¢ B. Also define the machine operation cost or manufacturing cost as m, and r as

the salvage cost of an item before its processing.

Assume that :F represents the information available from observing the specification

X. Define v as the function of the information taking values 0 or I, with 1 meaning that

we process the item. Note that v depends on the information :F, therefore it is a random

variable (decision rule). Thus, the total cost for such a decision can be written as

C =R + mE[v} - VE[vl{x E A}I{z E B}} + rE[I- v} + dE[v(I-I{x E A}I{z E B})}

520

= (R+ r) + E[II((m + d - r) - (V + d)l{x E A}I{z E B})]

= (R+ r) + E[E[II((m + d - r) - (V + d)l{x E A}I{z E B})IF)]

= (R + r) + E[II[(m + d - r) - (V + d)P[x E A,z E BIF],

where we used the assumption that the quality control rule depends on the information Y

in the last equality (for the conditional expectation, see Ross (1985)). To minimize C, 11

should be one on the set of x such that

m+d-r
P[X E A,Z E BIF) > V +d '

and zero otherwise. However, since

P[X E A,Z E BIF) = {~{Z E BIX E A}, on X E A,

otherwise. '

therefore, the optimal quality control policy of this work station is: if the specification of

an item yields X tf. A , discard it; if X E A, then process it only when the conditional

probability P{Z E BIX E A} ~ m:!~l. Hence one can define the acceptance region

m+d-r
S={xIP{ZEBIX=x}~ V+d }.

3.2.1. Illustrative example: Consider the peg-and-hole problem with X being the
radius of the peg and Z being the clearance between the radii of the peg and the hole.

Assume that the joint density of specifications of X and Z follows a bivariate normal

random variable, where 1-'1 and 1-'2 are the mean of the radius of the peg and and mean

of the clearance between the two radii of the peg and the hole, and 0:'1 and 0:'2 are the

variances:
 521

then it is known that the conditional distribution function of the specification Z, given

that X = x, is a normal random variable N(1A2 + p~(x - lAd, ar~(1 - p2» with density

function

This result facilitates the quality control decision making. Since the conditional probability

value of (4) can be easily found, given the cost structure and the tolerance A and B, the

decision rule is straight forward. In fact the acceptance region S maybe computed explicitly

in the bivariate normal case. If B is the region that the tolerance is within say 8 of its

mean value, then what is needed is only to check the standard normal table to find the x

values that satisifies

where <1>(.) is the cumulative distribution function of the standard normal random variable.

3.4. Flow lines with interdependencies and partial inspections

In the previous section, it is assumed that an incoming production batch to a certain

work station is fully inspected (known quality). Nevertheless, for some situations, it may

be cost prohibitive to inspect every item (for example in cases quality cannot be inspected
except by destructive tests). In such flow lines the portion of the incoming material to be

inspected and the acceptance specification limit are both decision variables. More clearly,

the type of flow lines considered here, is similar to the one previously examined, but with
partial inspection between the work stations.

Again we focus on two successive work stations in the flow line. Assuming that the

portion to be inspected between two successive stations is u and the decision regarding

accepting or rejecting an inspected item is II. Assume further that .p( u) is the cost for
522

inspecting a fraction u of the incoming flow, and other cost assumptions are the same as

in the previous section. Then based on the analysis of last two sections, for acceptance

policy II, the total cost C( u) of inspecting fraction u of the incoming flow can be expressed

as

C(u) = R+ u«R+ r) + E[II[(m + d - r) - (V + d)P[x E A,z E BIF])

+V(l- u)(l- P[X E A,Z E B])] + d(l- u)(l- P[x E A,z E BIF]) + .p(u)

Applying the same approach as in Section 3.2, the cost function can be simplified to

C(u) = {R + VP[X E A,Z E B] + d(l - P[X E A, Z E B])}

+u(r + 1«m + d - r) - (V + d)P[y E Zlx])dx - VP[X E A,Z E B]

-d(l - P[X E A, Z E Bm + .p(u)

The first term in the cost function does not depend on the quality control decision

rules . Hence we will only consider the second part. Given the tolerance A, B, and the

interdependency of the two specifications (assuming bivariate normal distribution), then

the acceptance region S and the probability P[X E A, Z E B] can be computed from the

previous section. Consequently one can determine the following expression

k= r+ 1 «m+d-r)-(V+d)P[y E Zlx])dx-VP[X E A,Z E B]-d(l-P[X E A,Z E B]),

therefore, the optimization problem can be rewritten as

min ku + .p(u)
s.t.
 523

Let us again denote by 1(.) the inverse function of tjJ( .). Thus the optimal policy is as

follows:

if I( -k) 2:: 1, it is optimal to inspect all the items;

if I( -k) ~ 0, it is optimal to not inspect;

if 0 < I( -k) < 1, it is optimal to inspect 100I( -k)% of the incoming items.

Depending on the above rules and the result from the previous section, one can decide

whether to process the item further or discard it .

One interesting phenomenon to observe is that when the inspection cost is linear, then

the optimal inspection policy is again either inspect all of them, or inspect nothing.

4. Concluding Remarks

In this paper we presented strategies for three types of flow lines: flow lines with inde-

pendent process capability; flow lines with dependent process capability and full interstage

inspection; and flow lines with interdependencies and partial inspection. The analysis

revealed that the optimum inspection policies assume a threshold form. In the studies,

these thresholds are found to only depend on the cost structure. The results of this work

answer the questions of where to inspect, the portion to be inspected, and where to set

the acceptance limit for the various work stations.

With the advances afforded by today's sensors and controllers this work can be ex-

tended to utilize the information provided by each inspection station to adjust the var-
ious processing parameters of the line stations via feedback and feedforward techniques.

Stochastic control methodology could be beneficial in this regard.

524

5. References:

Abdel-Malek, L. and Benes, V.E., Nonlinear filtering and stochastic control in the design
of quality assurance systems. Unpublished IE Technical Report, New Jersey Institute of
Technology, Newark, NJ, 1991.

Banks, J. Principles of Quality Control, John Wiley & Sons, NY, 1989.

Benes, V.E. and Karatzas, I., "Estimation and Control for Linear, Partially Observable
and Optimization", 14, No. I, pp. 62-84, 1976.

Bhat, U.N., Ram, L. and Karunaratne, "A Sequential Inspection plan for Markov Depen-
dent Production Processes", lIE Transactions, 22, 1, pp. 56-64, 1990.

Bhat, U.N. and Ram, L., "Attribute Control Charts for Markov Dependent Production
Processes," lIE Transactions, 22,2, pp. 181-188, June 1990.

Bulgak, A.A. and Sanders, J.L. An analytical performance model for assembly systems with
automatic inspection stations and repair loops, Computers and Industrial Engineering, 18,

3, 373-380 (1990).

Kennedy, C.G., Hoffman, E. and Bond, S., Inspection and Gaging, Industrial Press, Inc.,
NY, 1987.

Lindsay, G.F. and Bishop, A.B., "Allocation of Screening Inspection Effort-A Dynamic
Programming Approach", Management Science, 10, 2, pp. 342-352, 1964.

Modarress, B. and Ansari, A., "Two New Dimensions in the Cost of Quality", The Inter-
national Journal of Quality & Reliability Management, 4,4, pp. 9-20, 1987.

Peters, M.H. and Williams, W.W., "Economic Design of Quality Monitoring Efforts for
Multi-State Production Systems", 19, 1, pp. 81-87, 1987.

Ross, S. Introduction to Probability Models, Third Edition, Wiley, 1985.

 525

Shtub, A. and Chakravarty, A.K., "Strategic Allocations of Inspection Effort in a Serial,

Multi-Product Production System", 19, 1, pp.l3-22, 1987.

Taguchi, G., Introduction to Quality Engineering, Tokyo, Asian Productivity Center, 1986.

Vining, G.G. and Myers, R.H ., "Combining Taguchi and Response Surface Philosophies:A

Dual Response Approach", Journal of Quality Technology, 22, 1, pp. 38-49, 1990.

White, L.S ., "The Analysis of a Simple Class of Multi-Stage Inspection Plans", Manage-
ment Science, 12, 9, 685-693, 1966.

White, L.S., "Shortest Route Models for the Allocation of Inspection Effort on a Produc-
tion Line" , Management Science, 15, 5, pp. 249-259, 1969.
VII. Neural Networks and Genetic Algorithms
 MUL TIPLE·STEP·AHEAD PREDICTION BY HIERARCHICAL NEURAL
NETWORKS

Tep Sastri, Associate Professor

Industrial Engineering Department
Texas A&M University
College Station, TX 77843-3131, USA

ABSTRACT:

Two methods for constructing a neural network map of stochastic systems from input-output time
series are presented. Extension of the iterated map to multiple-step predictions outside the training
data set is through use of a novel hierarchical architecture, based on forward time-shifting of the
general NARMAX model which is subsumed by the Werbos' time-lagged recurrent network in each
level of the hierarchy. The proposed methodology only requires a partial knowledge of the system
model orders. A number of numerical examples is given, one of which is on multiple-step-ahead
forecasting of an hourly municipal water consumption time series. The test cases demonstrate that the
proposed hierarchical mapping idea is valid.

1.INTRODUCTION

Identification of dynamical systems from input-output time series data can be accomplished by multi-
layered neural networks using a derivative-based learning rule [e.g. Rumelhart et al. 1986, Werbos
1990]. The majority of existing neural networks research, however, utilizes the black-box approach
to input-output mapping. Many published results show that they are quite successful for non-linear
functional interpolation within a training-sample set; the mapping accuracy, in some reported works,
can be much higher than conventional interpolation functions in sciences and engineering. At the
present time, out-of-sample neural-networks for multiple-step prediction of non-linear time series are
still not accurate, due to lack of generalization capability of the black-box mapping approach.

One way to improve upon the network multiple-step prediction accuracy is to design an appropriate
network structure which captures the time evolution of the stochastic process that generated the time-
series data. This paper presents a novel hierarchical neural network design based on the time-lagged
recurrent architecture proposed by Werbos, which encompasses the non-linear ARMAX stochastic
difference equation, for pelfOlming multiple-step prediction beyond the training set.

Following a blief inu·oduction of background information on the prediction problem and related
concepts, two neural network architectures for one-step-ahead prediction are presented in Section 2.
The novel hierarchical neural network architecture for multiple-step-ahead prediction is introduced in
 530

Section 3. Numerical examples of an application of the proposed idea are also presented: two
simulation results and one report on multiple-step-ahead prediction of a real time series are given. The
final section concludes the paper.

The Prediction Problem

Predicting future values of an output time series, {Xt}' from a stochastic dynamical system is a
useful activity that is relevant to many science and engineering disciplines. In practice an optimal
predictor may be derived from a known system model, if available. A model of the system, on the
other hand, may be determined through a data analysis of historical input and output observations.
However, a good prediction model is not necessarily the same as a theoretical input-output model
fitted to data by means of a parameter estimation method. The reason is that the prime concern of
parameter estimation is on how close the estimated parameters are to the true values, whereas in
prediction the real measure of performance is on how small the prediction errors are.

The most widely used forecasting performance measure is the mean square error of multiple-step-
ahead prediction. Using past observations Xt.Xt-I,... ,Xt-n to predict Xt+m ,the optimal (minimum
mean-square error) predictor is the conditional expectation of Xt+m, E[Xt+m1xt.xt-I,. .. Xt-n].
(Extension of the defmition to a multivariate system with input variables is straightforward.) This
theoretical expectation is applicable if we either know the joint distribution for {Xt }, or have a finite
parameter model for {Xt } from which the conditional expectation can be evaluated by a computation-
efficient recursive algorithm.

In the case {Xd is a Gaussian process the conditional expectation, Xt+m = E{Xt+m I xt>Xt_!> ... }, is
lin~arand may be expressed by a linear combination
Xt+m = aOxt+alxt-l+" .+anx t - n , (l)
and the multiple-step-ahead prediction problem is to fmd the values of the constants which minimize
'" 2
Mm = E {Xt+m-Xt+m} . The main advantage of Gaussian assumption is that the least-squares
solution is a linear forecast, comprising a linear combination of past observations. For non-Gaussian
series it is possible that a non-linear forecast might be superior, in the least squares sense, to a linear
one. But this possibility is not a necessity.

The mean square error is also useful for evaluating prediction models in order to search for the best
one from a set of tentative models. By a simple analysis the mean square error can be broken into a
bias contribution and a variance contribution. To reduce bias one has to employ a model with more
parameters; but the prediction variance will increase with the number of estimated parameters. The
best model structure is therefore a tradeoff between flexibility (greater explanatory power) and
parsimony (minimal model orders). This important point is demonstrated by Kashyap and Rao
 531

[1976] in their riverflows modeling study; a model with 25 parameters which gives a much better fit
to the data does not forecast as well as a simple autoregressive model with only two parameters.

Conventional Approaches

Linear prediction, albeit being convenient and useful in many applications, cannot explain
nonhomogeneities such as local non-stationary patterns in a time series caused by time-varying
disturbances. For example, transient effects of sporadic rainfall on riverflows and persistent effects of
hot days on air-conditioning demand for electricity cannot be accurately predicted by linear models.
Prediction models that can account for nonhomogeneous patterns have to be more complex; the use of
indicator variables in statistical analysis and non-linear state-dependent or threshold models of time
series has led to significant improvements in forecasting accuracy [Priestley 1988].

For a stochastic dynamical system exhibiting non-linear and non-stationary patterns, a single model
will not be adequate in describing its behavior under various input conditions. Multiple model
approaches [e.g. Stengel 1986, Landau 1978, Athans and Willner 1973] which employ a number of
candidate models developed on theoretical as well as on empirical grounds are very viable. The
multiple-model approach to the system identification problem may be described as follows: Given a
number of possible values of the system parameters, the state, and the noise characteristics, a set of
models can be identified. To each model in the set, a hypothesis is defined in terms of specific system
parameters and the con-esponding initial conditions on the system state. To identify the best model
from the set, the Bayes' rule is used to estimate the likelihood that a specified hypothesis (and the
associated model) is con-ect for the observed data. The model having highest likelihood value is then
selected.

Success of the multiple-model identification approach is dependent not only on the initial choice of
likely models, but also on the criterion chosen for a hypothesis test. The greatest strength of the
multiple-model approach is that it allows alternate models to be developed on theoretical as well as on
empirical grounds; thus valuable physical understanding as well as empirical knowledge of or
experience about the dynamic system can be used to assist model identification. A major difficulty is
that the required probability density functions are not known and have to be estimated; if there are
several model and many parameter choices a large number of parallel models would be required and
will demand a great deal of computation.

Parallel, distributed processing devices like neural networks seem most suitable in the multiple-model
environment. Further the neural network alternative can be applied to non-stationary process
monitoring and change detection as well to model identification [Sastri 1991].

Neural Networks for Input-Output Mapping

 532

Neural networks for input-output mapping of non-linear dynamical systems, sometimes called
neuro-identifiers, are the multi-layer perceptrons with derivative-based weight adaptation schemes
[Rumelhart et al. 1986, Werbos 1990], the functional link net, the group method of data handling,
and the radial-basis function type of neural networks. Essentially, the multilayer architectures and
activation functions employed by these methods, in particular the sigmoid function, determine the
kind of non-linear approximation of the dynamical systems. The number of processing elements
contribute to the degree of accuracy required in the approximation of the desired output values.

The goal of a neuro-identification method is to indirectly derive an unknown functional relationship

from the given training samples such that the network's output is as close to the true output as
possible in the mean-square error sense. Many authors have demonstrated that learning in artificial
neural networks (ANNs) can achieve the same goal of the traditional system modeling, i.e. to obtain
an adequate input-output map of a system from given data. Nevertheless, the utility of an ANN
mapping is still limited if it does not generalize to new input values outside the training set In
addition, the danger of model misspecification is imminent, especially when the neural networks
analyst is not familiar with the dynamical system at hand.

To some extent, a priori information on the model orders is needed for the design of a neuro-
identifier. Similar modeling problems which are typical in system identification [Ljung 1987], e.g.
model misspecification errors, may also result if incorrect lagged input and output observations are
input to the neural network [Billings et al. 1992]. In practice complete information is not available for
most system identification tasks; therefore uncertainties and specification errors will lead to an
inadequate representation. One way to alleviate model misspecification errors is to incorporate
information about alternative models and model validation capability into the network structure to
facilitate model identification experiment on the computer [Sastri 1991]. Such a multifunctional neural
network should be able to select the best model from a pre-determined model set, given that adequate
amount of input-output data is submitted to ensure adequate learning. Existing multi-layer neural
networks do not yet have such capability.

Out-ol-Sample Generalization

Multiple-step prediction outside the training sample of a trained neural net, known as extrapolation in
the literature of dynamical systems, will be called out-oj-sample generalization in this paper. The
present interest of the neural networks research in non-linear time series mapping, however, is mainly
on interpolation; the term generalization in the literature is synonymous with "interpolation along the
best fit surface which generates the map" [Lowe and Webb 1990]. In other words, most current
research works are concerned with in-sample generalization. Categorically, longer than one-step-
ahead predictions are seldom discussed because the traditional purpose of dynamical system
identification has been in iterated maps for producing synthetic data with the characteristics of true
data.
 533

Extrapolation perfonnance of most existing neural networks are poor and they tend to become
unstable, when applied to non-linear and non-stationary time series. Non-linear prediction methods,
according to Lowe and Webb [1990], are "not likely to be significantly better than simple linear
methods", unless the data generator is static (i.e. stationary and detenninistic).

2.0NE-STEP-AHEAD NEURAL NET PREDICTORS

Two types of neural network architectures for input-output mapping of a dynamical system are
presented in this section: i) direct mapping, and ii) indirect mapping. Both architectures are
extendable to non-linear, non-stationary stochastic systems as well. Out-of-sample neural network
generalization, based upon these two architectures, will be introduced in Section 3.

Direct-mapping Neural Networks

In direct mapping, also called direct learning (DLN) in [Sastri 1992], the idea is to directly transmit
predetennined knowledge of process models into a neural network structure and connection weights,
thus bypassing the time consuming weight training and iterative computer experimentation. Through
direct learning, infonnation about the model orders (i.e. model structure) is used to specify the
number of nodes in a single-layer neural network and the model parameter values (or estimates) are
used directly as connection strengths. As a consequence, the network's output is exactly equal to the
conditional expectation, provided that the input-output model is linear in the parameters. Justifications
for the direct learning neural network are given in [Sastri 1992].

There are several useful non-linear models that can be represented by the DLN approach; the
polynomial nonlinear autoregressive-moving average with external inputs (NARMAX), the threshold
autoregressive (TAR), the bilinear and the quadratic Volten'a models [Priestley 1988] are all linear in
the parameters. FIGURE 1 illustrates a bivariate AR threshold DLN with m members in the model
set. The process may switch from one model in the set to another according to the following
description:
yet) = ag)+a\i)y(t-I)+ay)y(t-2)+ ... +a~i)y(t-k)+ e~i) (2)
if {Yt-l ,... ,Yt-k} E R (i) = i-th region in the k-dimensional space Rk, {e~i)} is a white noise process,

and i=I,2, ... ,m. The switching time is assumed random, because at a given time (e.g. the process
sampling instant) it is not known which model in the set is the best description of the process. Given
that on-line data windows are submitted sequentially, the neural net can be used as a process monitor
in the following manner: at each sampling time it computes one-step-ahead prediction errors
(innovations) and perfOlms multiple statistical hypothesis tests (in parallel) on the innovations
sequences [Sastri 1991]. The output node that produces uncorrelated errors (white-noise innovations)
534

with a minimum variance will identify the best model for the most recently submitted input vector.
For a bivariate time series, a two-dimensional moving data window of an appropriate number of
observations must be specified.

Binary output

Lateral
competition

Innovations
computation

Modell

Input Slab
Target vector

FIGURE 1: MULTIMODEL IDENTIFICATION BY DLN

Each submodel in FIGURE I is a bivariate AR representation, and the network output from a typical

t
subnetwork is the conditional expectation on the input data up to time t-1. For example, if k=2 the

:i:ru (r~:T is iV;:~ ~;;Oi~~~~;:g sup=ript foc conveme~) (3 )

It is clear that the direct-mapping neural network yields the exact computation of the conditional
expectation, given precise knowledge of the underlying process models and parameters. Extensions
to the general multivariate polynomial NARMAX and other models that are linear in the parameters is
straightforward [Sastri 1992].

Indirect-Mapping Neural Networks

 535

In the situation where only a partial system knowledge is available, but plenty of process
measurements is at hand, a "gray box" mapping neural network is a viable alternative to tedious
multivariate time series analysis. Such input-output map provides a universal approximation function,
depending on the non-linear activation function used at each processing element. For example, when
the transformation function is sigmoidal the resulting map yields a non-linear polynomial function of
all inputs to the neural network [Billings et al. 1992].

Under fairly weak conditions on the functions f( .) and g(.) in Eqn.(4) below, a multi-layer neural
network can be constructed to approximate the mapping over a training set,
y(t+ 1) = f[y(t),y(t-l ),oo. ,y(t-n+ I)] + g[u(t),u(t-l),oo .,u(t-m+ 1)] (4)
where m ~ nand {u(t),y(t») is an observed input-output pair of the single-input single-output (SISO)
deterministic system at time t [Narendra and Pathasarathy 1990]. The gray-box mapping may also be
defined by a single function as follows :
y(t+ I )= F[y(t),y(t-l),oo ,y(t-n+ l),u(t),u(t-l),oo,u(t-m+ 1)] (5)
The neural network becomes a universal approximator of the dynamical system similar to a
polynomial multiple regression model fitted, based on the least squares principle, to a given data set.
This type of modeling exercise is normally performed during a preliminary system analysis, when
knowledge about the underlying process is incomplete. The task of weight estimation is delegated to
the neural network training procedure (e.g. the enol' back-propagation). The learning performance
criterion is in minimizing the mean-square elTor based on the one-step-ahead prediction.
Categorically, the resulting indirect map cannot perform adequately as a multiple-step-ahead predictor
beyond the u'aining set.

Two Architectures for Indirect Mapping

A neural network map of a NARX process with a linear noise model is proposed in Billings et
al.[1992] and shown in FIGURE 2. The partial knowledge required is the model orders n ,n ,and
u y
n conesponding to the input, the output and the enor terms, respectively. The authors neither
e
explain how the enol' inputs for training the network might be obtained, nor whether the network
might be extended to include non-linear noise terms.

It is easy, on one hand, to extend the Billings et al. architecture to polynomial NARMAX simply by
connecting the enor inputs to the hidden nodes where sigmoidal transformation is performed. On the
other hand, computation of the time-lagged enor inputs will be more involved than that of the direct
learning network [Sasu'i 1992]. An iterative procedure for calculating the innovations and storing
their time-lagged values is deemed necessary for both cases. But, unlike the direct mapping approach,
the indirect map may not produce a stable enor sequence since the indirect estimates of the connection
weights may not meet a stability criterion for parameter convergence [Goodwin and Sin 1984].
536

One-step-ahead
prediction

y(t-l) u(t-l) e(t-l) e(t-ne>

FIGURE 2: AN ARMAX NEURAL NETWORK WITH LINEAR ERRORS

(BILLINGS ET AL. 1992)

Werbos et al. [1992] propose a time-lagged recurrent network (TLRN) architecture which is able to
encompass the NARMAX model. By time-lagged recurrence, a PE at time t is allowed to receive the
output from any other PEs at time..!-1. The vector of inputs to the network at time 1, X(t), is made up
of the n-vector output at time t-l, Y ( t -1) , and an exogenous m-dimensional input vector, u(t). The
network structure is based on the multi layer perceptron (MLP) with the total number of inputs equal
to N + n + 1. The general TLRN is described by the following equations:

IS i Sm

~WijXj(t) +!W'ijXj(t-l), m+lS i SN+n (6)

j =0 j =i

x;(t) = f[v;(t)) , m+lSiSN+n

~

Y i (t) = Xi+N(t), 1 S i Sn

where f denotes the sigmoid function, and Xo represents the constant 1.0. The authors suggest that
the weight adaptation may be done by minimizing square error exactly as in the error back-
propagation scheme. The network architecture is depicted in FIGURE 3.
 537

n output PEs

Fully connected
/eed10rward with
ime-lagged
edbacks to Ihe
N -m hidden PEs in I slab

FIGURE 3: THE TIME LAGGED RECURRENT NEURAL NETWORK (TLRN)

It may noted that the parameter N is left to the user to specify, which means that the topology is
dependent on a predetermined numbers of inputs and hidden nodes for a given problem. The N + n +
1 input nodes (including bias) are fully connected to each of the N - m PEs in the hidden slab and also
to each of the n PEs in the output slab. The input slab (or buffer) is composed of m external inputs at
time t and N+n-m time-lagged internal inputs (from the hidden and output layers). The total number
of nodes is therefore 2(N+n)-m+ 1. The following examples are intended to explain how the
parameter N may be determined for a given problem.

Example 2.1

Consider a general single-input single-output NARX model y(t)= F[y(t- l), y(t-2), u(t-I), u(t-2)] for
which a TLRN is to be developed to perform the one-step-ahead prediction to time t+ I, given
observed data up to time t. F is an unknown non-linear function. The partial knowledge about this
input-output model is n = I, m = 2, and that there are five input nodes in addition to the bias and the
lag-2 nodes. It follows that N+n = 4 and N = 3. Therefore, the number of hidden nodes is 1. This
network is shown in FIGURE 4a. Outputs from nodes 3 and 4 are delayed by one time period,
538

Extension of TLRN to NARMAX mapping

The TLRN architecture may be extended to include time-lagged terms other than t-l by simply adding
the following:

I!
k=2 j =i
W(k) ijXj (t-k) , m+l::; i ::;n+N

where ny is the largest lag value for y. Similarly, time-lagged inputs x.(t-k)
1
for, i =1, ... ,m and k =
1,2, ... ,n can be included in the TLRN formulation, Eqn.(6). Since time-lagged error inputs may be
u
interpreted as a special case of external inputs to the dynamical system, i.e. m = n + n , it then
u e
follows that the TLRN network can be made general to subsume NARMAX models [Werbos et al.
1992].

Example 2.2

Suppose that we are dealing with a multivariate NARX system having n = 2 and m = 6. It is desired
to c:nstruc(t~(~L1RN for the one-step-ahead prediction,

Y(t) = '" =fI Y(t-l),u(t),u(t-l)]

Y2(t) ,
The input vector for this network consists of 6 components Ul(t), U2(t), U3(t), Ul(t-l), U2(t-l), and
U3(t-l), an unknown number of time-lagged signals from the hidden slab, and 2 time-lagged output
values from the network Y1(t-l), and Y2(t-l). Choosing a single hidden node, we have N - m = 1 and
N = 7. Hence the total number of nodes is 13. This TLRN structure is given in FIGURE 4b.

A Modified TLRN Structure

There is an important advantage that TLRN can offer over the alternative architecture by Billings et
al.; the feed-forward signals present in any layer in a TLRN essentially comes from all the layers
below. In other words, anyone layer is fully connected to all layers above it, not just to the layer
immediately above as in ordinary MLP networks. Such special connections provide multiple nesting
of the sigmoidal transformation; this is equivalent to having many extra cross-product terms for all
signals within the neural network. For some highly non-linear problems, these extra connections
should increase the interpolation accuracy.

In the following section only TLRN will be used because of its greater flexibility. However, the
numerical examples of the present study have not implemented the feed-back connections from the
hidden PEs; a commercial software package used in this study does not support the TLRN structure.
 539

FIGURE 4A: A TLRN FOR A UNIVARIATE NARX

One-step-ahead
prediction

1.0
Yl t-

FIGURE 4B: A TLRN FORA BIVARIATE NARX

540

A useful modification has been made to the original TLRN in this study; the purpose is to specify the
number of input and hidden units directly in terms of m and n. The idea is as follows: create lag-I
delayed inputs from those at time t and replace the inputs to nodes #(m+ I) to #(N+n) in FIGURE 3
by these signals. As the consequence the toal number of input nodes would be 2m + I. Equating the
original number of inputs, N + n, to 2m and solve for N, it follows that N =2m - n. Now it is easy
to determine the number of hidden PEs, which is equal to (2m - n) - m =m - n, and the total number
of nodes in the modified TLRN becomes 3m + 1.

Consider for example the following non-linear ARMA (NARMA) model

y(t) =a1y(t-l) + a2y(t-l)e(t-l) + a3y(t-1)y(t-2) + e(t) (7)
where {e(t)} is a white noise process with mean Il and variance ~. The one-step-ahead least-squares
predictor is given by
Y(t+ 1) = Il + a1y(t) + a2y(t)e(t) + a3y(t)y(t-l)
=fl [y(t) , y(t-I), e(t)] (8)

The modified TLRN for one-step-ahead prediction of this NARMA time series would have m = 3, n
= I, two hidden PEs and the total numer of nodes is 10. The 3 lag-I delayed input nodes would
correspond to [y(t-I), y(t-2), e(t-I)], and the redundant y(t-I) input node may be deleted to further
simplify the network.

3. MULTIPLE-STEP-AHEAD NEURAL NET PREDICTORS

Recall that the input-output map underlying the direct learning network (DLN) can only produce the
one-step-ahead prediction of the output time series. To develop a DLN for a least-squares multiple-
step-ahead prediction the network architecture must be built such that the map directly computes the
multiple-step conditional expectation for a given input vector.

The two-step-ahead and three-step-ahead minimum mean-square predictors for the process in Eqn.
(7) can be shown to be
y(t+2) =(1l+a2S2) + (al + a21l)y(t+ I) + a3y(t+ I)y(t)
=f26(t+ I), y(t)] (9)
y(t+3) = (ll+a2s2+ala3cr2-a2a3m3) + (al + a21l)y(t+2) +al a3 cr2 y(t+I)+
a3 y(t+ 1)y(t+2) + aj(1l 2+cr2)y(t)
(10)
=f36(t+2), 9(t+ I),y(t)]
Comparing (9) to (10), it is evident that the DLN for computing an arbitrary k-step-ahead prediction
of an NARMA process has a simple time-lagged relationship with the (k-I)-step-ahead prediction.
This forward chaining of predictions, which is due to the nature of stochastic difference equations,
can be utilized for designing the hierarchical structure illustrated in FIGURE 5 for performing
recursive calculation of multiple-step-ahead predictions. Here Slabl and SIab2 of each level in the
 541

hierarchy provide the necessary inputs to the hidden and the output PEs in Slab3. The connections for
the indirect map are the same as discussed previously.

o
bias Slab 1 Slab 2

00 Slab 3
Yt-l
1\

Yt+l

00
bias
6 60

1\

Yt+2
The bias is connected
to all hidden
0
and output PEs. bias

00
Yt+3

FIGURE S. A HIERARCHICAL TLRN: THREE-STEP

PREDICTION OF NARMA(2,1)

The hierarchical
.
architecture can be extended to a general NARMAX (n y, nu,ne). The idea is
applicable to both direct and indirect maps. However, it is much easier to do multiple-step prediction
using indirect mapping of a complex NARMAX model; one only needs to know the model orders n ,
y
nu, and ne for a stochastic dynamical system of interest.

The hierarchical neural network architecture for the general NARMAX model is a straightforward
extension of the NARMA network in FIGURE 5. The number of levels in the hierarchy would
correspond to the maximum prediction leadtime (number of steps or time periods) to be made from a
 542

given time t The inputs to the frrst slab of any level would be the measurements from a process
monitoring system plus time-lagged predictions and other data from the level direcly below. The local
memory of a processing element can be used to store the most current lagged data and innovations
computed on-line.
Numerical Example 3.1

The NARMA model in (7) was used to generate a synthetic time series of 400 data points(Data Set A)
using the following parameter values: a l = 0.4, a2 = 0.5 and a3 = -.0025. The white-noise sequence
had a non-zero mean =2.0 and a variance of 0.1. The conditional expectations were computed using
(8) - (10). These values would be the same as the forecasts from the hierarchical direct-mapping
neural network with the same structure as shown in FIGURE 5; however, the connection weights
would be set equal to the parameter estimates and product inputs would be used.

The optimal least-squares predictions for the leadtimes equal to one and three are shown in FIGURE
6.1 and FIGURE 6.2. They indicate that the forecast rapidly deteriorates beyond the one-step-ahead
prediction. This is as expected since the time series memory span is essentially one time period.

The frrst half of the data set (200 data points) was used to make the training samples and the
hierarchical neural network was trained by the back-propagation method. The hierarchical TLRN's
performance for in-sample predictions of one and three steps ahead at every time t is given, for
comparison purpose, in FIGUREs 6.3 an 6.4. The same short-memory forecasting behavior is
exhibited but the rate of decay toward the mean leavel of the process is much faster than the
theoretical one. Out-of-sample prediction performance for the leadtime of three steps ahead (not
. shown) is not much different from the process mean.
Numerical Example 3.2

A NARMAX model in Chen and Billings [1989], shown in (11), was used to generate an output
and an input time series (Data Set B) of 700 observations. The underlying non-linear polynomial
model is linear in the parameters:
yet) = 0.7578y(t-l) +O.389u(t-l) -0.03723y2(t-l) +O.3794y(t-l)u(t-l)+O.0684u 2(t-l)
+0. 1216y(t-l)u2(t-l) +0.0633u 3(t-1) -0.73ge(t-1) -0.368u(t-1)e(t-1) +e(t) (11)
where (u(t)} is a uniformly distributed random sequence, the distribution function being a Uniform (-
0.8,1.2). The white-noise error sequence, (e(t)}, is a N(O,O.OI) Gaussian process. The NARMAX
model is stable, but it has a very low order and small noise variance. A statistical test indicates that the
time series (y(t)} is a white noise series. Such characteristics are typical of most published results in
the nonlinear system identification literature.

The hierarchical TLRN for this map is almost identical to FIGURE 5, with the exception that the
inputs to the network are based on the following mapping functions:
y(t+1) = f1[y(t),u(t) ,e(t)]; y(t+2) = f2@(t+1),u(t+1)], and y(t+3) = f3@(t+2),u(t+3)] (12)
 543

There are 2 hidden PEs and one output PE for each of the three hierarchical levels. The prediction
capability of the hierarchical neural network is very well demonstrated in FIGURE 7.1 and FIGURE
7.2. The out-of-sample generalization to three steps ahead is extremely accurate, even though the
ouput series is not statistically different from a white noise sequence. This excellent predictive
capability is due to the persistently exciting inputs u(t), u(t+ i), and u(t+3) employed by the neural
network. This situation is unlike the NARMA case in Example 3.2 where the only inputs are
stochastic and /lot measurable.

Numerical Example 3.3

An hourly water consumption record from the city of Arlington, Texas, provides the real time series
used in this example. The water use record is for the month of April, 1989, and consists of 720
observations (Data Set C) . The time series is characterized by irregular periods of hot and cold days
mixed with sporadic rainfalls. A preliminary analysis of the water use data over the whole year of
i989 has shown that the underlying process is seasonal, stochastic and non-stationary. The non-
linear and non-stationary effects of rainfalls, cold weather fronts and summer heating are difficult to
formulate mathematically. In this situation one cannot use the direct learning approach.

The hourly obselvations exhibit a daily cycle of 24 hours, one annual seasonal cycle with a peak
during the Summer. Nonhomogeneous behaviors like time-varying noise variance, randomly
fluctuating amplitudes, and outlier observations are seen throughout the year. The sample
autocorrelation functions of the data show largest correlations between the hourly consumptions 24
hours apart and i hour apcu1. Limiting the a priori information about the stochastic process to just the
historical water use record, the objective of this experiment is to test the validity of the hierarchical
neural network concept.

The correlation analysis suggests that the following NARMA mapping function might be appropriate
yet) = F[y(t-i), y(t-23), y(t-24), y(t-25), e(t), e(t-I), e(t-23), e(t-24), e(t-25)]. (13)
The time lags of one and twenty-four periods correspond to the two largest sample autocorrelations;
the interaction between the memories of one and twenty-four hours apart implies that the observations
y(t-23) and y(t-25) should be included. The innovations inputs are needed to account for residuals
autocorrelation, which would be present if only time-lagged data values are used.

A hiercu'chical TLRN for this NARMA map for forecasting one to twelve-hours ahead, consisting of
twelve hierarchical levels, has been constructed from (13) above. The 2m inputs to the first level for
one-step-ahead prediction of y(t+ 1) are [yet), yet-I), y(t-22), y(t-23), y(t-24),y(t-25), e(t), e(t-l), e(t-
22), e(t-23), e(t-24), e(t-25)], after all redundant nodes have been eliminated. The modified TLRN
architecture has m = 6 and n =i; thus there are m - 1 = 5 hidden PEs for each level of the hierarchy.
The inputs to the second level are simply the one-step forward time-shifting of the first level:
[y (t + 1),y(t),y(t-2i ),y(t-22),y(t-23),y(t-24),e(t),e(t-2l),e(t-22),e(t-23),e(t-24) 1
This simple scheme may be repeated for all levels up to the last level of the hierarchy.
 544

Randomly generated error inputs and lagged values, having the same statistical characteristics as the
innovations associated with the same month of the previous year, have been used as pseudo error
inputs for training the network and for the out-of-sample generalization test (over the last two weeks
of data). The training samples are from the fIrst two weeks of the data and the test data set is based on
the last two weeks of the time series.The in-sample and out-of-sample predictions for the leadtime of
12 hours ahead are depicted in FIGUREs 8.1 and 8.2, respectively.

The multiple-step forecast performance beyond the training set is quite good; the hierarchical neural
network is able to predict new cyclical patterns, with respect to the times of the peaks and the
troughs, in the last two weeks of the time series. It does not completely map the random fluctuations
in the training data, a desirable characteristic of a good fIlter; these extreme values are either outliers
or transient responses to extreme weather conditions.
The neural network does not very well predict the last two 24-hour cycles of the data. A closer
examination reveals that the hourly water use measurements during the last two days of the month,
which correspond to a weekend, behave uniquely different from the weekday patterns. The last
weekend water use pattern happened to increase signifIcantly from the level of water consumption
during the fIrst two weeks used for the network training. Obviously, the neural network cannot
respond to this "untrained" behavior.

CONCLUSIONS

Two different neural network modeling approaches for system identification from input-output time
series are presented. In the rust situation a complete knowledge of the underlying model orders and
parameter values are required; the resulting neural network map is called direct mapping, or direct
learning, neural net. The second case uses only a partial knowledge of the input-output system in
terms of the time-lagged orders of the variables involved. Both neural network maps aim at one-step-
ahead prediction of the output variable of a discrete-time dynamical system.

Out-of-sample generalization to multiple-step forecasting beyond the training data can be achieved by
a simple forward time-shifting operation on the neural-net map of an NARMAX process. The time-
lagged recurrent network (TLRN) ofWerbos is used as the building block for contructing a multi-
level hierarchical neural-net architecture. The hierarchical architecture has been shown to be valid,
through empirical analyses, given that the inherent time-lagged relationship in the given time series is
much greater than one period.

The necessary time-lagged relationship can be discovered by means of a simple correlation analysis of
a given time series. The generalization peIformance can be drastically increased if future values of a
persistently exciting input variable are available.

ACKNOWLEDGEMENTS
 545

This material is based upon work supported by the National Science Foundation under Grant
No. ECS-9014109. The government has certain rights in this material.

REFERENCES

Athans, M., and D. Willner,"A practical scheme for adaptive aircraft flight control systems," Symp.
on Parameter Estimation Techniques and Applications in Aircraft Flight Testing, NASA FIt. Res.
Ctr., Edwards AFB, April 1973.

Billings, S.A., H.B. Jamaluddin, and S. Chen, "Properties of neural networks with applications to
modelling non-linear dynamical systems," Int. 1. Control, 45, 1, 193-224, 1992.

Chen, S. and S. A. Billing, "Representations of non-linear systems: the NARMAX model," Int. 1.
Control, Vol. 49, No.3, 1013-1032, 1989.
Goodwin, G.c., and K.S. Sin, Adaptive Filtering Prediction and Control, Prentice-Hall,
Englewood Cliffs, New Jersey, 1984.

Kashyap, R.L., and A.R. Rao, Dynamic Stochastic Models from Empirical Data, Academic Press,
N.Y. 1976.

Landau, 1.0., "Elimination of the real positivity condition in the design of parallel MRAS," IEEE
Trans. Auto. Control 23, 1015-1020, 1978.
Ljung, L. System Identification Theory for the User. Prentice-Hall, Englewood-Cliff, NJ. 1987.

Lowe, D. and A.R. webb, "Time series prediction by adaptive networks: a dynamical systems
perspective," lEE Proceedings-F Vol. 138, No.1, 17-24,1991.

Narendra, K.S., and K. Parthasarathy,"Identification and control of dynamical systems using neural
networks," IEEE Transactions on Neural Networks, Vol. 1, No.1, pp. 4-26, 1990.

Priestley, M.B., Non-linear and Non-stationary Time Series Analysis, Academic Press, N.Y., 1988.

Rumelhart D. E., and J. L. McClelland, Parallel distributed processing: explorations in microstructure

of cognition, MIT Press, 1986.
Sastri, T. "Neural networks for detection of process change in manufacturing systems," Proceedings
of the Joint US/German Conference on New Directions for Operations Research in Manufacturing,
National Institute of Standards and Technology, Gaithersburg, MD, July 30-31,1991.

Sastri, T. and CO. Malave, "Direct learning and knowledge updating in neural nets", Proceedings of
the Artificial Neural Networks in Engineering, November 15-18, 1992, St. Louis, MO.)
Stengel, R. F., Stochastic Optimal Control Theory and Application, Wiley, N.Y. 1986.

Werbos, PJ., "Neurocontrol and related techniques," in Maren, AJ., CT. Harston, and R.M. Pap
(Eds.), Handbook of Neural Computing Application. Academic Press, New York [1990].

Werbos, PJ., T. McAvoy, and T. Suo "Neural networks, system identification and control in the
chemical process industries," in White, D., and D. Sofge (Eds.), Handbook of Intelligent Control.
Van Nostrand, New York 1992.
546

FIGURE 6-1: THEORmCAL. ONE-STEP-AHEAD. CONDmONAL IXPECTAnON

(DATA SET A)

208

200

192

184

178

~
188
~
180
152

144

136

128

120
0 100 150 200
nIlE

FIGURE 6-2: THEOREnCAL THREE-STEP-AHEAD. CONDInONAL EXPECTAnON

(DATA SET A)
Prediction
208 ----- Actual
;
200 ••
•••
192 ""
" . ..:.
.
II '.
184 '. "' I
I,
'..'I,
''.'
176
,," ":'

a
.188
180

152
""
..
N
"
"" ,, '' .. "" ""
144 ~ " N
" ""
"" "
~ ""
~
I ""
!,
138 ~"

128
 547

FIGURE 6-3: HIERARCHICAL TLRN'S ONE-STEP, IN-SAMPLE PREDICTION

(DATA SET A)
PredlcUOIl
208 ••••• Actual

200

192

184
176

168
~
c
CI
160

152
..., ,.
,
144 ~

136

128

120
0 50 100 150 200
TIME

FIGURE 6-4: HIERARCHICAL TLRN'S THREE-STEP, IN-SAMPLE PREDICTION

(DATA SET A)
PredlcUoll
• •••• Actual
;
••
•••

....
""
""
't II
,
" I,

....
I, "
'1' 1

""" '''
,, ""
1 "
f ::

...., ......
,
, ,
,'
"" .,
""
~"

80 120 160 200

TlO
548

FIGURE 7-1: HIERARCHICAL TLRN'S ONE-STEP, OUT-OF-SAMPLE PREDICTION

(DATA SET B)
----- Aclual
2.20 - - Nelwork PredicUOD

2.00

·.·
1.80
1
I

1.60
.,i "
I

1.40
1.20
~ 1.00
= 0.80
a
..
.,
II

0.60 I I
1•
0.40
0.20
-0.00
-0.20
-0.40
0 50 100 150 200
ma:

FIGURE 7-2: HIERARCHICAL TLRN'S THREE-STEP, OUT-OF-SAMPLE PREDICTION

(DATA SET B)
----- Aclual
2.20 - - Nelwork PredicUOD

2.00
1.80
1.60
1.40
1.20
1.00
= 0.80
~
a

0.60
0.40
0.20
-0.00
-0.20
-0.40
0 50 100 150 200
TIME
 549

FIGURE 8-1: HIERARCHICAL TLRN'S 12-STEP-AHEAD IN-SAMPLE PREDICTION

(DATA C)
Network Prediction
Actual
2.30

2.10

.." ,..
1.90 ,,.
1.70 ~:

1.50 ."" .,,,

'

.:
.: 1.30
I-
Q
1.10
0.90

0 .70 I

0.50 ." ' j

,
01
!
0.30

0. 10
0 50 100 150 200 250 300 350
TIME

FIGURE B-2 : HIERARCHICAL TLRN'S 12-STEP-AHEAD OUT-OF-SAMPLE PREDICTION

(DATA SET C)
Network Prediction
Actual
2.30

2 . 10
,l ,
1.90 " ,:
, ,"
I
1.70
1.50

~ 1.30
1.10
0.90

0.70 'II'
:.1
0 .50 ':" ~!
I
,,~
I
I I I
I I
I I
0 .30 f
I
I
!
0.10
340 390 440 490 540 590 640
TIME
 Integrating Neural Nets, Simulation, and
Genetic Algorithms for Real-time Scheduling

by

Albert Jones
National Institute of Standards and Technology
Gaithersburg, Md 20899
USA

Luis Rabelo
Ohio University
Athens, OH 45701
USA

1. Introduction

In this paper we briefly review the generic architecture for intelligent controllers proposed in DAVIS et
al. (1992). We then describe an approach for carrying out the scheduling functions contained within
that architecture. This approach integrates neural networks, real-time Monte Carlo simulation, and
genetic algorithms.

2. The Generic Control Architecture

The generic architecture for the intelligent controller proposed in DAVIS (1992) is shown in Figure 1.
It performs four major functions which expand and generalize those developed in JONES/SALEH
(1990): assessment, optimization, execution, and monitoring. These functions are described in the
following sections.

2.1 Assessment

The Assessment Function (AF) formulates all "planning and scheduling" problems based on the
current state of its subordinates and input from its supervisor. Three sets of inputs are used in this
problem formulation task. The first is a list of assigned tasks with due dates and other constraints
which are specified by the controller's supervisor. The second is the feedback information from the
Monitoring Function describing 1) the current state of the system, and 2) the projected system
 551

~
Feedback Task. and time.
fo superuisor from Superulsor

1 &fIfffJfIf1J 0(t (t@fffJ'iJf1J@[l,{l, fIf1J

.
.•
RSSESSMENT
FUNCTION Problem

-....
~ formulations
~

c: Performance ,
-
a;,
~

~
c:
Q
I..l
.... StBfisitics
PPTlMllRTIO'"
FUNCTION Selected
Control Law

•
Constraint Evaluation

....
,
MONITORING ... EHECUTION
FUNCTION Current Control Schedule FUNCTION
t ,
INTERFRCE
MOOU ES

Rs"gned Ta.k. Feedback

and limit time.

SUBORO INRTE
PROCESS(ES)

FIGURE 1. Generic Control Architecture

response under the current control law. The third comes from the Optimization Function which
summarizes its current success in finding an optimal solution to the problems it has been given to
solve. In formulating these optimization problems, the AF ftrst speciftes the constraints. There are
two types: hard and soft.
552

Hard constraints are those that cannot be violated either by the optimization and execution functions in
the same controller or by subordinate controllers. These constraints are typically of three types:
external, internal, and those related to the physical dynamics of subordinates. We give several
examples of each. Externally imposed hard constraints come from either the supervisor or the process
planner. The former are usually in the form of due dates, priorities, maintenance schedules, etc. The
latter results in a set of candidate process plans that can be used to manufacture each product. If
multiple process plans exist, the AF will screen them to eliminate infeasible plans based on the current
state of the subordinates. For example, any plan requiring a "downed"-machine must be eliminated.
The remaining alternatives are passed as hard constraints to the Optimization Function which will
determine the actual run-time production plan. This run-time plan together with collections of
scheduling alternatives, which are also based on the state of the system, are passed to the Optimization
Function a~ hard constraints to be used in determining the optimal schedule. The third type of hard
constraint is derived from the physical layout of subordinates. For example, the AF must account for
the times that will be required to transfer a job from one process to another. It must also consider the
number of jobs that can reside at each subordinate location. There are also certain physical
assignments of jobs that could result in deadlock which must be avoided.

Soft constraints provide an additional mechanism for the AF to control the evolution and behavior of
its subordinates. In most cases, these can be violated without violating the hard constraints. We give
several examples. While the due date for a given job may be a hard constraint, at each layer there
may be start and finish times for some of the tasks that can be viewed as soft constraints. Even though
a subordinate may be capable of storing several jobs, the AF may decide to impose a Kanban strategy
which significantly reduces the number of jobs that will ever be queued before a given process. The
AF may also impose soft constraints on utilization rates for subordinates. Unlike hard constraints, soft
constraints are not e~ential in the basic definition of the projected response of the subordinates.
)
Nevertheless, their imposition will further constrain the evolution of subordinates in the desired
manner.

The Assessment Function also specifies the performance criteria and all job priorities for each
optimization problem to be considered by the Optimization Function. These criteria can be related to
the "performance" of individual subordinates or jobs, or the entire collection of subordinates or jobs.
Examples for subordinate performance include utilization, and throughput. Examples for job
performance include lateness, tardiness, and makespan.

2.2 Optimization

The Optimization Function (OF) has two major responsibilities. First, it solve the real-time
optimization problems posed by the AF. Second, it restores feasibility (optimality if possible)
whenever 1) the Execution Function cannot restore feasibility using the current hard constraints or 2)
the AF has determined that the current projected system response is unacceptable based on either the
 553

perfonnance criteria or soft constraints. Sub-gradient optimization techniques (GEOFFRION 1970)

are commonly used to solve this type of problem when everything is assumed detenninistic. Since
manufacturing systems are not detenninistic, this approach cannot be used. Our approach is described
in section 3.

For each assigned job, the output from the Optimization Function consists of a process plan, estimated
processing durations (pessimistic, optimistic, and average) and a scheduling rule. These are all used
by the Execution Function to detennine the actual start and finish times for the tasks in that plan.

2.3 Execution and Monitoring

The Execution Function (EF) addresses the following two tasks: it compute limit times for each
subordinate, and it restores feasibility (see below) when violations of those limit times arise from the
actions of the subordinate subsystems.

The Monitoring Function also has two major responsibilities: updating the system state, and
evaluating proposed system responses against the hard and soft constraints set by the Assessment
Function. The system state summarizes the operational status, buffer status, and job status of
subordinates. The exact definition changes as one moves from one layer to another. The Monitoring
Function generates this state infonnation by aggregating the detailed feedback data from the various
subordinates. The Monitoring Function does not communicate directly with subordinates. It
communicates through an Interface module. This module perfonns all of the functions necessary to
make meaningful communication possible between a supervisor and its subordinates. This updated
system state is then broadcasted to all of the other functions, because it is the starting point for all of
the real-time analysis perfonned by the generic controller.

The Monitoring Function also evaluates each decision against the current set of constraints. Recall that
these constraints are of two types: hard (which cannot be violated) and soft (which are desirable
goals). Hard constraints are imposed by the supervisor on Assessment, by Assessment on
Optimization, by Optimization on Execution, and by the Execution Functions on the controller's
mbordinates. A series of refinements takes place in the sense that the hard constraints imposed by the
l\ssessment Function on the Optimization Function are a subset of those imposed on the Assessment
function by the supervisor. The same is true for the Optimization and Execution functions. Each time
)ne of these functions makes a decision, the monitoring function will detennine if any of these hard
:onstraints have been violated. Each time a violation occurs, the Monitoring Function will identify the
:onstraint(s) and initiate an error recovery strategy (see below).

:;'or the soft constraints, the situation is a little different. These are tracked continually and transmitted
o the assessment function where several things can happen. Consider a soft constraint on System
Jtilization (SU > 90%). The Assessment Function can choose to ignore the violation when it is small
554

(SU = 85%). On the other hand, the Assessment Function could simply request the Optimization
Function to generate a new schedule or pick a new plan, whenever this violation is outside some
threshold. We allow the threshold to a two sided limit (perhaps 70% < SU < 95%). This allows the
Assessment Function to react when the utilization is too small (i.e. machines sitting idle) or too large
(i.e. bottlenecks may be occurring).

2.4 Error Recovery (Rescheduling)

As noted above, the Monitoring Function recognizes when a constraint has been violated. DA VIS et
al.(1992) described an error recovery process which is shown in Figure 2. In Figure 2(a), the hard
constraints specified by the Assessment Function are depicted by the outermost solid lines. The limit
times imposed by the Execution Function form a tolerance band and are depicted as the shaded region
in the figures. Based on these limit times, the subordinates will choose their own trajectories. As
subordinate systems evolve in time, deviations from their computed trajectories are bound to occur.
Any deviations leading to a trajectory which 1) stays within the tolerance band and 2) leads to the
same final state is acceptable. Any deviation which yields a state which falls outside the tolerance
band triggers the error recovery mechanism discussed below.

Figure 2(b) indicates that a disturbance has occurred at one or more of the subordinates leading to a
state outside the tolerance band. Assume that it is only one subordinate. This means that this
subordinate proposes to follow a new trajectory (schedule) which will cause one or more of the
original limit times to be violated. This triggers the Execution Function to generate new limit times.
The objective is to restores feasibility with respect to all hard constraints, and get back to the original
trajectory as quickly as possible. Restoration would be achieved by updating some limit times or by a
minor resequencing of tasks. The idea is to first try to restore feasibility by maintaining the original
hard constraints. If this can be done, then the Execution function would generate a new set of limit
times using this new plan and/or scheduling rule. The resulting solution may have the same limit times
and a match-up form, as depicted in Figure 2(b), or another set of feasible limit times and optimal
trajectory such as the one shown in Figure 2(c). To the best of our knowledge, such an approach has
not been reported in the literature.

Whenever the Optimization function cannot restore feasibility and maintain the original set of hard
constraints the Monitoring function will inform the Assessment function. The Assessment can impose
a new set of hard constraints or change the weights of the performance criteria. This may lead to a
new feasible solution or it may not. Success would again be depicted by pictures similar to those
shown in 2(b) and (c). Failure is shown in Figure 2(d). In this case, the Assessment function must
negotiate with the control module's supervisor to change one or more of the externally imposed hard
constraints. Once this is done, the entire process would begin again, until a solution feasible with
respect to these new constraints was found.
 555

""""""
lal DuIr<4Sys_T~wlJAT _ _ /.JmJJl

1<) UpdtlldSys_~_.N_OptitrtDJ
~tsb<_

----.........~~ ---

-
" '....._ ... Iiool

FIGURE 2. Conceptual View of Error Recovery

556

It is important to note that each function is limited in its response to a disruption. That is, the
Execution Function first reacts by using the current control law (i.e., the schedule) to satisfy all of its
hard constraints (i.e., limit-times.) If it can, the error recovery procedure stops. If it cannot, the
Optimization Function then attempts to modify the control law (i.e. change the scheduling rule or
production plan) to restore feasibility. If it can, the procedure stops. If it cannot, the Assessment
Function will specify a new formulation by 1) adding new candidate plans, or 2) adding new
scheduling rules, or 3) negotiating with the supervisor to change constraints. If this cannot be done,
then the problem is passed up to the supervisor.

3. Implementation Strategy

Real-time scheduling and rescheduling in the event of errors are critical functions within the
aforementioned architecture. To implement these functions, we propose a three step refinement
process. The first step is to generate a set of candidate scheduling rules from a much larger list of
scheduling rules. Our approach is based on neural nets and is discussed in section 3.1. Step 2 is to
rank these rules using a detailed real-time simulation based on the current "system state" and desired
performance measures. This approach is discussed in section 3.2. The last step is to use that ranking
to determine the "best schedule". Our approach is based on genetic algorithms and is discussed in
section 3.3.

3.1 Neural Networks

The first step in this process is to select a small list of candidate scheduling rules from a larger list of
such rules (see Figure 3). For example, we might want to find the best five dispatching rules from the
list of all known dispatching rules. These five could be chosen so that each one maximizes (or
minimizes) at least one of the performance measures, with no regard to the others. In addition, we
would, whenever a problem occurs, like to use to select those rules most likely to result in a quick
recovery. To carry out this part of the analysis, we propose to use a neural network approach.

Neural networks have shown good promises for solving manufacturing scheduling problems
(LO/BA VARIAN 1991). Their applications have particularly been dedicated to NP-complete
problems, and in specific to the Traveling Salesman and Job Shop Scheduling problems. However,
the direct application of neural networks for scheduling problems have not produced consistently
optimal/real-time solutions for realistic problems due to the limitations in hardware and the
development of algorithms. These neural network implementations have emphasized the utilization of
neural networks based on relaxation models (i.e. pre-assembied systems which relaxes from input to
output along a predefined energy contour). The neural networks are defined by energy functions in
these approaches. The lowest energy state of the network corresponds to the
 557

CANDIDATE RULES

SYSTEM STATE

NEURAL SYSTEM

SCHEDULING POLICY #1

SCHEDULING POLICY #2

SCHEDULING POLICY #3

•• MANAGEMENT
• OBJECTIVES
SCHEDULING POLICY #n

FIGURE 3. Neural System for Candidate Rule Selection

"best schedule". LO/ BA VARIAN (1991), formulated the objective function which minimizes the
time required to complete all the jobs as

Et = 1/2 Lj=l Li=lLl=l

where i, j, k, I, ro, n are indices;

558

Ck is a scaling factor;
Vijl is the output of neuron ijl
and Tij is the time required for by jth machine to complete the ith job.

FOO(fAKEFUJI (1988) and ZHOU et al. (1990) have applied stochastic Hopfield networks to solve
4-job 3-machine and to-job to-machine job shop scheduling problems respectively. These
implementations frequently produce constraint violating solutions and the computational cost limits
their consideration as viable solutions. LOIBA VARIAN (1991) extended the two-dimensional
Hopfield network to 3 dimensions to represent jobs, machines, and time. Another implementation
based on stochastic neural networks applied to scheduling can be found in ARIZONO et al. (1992).
However, due to a large number of variables involved in generating a feasible schedule, it has been
difficult for these approaches to formulate realistic and complex scheduling problems and scheduling
objectives. Because of the distributed, hierarchical nature of the controllers we intend to use, the size
and complexity of our scheduling problems are greatly reduced. In addition, we will emphasize on the
utilization of the "learning" neural network models which are more developed than the relaxation
models (RUMELHART et al. 1988).

3.2 Network Paradigms for Selecting Candidate Rules

The optimal neural network paradigm for "candidate rule selection" should be assessed on the
following factors:

1. Trainability: The impact of the initialization and learning parameters in the probability of the
network learning training sets for candidate rule selection.

2. Generalization: The ability of the network to model the behavior of the specific FMS
relationship. This is tested utilizing unseen cases.

3. Computational characteristics: Storage needed for programs and data files, training time,
and execution time.

4. Re-Trainability: Ability to be retrained when the training set has been incremented or
modified.

Several paradigms are possible choices for neural network structures to generate these Candidate
Rules. In this research, we will focus our initial efforts on Backpropagation and Fuzzy ARTMAP.
These paradigms have been selected due to their supervisory learning mode, non-linear and fuzzy
modeling capabilities, and generalization and automation potential.
 559

Backpropagation (RUMELHART et al. 1988) is a supervised learning paradigm that learns adequate
internal representations using deterministic units to provide a mapping from input to output. The
backpropagation neural network structure is a hierarchical design consisting of fully interconnected
layers or rows of processing units. Each unit itself comprised of several layers or rows of processing
elements (see Figure 4). In this network, the input might represent the state of the FMS, the process
plans, and the L performance indices. The output might represent the available scheduling policies
from which a quantitative evaluation is going to be obtained.

OUTPUT UNITS
Scheduling policies

<FMS state, Process Plans, L Performance Indices>

BIAS
INPUT UNITS

FIGURE 4. Back Propagation Neural Network for Candidate Rule Selection

Another attractive option is the Fuzzy ARTMAP (CARPENTER et al. 1991) which provides a
supervised learning of recognition categories in response to arbitrary sequences of analog or binary
inputs. It incorporates two Fuzzy ART modules (networks which utilize the fuzzy set theory
conjunction in the learning rules in order to develop stable categories) that are interlinked through a
mapping field Fab (see Figure 5). The inputs to the Fuzzy ARTa Module (also outputs for the entire
neural structure if Fuzzy ARTb is considered the input Module) could be the state of the FMS, the
process plans, and L performance indices. The inputs to the Fuzzy ARTb Module (also outputs for the
560

entire neural structure if Fuzzy ARTa is considered the input Module) might represent the available
scheduling policies.

INTER-ART
ASSOCIATIVE
MEMORY

II

FuzzyARTa Fuzzy ARTb

a b

<FMS state, Process Plans, L Performance Indices> <Scheduling Policies>

FIGURE 5. Fuzzy ARTMAP Neural Network for Candidate Rule Selection

3.3 Input Feature Space

In order to design a system for the "Candidate Rule Selection" process based on neural networks an
appropriate representation of the "state" of the FMS, should be developed. The state contains
aggregate information about all in-process jobs, processes, and the active schedule. This aggregate
state information is derived from from data for each job includes job ID, current location (buffer,
transporter, or process), due date, expected completion time (if it has previously been scheduled),
shop floor release time, current active routing and expected/actual start and finish time at each process
in that routinl!. The state of the FMS orocesses should indicate accordinl! to their tvoe the iobs. batch
 561

size, the planned start and delivery time of the jobs, and infonnation about the transporters. The
current schedule and its variables. The process plans infonnation is provided by the processing times
and routings. The L performance indices such as Average Tardiness, Job Flow Time, Job
Productivity, and Process Utilization will be required. All of this infonnation should be translated
into a suitable representation for the neural networks to provide the desired evaluation of the different
scheduling rules available in the system. Previous studies have indicated guidelines for this task
(RABELO/ALPTEKIN 1989 and RABELO 1990).

4. Real-Time Simulation

After these R candidates have been detennined, each of them will be evaluated using an analysis
technique tenned real-time Monte Carlo simulation (DAVIS/JONES 1989). The outputs from these
simulation trials yield the projected schedule of events under each scheduling rule. These schedules are
then used to compute the values of the various perfonnance measures and constraints imposed by the
Assessment Function. Since R rules must be considered, we plan to run R real-time simulations
concurrently to avoid unacceptable timing delays in the analysis. At this time, we plan to follow the
approach described in DAVIS et a!. (1991) and use a network of Personal Computers, with each one
simulating a single scheduling rule. We hope to move on to a parallel processing machine in the near
future. The length of each simulation trial for a given simulation engine is determined by the length of
the planning horizon of the controller.

This fonn of Monte Carlo simulation differs considerably from traditional discrete-event simulation in
two ways. First, each simulation trial is initialized to the current system state as updated in real-time
by the Monitoring Function. This results in the second important distinction - the simulations are
neither terminating nor steady state. They are not terminating because the initial conditions may change
from one trial to the next. Furthennore, they are not steady state because we are specifically interested
in analyzing the transient phenomena associated with the near-tenn system response. The question is,
from a statistical perspective, does this really matter. Early experiments conducted by DAVIS et
a1.(1991) indicate that the answer is a definite maybe. That is, the inclusion or exclusion of new
events corresponding to changes in the initial state can bias the statistical estimates of certain, but not
all, perfonnance measures.

Given the last I ()() simulation trials for each of the R alternatives, a series of statistical calculations are
made for the given perfonnance criteria for each individual trial: the empirical probability density
function associated with each criterion, means, variances, and the correlations between each pair of the
L perfonnance measures. DAVIS et al.(1991) have found cases where the correlation is positive,
zero, and negative. These correlations are useful since they provide the statistical tradeoffs that exist
among the various perfonnance measures.
562

The computed statistics and associated empirical probability density functions are used to select that
rule which provides the best statistical compromise among the perfonnance criteria. In DAVIS et
al.(1991) this compromise analysis has been fonnulated to address the maximum expected value of a
utility function that would be achieved if a given rule was adopted. In the next section, we discuss a
new approach to this compromise analysis, genetic algorithms.

5. Genetic Algorithms

No matter how the utility function described above is constructed, only one rule from the candidate list
can be selected. This causes an undesirable situation whenever there are negatively correlated
perfonnance measures, because no one rule can optimize all objectives simultaneously. Conceptually,
one would like to create a new "rule" which 1) combines the best features of the most attractive rules,
2) eliminates the worst features of those rules, and 3) simultaneously achieves satisfactory levels of
perfonnance for all objectives. Our approach does not deal with the rules themselves, but rather the
actual schedules that result from applying those rules. Consequently, we seek to generate a new
schedule from these candidate schedules. To do this, we propose to use a genetic algorithm approach.

Genetic algorithms are an optimization methodology based on a direct analogy to Darwinian natural
selection and mutations in biological reproduction. Instances of a concept description correspond to
individuals of a species, and induced changes and recombinations of these concepts are tested against
an evaluation function to see which preserve in the next generation. In principle genetic algorithms
encode a parallel search through concept space, with each process attempting coarse-grain hill climbing
(GOLDBERG 1988).

The use of genetic algorithms requires five components:

1. A way of encoding solutions to the problem on chromosomes (e.g. string of symbols of a

fixed length).

2. An evaluation function that returns a rating for each chromosome given to it (i.e. a measure
of how well the chromosome perfonn as a domain solution)

3. A way of initializing the population of chromosomes.

,
4. Operators that may be applied to parents when they reproduce to alter their genetic
composition such as crossover (i.e. exchanging a randomly selected segment between
parents), mutation (i.e. gene modification), and other domain specific operators.

5. Parameter setting for the algorithm, the operators, and so forth.

 563

The basic execution cycle for a genetic algorithm is explained below:

1. The population is initialized using the specific procedures. This initialization process will
result in a set of chromosomes.

2. Each member of the population is evaluated, using the objective function.

3. Select candidates according to strength - relative ranking according to the objective

function. Candidates with the highest fitness rating have a greater probability of
reproduction.

4. The population undergoes reproduction until a stopping criteria is met. Reproduction

consists of a number of iterations of the following three steps:

(a) One or more parents are chosen to reproduce. Selection is stochastic.

(b) Genetic operators are applied to the parents to produce the offsprings.

( c) The performance of the new population is evaluated.

5. Iterate the process.

Genetic algorithms have been utilized in job shop scheduling implementing stochastic search. DAVIS
(1985) applied genetic algorithms directing stochastic search in job shop scheduling problems.
DAVIS/RITTER (1985) utilized several techniques in order to avoid the "epistatic" problem. Recently
)ther researches have applied directly genetic algorithms to job shop scheduling problems
~ BIEGEL/DA VERN 1990). However, their examples of applications are reduced to uncomplicated
;cheduling problems. In addition, genetic algorithms have been applied to job shop scheduling by
mpporting a simulated annealing scheduler (DA VIS/RITTER 1987).

When a genetic algorithm is run using a representation that encodes solutions to a problem such as
;cheduling and operators that can generate better offsprings, the methodology can produce populations
)f better individuals (e.g., schedules), converging finally on results close to a global optimum.
Knowledge of the the domain can often be exploited to improve the genetic algorithm's performance
hrough the incorporation of new operators or the utilization of integrated reasoning architectures (e.g.
nessy genetic algorithms, edge recombination, ordering crossover). Examples of these approaches
:an be found in WHITLEY et al. (1989), GOLDBERG et al. (1989), and LIEPINS et al. (1987).
564

The compromise analysis process carried out by a genetic algorithm can be thought of as a complex
hierarchical generate and test process. The generator produces building blocks which are combined
into schedules. At various points in the procedure, tests are made that help weed out poor building
blocks and promote the use of good ones. To reduce and support uncertainty management of the
search space, the previous two steps (candidate rules selection and parallel simulation with statistical
analysis) provide partial solutions to the problem of compromise analysis. Reducing uncertainty
makes the search process more effective, with the complexity of the scheduling problem becoming
more manageable in the process (see Figure 6).

.. cr... •
cc •• __ •
I OPERATOR 2

I I II I I
•
•
•

" OPERATOR 1

~ I I I
co:. :n

....
I I I I FINAL
---I.~ I I
SCHEDULE
•
•
•
_--'-, I I I I I
REPRODUCTION
INITIAL
AND
SCHEDULES
EVALUATION

FIGURE 6. Genetic Algorithm for Compromise Analysis

4. Summary

In this paper, we have specified an generic architecture for a controller which provides the building
block for an integrated shop floor control system. This controller contains a generic set of functions
that provides a high degree of intelligence and the ability to respond to disturbances in real-time.
Those functions are Assessment, Optimization, Execution, and Monitoring. They effectively integrate
 565

both decision-making and control. The decisions are formulated by the assessment function and
solved by the optimization function. The execution and monitoring functions provide the most of the
control. The paper also discussed a new integrated framework for addressing the real-time analysis a
problems essential to successful implementation of that architecture. It consist of neural networks to
provide candidate solutions, real-time Monte Carlo simulation to carry detailed performance analysis of
each of those candidates, and genetic algorithms to complete the compromise analysis.

s. References

DAVIS W., JONES A. and SALEH A., "A generic architecture for intelligent control systems",
Computer Integrated Manufacturing Systems, Vol. 5, No.2, 105-113, 1992.

JONES A and SALEH A., "A multileveVmultilayer architecture for intelligent shop floor control",
rnternational Journal of Computer Integrated Manufacturing special issue on Intelligent Control, Vol.
3, No.1, 60-70, 1990.

GEOFFRION A, "Elements of large-scale mathematical programming", Management Science, Vol. 16,

No. 11,652-691, 1970.

LO Z. and BAVARIAN R, "Scheduling with Neural Networks for Flexible Manufacturing Systems,"
Proceedings of the 1991 IEEE International Conference on Robotics and Automation, Sacramento,
California, 1991, pp. 818-823.

FOO Y.and TAKEFUJI Y., "Stochastic Neural Networks for solving job shop Scheduling",
Proceedings of the IEEE international Conference on Neural Networks, published by IEEE TAB,
1988, pp. 11275-11290.

lHOU D., CHERKAS SKY V., BALDWIN T., and HONG D., "Scaling Neural Network for Job Shop
Scheduling," Proceedings of the International Conference on Neural Networks, 1990, Vol. 3, pp.
~89-894 .

A..RIZONO I., YAMAMOTO A, and OHTA, H., "Scheduling for Minimizing Total Actual Flow Time by
\feural Networks," International Journal of Production Research, 1992, Vol. 30, No.3, pp. 503-
511.

~UMELHART D and the PDP Research Group, Parallel Distributed Processing: Explorations in the
l1icrostructure of Cognition, Vol. 1: Foundations, Cambridge, MA: MIT PresslBradford Books,
1988.

:ARPENTER G., GROSSBERG S., and ROSEN D., "FUZZY ART: Fast stable learning and
:ategorization of analog patterns by an adaptive resonance system" CAS/CNS-TR-91-015, Boston
)niversity, 1991.

~ABELO L. and ALPTEKIN S., "Synergy of neural networks and expert systems for FMS scheduling",
'>roceedings of the Third ORSAITIMS Conference on Flexible Manufacturing Systems: Operations
?esearch Models and Applications, Cambridge, Massachusetts, Elsevier Science Publishers R V.,
161-366, 1989.

~ABELO L., "A hybrid artificial neural network and expert system approach to flexible manufacturing
:ystem scheduling", PhD Thesis, University of Missouri-Rolla, 1990.

)AVIS W. and JONES A., "Issues in real-time simulation for flexible manufacturing systems",
>roceedings of the European Simulation Multiconference, Rome, Italy, June 7-9, 1989.
566

LAW A. and KELTON W. Simulation, Modeling and Analysis, McGraw-Hill, New York, 1982.

DAVIS W., WANG H., and HSIEH C., "Experimental studies in real-time Monte Carlo simulation",
IEEE Transactions on Systems, Man and Cybernetics, Vol. 21, No.4, 802-814,1991.
GOLDBERG D., Genetic Algorithms in Machine Learning, Addison-Wesley, Menlo Park, California,
1988.

DAVIS L., "Job Shop Scheduling with Genetic Algorithms," Proceedings on an International
Conference on Genetic Algorithms and Their Applications, Carnegie-Mellon University, 136-140,
1985.

DAVIS L. and RITfER F., "Applying Adaptive Algorithms to Epistatic Domains," Proceedings of the
Ninth International Joint Conference on Artificial Intelligence, 162-164, 1985.

BIEGEL J. and DAVERN J., "Genetic Algorithms and Job Shop Scheduling", Computers and Industrial
Engineering, Vol. 19, No. I, 81-91,1990.
DA VIS L. and RITTER F., "Schedule Optimization with Probabilistic Search," Proceedings of the
Third Conference on Artificial Intelligence Applications, 231-236, 1987.
WHITLEY D., STARKWEATHER T., and FUQUAY D., "Scheduling Problems and the Traveling
Salesman: the genetic edge recombination operator," Proceedings of the Third International Conference
on Genetic Algorithms, 133-140, 1989.
GOLDBERG D., KORB B., and DEB K., "Messy Genetic Algorithms: Motivation Analysis and
First Results," Complex Systems, Vol. 3,493-530, 1989.

LIEPINS G., PALMER M., and MORROW M., "Greedy Genetics," Proceedings of the Second
Intematinal Conference on Genetic Algorithms, 90-99, 1987.
 A Genetic Algorithm for Scheduling with
Resource Consumption
T. Starkweather and D. Whitley
Computer Science Department
Colorado State University
Fort Collins, CO 80523.
B. Cookson
Adolph Coors Company
Golden, CO 80401

Abstract

A production/shipping scheduler is described which utilizes a genetic algorithm

to optimize the sequence in which customer orders are filed . Data is collected by
a warehouse tracking system that provides detailed information about production,
product movement and inventory. Warehouse operation is sensitive to the mean
time at dock for trucks and rails as shipments are being loaded, as well as to the
average inventory. These measures are examined individually and in combination in
order to evaluate scheduler performance. Experimental results show how the genetic
algorithm is sensi tive to selective pressure and population size; these parameters of
the genetic algorithm can be tuned to generate good solutions more quickly, or better
solutions when the algorithm is allowed to run longer. A distributed parallelizable
implementation of the genetic algorithm is also presented which provides a practical
means of generating solutions at least one and perhaps two orders of magnitude
faster on parallel hardware.

1 Problem Introduction

The sequence in which the customer orders are filled can greatly affect the efficiency of

plant operations related to production , shipping and warehouse inventory. It affects the

degree to which is it practical to load product directly from the production line to a
568

shipping dock, the mean time at dock per order, and the amount of product that must
be maintained in inventory.
In the current study, we look at the problem of sequencing customer orders at the
Coors brewery in Golden, Colorado. In a given day numerous customer orders must be
filled, but there are a limited number of loading docks. If an order is placed on one of
the loading docks and some of the product the order requires is not produced for many
hours, that loading dock cannot be used to fill other orders. Also, if product is being
produced which is not needed by an order at one of the loading docks, that product must
be moved to inventory, which is limited in capacity. For this problem, we also need to limit
the amount of time that product spends in inventory. At present, the decision regarding
which order should be placed at a vacant loading dock is done manually; decisions are
made based on which order appears to have the "best" chance of getting filled quickly.
No effort is made to consider all of the orders at once and make decisions based on global
considerations. The ability to make decisions about the sequence of customer orders based
on global concerns is the motivation for this research.
Up to 16 production lines create products whose type and amount are known in
advance. There are approximately 500 of these beer types that can be produced, based
on the type of beer and different packages and labels. Each customer order requires
some amount of a subset of these beer types, and many of the orders have overlapping
requirements in terms of product. The product requirements to fill these orders is also
known in advance. In addition, the method of shipping is known in advance. This can
occur either on a truck or a rail car.
There are a finite number of loading docks at the plant for these orders. For a given
24 hour time period, there can be 200 orders which need to be filled, and there are only
approximately 40 loading docks for trucks and 18 loading docks for rail cars. The problem
becomes twofold: what is the best sequence in which to place orders on the loading docks,
and among the orders on the loading docks, what is the best way to allocate product to
these orders.
What is meant by the best sequence or allocation scheme? We have modeled several
optimization criteria which are discussed later in this paper. One criteria which is im-
portant to plant operations is the minimization of average inventory in the plant. This
 569

poptIiiOD
popaliioD

1----1-

···-1
'I------l
~--i-
~- 1
iDsert
eva/uale
oeWiDdi~du~

............................~~~~~~~~~~~~ ...........................................

Figure 1: A Steady State Genetic Algorithm

implies trying not to move product twice, and also reflects demands on cold storage. An-
other criteria is minimization of the mean time at dock for the orders. Later in this paper,
a method which appears effective for blending these dual concerns is presented.

2 Genetic Search Method

In this section, the mechanics of the genetic algorithm used in this research are described.
Readers familiar with genetic algorithms can skip over this section. Although this algo-
rithm seems simple, it has been demonstrated that the genetic method results in a global
search through the sampling of hyperplanes in an n-dimensional hypercube, where n is
the dimension of the problem space.
The genetic algorithm used for this problem is generally called a steady-state genetic
algorithm, as opposed to generational approaches[l]. Initially, a population is created
in which each individual in the population represents a solution to the problem to be
solved. Each individual (or solution encoding) is then evaluated and given a value based
on the objective function of the problem. The population is kept in sorted order based
570

Probability
of
selection

best in population worst in population

Figure 2: Function for Linear Selective Pressure

on their evaluation so that the individuals are ranked from best to worst. In the experi-
ments described in this paper, this initial population is randomly created. The algorithm,
illustrated in figure 1, then proceeds as follows:

• Two individuals are selected using a random number generator in conjunction with
a function which lends a slight "bias" toward individuals in the population which
are the better performers. This "bias" is also referred to as the selective pressure.
In these experiments a linear bias was used corresponding to the function shown
in figure 2. Appropriate values corresponding to the slope of this function will be
explored later in this paper.

• The two selected individuals are then recombined to form a new individual which
inherits information from each of the two selected individuals. This recombination
is also referred to as crossover.

• This new individual is then evaluated using the problem's objective function and
inserted back into the population. This insertion maintains the population in sorted
order and displaces the poorest performing individual in the population.

• This process continues for a preset number of trials. Each trial consists of the
generation and evaluation of a new individual.
 571

3 Representation and Evaluation

To represent the production/shipping optimization a symbolic encoding was used for each
individual in the population. Each individual is a string (chromosome), the individual
parts of which are discrete integers (genes) corresponding to customer orders. Thus an
individual in the population is a permutation of customer orders.
The evaluation function, which is used by the genetic algorithm to rank individuals in
the population, is a simulation of plant operations. This simulation performs operations
in the plant based on the sequence of orders in the chromosome. As a result of the
plant simulation , a value can be returned (the chromosome's evaluation) such as average
inventory in the plant for the given time period, mean time of orders at the docks , or the
number of orders filled in a given time period.
Because the evaluation function must be executed many times during the course of
genetic search, an effort was made to keep the simulation as simple as possible while still
incorporating necessary aspects of plant operations. The decision regarding which orders
to place on a loading dock is made based on the sequence of orders in the chromosome
passed for evaluation; the first order in the sequence is the first to get placed on an empty
dock, and the last order in the sequence will be the last order placed on a dock during
the simulation. Because plant operations are continuous, for a given data-set some of the
orders will always start out partially filled at some docks. The remaining docks are then
filled based on the orders starting at the beginning of the chromosome. When an order
gets filled during the simulation , the order to place on the dock next gets chosen based
on the next element in the sequence. Up front, the simulator analyzes the production
schedule from the production lines and creates a time-sorted list of pallets available for
each beer type produced . During time-steps in a given simulation, the components of
this list are allocated to orders that exist at docks based on the sequence in which these
orders appear on the chromosome being evaluated. Thus, the position of an order in the
chromosome determines both when that order will be placed on a loading dock and the
priority that order has for product once it is placed on a dock.
572

4 Data and Results

Preliminary results of this approach toward production/shipping scheduling have been
reported in [2)[3). Since that time, a computerized warehouse tracking system has become
active which enables detailed information on pallet movement from production, to orders,
and to and from inventory. Because of this development, the problem specification has
become much more precise and stronger conclusions can be drawn. The following analyses
of population size and selective pressure are based on a 24 hour segment of data and involve
the sequencing of 146 customer orders.

4.1 Operator Analysis

In previous studies it was demonstrated that recombination operators which emphasize

information having to do with the relative order of the elements in a chromosome work well
for this scheduling application[2). This makes sense in cases where resource consumption
is involved. In this particular application, consider the case of order X and order Y.
Assume order X requires 10 pallets of beer type xx, as does order Y. For simplicity, also
assume that these are the only orders which require type xx. Now assume that at some
given time only 15 pallets of type xx are produced and no more of that type is produced
during the given time period. Provided there is no product of this type in inventory, only
order X or order Y can possibly be filled. Given the way our simulation works, the first
of these orders to appear on the chromosome will be able to fill that part of its order, but
the other order will never be filled no matter where else it appears on the chromosome.
The operator which we have found the best results has been referred to as order#2
crossover [2) , developed by Syswerda[5). This recombination operator is used in all ex-
periments reported here. One of the parents to undergo recombination is "selected" at
random. Several points in the selected parent are chosen randomly and the order in
which these elements appear in the selected parent is imposed on the other parent. In
these experiments, the number of selected points averages half the length of the string.
Beginning with the initial item in the un selected parent, if an element is in the list of
selected elements it is replaced with the first element from this list. If an element is not in
this list it is inherited from the unselected parent from the current position. This process
 573

Parent 1: abc d e f g h i
Parent 2: b d a e c g f i h
Cross Pts: (Parent 1 selected)
** **
Offspring: a b d c e g f h i

Figure 3: Order Crossover #2 example

continues until the offspring is complete.

In figure 3, positions 3, 4, 6, and 7 have been selected as has Parent 1. The elements
in these positions are c, d, f and g. Beginning with P2[1] (a) , elements are checked to see
if they are in the list of selected elements. This is not the case for P2[1]' but P2[2] is a d,
which has been selected. Therefore, the offspring inherits this element in the position of
the first selected element: Off[3] = d. Since it is also the first element in the select list it
is inherited by the child: Off[l] = Pl[l] = a. P2[5] = b is the next element in the selected
list of elements, so it is copied to Off[4]. This continues until all of the selected elements
have inherited their order from P2. All remaining elements are directly copied from PI.

4.2 Population Size and Selective Pressure

Population size and selective pressure are the only parameters which need to be tuned with
the steady state genetic algorithm described earlier. In this implementation of a genetic
algorithm, the population is maintained in sorted order; as population size increases the
insertion cost of new individuals goes up. However, smaller populations may not have the
necessary diversity to sustain search. Generally, a "good" value for this parameter must
be arrived at empirically for a given problem.
Figure 4 shows the effects of population size on the performance of the genetic sched-
uler. These results are averaged over 10 experiments. Although at 100,000 recombinations
the population of size 500 yields a superior solution, the population of size 200 is quite
close, and good intermediate results can be achieved sooner. It can also be seen that very
rapid initial results can be achieved with a small population, but the loss of population
574

pop=50 -
940000 pop=100 -_.
pop=200 - --
pop=500 ..... .

Ci 920000 l\
3
iji
900000
.S

8800 00

860000

840000 ~ ________ ~ ________ ~ __________L -________-L________- J

o 20000 40000 60000 80000 100000
recombinations

Figure 4: Population size comparisons

diversity does not allow continued optimization.

Figure 5 shows the effects of varying the amount of "bias" toward good individuals
with a population size of 200. Figure 2 shows the linear bias function which we used.
A bias of 1.0 indicates no selective pressure, or a flat line in the figure. A bias of 2.0
means that the best individual in the population would receive twice the probability of
being chosen as the individual at the median of the population. Note that the effects are
similar but less pronounced than population size. A lower bias produces a slower, but
more sustained search that often produces superior results given sufficient time.

5 Alternate Evaluation Functions

Because the simulator can return any number of metrics based on plant operations as
the evaluation criteria for an individual, we have explored optimizing alternatives to the
average inventory optimization of the previous experiments. For these experiments, we
used data which covers a 72 hour period of plant operations. The number of orders to be
filled in this time period was 525. A population size of 500 was used, with a bias of 1.1.
 575

bias=1.1 -
bias = l.3 --"
940000 bi as :1 . 5 - - -
bias =l.7 -....
bias=1 .9 _.-

920000

C
Sc
w 9 00000
>
c
.~

w
'"ro
~
w
>
ro 880000

860000

840000 L-__ ~ ____- L____ ~ ____ ~ ____L-__ ~ ____ ~ ____ ~ ____L-__ ~

o 1 0000 20000 30000 400 00 50 000 60000 70 00 0 8 0000 90 000 10 0000

recombinat ions

Figure 5: Effect of various biases

The results are averaged over 5 experiments.

Figure 6 shows the results of optimizing average inventory with this larger data-set.
Note that the mean time at dock measure of performance improves very little when average
inventory is the optimization criteria. Figure 7 shows the effects of using mean time at
dock as the evaluation function. When mean time at dock is being optimized, the average
time at dock for an order is less than half of the value it reaches when average inventory
is being optimized. The lower graph in figure 7 shows that when mean time at dock is
the optimization criteria, average inventory actually increases as mean time decreases.

5.1 Combining Evaluation Criteria

Because keeping inventory low and having small dock times for orders are both important
for plant operations, we have investigated methods for combining these factors into a
single evaluation function for the genetic algorithm to optimize.
The scales for mean time at dock and average inventory are very different ; average in-
ventory has values in the range of hundreds of thousands, whereas mean time at dock is in
576

average inventory -

700000

i:'
..,0
c:

.::">
".'"
"
"
> 500000
"

300000 L -____~____~______~____~_____L_ _ _ _~~_ _ _ _~_ _ _ __L_ _ _ _~_ _ _ _~

o 10000 20000 30000 40000 50000 60000 70000 80000 90000 100000
Evaluations

mean time - - .

.><
II
0
900
..,
."

"
..,"
.~
c:
""e
600

300 L-____ ~ ____ ~ _____ L_ _ _ __ L_ _ _ _ ~ ____ ~ ____ ~~ ____ ~ ____ ~ _ _ _ _J

o 10000 20000 30000 40000 50000 60000 70000 80000 90000 100000
Evaluations

Figure 6: Average Inventory Optimization: The upper graph shows the optimization of
average inventory, and the lower graph shows how this optimization affects mean time at
dock.
 577

me~ n ti me -_.
1200

-'"
u \

,
0 \

'"u
900 \

'" ,
\
\
.~u \
,
c \
\
'E"
ID \

" ,,
600
,,
'-,
----
~-

--------------------------------

300 L-____ ____ ____ _____ L_ _ _ __ L_ _ _ _ ____ ____ ___L___ _

o 10000 20000 30000 4 0000 50000 60000 70000 8 0000 90000 100000
Evaluations

700000

C
3c
ID
>
.~
ID
'"
">'"
ID
500000
'"

300000 L-____ ~ ____ ~ ______L-____ ~ _____L_ _ _ _ ~ _ _ _ _ _ _L__ _ _ _ ~ ____ ~ _ _ _ __ J

o 10000 20000 30000 4 0000 500 00 60000 70000 80000 90000 100000
Eval ua t i on s

Figure 7: Mean Time at Dock Optimization: The upper graph shows the optimization
of mean time at dock, and the lower graph shows how this optimization affects average
inventory.
578

the range of hundreds of minutes. Simply adding these two evaluation measures together
would be ineffective. Because of this, the mean time at dock measure was normalized
with respect to average inventory. When the initial population of the genetic algorithm
was created, an average of mean time at dock and average inventory was calculated for
all (500) of the individuals in the population. A normalization factor was then calculated
based on these averages and a simple additive function of normalized mean time at dock
and average inventory was returned as the evaluation function.
Figure 8 shows that both average inventory and mean time at dock are minimized.
Neither function is minimized to the extent that they were when optimized separately,
but a blending has occurred where both criteria are taken into account.
The final sequence discovered using this dual evaluation function was analyzed using
a detailed simulation of the plant. This detailed simulation models all relevant plant
operations such as the movement of pallets on the factory floor and the time it takes to
move trucks in and out of the docks. A given simulation using this detailed model takes
several minutes on a high-speed workstation for a single sequence of customer orders,
whereas the actual evaluation function requires less than a tenth of a second. This detailed
model allows us to analyze the effects of a given sequence of orders. Because the data is
derived from actual plant operations, the sequence in which the orders were actually filled
during the given time period can be analyzed with this tool for comparative purposes ..
Using manual decision making at the plant for the data set of 525 orders, the filling
of these orders took 72 hours before they were completed. Using the sequence resulting
from genetic search with the dual evaluation function, all of the orders were completed
in 69 hours, and for the 6 hours before the schedule completed there were only 7 orders
on loading docks. This is roughly a 10% savings in the time needed to fill orders over a
three day period.

6 A Parallel Implementation for Increased Perfor-

mance
This parallel implementation of a genetic algorithm is also described in [6][4J. It uses
multiple genetic algorithms in a distributed fashion. Each genetic algorithm functions
 579

av erage inventory ---

700000

>,
H
S
C
ID
>
.~
ID
'"'"
H
ID
~ 50000 0

300000 L-____ ~ _____ L_ _ _ _ ~L_ ____ ~ _ _ _ __ L_ _ _ _ ~ ______ ~ ____ ~ _ _ _ __ L_ _ _ _ ~

o 10000 20000 30000 40000 50000 60000 70000 80000 90000 100000
Evaluations

mean time - _.
1200

I
,,
"ou
'0 900
\
I
\

'" ,
I

V'
I'
'-
600

300 ~ ____ ~ ____ ~ ______L -____ ~ ____ ~ ____ ~ ______ ~ ____ ~ ____ ~ ____ ~

10000 20000 3000 0 40000 50000 60000 70000 80000 90000 1000 00
Evaluations

Figure 8: Dual Evaluation Function: The upper graph shows the effects of the dual
evaluation function on average inventory, and the lower graph shows the effects of the
dual evaluation fun ction on the mean time at dock measure of performance.
580

autonomously, but is also provided with 2 additional parameters which determine how
the subpopulations exchange information. These are the number of trials between in-
formation exchange or swapping (the epoch), and the number of individuals exchanged
(nswap). Information exchange occurs in the following manner: when the first time ex-
change occurs, processor-l sends some number (a parameter) of the best individuals from
the subpopulation to processor-2, processor-2 sends to processor-3, etc. At the next epoch,
processor-l sends some number of the best individuals to processor-3, processor-2 sends
to processor-4, etc. Thus, in log2(P) number of exchange intervals, where P is the number
of processors, a "good" solution can migrate to all of the subpopulations. Experiments
indicate that swapping too often or not often enough degrades performance. We have
also found that swapping a small number of strings appears to give better results than
swapping a single string or a large number of strings.
For these experiments, the same data for a 24 hour time period was used as the first
set of experiments in this paper. The overall population size was fixed at size 1000, so
that when 10 subpopulations are used, a single subpopulation has 100 individuals. Fur-
thermore, the total number of recombinations was fixed . The experimental parameters
for these experiment for 1, 2, 5, 10, and 50 subpopulations are shown in table 1 "Subpop
Count" refers to the number of subpopulations used, while "Subpop Size" refers to the
number of strings in one subpopulation. "Trials" refers to the number of evaluations that
occur in each subpopulation (not counting the evaluation of the initial subpopulation).
Note that when Subpop Size or Trials are multiplied by the Subpop Count that the total
population size and number of evaluations are the same across all experiments. "Swap
Interval" indicates how many recombinations (or evaluations) occur before swapping oc-
curs, and "Number Swapped" indicates how many strings are passed to the appropriate
neighbor.
Figure 9 shows the effects of running this parallel implementation on the produc-
tion/shipping scheduler, averaged over 5 experiments. Each experiment with different
numbers of subpopulations achieves very similar optimization, although subpopulations
of 2,5, and 10 actually outperform the single population. This suggests that operation of
the genetic scheduler can obtain 1, maybe 2 orders of magnitude in speedup using parallel
hardware that is currently available.
 581

Subpop Subpop Swap Number

Count Size Trials Interval Swapped
1 1000 250000 n.a. n.a.
2 500 125000 12500 10
5 200 62500 5000 5
10 100 31250 2500 5
50 20 15625 500 2

Table 1: Subpopulation Experiment Data

860000 r---------,-----~--~----------r_--------,_--------_,
1pop -
2pop -_.
5pop - --
10pop .....
50pop _ . -

850000

50000 1 00000 150 000 200000 250000

recombinations

Figure 9: Performance of genetic scheduler with various subpopulation sizes

582

7 Future Work
As mentioned earlier, an effort was made to keep the scheduling simulator as simple as
possible so as to speed the evaluation function used with the genetic algorithm. We have
now finished construction of a detailed simulation model which precisely models the flow
of pallets in all parts of the plant. The first validation step has been to make sure that
this detailed model accurately models plant operations. We are now engaged in validation
of the genetic scheduler using this detailed model for feedback. Sequences of customer
orders achieved through genetic optimization are being analyzed with this tool in order
to demonstrate that the technique is effective in the real world. The detailed model also
allows analysis of the effectiveness of the manual techniques currently used at the plant,
i.e. the way things actually happened during the time period of the data-set. Although
testing needs to occur across a wider range of data-sets, preliminary results indicate
that this system will allow more orders to be filled in the same time period, resulting in
substantial savings.

8 Summary and Conclusions

New results on detailed data-sets indicate that a genetic scheduling system for a produc-
tion/shipping application is both viable and can result in substantially improved sched-
ules. For applications in which the relative order of elements in a sequence is important,
the recombination operator described as order #2 appears to work very well. As observed
in many other application areas, for this problem fast initial search can be achieved with
small population sizes, although larger populations eventually evolve superior results. Be-
cause this approach to scheduling involves a plant simulation, numerous evaluation criteria
can be compared. In addition, it is possible to blend different evaluation criteria together,
thus taking more than one factor into account during optimization. Substantial speedup
of the genetic scheduler can be achieved by using a distributed algorithm with parallel
hardware; this consideration may be important for real-time operation of the scheduler.
Finally, the first stage of verification using a detailed plant simulation model has demon-
strated that savings of up to 10% in the overall length of a schedule are possible using
this method.
 583

References
[1] Lawrence Davis , editor. Handbook of Genetic Algorithms. Van Nostrand Reinhold ,
1991.
[2] Timothy Starkweather, Susan McDaniels, Keith Mathiask, Chris Whitley, and Dar-
rell Whitley. A comparison of genetic sequencing operators. In Proceedings of the
Fourth International Conference on Genetic Algorithms, pages 69-76. Morgan Kauf-
mann Publishers, Inc., 1991.
[3] Timothy Starkweather, Darrell Whitley, Keith Mathias, and Susan McDaniels. Se-
quence scheduling with genetic algorithms. In Lecture Notes in Computer Science.
Springer/Verlag, 1991.

[4] Timothy Starkweather, Darrell Whitley, and Keith Mathiask. Optimization using
distributed genetic algorithms. In Lecture Series on Parallel Computing in Nature.
Springer /Verlag, 1990.

[5] Gilbert Syswerda. Schedule optimization using genetic algorithms. In Lawrence Davis,
editor, Handbook of Genetic Algorithms, chapter 21, pages 322-349. Van Nostrand
Reinhold, 1991.
[6] D. Whitley and T. Starkweather. Optimizing small neural networks using a distributed
genetic algorithm. In Proceeding of the 1990 International Joint Conference on Neural
Networks, 1990.
 Genetic Algorithms in Problem Space for Sequencing Problems

by

Robert H. Storer
S. David Wu
lnKyoung Park

Department of Industrial Engineering

Lehigh University
Bethlehem, P A 18015, USA

1. Introduction

In this paper, genetic algorithms will be developed for sequencing type problems
of importance in manufacturing systems. The proposed algorithms are based on an
auxiliary problem domain called "problem space" [15, 17]. Problem space provides a
framework in which problem-specific information can be incorporated explicitly into
local search heuristics. The proposed space has been found to be well suited for search
by genetic algorithms perhaps because standard crossover can be used. In this paper,
properties of problem space will be discussed, then three test cases will be presented
which illustrate the usefulness of the method. The test problems we present are 1) the
number partitioning problem, 2) the classic job shop scheduling problem (JSP) with
minimum makespan as the objective, and 3) the "standard cell placement problem"
which arises in the design of VLSI circuits.

2. Background
Local search heuristics have received much recent attention spurred by new
probabilistic algorithms such as tabu search, simulated annealing, and genetic
algorithms. These methods provide mechanisms for the search to escape local optima
thus overcoming an important obstacle for local search. Some impressive results have
been obtained, leading the Committee On the Next Decade of Operations Research to
call these methods "extremely promising" [5]. Other authors are less optimistic, and
some other recent results have been discouraging [8, 9] .
While much effort has been devoted to search methods, relatively little research
exists on defining the neighborhoods which serve as the basis for local search. For
sequencing type problems, local search typically relies on "swapping neighborhoods."
For example in scheduling problems a neighborhood is usually defined as all schedules
 585

that can be obtained by swapping two jobs in the incumbent sequence. Variations on
this theme include adjacent job interchange and all-pairs interchange. In the
application of genetic algorithms to sequencing problems, the analogous procedure is the
use of the sequence itself as the chromosomal encoding of a solution. The selection of a
search neighborhood (and solution encoding for genetic algorithms) is crucial to the
performance of local search, and has been termed "an art which is usually guided by
intuition" [13].

In this paper we focus on an alternative search neighborhood. An inherent

weakness in typical neighborhood definitions is the failure to incorporate problem-
specific information to help identify promising candidate solutions. For example, there
is no way to determine, a priori, a promising swap from a bad swap. In the job shop
scheduling problem (JSP), many swaps will even yield sequences which violate
precedence constraints. The approach taken in this paper is to use problem specific
information to identify promising neighbors at the expense of more efficient evaluation.
This is achieved by incorporating known, constructive heuristics into the definition of a
search neighborhood.

3. Problem Space

The key to problem space is the utilization of a fast, problem-specific heuristic as

a basis. Problem space and the resultant search neighborhood is based on the fact that
such a heuristic is a mapping from a problem to a solution; H(P)=S. Thus a heuristic,
problem pair (H,P) is an encoding of a solution sequence. Formally, the problem space
optimization problem can be defined as follows:

Min V(H(PO + 6 ) j PO)

6{D

Where:
V(SjP) : is the objective function value of solution sequence S to problem (data) P
Po: an Nxl vector containing the original problem data
H : a base heuristic which generates feasible solutions given problem data
6: an Nx1 perturbation vector, where a perturbed problem is P = Po +6
D : the domain of perturbation vectors; a subset of ~N
This optimization problem is defined on the domain of problem perturbations bfD, or
equivalently, on the space of "dummy problems" P=P 0+6 thus the name problem
space. Note also that the original problem data Po is always used to evaluate the
objective function.
586

Using problem space as a basis, local search can be applied to perform the
optimization. Since the space is defined in RN many metrics suggest themselves as a
means to define neighborhoods. A straightforward approach is to define a neighborhood
as all perturbed problems within a certain distance (t) from the current incumbent P'.
A neighborhood (N p ) of problems close to P' can be defined as:
Np = { all P'+6 such that "6,, <t }
where " . 11 is the Euclidian norm. A corresponding neighborhood of solutions Ns results
by applying the base heuristic H to all elements of Np:
Ns = { all S=H(P) such that PtN p }

4. Properties of Problem Space Search

Problem space search has several desirable properties worthy of note which are
summarized in this section.

A fundamental assumption III the use of the problem space is that the base
heuristic performs well on the problem at hand since it has been designed to exploit
problem-specific knowledge and structure. Thus problem-specific knowledge IS

explicitly incorporated into the local search algorithm through this base heuristic.

Every solution sequence visited by the method is a solution generated by the

base heuristic H. If only slight perturbations to the original problem data are made it
seems reasonable that "good" solution sequences will be generated by the base heuristic.
That is, we expect solutions in the vicinity of the original problem to be good solutions.
Thus the search may be restricted to an area (that is a subset D of RN) where
promising solutions should be concentrated.

An additional benefit of the embedded base heuristic is that it guarantees the

feasibility of every solution visited by the search (presuming of course that the base
heuristic generates feasible solutions). Conversely, when swapping is used to generate
neighboring sequences, feasibility is often violated, especially in highly constrained
problems. Thus for many problems, algorithm efficiency may be greatly enhanced by
restricting the search to feasible solutions. For this reason we expect problem space
search to be an attractive practical approach for dealing with problems with many
complex constraints.

A significant property of problem space is that it is particularly well suited for

genetic algorithms. A major obstacle in the application of genetic algorithms to
sequencing problems has been the difficulty of producing a well defined crossover
operator. When the usual sequence encoding is used, the application of standard
crossover invariably yields illegal offspring solutions thus mandating unwieldy operators
 587

such as "PMX" [6], "cycle crossover" [7], "edge recombination" [18], etc. By operating
in problem space, standard crossover operators are trivially constructed and applied:
form an offspring problem by taking part of the problem data from one parent and the
rest from the other. The base heuristic is then applied to the offspring problem to
produce a solution guaranteed to be a feasible sequence.

In some problems, additional advantages can be achieved by modifying the

neighborhood structure and/or the perturbation method to capitalize on problem
attributes. In our initial implementations we have simply added random numbers
(uniformly distributed over a restricted range) to the data. In future applications, more
intelligent means of perturbing the problem data may prove useful. In section *. we
apply problem space to the number partitioning problem. For this problem, a
somewhat modified perturbation scheme is required to achieve best results. Another
good example is when "rescaling" the problem data causes the base heuristic to produce
a different solution. An example is the case of the space filling curve heuristic for the
planar TSP [3]. If the cities are simply rotated or translated in the plane, the space
filling curve heuristic will produce different solutions. Thus the heuristic can be forced
to generate alternative solutions without changing the fundamental nature of the
problem.

Under quite general conditions it can be shown that all possible solutions
(including of course the optimal solution) are contained in problem space. Sufficient
conditions are: 1) if for any solution S the base heuristic can be "reversed" to find a set
of problem data that will cause the base heuristic to generate S, and 2) the method used
to perturb the problem can (perhaps after repeated application) generate any instance
(of a given size N) of problem data. Under these conditions, it is easy show that all
solutions can be reached by a local search of problem space. By 1), for every solution S
there is a set of problem data P S which, when plugged into the base heuristic, will yield
S. By 2), the perturbation method is capable of generating P S'

A problem space local search method will of course require the application of the
fast, base heuristic at each iteration of the search. As discussed above, applying this
heuristic at each iteration provides a means to incorporate problem-specific knowledge
into the search, and assures the feasibility of each solution generated. Since applying
the heuristic at each iteration entails significant computational effort, fewer solutions
will be investigated as compared to swapping based methods. The general effectiveness
of the problem space search will depend, in part, on the relative computational effort
required at each iteration as well as the overall quality of the base heuristic and the
search space it induces. For example, problem space methods seem attractive in the
case of the JSP where feasibility checking and solution evaluation are costly even for
588

simple swapping. Conversely, in the traveling salesman problem, the effect of swapping
two arcs in the tour can be quickly computed locally by considering only the two old
arcs and the two new arcs. Further, infeasible swaps are easily avoided, thus problem
space methods may be less attractive for the TSP.

5. Problem Space Genetic Algorithms

The means by which problem space can be most effectively searched has yet to
be specified. We have experimented with several search strategies in [16]. Among the
many methods tested, the most effective was, to our surprise, the genetic algorithm. As
a result of this success, further experiments have been performed to tune the problem
space genetic algorithm (PSGA). The resultant algorithm, which we outline below, has
performed well in several applications without additional tuning.

In GA's, the fitness evaluation function, mutation rate, population size, crossover
rate, and a method for generating the initial population must be specified. The fitness
evaluation function assigns higher selection probabilities to better solutions.
Specifically, for each solution (i) in the population P, the selection probability is:

K- 1 (~ax [V(j)] - V(i)

JlP
Y
where V(i) is the objective function value of solution i, and ,.. is a tuning parameter. K
is a proportionality constant calculated so that the probabilities summed over the
population add to one. The power,.. determines the selectivity of the fitness function.
With a high value of ,.., only the fittest few solutions are likely to be selected, diversity
is quickly lost, and the algorithm will quickly converge to a population of identical
solutions. Thus,.. must be chosen to balance convergence and diversity. Empirical
testing showed that ,..=4 provided good results.
The mutation operator simply adds a Uniform(-9,9) random perturbation to
problem data elements selected for mutation. The parameter 9 controls the magnitude
of the perturbations. The same method is used to generate the initial population;
random perturbations are added to each element of the problem data vector.
Parameter settings that have produced good results include a high mutation rate of
15%, a population size of 100, and a crossover rate of 80%.

In problem space, a chromosomal representation is easily achieved using the

vector of perturbed problem data. The standard crossover scheme can be used to
combine parental solutions to produce an offspring solution.
 589

6. Case 1: Number Partitioning

In a recent paper by Johnson et al. [9], simulated annealing was subjected to a.

rigorous series of experimental tests. One particular experiment, the application of
simulated annealing to the number partitioning problem, has inspired our first
experiment. While number partitioning has no apparent relevance to manufacturing, it
illustrates clearly the ease of use, and the potential for success of local search in problem
space.

In their application of simulated annealing to number partitioning, Johnson et al.

found it to be generally inferior to the "differencing" algorithm of Karmarkar and Karp
[11]. Since the differencing algorithm runs in O(nlogn), the performance of simulated
annealing was apparently quite poor. Johnson et al. explain these poor results quite
eloquently:

"The challenge of this problem (number partitioning) is that the natural

"neighborhood structures" for it, those in which neighboring solutions differ
as to the location of only one or two elements, have exceedingly
"mountainous" terrain, in which neighboring solutions differ widely in
quality. Thus, traditional local optimization algorithms are not competitive
with other techniques for this problem, in particular the "differencing
algorithm" of Karmarkar and Karp (1982)."
They go on to state:
"There remains the question of whether some other neighborhood structure
for the problem, perhaps using different notions of solution and cost, might
prove more amenable to annealing. We do not rule out this possibility,
although at present we see no reasonable alternatives."

The apparent failure of swap neighborhoods and the existence of a fast, high-
quality heuristic (the differencing algorithm of Karmarkar and Karp [U]) would seem to
provide the ideal setting for problem space search.

6.1. The Number Partitioning Problem

Given N real numbers aI ' a2' .... . , an III the interval [0,1], the number
partitioning problem is simply to divide the numbers into two sets Sl and S2 so as to
mllllmlze:

There are no restrictions as to the cardinality of SI or S2'

The differencing algorithm for the number partitioning problem can be

590

summarized as follows (a more detailed description appears in Johnson et al.):

O. Initialize a set of n nodes labeled with the values ai and declared "live".
1. Among live nodes, find those with the largest (u) and second largest (v) label.

2. Join u and v with an edge, relabel u with lu-vl, and declare v "dead".

3. If more than one live node remains, go to 1.

Otherwise, two color the resulting tree to obtain the partition.

6.2. Problem Space for Number Partitioning

The success of the differencing algorithm can be attributed to the fact that at
each iteration, two nodes with large labels are replaced by one node with significantly
smaller value. This allows for a "fine tuning" effect as the algorithm proceeds. The
obvious problem space for this problem would be to perturb each ai' apply the
differencing algorithm to the "dummy ai's" to obtain the partition, and then evaluate
the partition using the original ai's. However by using perturbed data throughout the
differencing algorithm, the fine tuning effect is lost, and results are poor. In order to
force the algorithm to generate alternative solutions while preserving the fine tuning
effect, we modified the obvious perturbation strategy slightly. The initial set of n
nodes in step O. are labeled with the perturbed ai's. However in step 2. when nodes are
relabeled with lu-vl, the original values of u and v are used. The result of this
modification is to vary the pairings of nodes during the early stages of the algorithm
while maintaining the fine tuning effect which occurs during the later stages.

6.3. Experiment

Following the method outlined in Johnson et al., test problems were created by
randomly generating n uniformly distributed numbers in [0,1]. Each number a1 to au
consisted of 36 digits to prevent errors induced by roundoff. Ten problems of size
n=200 were generated.

In Johnson et al. [9], a problem generated by this method was tested.

Application of the differencing algorithm resulted in a partition with an objective
function value close to 10-8. Several long runs of simulated annealing and local search
with random restarts were also performed based on the swap neighborhoods. Rarely
was the differencing algorithm beaten, and the best solution found in all the runs was
greater than 10-10. In contrast, the results of 20 generations of PSGA on the ten test
problems appear in Table 1. Clearly, by creating a search space which explicitly
exploits know problem specific heuristics, good results are obtained.
 591

Table l.
PSG A Results for 200 element Number Partitioning Problems

Differencing
Problem Algorithm Result PSGA Result
1 1.09x10- 9 8.95x10- 13
2 2.13x10- 9 9.31x10- 15
3 3.79x10- 9 3.91x10- 13
4 5.52x10- 10 1.90xlO- 13
5 6.03x10- 9 2.46xlO- 13
6 1.46x10- 9 2.65xlO- 13
7 9.85x10- 9 3.28xlO- 13
8 2.29x10- 9 2.61xlO- 14
9 2.77x10- 9 4.33xlO- 13
10 5 . 17x10- 10 1.65x10- 13

7. Case 2: Job Shop Scheduling

The job shop scheduling problem with mlllimum makespan objective is a

classical combinatorial optimization problem of great interest from both a practical and
theoretical standpoint. The JSP is well known to be NP-hard. Even within this class,
the JSP seems to be especially notorious in that even relatively small problems cannot
be solved by classical implicit enumeration methods. Sophisticated branch and bound
schemes ([1], [4]) have been developed recently which, given extensive computation
time, can solve problems on the order of 100 operations. Since practical problems are
typically much larger, heuristic solution procedures become attractive.

7.1 The Base Heuristic

The fast problem-specific JSP heuristic used to generate the problem space was
first used by us in [15] and is based on a hybrid version of the well known concepts of
active and non-delay schedules combined with SPT dispatching. The hybrid method
generates schedules that lie between active and non-delay schedules. The algorithm is
parametrized by 0, which represents the maximum time a machine may be kept idle
when jobs are immediately available for processing. Keeping the machine idle may be
beneficial if it is waiting to process a more critical operation which will soon become
ready. When 0,=0 , the search space consists of only the set of non-delay schedules. As
0, increases from 0 to infinity, the cardinality of the set of schedules also increases until
the search space contains all active schedules. Details of this algorithm may be found in
[15].
592

Since SPT dispatching is used to prioritize waiting jobs, the obvious problem
space strategy is to perturb the processing times of operations. Specifically, if (iJ)
indexes operation j of job i, the processing times of operations P(iJ) are perturbed by
adding a Uniform( -/J,/J) random variable to yield "dummy" processing times DP(iJ).
These dummy processing times are used at the scheduling decision points in the
application of the SPT dispatching rule to force the scheduling algorithm to produce
different sequences. When operations are actually scheduled, real processing times are
used to construct the schedule and compute the makespan.

In summary, the genetic algorithm for the JSP uses a list of dummy processing
times as the encoding, and standard crossover is applied. The perturbation method just
described is used to generate the initial population, and to perform mutation. The base
heuristic decodes the chromosome and computes makespan.
Our initial application of the PSGA consisted of a single run of 2000 generations
for each problem. On the smaller problems, little improvement was observed after 100
generations. Our second application of the PSGA thus entailed 20 independent runs of
the PSGA over 100 generations. Results of both experiments appear in Table 2. under
"Ix2000" and "20xIOO".

7.2. Test Problems

Twenty job shop scheduling problems were used for testing. In all problems,
each job must visit each machine exactly once. Processing times are uniformly
distributed on the integers between 1 and 100. Fifteen of the problems, ten consisting
of ten machines and ten jobs (lOxlO) and five (I5xI5) were taken from [2], and have
appeared before in the literature. For these problems, the order in which jobs visit
machines was generated randomly. The other five problems were (IOx50) and were
generated by us in a manner similar to that used in [12]. This problem generation
scheme creates tough problems by placing special restrictions on precedence constraints.
Specifically for a problem with 10 machines, all jobs must visit the first 5 machines
before visiting the second 5 machines. Within the two sets of machines, routings are
random.

7.3. Algorithms Used for Comparison

In order to assess the effectiveness of the PSGA heuristic on the JSP, we

compared our results to variations of the shifting bottleneck algorithm first described in
[1]. This paper ([1]) presented results for five of the IOxlO and all of the I5xI5
 593

problems used here. Results for two algorithms were given. The first algorithm, SBI, is
a sophisticated heuristic based on solving a series of interrelated one-machine problems.
The second algorithm, SBII, augments the SBI heuristic with a branch and bound
search in an attempt to improve the solution.

In a more recent paper [2], Applegate and Cook describe an implementation of

the SBI algorithm as well as their own modification which also uses a branch and bound
scheme to improve on the original SBI solution. Applegate and Cook refer to their
implementation of SBI as "Bottle" and their augmented algorithm as "Shuffle". We
have been fortunate to receive code for the "bottle" and "shuffle" algorithms from Cook,
which in turn allow us to test PSGA on some of the larger lOx50 problems.
Unfortunately, the bottle algorithm was able to solve only three out of the five lOx50
problems in under 24 hours on a Sun 4/280 computer. The shuffle algorithm could
solve none in the same time period. We observed this same behavior in our own
previous attempts at implementing the SB algorithm on larger problems.

Results of the experiment appear in Table 2. The first 15 problems are labeled
as III [2]. The PSGA results appear competitive with the SB methods. It seems that
the relative performance of PSGA improves as the problem size increases.

8. Case 3: A Standard Cell VLSI Design Problem

The layout of VLSI circuits consists of the placement of circuit elements (or
"cells") and connections of elements (or "wires") on a layout area. In the most general
case, a set of rectangular cells of variable size and shape are given, along with
connectivity information or a "net list". Each net specifies a set of cells which must be
connected (wired) to each other to form the signal net. Thus each cell will typically be
a member of several nets, and each net will connect several cells. Three principal
design styles are used for cell based design, the gate array, the standard cell, and the
macro cell approach.

A gate array circuit is composed of cells of exactly the same size and shape laid
out in a lattice arrangement. Vertical and horizontal "channels" are placed between
each row and column of cells. These channels provide space for the "wires" used to
connect cells.

Standard cells have a fixed height but different widths, depending on their
functionality. Terminals are located along the cell's top and bottom edge. A standard
cell layout consists of cells placed next to each other in rows with channels included
between rows to accommodate wires. If wiring across a cell row is required, "feed-
594

Table 2.
PSGA Results for Job Shop Scheduling Experiments
Optimum Ix2000 20xIOO A&C A&C
Prob Size Sol'n PSGA PSG A SBI SBn Bttle Shfle
---------------------------------------------------------------------------------------------
La16 10x10 945* 1008 981 1021 978 1076 946
La17 10x10 784* 801 794 796 787 829 784
La18 10x10 848* 860 860 891 859 855 848
La19 lOx10 842* 871 860 875 860 863 846
MT10 lOx10 930* 954 956 1015 930 952 930
Orb1 10xlO 1059* 1070 1070 1176 1070
Orb2 lOx10 888* 910 898 927 890
Orb3 10x10 1005* 1046 1031 1090 1021
Orb4 10x10 1005* 1037 1031 1065 1019
Orb5 10x10 887* 935 903 939 896
La36 15x15 1268* 1325 1305 1351 1305 1326 1326
La37 15x15 1397* 1463 1458 1485 1423 1471 1439
La38 15x15 unkn 1244 1239 1280 1255 1307 1295
La39 15x15 1233* 1281 1258 1321 1273 1301 1300
La40 15x15 1222* 1281 1258 1326 1269 1347 1286
S\,l1 10x50 unkn 3134 3179 3423 unkn
SI.'12 lOx50 unkn 3130 3240 unkn unkn
S\'13 10x50 unkn 3220 3339 3508 unkn
SI.'14 10x50 unkn 3073 3131 unkn unkn
S\'15 10x50 unkn 3056 3154 3280 unkn
---- - ---------------------------------------------------------------------------------------
Notes:
1. SBI and SBn results available only for problems tested in [IJ
2. lOx50 problems which A&C algorithms failed to solve in under
24 hours on a Sun 4/280 are labeled unkn.
--------------------------------------------------------------------------------------------

through" cells may be added, but this alternative is considered undesirable. It is the
standard cell placement problem that we address in this example.
A macro cell is rectangular but otherwise unrestricted 10 shape, and has
terminals located along all four edges. The layout of a macro cell circuit is the most
challenging of the three problems, and bares resemblance to the facilities layout
problem often encountered in manufacturing.

A two phase approach is typically taken in VLSI design. In the first phase,
"placement", cells are laid out with the goal of minimizing the total required length of
wiring needed to form the nets. That is, cells which are highly interconnected will be
placed close together so as to minimize the required wiring. Once cells are placed, the
second phase, "routing", lays out the wires in the channels to form nets.
 595

The mam objective of placement and routing is to minimize the total layout
area. It should be noted that this objective can only be exactly evaluated after both
design phases are complete. However, total area is closely related to objectives which
can be directly computed during placement. In this experiment we use the "bounding
box" objective. Given a placement, one can determined for each net the smallest box
that contains all terminals in the net. Summing the height and width of each box over
all nets yields the bounding box objective function value. In this experiment we
attempt to solve the standard cell placement problem with the bounding box objective.

The most popular current approach to solving this problem is simulated

annealing based on swapping the positions of cells in their layout rows. In particular,
the TimberWolf package [14], which is widely used in VLSI design, is based on this
simulated annealing approach.

8.1. The Base Heuristic for Standard Cell Placement

Following the approach of previous researchers (d. [10]), our base heuristic
consists of first producing a single row layout of all the cells, and then "folding" this row
so as to fit the desired layout shape. The single row layout problem closely resembles a
maximal weight Hamiltonian path problem. Our heuristic can be summarized as
follows:
l. Create a complete graph with a node for each cell. The weight W(i,j) on arcs
between cells i and j is equal to the number of nets which include both cells. Thus
the arc weight represents in some sense the degree of connectivity of the cells.

2. Solve the maximum weight Hamiltonian path problem on this graph. We

replace W(i,j) with bigM-W(i,j). We then apply the minimum spanning tree
heuristic to find the path. The path, in turn, specifies the linear ordering of the
cells in the row.
3. Fold the row of cells to fit into the layout area. Folding is done in the obvious
way, i.e. following rows left to right, right to left, left to right.. .. will reproduce the
linear layout.
The complete graph formed in step l. is actually quite sparse, that is most W(i,j)=O.
Our problem space strategy is to perturb the W(i,j) which are initially non-zero. Zero
weight arcs are not perturbed, and in fact are ignored so that data structures for sparse
graphs can be used.

8.2. Results

Three example circuits were used to test the problem space algorithm. We
compare our results to those produced by the TimberWolf 3.3 simulated annealing
package [14]. The results appear in Table 3.
596

Table 3.
Standard Cell Placement Problem Results

Circuit Problem Space TimberWolf

---- ... ---------------------- ...... --- .. -----... _---------- .. _--- .... ------------ ...... _.... _.. --------- .... __ ...... _--------------------------
Wire Length Number of Wire Length Number of
Cells Nets (Bounding Box) Configurations (Bounding Box) Configurations
100 213 56K lOOK 42K 1.17M
183 254 90K lOOK 95K 2.14M
469 495 46K 72K 42K 5A9M

While the problem space results are in the same ballpark as simulated annealing, the
results are somewhat disappointing. In retrospect it seems clear that standard cell
placement is an example of a problem which is well suited for swapping neighborhoods
thus putting problem space at a relative disadvantage. The effect that swapping two
cells has on the bounding box objective can be quickly determined by examining only
nets which contain the two cells. Thus, many more layouts can be investigated using
the swap neighborhoods. The total number of configurations investigated by each
method (for runs of roughly the same CPU time) is also shown in Table 3.

9. Conclusions

In conclusion, we feel that problem space based methods show promise for some,
but certainly not all, combinatorial optimization problems. They seem particularly
attractive in cases which usually prove difficult for traditional methods. In particular,
problem space seems to be a good heuristic approach for large problems with complex
constraints, and with objective functions which cannot be locally computed for swap
neighborhoods. Further, the inherent simplicity adds to the appeal. While genetic
algorithms have worked fairly well as a search procedure, we feel that there is still much
room for improvement in this area.

10. References
1. Adams, J., E. Balas and D. Zawack, (1988). "The Shifting Bottleneck Procedure for
Job Shop Scheduling", Management Science. M(3), 391-401.
2. Applegate, D. and Cook, W. (1991). "A Computational Study of the Job-Shop
Scheduling Problem", .QRSA Journal on Computing. a(2), 149-156.
3. Bartholdi, J. J., and Platzman, L. K., (1985). "Heuristics Based on Spacefilling
Curves for Combinatorial Problems in Euclidian Space", Management Science. M(3),
291-305.
 597

4. Carlier, J. and E. Pinson, (1989). "An Algorithm for Solving the Job-Shop Problem,"
Management Science, ~ 164-176.

5. Committee On the Next Decade of Operations Research (CONDOR), (1988).

"Operations Research: The Next Decade," Operations Research, ~(4).

6. Goldberg, D. E., and Lingle, R. (1985). "Alleles, Loci, and the Traveling Salesman
Problem", In J . Greffenstette (ed.) Proceedings Q{ ~ International Conference on
Genetic Algorithms and Their Applications, Carnegie Mellon University.

7. Greffenstette, J. J ., Gopal, R., Rosmaita, B., and Van Gucht, D. (1985). "Genetic
Algorithms for the Traveling Salesman Problem", In J. Greffenstette (ed.) Proceedings
of an International Conference QD. Genetic Algorithms and Their Applications, Carnegie
Mellon University.

8. Johnson, D.S., Papadimitriou, H., and Yannnakakis,M. (1988). "How Easy is Local
Search ?", Journal ill' Computer and System Science, ~(1), 79-100.
9. Johnson, D.S., Aragon, C.R., McGeoch, L.A., and Schevon, C. (1991).
"Optimization by Simulated Annealing: an Experimental Evaluation; Part II Graph
Coloring and Number Partitioning" , Operations Research ~(3), 378-406.

10. Kang, S. , (1983). "Linear Ordering and Application to Placement", Proceedings of

the 20th Design Automation Conference, pp 457-464.

11. Karmarkar, N. and Karp, R.M. , (1982) . "The Differencing Method of Set
Partitioning". Report No. UCB/CSD 82/113, Computer Science Division, University of
California, Berkeley.

12. Muth, J .F. , P. Thompson and P . Winters (eds.), (1963). Industrial Scheduling,
Prentice-Hall, Englewood Cliffs New Jersey.

13. Papadimitriou, C.H. and K. Steiglitz, (1982). Combinatorial Optimization

Algorithms and Complexity, Prentice Hall Inc., Englewood Cliffs, New Jersey.

14. Sechen, C. and Sangiovanni-Vincentelli, A. (1985). "The TimberWolf Placement

and Routing Package", IEEE Journal ill' Solid-State Circuits, 20(2), 510-522.
15. Storer, R.H., Wu, S.D. and Vaccari, R. (1990). "New Search Spaces for Sequencing
Problems with Application to Job Shop Scheduling", to appear, Management Science.
16. Storer, R.H. , Wu, S.D. and Vaccari, R. (1991). "Local Search in Problem and
Heuristic Space for Job Shop Scheduling", Working Paper, Dept. of Industrial
Engineering, Lehigh University, Bethlehem, PA.
17. Storer, R.H. and Wu, S.D. (1991). "Local Search in Problem and Heuristic Space for
Job Shop Scheduling Genetic Algorithms", to appear in Proc. Joint US-German Conf.
QD. New Directions ill' OR in Manufacturing.

18. Whitley, D., Starkweather, T. and Fuquay, D. (1989). "Scheduling Problems and
Traveling Salesmen: The Genetic Edge Recombination Operator" Proceedin~s 2f ~
Third International Conference QD. Genetic Algorithms, George Mason University,
Morgan Kaufman Publishers, San Mateo CA.
List of Sponsors

Fernuniversitat Hagen

Deutsche Gesellschaft fUr Operations Research (DGOR)

Deutsche Forschungsgemeinschaft (DFG)

Minister fUr Wissenschaft und Forschung des Landes Nordrhein-Westfalen

National Science Foundation (NSF)

Operations Research Society of America - Special Interest Group on

Manufacturing Management (ORSA-SIGMA)

Volksbank Hagen